Satya Nadella has given the cost a name. In a July essay, he calls it the Reverse Information Paradox: a buyer of hosted intelligence pays twice, once in cash and again in the proprietary knowledge the model absorbs to become useful. Every prompt and correction teaches the provider’s system something about the customer’s business. Run the model on hardware you control, and both the record and the learning stay inside your gates.

This is why enterprise demand is moving toward open-weight models, the ones whose trained parameters can be downloaded and run on hardware a company owns. The hardware arrived this year at every scale, from laptops running 120-billion-parameter models on Nvidia’s RTX Spark to corporate GPU servers running larger systems with no hyperscaler in the loop.

The problem sits in the AI scoring leaderboard. On the Artificial Analysis Intelligence Index in mid-July, Moonshot’s Kimi K3, a 2.8-trillion-parameter model with open weights due July 27, ranked fourth in the world, behind only Anthropic’s Fable 5 and two configurations of OpenAI’s GPT-5.6.

Z.ai’s GLM-5.2 leads the models whose weights are published, ahead of DeepSeek’s V4 and MiniMax’s M3. Release by release since 2025, the strongest open-weight models have come from Chinese labs.

Closed frontier vs. the open tier. Source: Artificial Analysis Intelligence Index (v4.1), mid-July 2026; company announcements.

That lead is now state policy.

Opening the World AI Conference in Shanghai on July 17, Xi Jinping urged countries to seize the opportunity of open-source AI, cast China as a supplier of AI public goods, and welcomed a new World AI Cooperation Organization that twenty-nine countries had set up a day earlier, headquartered in Shanghai. Ben Thompson reads the move as commoditizing your complements: give the model away, and value flows to the physical industries (robotics in particular) built on top of it, where China leads.

Most companies never needed a frontier generalist AI model anyway. A logistics firm’s model does not have to design rockets. Fine-tune a smaller open model on your own data and you get a system that beats a hosted generalist at your actual work, fits on hardware you own, and keeps the training signal inside the company. The Western side is moving there too. Mira Murati’s Thinking Machines released Inkling, a 975-billion-parameter open model under Apache 2.0, the largest Western open-weights release to date, but it still significantly trails the Chinese flagships. Palantir’s Alex Karp told CNBC that enterprises are shifting to open weights for cost and for control over their models, data, and compute.

Western controls do not reach any of this.

Germany, Italy, Australia, and several US agencies moved against the DeepSeek app because it sends what users type to servers in China. An open model running on a company’s own hardware sends nothing anywhere.

Export controls target chips entering Chinese data centers. How they apply to a Chinese model file sitting on a server in Frankfurt is unsettled. With the new cooperation organization, Chinese open models now travel through diplomatic channels that no app-store ban touches.

What remains, once data movement is off the table, is the model itself: its provenance and lineage. Lineage runs through distillation, which is training a new model on the outputs of a stronger one. Western frontier labs forbid distillation in their terms of service, which cuts Western open-source AI model builders off from the best teachers. In contrast, Chinese labs distill the American frontier anyway and face little enforcement. Thinking Machines has disclosed that Inkling’s training drew on Moonshot’s Kimi K2.5. The chain now runs Western open models learning from Chinese ones, which learned from an American frontier that Western builders are barred from using.

Read through the trust-intelligence-power triangle, the shift is structural. The West competed on Intelligence, treating the best model as the prize. Self-hosted inference ends that scarcity, and advantage drains toward the layers that stay scarce: the Power under the model and the Trust the model carries. The West leads on Power and on the closed AI frontier. It has not governed the Trust layer of the open models that companies can own outright, and those are Chinese.

The response works across all three layers.

On Intelligence, compete for the self-hosted tier: legislation making training data fair use, and barring terms of service that forbid distillation, would let Western open models learn from the Western frontier at the source.

On Trust, govern provenance where models are distributed and bought: platform catalogs such as Microsoft’s Foundry Local can disclose what a model was distilled from and who modified it, with CAISI authoring the standard and binding directives keeping unvetted weights off federal systems.

On Power, move safeguards from the network to the model: self-hosted inference leaves little traffic to inspect, so evaluation and red-teaming of the weights become the checkpoint, work CAISI and the United Kingdom’s AI Security Institute already run.

For years, sovereign AI was governed at the central facilities: a data center you can locate, a chip you can license, a cloud account you can subpoena. A growing share of intelligence now runs past all three, on hardware that companies own precisely to keep their data and their learning to themselves. The motive is sound. The gap is that the models best suited to fill it carry a provenance the West has not measured.

Every organization will own its AI. The West should make sure that what gets owned is something it helped build.

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sources

• Judge orders OpenAI to produce 20M ChatGPT logs (Bloomberg Law)

• Nadella, “The Reverse Information Paradox” (essay)

• Nvidia RTX Spark and Windows AI runtime (Windows Blogs)

• Artificial Analysis Intelligence Index (leaderboard)

• Moonshot Kimi K3 closes gap with US rivals (Bloomberg)

• Alibaba previews open-weight Qwen3.8-Max (The Decoder)

• Xi promotes open-source AI at WAIC Shanghai (US News)

• World AI Cooperation Organization, 29 founding countries (Quartz)

• Thompson, “Who’s Afraid of Chinese Models?” (Stratechery)

• Thinking Machines releases Inkling open-weights model (SiliconANGLE)

• Karp: enterprises shifting to open-weight models (CNBC)

• Palantir + Nvidia Nemotron for sovereign environments (Business Wire)

• Germany asks Apple, Google to block DeepSeek app (CNBC)

• CAISI evaluation of DeepSeek finds shortcomings and risks (NIST)

• The TIP (trust-intelligence-power) stress-test framework (Sovereign Compute)

A two-stack world has two points of failure and two very different sets of values baked into the systems people ask for answers. Models carry the politics of their training data. A Chinese-origin model frames questions of governance and rights differently from an American one. A resilient world needs a plurality of models trained on local data and local loops: Falcon for the Emirati worldview, the American frontier labs for Washington’s, Qwen and DeepSeek for Beijing’s. A third pole, large, democratic, and aligned with neither bloc, would keep that plurality standing.

India is the only candidate with the scale. It ranks third on Stanford’s Global AI Vibrancy Index, behind the United States and China and up from seventh a year earlier on the strength of talent, research output, and a national program with intent behind it. It has 1.4 billion people, one of the world’s largest developer communities, and a foreign policy that refuses capture by either side.

Run India through the Trust-Intelligence-Power (TIP) scoring framework and the problem becomes precise. Of the six jurisdictions the framework scores, India is the only one without a structural asymmetry. Its problem instead is size. None of the three vertices rises above the 3.5 score. Every other AI pole has one weakness to fix; India has the whole triangle to grow, and growth requires the two inputs the USA-China duopoly poles gatekeep.

The first input is capital.

Four US hyperscalers plan roughly $725 billion in capital expenditure for 2026, and analysts project the total crosses $1 trillion in 2027. The IndiaAI Mission’s budget is $1.25 billion over five years. Sovereign Capital is India’s weakest Power sub-score at 2.5, and no domestic pool closes a gap of that width while each model generation demands more compute than the last.

The second input is compute.

Access to it is political. India faces no chip restrictions today, but that access rests on goodwill, with no binding commitment behind it. The record shows how goodwill gets priced. Washington’s tariffs on Indian goods reached 50 percent in 2025 over discounted Russian crude purchases, the highest rate among major US partners. The February 2026 deal cut them to 18 percent, reportedly in exchange for India winding down those purchases. The relief was paid for partly in barrels of foreign-policy autonomy.

The TIP framework prices the bind. India’s Trust score of 3.5 sits below the American and European 4.5 because non-alignment buys independence from both blocs but costs alignment credit with each. The property that makes a third pole worth building, an AI stack answering to neither Washington nor Beijing, is the property that starves it of inputs. The TIP stress tests say the same thing numerically. A tightening of export controls hits India’s Intelligence score harder than Europe’s, because Indian AI depends on American model and chip access. At the same time, in the redistribution scenarios, where hyperscaler capital diversifies away from concentrated hubs to reduce risk, India is the single largest beneficiary. In other words, India is the highest-beta jurisdiction to any crack in the duopoly, in both directions.

The problem is, India will not outspend anyone. Its potential route is capability per dollar: the efficiency path DeepSeek and Z.ai proved viable, domestic and pooled compute that waits on no license, open weights no directive can recall. What decides the outcome is capital that does not demand alignment as its entry price: Indian conglomerates, global investors backing applied and open AI, and Gulf co-investment structured as a partnership of equals. If India secures those inputs during the buildout, the 2030s get a third pole. If it does not, expect sovereignty to fragment downward, with dozens of states downloading open weights (mostly Chinese) and training them locally, buying through decentralization the resilience a third pole would have provided through scale.

Sources

Tom’s Hardware (FT compilation): hyperscaler capex plans reach $725 billion in 2026 – https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion

Yahoo Finance / Goldman Sachs: hyperscaler capex trajectory toward $1 trillion in 2027 – https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html

Press Information Bureau: IndiaAI Mission, Rs 10,372 crore over five years – https://www.pib.gov.in/PressReleasePage.aspx?PRID=2012375

White House: United States–India Joint Statement, February 2026 – https://www.whitehouse.gov/briefings-statements/2026/02/united-states-india-joint-statement/

Al Jazeera: US tariffs on India cut from 50 percent to 18 percent – https://www.aljazeera.com/economy/2026/2/2/trump-to-slash-us-tariffs-on-india-from-50-percent-to-18-percent

Stanford HAI: Global AI Vibrancy Tool, 2025 edition – https://hai.stanford.edu/ai-index

Press Information Bureau: GSI rare earth ore resources and REPM manufacturing scheme – https://www.pib.gov.in/PressNoteDetails.aspx?NoteId=157165&ModuleId=3

Bloomberg: US drafts rules requiring approval for global AI chip shipments – https://www.bloomberg.com/news/articles/2026-03-05/us-drafts-rules-for-sweeping-power-over-nvidia-s-global-sales

Discovery Alert / Indian trade data: 93 percent of permanent magnet imports sourced from China, FY2024-25 – https://discoveryalert.com.au/rare-earth-india-china-dependence-processing-magnets-2026/

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

A nation’s standing in the AI race is decided below the model. It rises when a state can run frontier capability on ample power, compute, and data it controls. Most cannot.

Every government can now get its hands on a frontier-grade model. Open weights put one within reach of any state that wants it. The contest that now sorts competitive nations from dependent ones is whether a state can run frontier AI as its own, on its own compute, its own power, its own data, under its own control. Most cannot. And that gap is widening.

Open weights became a working default in 2026. The strongest open model, which is Chinese, now ranks fourth overall on the Artificial Analysis intelligence index, with only two proprietary systems above it that anyone can actually call after the US ban, and it reaches that level at a fraction of their price. On real-world task scores, it runs on par with the American frontier model most enterprises use, and within a few points of the leaders on coding and reasoning. When a capability that once cost a fortune runs for the price of the electricity behind it, its worth as a product falls sharply toward that floor.

Still, full commoditization arrives only when enough tier-one open models exist that no single vendor can lock a nation in, so it can switch the day a model closes or disappears, and when moving between them is painless, enabling a nation to carry its own work across without loss.

That specific work is the “loop”: the live, self-improving system a country builds on top of a base model. It’s the magic formula defining a country, the thing that turns its own data and experience into capability that sharpens the more it is used. And portability is the test. If switching means starting over, the model still owns the nation. The day the switch is clean, the model becomes a commodity, and the advantage moves somewhere else.

So the question is, can a nation move its loop to a new model today? Mostly yes. The valuable material a country builds up, its records, the history of how people used the system, its own tests for what counts as a good answer (in a democracy and elsewhere), and its store of national knowledge, does not belong to any one model. It sits in the country’s own hands and can be reused, if it’s been collected. What does not carry over is the expensive training of the model itself, because training done on top of one model cannot simply be pasted onto another. Instead, the country rebuilds that skill on the new model, feeding in the same records it kept and checking the result against its own tests until it matches.

How hard is that retraining? It’s expensive, though it is routine engineering rather than fresh invention, closer to repaving a road than designing the car. The effort grows with how much a country had customized: a light touch is quick, while years of deep tuning and carefully fitted behavior take time to redo and re-test. The winning nations have been customizing their models heavily, with the cost of retraining equal to computing power and skilled people. So the difficulty does not close the door on portability, but puts a steep price on it. A country with the data, the engineers, and the compute can switch whenever it needs to. A country short of any one of those is stuck.

So if the model itself is becoming a commodity, the next question is where sovereign advantage goes instead. That is what my Trust-Intelligence-Power framework (TIP) was built to answer. The interactive TIP triangle reads any sovereign strategy down three layers. Trust is who controls the rules, the data, and the alliances that decide whether you can even get the parts. Intelligence is who controls the models, the data, and the talent. Power is who controls the physical base: the grid, the firm electricity, the chips, the capital, the security of the buildings themselves. The three do not trade against each other. Strength in one cannot cover a hole in another, and the hole is where sovereignty breaks.

When grafting the commoditization question into the interactive TIP triangle, the answer is immediate. When the model goes free, Intelligence loses the scarcity that made it a moat, and advantage shifts to the layers that remain scarce: Power, and the parts of Trust and Intelligence that cannot be downloaded. The frontier labs feel it already and are remaking themselves into infrastructure and application companies. The prize has moved from the weights to everything the weights still need.

The moat moves. Advantage relocates from the Intelligence corner of the framework to the Power corner.

For nations, that shift produces an inversion: A government can download a frontier model for nothing, but it still cannot adapt it, secure it, or serve it to millions of citizens without chips, firm power, and capital it may not have. The gap drops to the layer that takes the longest to build and costs the most to own.

That reading is bullish for the Gulf

Saudi Arabia and the UAE have put serious sovereign capital behind a single idea: turn their cheapest asset, firm low-cost energy, into compute at scale. HUMAIN, the Stargate campus rising in Abu Dhabi, and partnerships with the largest American players carry that bet, and in a world where the model commoditizes, it is the correctly placed one. The scarce input is shifting toward the layer the Gulf can supply in abundance, and the imported-Intelligence dependency my framework scores as the region’s fragility shrinks as the models (closed included) go cheap.

It does not vanish, though. Two dependencies remain, and both sit one layer below the model. The first is capital, which the Gulf has in depth. The second is the compute itself: the advanced chips still come from the United States, and they arrive by conditioned export license rather than as of right. Saudi Arabia and the UAE are the only two states outside the top US tier cleared to buy Nvidia’s leading Blackwell parts, under security and reporting terms that Washington monitors on an ongoing basis. And at the very top of the AI reasoning curve, the hardest work still tends to run on closed American models, rather than on systems the region owns. So commoditization vindicates the energy bet and narrows the model dependency, while moving what remains of it onto the chip and the export relationship the Gulf has worked to secure.

The Harder Case

AI frontier access is being gated at the source, by the governments that produce it.

This month the United States pulled one lab's most capable models offline worldwide under an export order, and asked another to stagger its newest release to a handful of government-approved customers, one approval at a time, under a new executive order that sets a classified bar for the most capable models.

The open frontier, meanwhile, stays concentrated in a few labs, much of it Chinese and already contested over intellectual property and trust. So a nation leaning on the best closed models can have them rationed in Washington, and a nation leaning on the best open ones depends on a supply that may re-price or re-close. Either way, the base layer is borrowed. The nations that own the loop and the energy beneath it carry on, because the loop transfers to whatever model remains and the energy is theirs regardless.

Then comes the part that decides whether democracies can sustain any of this.

Nadella’s recent warning to firms was that if a handful of models capture most of the value, the political economy will not tolerate it, much as an earlier wave of globalization hollowed out industrial regions while national totals still looked healthy. Inside a company, that is a question of competitiveness. Inside a democracy, it becomes a question of consent. The evidence already sat in my “Three Models” piece: the spreading data-center backlash, the moratorium proposals, the climbing residential power bills, the local recall campaigns against approved projects.

So the same commoditization cuts two ways at once. It frees a nation from depending on anyone else’s model, but it concentrates the real value in whoever owns the energy and the compute, while straining the consent a democracy needs to build them. Owning a model is the cheap part, and on its own it is now close to worthless. Sovereignty is the loop, the loop runs on power, and a democracy has to share the proceeds to keep the right to build the thing that produces them. For an investor or a minister, three questions now settle the matter: do you own a loop, do you own the power and compute beneath it, and are you spreading the gains fast enough to sustain it.

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

For three years, the AI story was an equity story. Companies with enormous cash flows bought chips, and investors rewarded the spending. That phase is ending. The six largest US hyperscalers are on track to spend around $700 billion on capital projects in 2026, by Moody’s estimate, and their combined capex is set to cross their operating cash flow around the third quarter of this year, according to Epoch AI. When spending exceeds the cash coming in, the balance gets funded with debt. US technology issuers sold roughly $200 billion of bonds in 2025, and AI-infrastructure borrowing alone reached about $121 billion, more than four times its five-year average, with most of it raised in the final quarter. In parallel, another $120 billion of data center borrowing was moved into off-balance-sheet vehicles by Oracle, Meta, xAI, and CoreWeave. The buildout became a credit story while most coverage stayed fixed on stock prices.

The duration mismatch: a chip’s frontier earning power collapses years before the debt raised against it matures.

Here is the mechanism that matters. The bonds and project loans behind this run long, often 15 to 30 years, matched to the 20-to-30-year life of a data center building. The chips inside that building are a different asset. Nvidia has moved toward a roughly annual release cadence, and the resale and rental value of last cycle’s hardware falls accordingly; H100 rental rates have dropped on the order of 70 to 90 percent since 2023. The hardware that secures much of this debt loses its frontier earning power in two to three years, while the debt itself sits on the books for decades.

The hyperscalers bridge that gap with an accounting choice. Amazon, Microsoft, and Alphabet have all extended the assumed useful life of their servers from three or four years to six, a change that lowers reported depreciation across the group by roughly $18 billion a year and keeps earnings, the thing that services the debt, looking healthy. Michael Burry flagged this in November, estimating $176 billion of understated depreciation across the industry through 2028 and comparing the setup to Cisco at the top of the dot-com cycle. The industry’s defense is the cascade: a chip earns its keep at the frontier for a year or two, then drops to cheaper inference, then to lighter serving work, generating revenue across all six years. Satya Nadella has been candid about the tension, saying he did not want to be stuck with four or five years of depreciation on a single generation.

That defense rests on inference staying a paid business as the chip ages. This is where GLM-5.2 enters as a financial event rather than a technical one. Z.ai released it on June 16 under an MIT license (a very permissive software license originating at the Massachusetts Institute of Technology (MIT) in the late 1980s that puts few restrictions on reuse and has high license compatibility), frontier-class on long-horizon coding work, priced through providers at around $4.40 per million output tokens against $25 for Claude Opus 4.8, which is now comparable in output.

A capable open model that anyone can run compresses the price of exactly the inference that is supposed to fill years two through six of a GPU’s life and repay the loans raised against it. The cascade still works mechanically: the chip can still do the work, but it is the margin on that work that comes under pressure, and the margin is what services the debt.

So the open-weight release reads as a credit event before it reads as a competitive one. It does not touch the equity first. Free cash flow at Microsoft or Google is large enough that the stock can hold even as a non-cash depreciation argument plays out, which is why the share price is the wrong gauge. The pressure shows up in credit. Issuer spreads for the most exposed names widened through late 2025 and into 2026, with Oracle and Meta the clearest cases, and Oracle now faces a bondholder lawsuit tied to its buildout. It shows up faster in the collateral. CoreWeave pioneered borrowing against GPUs, pledging the chips and the customer contracts as security; its debt has grown past $21 billion, and one earlier facility began repayment in January at around 11 percent as the collateral’s market value was falling.

Two protections sit behind this debt, and a falling price for inference weakens both at once. The first is the hardware pledged as collateral. If a borrower defaults, the lender can seize the GPUs and sell them, but the resale market for used data-center GPUs is thin, and prices slide as each new Nvidia generation ships, so the amount actually recovered is uncertain and shrinking. The second is the customer contract. Much of this borrowing is repaid through take-or-pay agreements, the same structure that underpins gas pipelines and LNG terminals: in this case, a large customer commits to pay for a fixed block of compute whether or not it uses it. Because the payment is contractual, the lender is in effect underwriting the customer’s promise to pay, and the chips serve as a backstop. Rating agencies have followed that logic, grading these deals on the strength of the hyperscaler standing behind the contract. A cheaper, capable open model puts pressure on both protections together. When the same capacity can be served for a fraction of the price, an above-market commitment becomes harder to honor, and the collateral securing it is worth less in the very moment the contract is tested.

For an investor or a lender, the watch list follows directly: Track the spread between the most AI-exposed issuers and the investment-grade index, not the equity. Track the GPU-collateralized securitizations that JPMorgan expects to reach $30 to $40 billion a year. Track the health of the take-or-pay anchors, OpenAI’s commitments to Oracle and CoreWeave chief among them. The buildout was underwritten on the belief that producing intelligence would stay expensive. A lab in Beijing has begun testing whether that belief survives, and the answer will print in credit before it prints in any stock.

There is also a larger question sitting behind this one. When intelligence becomes close to free and AI-embedded robots take over much of the work, money stops being a claim on human labor and becomes a claim on the two things still worth rationing, energy and compute. That shift likely runs in two phases, and what happens to money is hard to call in both. In the first, energy and compute are still scarce, and they become the anchor that human labor used to be. In the second, energy and compute stop being scarce at all, and the question turns to what money is a claim on when almost nothing is rationed. The second phase is the harder one to foresee, and it is where the meaning of money comes loose. I may take this up in a future piece.

This newsletter represents my personal analysis and does not reflect the views of any organization with which I am affiliated.

Sources

Hyperscaler 2026 capex (~$700B, Moody’s) and GPU-securitization outlook (JPMorgan). https://www.theaiconsultingnetwork.com/blog/ai-data-center-gpu-debt-financing-insurance-cre-investors-2026

Capex set to cross operating cash flow in Q3 2026 (Epoch AI). https://fourweekmba.com/hyperscaler-capex-exceeds-cash-flow-crossover/

2025 tech bond issuance (~$200B) and Oracle spread pressure (Barclays Private Bank). https://privatebank.barclays.com/insights/market-perspectives-february-02-2026/ai-prompts-the-big-corporate-bond-boom/

AI-infrastructure debt ~$121B and spread widening for Oracle and Meta. https://www.ainvest.com/news/ai-data-center-boom-structural-shift-corporate-debt-credit-markets-2602/

~$120B of off-balance-sheet SPV financing (Meta Hyperion, Oracle Abilene). https://www.cryptopolitan.com/oracle-xai-meta-coreweave-off-balance-sheet/

GPU useful-life extension to six years, Burry’s $176B estimate, and the Nadella remark (CNBC). https://www.cnbc.com/2025/11/14/ai-gpu-depreciation-coreweave-nvidia-michael-burry.html

GPU-collateralized lending and CoreWeave’s debt structure (Quartz). https://qz.com/gpu-collateralized-debt-ai-neocloud-coreweave-financing-risks-050526

AI data-center financing structures and collateral-valuation risk (Quinn Emanuel). https://www.quinnemanuel.com/the-firm/publications/client-alert-emerging-litigation-risks-in-financing-ai-data-centers-boom/

Oracle sued by bondholders over losses tied to the AI buildout (Reuters). https://www.reuters.com/legal/oracle-sued-by-bondholders-over-losses-tied-ai-buildout-2026-01-14/

GLM-5.2 open weights, MIT license, and pricing (Z.ai). https://z.ai/blog/glm-5.2

Source: Optro (formerly AuditBoard), 2026 Risk Intelligence Report, “The AI Oversight Gap,” Mar 2026; n = 800+ global GRC and IT decision-makers.

That same structure now operates at the level of the state, where the stakes move from a compliance fine to a strategic miscalculation. A national AI strategy is a declaration made where leaders can see chips secured, data centers permitted, grid reinforced, and export licenses issued. But the AI models actually executing inside those buildings sit one layer down, where no government keeps a map: A country can hold every visible instrument of sovereignty and still run its intelligence on foundations, on AI models, which it did not build and does not track.

The United States is the proof case because it tried the hardest. Washington has run four rounds of export controls since 2022, covering two dozen categories of semiconductor equipment and well over a hundred entities, with the explicit goal of denying China the compute to reach the AI frontier. The wall at the hardware layer is real. The entry at the software layer is still open.

A widely repeated figure held that 80 percent of US AI startups run Chinese open-source models. The number is wrong, and the correction is the more interesting fact.

Martin Casado of Andreessen Horowitz, the original source, clarified that he meant 80 percent of the 20 to 30 percent of startups that build on open source at all, which nets to roughly 16 to 24 percent of the startups his firm sees. The inflated version traveled anyway. It reached the U.S.-China Economic and Security Review Commission, the congressional body that reports to Congress on exactly this question, which wrote in its March 2026 report that "some estimates suggest that around 80 percent of U.S. AI startups now use Chinese open-source models." A venture capitalist's offhand pitch-meeting estimate became a hedged line in a national-security assessment, uncorrected. That is the provenance gap stated openly: the arm of government built to track Chinese technological dependence borrowed a number, and borrowed the wrong one.

Sources: Martin Casado / a16z, correction posted Nov 2025 (“20–30% use open source; of those ~80% Chinese-based; closer to 16–24%”); USCC report, Mar 24 2026; Reuters, Mar 24 2026.

Even the corrected number is still remarkable. A fifth of new US ventures, inside the regime built to deny China AI capability, choose to build on weights that come out of China. The cohort that does so clusters where iteration speed decides outcomes. Shayne Longpre, an MIT researcher who tracks model release patterns, points out that Chinese labs ship open models on a weekly or biweekly cadence against the semi-annual rhythm of American labs, and that volume drives familiarity and adoption. On OpenRouter, one of the largest model-routing platforms, Chinese models reached about 61 percent of token consumption among the top ten models in February 2026, and the share has widened since.

Sources: Shayne Longpre / MIT, via implicator.ai (Nov 2025); OpenRouter token data via digitalinasia (Feb 2026); Hugging Face Spring 2026 State of Open Source (Qwen cumulative downloads surpassing Llama).

Visible-layer figures are slow-moving; provenance-layer figures verified at time of writing, June 2026.

The obvious objection is that open weights are inert. When a startup runs Qwen on a server in Virginia, no data travels to Hangzhou; the weights are a file. But the exposure was never exfiltration. It runs along two other lines. First, behavior travels with the weights: documented value-aligned outputs that activate on sensitive topics arrive embedded in the model and surface in the application built on top, a point the US-China commission flagged directly. Second, roadmap dependency: when a fifth of a startup base standardizes on a foreign release cadence, the national innovation pipeline inherits another state’s schedule and priorities.

Sources: USCC report and public remarks by Vice Chair Michael Kuiken, Mar 2026; implicator.ai analysis of embedded model behavior (Nov 2025).

Justin Lin, who leads Alibaba’s Qwen models, has put the odds of a Chinese lab leapfrogging the frontier in the next three to five years below 20 percent. That is precisely why distribution, not capability, is the play. An open model does not need to be the best to become the default, and the default is what shapes a generation of downstream applications. Winning the substrate is the rational move for a player who expects to lose the frontier.

Source: Bloomberg interview with Justin Lin, Alibaba Qwen (Jan 2026).

A jurisdiction can score perfectly on physical compute and run its inference on foreign foundations, and nothing on the official dashboard will show the divergence.

Two practical implications follow. For policymakers: map provenance before legislating the buildout. An export-control regime aimed only at chips leaves the software layer unwatched, which is where the dependence is actually accumulating. For investors: when foundation models are free, the value moves up a layer, to deployment, integration, and the proprietary data and physical systems built on top of the weights. That is where returns concentrate, and where provenance matters least.

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Governments from Paris to Riyadh to New Delhi have committed tens of billions to domestic compute. The IEA expects data-centre electricity demand to reach roughly 945 TWh by 2030, slightly above Japan’s entire consumption today. The siting decisions being made right now will bind national capability for thirty years.

Almost every tool guiding those decisions measures the world at rest. Readiness indices benchmark what exists today: compute stock, talent pools, regulatory quality, capital. They are snapshots. They tell you who leads. But they do not describe what happens to that lead when a strait closes, an export control lands, or a drone finds a data centre.

In the first quarter of 2026, that stopped being hypothetical. Military action in the Middle East began on February 28, the Strait of Hormuz moved to a de facto closure within days, and Iranian drones struck operational data centres in the Gulf.

Today I am publishing what I have been building and testing privately since January: the TIP Triangle, a framework that scores sovereign AI readiness and then subjects the scores to stress. The interactive model is live at the link below. You can run the scenarios, change my assumptions, and break it.

[→ Open the interactive TIP Triangle demo]

The framework scores six jurisdictions (United States, Gulf, India, ASEAN, European Union, China) on a 1 to 5 ordinal scale across three layers. Trust covers institutional reliability and geopolitical alignment: can partners depend on this jurisdiction’s commitments, data regimes, and export behaviour. Intelligence covers talent, research depth, and model capability across five subdimensions. Power is the physical layer, and the reason the framework exists: grid headroom, energy diversity, cooling, physical security, connectivity, sovereign capital, permitting. Most assessments treat this layer as a footnote. It is the binding constraint on everything above it.

Every score is documented. Where hard data exists, it is cited in my methodology paper (available to premium subscribers). Where structured judgment fills a gap, the methodology says so. The scales are ordinal, so the outputs are directional and comparative. A 4.0 beats a 3.0; it is never “33 percent better.” That kind of precision on questions like these, should always be questioned.

The baseline rendered above already paints strategic realities. The US runs strong on every layer, with one exception buried inside Power: permitting, scored 1.5, the worst of any jurisdiction except the EU. China pairs the deepest build capacity in the set (a Power composite of 4.07, the highest of the six) with the lowest external trust (2.0). The Gulf holds the global maximum on sovereign capital and the fastest permitting anywhere, alongside concentrated exposure in energy diversity, cooling, and physical security.

That last profile matters, because fragility lives in the gap between a jurisdiction’s strongest and weakest layer. Which is where the framework does its real work.

The core of the TIP Triangle is a cascade engine: 41 documented transmission rules that propagate a shock across layers and across borders. Shock one dimension of one jurisdiction and the engine follows the consequences along causal chains the methodology maps explicitly: gas-supply cascades, the water-cooling trilemma, insurance freezes, capital redistribution, cable severance, sovereign-capital diversion.

The signature scenario is a four-stage Strait of Hormuz escalation, from tanker harassment to prolonged closure. Run it and the Gulf’s Power composite falls from 3.29 to 3.17 to 3.03 to 2.60 to 2.30. Then watch what a single composite number hides. The engine propagates the Power shock into Gulf Trust (investment review, bilateral fracture) and Gulf Intelligence (construction freeze, talent flight), then across borders into India, ASEAN, China, the US, and the EU. Eight cascades fire at Stage 3 alone.

The model separates five vulnerability types that composite indices collapse together: price exposure, physical disruption, strategic dependence, continuity of operations, and ability to keep scaling. Under Stage 3, Gulf physical disruption deteriorates to 4.7 out of 5 while price exposure moves only to 3.3. Those are different problems with different fixes, and a ranking cannot show you the difference. A jurisdiction can absorb a price shock and still lose its ability to scale. Another can hold institutional trust and prove physically fragile.

The model also declines to pretend that weighting is neutral. It ships with four selectable Power-layer weighting modes, including a system-breaker mode built on a simple observation: a collapse in physical security ends operations, while a slow permitting regime only delays them. Some constraints break systems and others degrade performance, and a framework should let you test both live. Switch modes and the Gulf’s Power composite moves from 3.29 to 2.84. The choice of weighting is an analytical claim, and the model makes you state it.

I finished the first implementation in early March 2026. The conflict’s first weeks coincided with the disruption the framework was designed to model, and I want to be precise about what that establishes. The framework had scored Gulf physical security at 2.5 before hostilities reached the region’s digital infrastructure. Iranian drones then struck operational data centres in the UAE and Bahrain, and on March 31 the IRGC named US technology firms operating in the Gulf as legitimate targets. The cascade model had mapped Gulf Power stress propagating into Trust and Intelligence; all three movements materialised within weeks. Brent started the year at $61 and closed the first quarter at $118 (EIA), the largest inflation-adjusted quarterly increase in the series going back to 1988, and the price-transmission channels the framework maps operated in real time.

That gives me directional consistency with a live stress episode.

The methodology has also been through independent expert review. The critique drove three changes you now see in the model: the weighting modes, the vulnerability taxonomy, and the explicit separation of static readiness from dynamic resilience.

The interactive demo gives you the full Hormuz escalation sequence, live sliders for every layer and Power subdimension, and the vulnerability-type analysis. The complete model (all scenarios, the four weighting modes, the transmission channels, the fragility analysis, the editable 41-rule matrix, and the custom-jurisdiction builder) is available to Sovereign Compute premium subscribers, and includes a custom-jurisdiction builder with four archetypes (allied importer, energy exporter, connectivity hub, non-aligned), so you can create and place your own jurisdiction / country inside the system and stress-test it.

[→ Open the interactive TIP Triangle demo]

Three requests. Run your own scenario and tell me where the model gets it wrong; the cascade weights are editable for exactly that reason. If you work in government, infrastructure, or capital allocation and want a jurisdiction modelled properly, reply to this email. And if the equal-weighting default does not suit you, switch modes and see how much your conclusion rests on pre-existing assumptions.

A longer report, The Power Behind the Model, is in preparation: twenty pages on the energy layer of sovereign AI, built on this framework and available to premium subscribers.

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

This is the backdrop against which Microsoft and Nvidia moved the location of AI this month. At Build on June 2, Microsoft introduced a Windows system that runs large AI models directly on the machine, and it paired that with Nvidia’s RTX Spark chip, which ships this fall in laptops from Surface, Dell, ASUS, HP, Lenovo and MSI. A model running on the device needs no data center and no internet connection. There is no server-side transcript, because there is no server in the loop. The exchange happens on the hardware and stays there.

Artificial Analysis Intelligence Index, June 2026. Closed frontier models run only in the cloud; open-weight models run on local hardware.

The consequence turns on which models can run that way. The systems at the closed frontier, Anthropic’s Claude, OpenAI’s flagship GPT, Google’s Gemini, are not available for download; they stay in the cloud behind an account, and they still carry most production and customer-facing work. The models that run locally are open-weight, their files public and free to install. By independent 2026 benchmarks, the leading open-weight models now come from Chinese labs: DeepSeek, Moonshot’s Kimi, Zhipu’s GLM and Alibaba’s Qwen, ahead of Meta’s Llama and the smaller open releases from Google and OpenAI. Two years ago Llama set the open-weight pace; the lead has since moved to Beijing’s labs.

This is not confined to private users. Regulated industries are the natural buyers of local inference, since finance, healthcare, legal and government-adjacent work increasingly cannot send sensitive data to an external model under GDPR, HIPAA and the EU AI Act. The survey evidence points the same way, with 41 percent of enterprises planning to expand open-weight use and another 41 percent ready to switch from closed models once performance reaches parity. On-device hardware carries that local layer from the server room to the endpoint, where the most sensitive material already lives.

The strategic turn is what this does to the West’s main instrument against Chinese models. The restrictions in place apply to the hosted service. In 2025 Germany’s Berlin data protection authority concluded that DeepSeek’s app unlawfully transfers German users’ personal data to servers in China, where national security law gives the state access, and asked Apple and Google to remove it from their stores. Italy’s regulator had already ordered the company to stop processing Italian users’ data, and Australia and several US agencies barred the app from government devices. Each action rests on the same fact: the app sends what users type to infrastructure in China. An open-weight model downloaded and run on a local machine sends nothing anywhere. The transfer these bans were written to stop does not occur when the weights execute on a closed device, so the policy tool the West has relied on does not reach the on-device case. What remains is provenance. The training, behavior and content rules of these models are decided in China, and a growing share of private and regulated inference inside Western institutions can run on a strategic competitor’s model with no traffic for a regulator to inspect.

For two years sovereign AI has been governed at the central facility: the data center that can be located, the chips that can be licensed, the cloud account that can be subpoenaed. On-device inference moves a real share of capability outside all three. It offers a way to run AI on sensitive material with no retrievable trail, which is the OpenAI ruling read in reverse. It also gives the question of national AI capability a harder form, since the most capable portable models are trained by a strategic competitor and travel on consumer hardware that no border check and no export license reaches. The campuses keep rising, and they still matter. The device has become the second place sovereignty is decided, and it is the place current policy was not built to govern.

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sources

OpenAI ordered to produce 20 million ChatGPT logs (Bloomberg Law) https://news.bloomberglaw.com/ip-law/openai-must-turn-over-20-million-chatgpt-logs-judge-affirms

Windows PCs accelerated by Nvidia RTX Spark (Microsoft) https://blogs.windows.com/windowsexperience/2026/05/31/introducing-a-powerful-new-chapter-for-windows-pcs-accelerated-by-nvidia-rtx-spark/

Model intelligence rankings (Artificial Analysis) https://artificialanalysis.ai/leaderboards/models

Germany asks Apple and Google to block DeepSeek (CNBC) https://www.cnbc.com/2025/06/27/germany-tells-apple-google-to-block-deepseek-ai-app.html

DeepSeek restrictions by country (Euronews) https://www.euronews.com/next/2025/06/27/german-official-urges-apple-and-google-to-ban-ai-company-deepseek-citing-privacy-concerns

Enterprise open-weight adoption trends (a16z, via Hatchworks) https://hatchworks.com/blog/gen-ai/open-source-vs-closed-llms-guide/

Europe pays to discard its cheapest power, then builds an AI strategy around the energy it says it lacks.

In 2025, wholesale electricity prices in France, Germany, the Netherlands and Spain dropped below zero in about one hour out of every sixteen, up from one in twenty-five a year earlier, according to the IEA. Germany alone recorded 573 hours of negative prices, a 25% rise. A negative price means there is more electricity on the market than the system can use, so producers that keep generating are effectively paying to offload their power rather than being paid for it. As this publishes, a June heatwave has pushed solar output high enough to send prices below zero again across parts of the continent.

Throwing that power away carries a significant bill. Aurora Energy Research puts the cost of curtailing renewables across Europe at €8.9 billion in 2024, with more than 10 terawatt-hours of clean generation discarded and German solar curtailment up 97% in a single year. Britain spent £1.47 billion paying wind farms to stop and gas plants to start. This is the cheapest electricity Europe produces, and the continent pays twice: once to build the turbines, again to switch them off.

The waste lowers no one’s bill because the power is stranded. Cheap renewable supply sits in the wrong place at the wrong moment, behind a grid that cannot move it. Average EU wholesale power ran near twice US levels in 2025 even with the surplus. The European Commission now estimates €1.2 trillion in grid investment is needed by 2040, up from a €584 billion figure for 2030. By one survey of sixteen countries, around 1,700 GW of renewable projects wait in connection queues, more than three times what the EU’s 2030 targets require, and cross-border lines average 5.6 years to permit. In the Netherlands, roughly 12,000 companies sit in line for a connection.

The capital is arriving, and where it lands settles the argument. At the Choose France summit in late May, SoftBank committed up to €75 billion to develop 5 GW of AI data center capacity in France, its largest AI infrastructure investment in Europe, with €45 billion and 3.1 GW in a first phase by 2031. Ardian and its platform Verne added up to €5 billion for a 500 MW campus near Paris. Both named the same draw: France’s low-carbon nuclear grid and the firm, steady power it supplies, with Schneider Electric among the industrial partners. The money goes where the power is firm and the connection clears.

That routing exposes a split inside Europe.

France can take gigawatts because nuclear baseload holds its prices steady and its grid clears. The curtailment sits in Germany, Spain, the Netherlands and the United Kingdom, where the surplus is wind and solar that the grid cannot move. Capital follows firm power to the market that already solved the question, while the markets richest in cheap clean electricity watch it pass. Brussels’ own InvestAI programme, about €20 billion for four or five training gigafactories, is a fraction of what a single foreign commitment now brings to France alone.

The assets Europe actually holds point in a different direction. ASML remains the only company on earth that builds the lithography machines advanced chips require, a chokepoint no rival can route around. Open-weight, data-resident models, Mistral among them, win regulated buyers like the French armed forces and HSBC, who pay for sovereignty even when a benchmark lags. And the stranded clean power is a siting advantage for compute that can run when the wind blows and pause when it drops. Batch inference and training checkpointing tolerate that rhythm; latency-critical work does not. The Gulf understood the principle early and priced energy as the foundation of its compute build. Europe sits on a comparable surplus, but it is not leveraging in the same way.

The opening is narrow and conditional. Flexible compute can monetise wasted power only where it co-locates behind the constraint, and only if connection and permitting reform arrives. That reform is the variable worth watching. For investors, the value accrues to whoever pairs interruptible compute with curtailment hotspots, to storage that captures the midday glut, and to the connection queue itself. Europe’s compute future runs through its grid, and the grid is the slowest thing it builds.

Sources

IEA Electricity 2026 (negative-price share, EU vs US wholesale): [link] Aurora Energy Research via Clean Energy Wire (€8.9bn curtailment, 2024): [link] Euronews (UK curtailment cost; heatwave negative prices): [link]

European Grids Package via GLOBSEC (€1.2tn by 2040): [link] European Commission / EuroHPC (InvestAI gigafactories): [link]

SoftBank (€75bn / 5 GW France commitment, May 2026): [link] DataCenterDynamics (Ardian / Verne €5bn / 500 MW): [link] Ember European Electricity Review 2026: [link]

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

The evidence is there. Materials Nexus designed a rare-earth-free magnet on an AI platform in 2024 at roughly 200 times the pace of conventional methods. The University of New Hampshire built an AI database of 67,573 magnetic materials and identified 25 previously unrecognized compounds that hold their properties at elevated temperatures (Nature Communications, 2025). Niron Magnetics, with DOE backing going back to a $17.5M ARPA-E grant in 2022, broke ground in September 2025 on a 1,500 ton/year iron-nitride magnet plant in Minnesota, operational in early 2027, and in March 2026 announced a second, $1.8B, 10,000 ton/year facility.

That race is real. It is also slower than the alternative race China is winning.

AI compresses one stage of substitution and leaves the others untouched. Discovery moves from decades to months. Qualification cycles for permanent magnets in industrial motors, defense, and aerospace still run five to ten years due to safety-critical regulatory requirements in defense and aviation. Scale takes longer. Niron’s 1,500 ton/year Plant 1 in 2027, against Chinese NdFeB output of roughly 260,000 tons in 2024, is well under one percent of Chinese capacity. The planned second plant would bring the U.S. total to 11,500 tons by the end of the decade. Replacing the Chinese magnet supply chain is a fifteen-to-twenty-year program, even with AI on the discovery side. China built it in three decades. Unwinding it will take comparable time.

This timeline asymmetry is what the United States has been paying for since 2025. After Beijing imposed export controls on seven heavy rare earths in April 2025 (dysprosium and terbium prices roughly tripled in Western markets) and extended jurisdiction extraterritorially in October to any foreign product containing 0.1% Chinese rare earth content, Washington negotiated a truce at Busan on October 30, 2025. Per the White House Fact Sheet, China suspended the October 9 controls for one year and issued general licenses for rare earths, gallium, germanium, antimony, and graphite. In exchange, the United States cut fentanyl-related tariffs from 20% to 10%, paused a threatened 100% additional tariff, suspended a sanctions-affiliate rule for a year, and extended certain Section 301 exclusions. The deal expires around October 2026.

The H200 chips were supposed to be Washington’s other card. Trump authorized H200 sales to China on December 8, 2025, with a 25% revenue share to the U.S. Treasury, and licenses were issued in January 2026 to around ten Chinese firms. As of mid-May 2026, no H200 has shipped. Beijing has declined to approve purchases, and the Trump-Xi Beijing summit in May ended with no movement on chip exports. The reason sits on the demand side. DeepSeek released V4 on April 24, 2026, with day-zero inference support on Huawei’s Ascend 950PR platform, having granted Huawei pre-release kernel access. Alibaba’s T-Head GPUs reached scaled mass production. ByteDance lifted 2026 AI capex to roughly $30B with a rising share going to domestic chipmakers. Nvidia’s footprint in China’s frontier AI datacenter deployments has fallen to effectively zero from approximately 95% before export controls tightened. The H200 ultimately carried far less strategic leverage than Washington assumed.

The early evidence is of two parallel sovereign stacks forming at different speeds. The United States is using AI to substitute the material layer, with first commercial capacity in 2027 and meaningful scale in the 2030s. China is using policy and capital to substitute the chip layer, with frontier models already running in production on domestic chips in 2026. Neither has solved the problem, but both have demonstrated the problem is solvable on a sub-fifteen-year timeline.

For investors and sovereign AI buyers, the read is precise. The transition window is roughly the next decade. During it, U.S. policy is constrained by rare earth dependency, and Chinese policy is constrained by chip dependency only at the leading edge. The trade Washington offered was a card Beijing chose not to play, because the cost of waiting was lower than the cost of accepting U.S. routing through a 25% surcharge. The currency of decoupling is shifting from chips to materials. The side that closes its gap first sets the terms of the next decade.

Sources

• IEA, Rare Earth Elements analysis and critical minerals commentary (2025–2026)

• White House Fact Sheet on the Trump-Xi trade deal (November 1, 2025)

• China Briefing, Trump-Xi Meeting Outcomes and Implications (November 2025)

• CSIS, China’s New Rare Earth and Magnet Restrictions Threaten U.S. Defense Supply Chains (October 2025)

• CNBC, U.S. clears H200 chip sales to 10 China firms (May 14, 2026)

• Artificial Intelligence News, The Nvidia H200 China deal survived the Trump-Xi summit (May 2026)

• Nvidia Form 10-Q, Q1 FY2026

• Itani, Zhang, Zang, The northeast materials database for magnetic materials, Nature Communications 16 (2025), DOI: 10.1038/s41467-025-64458-z

• Beeson et al., C16 Phase High Entropy Borides, Advanced Materials (2025), DOI: 10.1002/adma.202516135

• Materials Nexus rare-earth-free magnet announcement (June 2024)

• Niron Magnetics Plant 1 groundbreaking (September 2025); Wood EPCM contract (January 2026); $1.8B second plant site selection (March 2026)

• Persistence Market Research and IDTechEx on permanent magnet market sizing (2025)

• Strategic Metals Invest retail benchmark prices, dysprosium and terbium (May 19–20, 2026)

• ChinaTalk, DeepSeek V4 (April 2026); Reuters on Huawei Ascend day-zero V4 support (April 24, 2026)

Sovereign Compute is an independent publication. Views expressed are those of the author and do not represent the views of any affiliated organization.

The coverage of that framework has centred on the $100 billion in frozen Russian assets, the Ukrainian army cap, and the territorial clauses around Donetsk. Four other clauses describe something different. Provision 12 sets up a Ukraine Development Fund to invest in “fast-growing industries, including technology, data centers, and artificial intelligence.” Provision 13 commits the United States to a “long-term economic cooperation agreement” with Russia in “energy, natural resources, infrastructure, artificial intelligence, data centers, rare earth metal extraction projects in the Arctic, and other mutually beneficial corporate opportunities.” Provision 14 funds a US-Russia investment vehicle from what remains of the frozen Russian assets after Ukrainian reconstruction. Provision 19 places the Zaporizhzhia nuclear plant, 5.7 gigawatts of capacity, under IAEA supervision with electricity shared 50:50 between Ukraine and Russia.

These four clauses describe three different compute propositions. They are worth distinguishing because Russia, Ukraine, and the Gulf each play a fundamentally different role in whatever post-war infrastructure stack actually gets built.

Ukraine, as I argued in an earlier piece, is assembling the inputs for European sovereign compute: nuclear baseload through Energoatom and the planned Westinghouse AP1000s, distributed generation that DTEK doubled to 5,144 installations in 2025, 200 MW / 400 MWh of grid-forming batteries dispersed across six sites, and 31 bcm of underground gas storage (the largest in Europe, third globally). On top of that hardware, Ukraine is building a credential layer through the Diia.Engine state platform and the Guarantees of Origin registry that will integrate into the EU Association of Issuing Bodies by 2027. That layer matters because it makes Ukrainian clean electricity exportable as auditable compute under European green-accounting rules. Ukraine is not a candidate for an American hyperscale corridor. It is being assembled for the European one.

Russia is the opposite proposition. It will not build frontier models. It does not have the labs, the talent base, or the capital. What it does have is the asset profile that any large data center developer recognises immediately: roughly 20 percent of the world’s natural gas reserves, the largest in the world; surplus grid capacity built for a Soviet-era industrial demand that no longer exists; Arctic minerals that bypass Chinese chokeholds on gallium, germanium, and heavy rare earth magnets; and, in the peace plan, joint operation of the largest nuclear plant in Europe. Where Ukraine is building a credentialed, green-accounting stack legible to European regulators, Russia is offering raw, un-vetted, brute-force scale.

The Gulf model is in execution and is the benchmark every other proposition is now measured against. Stargate UAE, a 5-gigawatt campus in Abu Dhabi at more than $30 billion, opens its first 200 megawatts in Q3 2026. Saudi Arabia’s HUMAIN, chaired by the Crown Prince, targets 1.9 gigawatts of data center capacity by 2030 and 6.6 gigawatts by 2034, with a 600,000-GPU agreement signed with Nvidia. The Atlantic Council GeoTech Center’s May 2026 issue brief by Trisha Ray and Raul Brens Jr. assesses the UAE as an “advanced AI power” at 17.5 out of 25, and recommends that the UAE itself co-invest in distributed compute capacity “across geographically dispersed locations outside the immediate region.”

The UAE has begun to operationalise that exact logic through G42’s Digital Embassies framework, launched at Davos in January 2026: a government-to-government legal construct that lets sovereign workloads run on UAE infrastructure while jurisdiction and control remain with the partner state. Sovereignty becomes portable. The infrastructure does not need to sit inside the customer’s borders.

Read against the Gulf, the Russian proposition in the peace plan offers structurally similar advantages: cheap energy, abundant minerals, sovereign-scale capital absorption through the investment vehicle in Provision 14, geographic distance from American grid constraints. The Gulf has executed. The Russian version is on paper.

The structural difference between the two corridors is the counterparty record and the American export control law. Under the Bureau of Industry and Security regime tightened repeatedly through 2024 and 2025 and codified again in the January 2026 final rule, frontier AI chips (the H100, B200, GB200 class) cannot be lawfully sited in Russia, which is a Tier 3 destination under the AI Diffusion Rule. Even a comprehensive peace settlement does not automatically reverse those controls; the BIS framework tracks not only where the rack sits but the nationality and headquartering of the end user, and applies worldwide licensing requirements to any firm headquartered in an arms-embargoed destination. The Russian compute corridor as written in Provision 13 cannot function as a standard hyperscale availability zone. To work at all, it would need a G42-type Greenshield-equivalent enclosure: physically air-gapped, US-managed, US-jurisdictional infrastructure where Russia supplies energy, security, and minerals, and nothing else. Washington may want the gas, but it cannot (under current law) trust the host with the chips.

A real-options reading of Provisions 13 and 14 is that the United States is not committing to build through Russia. It is reserving the right to. The clauses cost nothing to include and created two strategic effects at once. They gave Russia an economic reason to accept the political terms of the plan, because the corridor only opens if the war ends and only opens to the extent sanctions and BIS controls are unwound. And they put a floor price under the Gulf. By sketching a theoretical second offshore corridor in a treaty draft, Washington signals to Abu Dhabi and Riyadh that their monopoly on sovereign-scale, fast-to-connect power is not mathematically permanent. The clauses introduce a benchmark even if no hyperscaler ever signs a letter of intent.

Whether any of this materialises depends on what Europe does next. Putin’s 9 May overture was not primarily directed at Washington. It was directed at Berlin, and through Berlin at the European capital flows that Russia actually needs. There is a logic for Europe to seek accommodation with a weakened Russia: access to natural resources in exchange for European capital and a settlement that stabilises the eastern flank. The E3 counter-proposal of November 2025 already stripped Provisions 13 and 14 out of the plan, replacing the US-Russia investment vehicle with a multilateral reparations framework. Brussels accelerated a separate €140 billion frozen-assets scheme that Friedrich Merz called the negotiating consensus across European capitals. The fight inside the peace plan is no longer between Washington and Moscow, but between Washington and Brussels over who reintegrates Russia and on what terms.

If Brussels wins that contest, Russia becomes an energy appendix to Europe’s green transition: gas and Arctic minerals flowing into European industrial systems, with the American compute corridor staying on paper.

If Washington wins, Russia is integrated into a transactional, bilateral compute-monetisation engine, and the Gulf gains a benchmark competitor that will tighten pricing on the next round of Stargate-scale commitments. Either outcome shifts where the world’s compute gets built and which jurisdiction the data lives under.

Infrastructure investors have priced the offshore American map as if the Gulf were the only option. A second corridor is being written into a treaty draft this quarter. Ukraine is building the European one in parallel, under fire, for entirely separate reasons. Statecraft is compute. The peace plan is statecraft. Provisions 12, 13, 14, and 19 are where the three meet.

SOURCES

28-point plan text and Provisions 12, 13, 14, 19: Axios (20 Nov 2025), Al Jazeera (21 Nov 2025), ABC News (21 Nov 2025), The Hill, Newsweek (21 Nov 2025), CSIS The Unfinished Plan for Peace in Ukraine (9 Feb 2026), Belfer Center A Flawed Path to Peace (10 Dec 2025), Ukrainska Pravda full text release (23 Nov 2025).

Putin Victory Day 9-10 May 2026, Schröder proposal, Zelensky meeting offer: Al Jazeera (11 May 2026), TASS (9 May 2026), Milli Chronicle on EU foreign ministers’ rejection (12 May 2026).

Atlantic Council GeoTech Center, The new playbook for AI leadership: The case of the United Arab Emirates, Trisha Ray, Ryan Pan, Raul Brens Jr. (7 May 2026).

G42 Digital Embassies and Greenshield framework: G42 / Core42 launch announcement (Davos, January 2026).

E3 19-point counter-proposal: Reuters via Kyiv Post (24 Nov 2025), Ukrainska Pravda.

Merz on the EU consensus on frozen assets: ARD interview (24 Nov 2025), Anadolu Agency (5 Dec 2025).

Zaporizhzhia 5.7 GW: France 24 (29 Dec 2025), Newsweek (31 Dec 2025), Kyiv Independent (24 Dec 2025).

Russia natural gas reserves, Arctic LNG 2 status: Oxford Institute for Energy Studies (Oct 2024), Real Instituto Elcano (Jun 2024), IEA Global LNG Capacity Tracker (17 Mar 2026), Moscow Times via Bloomberg (1 Oct 2025).

Ukraine compute inputs: Sovereign Compute, Ukraine Is Assembling the Inputs for European Sovereign Compute (28 Apr 2026), and underlying sources cited there.

Stargate UAE and HUMAIN: The National (26 Jan 2026), G42 press release (May 2025), Middle East Institute From Crude to Compute (6 Feb 2026).

BIS export controls on advanced AI chips: Federal Register final rule 2026-00789 (15 Jan 2026), Mayer Brown briefing (22 Jan 2026), Morgan Lewis (16 Jan 2026), Introl AI export controls global tracker (Jan 2026), SemiAnalysis on AI Diffusion Rule tiering.

A NOTE ON INDEPENDENCE

All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sovereignty has always required an instrument the state could capitalise at scale. Land, then navies, then reserves, then deterrents. Compute is the next one.

A state that does not control the AI infrastructure its citizens use does not control six things at once.

It does not control the model weights, which means it cannot guarantee the system will behave the same way next month as it does today. It does not control the training datasets, which means the priors baked into the model on history, religion, geography, and law were chosen elsewhere. It does not control the outputs, which means decisions affecting taxation, healthcare, border control, and education are made through a black box owned by a foreign vendor. It does not control where the data is processed or where the analytics are run, which means citizen records, financial flows, and intelligence inputs cross jurisdictions every time a query fires. It does not control the cultural outputs, which means what a child reads, what a journalist drafts, and what a civil servant summarises is shaped by a model trained to a different national context. And it does not control the substrate, which means in a disruption it cannot fall back on the distributed compute already in citizens’ pockets, their phones, because the orchestration layer sits on someone else’s cloud.

That is the inventory of what is at stake. The Anthropic-Pentagon dispute makes it concrete.

In late 2025, the United States Department of Defense pushed Anthropic to drop two carve-outs in its usage policy: a prohibition on mass domestic surveillance, and a prohibition on fully autonomous weapons. Anthropic refused. In February 2026, the Secretary of War designated Anthropic a “supply chain risk,” a label previously reserved for foreign adversaries. The Court of Appeals for the DC Circuit, in April 2026, declined to lift the designation while noting that doing so “would force the United States military to prolong its dealings with an unwanted vendor of critical AI services in the middle of a significant ongoing military conflict.” OpenAI moved within hours of the conflict’s start to secure a Pentagon agreement without Anthropic’s carve-outs. Treasury, State, and Health and Human Services were directed to move off Claude.

The Pentagon is the sole hyperpower’s defence ministry. Anthropic is one of three frontier US AI labs. They sit inside the same jurisdiction and the same political alliance. They still hit an impasse over use cases the state considered essential and the lab considered unacceptable, in the middle of an active war. The state had to find another vendor at speed.

Now translate that to any other state. A foreign government using Claude, ChatGPT, Gemini, or Grok for citizen-facing services has no equivalent recourse. It has a terms-of-service agreement. In October 2025, OpenAI updated its usage policy to ban “tailored advice that requires a license,” covering legal, medical, and financial guidance. Any government that had built citizen-facing public service applications on top of that API woke up to find entire categories of use potentially newly prohibited. APIs can be revoked. Cloud regions can be quarantined under extraterritorial law. Model versions can be retired without notice. Usage policies can change unilaterally.

This is what a state actually loses when its AI runs on someone else’s stack: operational continuity, jurisdictional control, cultural authorship, and the ability to shape its own citizens’ interactions with the technology that increasingly mediates public life.

The states that have internalized this have moved from regulating AI to building it.

When the Crown Prince of Saudi Arabia chairs the board of an AI company, the company is not a portfolio holding. HUMAIN was launched in May 2025 as a wholly owned subsidiary of the Public Investment Fund, with Mohammed bin Salman as chair and Saudi Aramco as a minority shareholder. The mandate runs across the entire AI stack: data centres, cloud, the ALLAM Arabic-language model, and consumer applications. Disclosed commitments include a 600,000-GPU agreement with Nvidia, a $5 billion AI zone with AWS, a $3 billion stake in xAI, and a target of three to six gigawatts of compute capacity that maps to between $90 billion and $300 billion of total infrastructure spend at full build-out.

The same architecture is showing up across the Gulf and Asia. Qatar’s QIA created Qai in December 2025 as a wholly owned AI company, then formed a $20 billion joint venture with Brookfield to develop integrated compute facilities, with explicit support from the Government of Qatar. Abu Dhabi’s MGX, with a $100 billion AUM target, is chaired by Sheikh Tahnoun bin Zayed, the UAE’s National Security Adviser. Singapore’s Temasek joined the AI Infrastructure Partnership in December 2025. Kuwait’s KIA joined six months earlier. Canada launched a $2 billion AI Sovereign Compute Infrastructure Program. The chair of the AI vehicle is, increasingly, a head of state or a national security official. The capital is sovereign capital. The mandate is written by the state.

Six things a state surrenders if it does not control its compute, scored across the major players.

Three large states have absorbed this and are paying for it accordingly. Each pays a different way.

The American model runs on hyperscaler balance sheets. Microsoft, Alphabet, Meta, and Amazon are tracking roughly $700 to $725 billion in combined 2026 capex, almost double 2025. Stargate, the announced $500 billion vehicle led by SoftBank and OpenAI with Oracle and MGX, sits on top, financed through SoftBank borrowing, JPMorgan project debt, and Gulf sovereign equity. The state directs through procurement, export controls, and tax policy. The Anthropic case shows the limit: when state and lab disagree on use cases, the state must find a different lab.

The Chinese model runs on a different instrument set. Beijing’s state venture capital guidance fund, announced in March 2025, targets nearly 1 trillion yuan over twenty years. Government-directed AI investment hit roughly ¥345 billion in 2026, with another ¥258 billion in corporate R&D at Alibaba, Tencent, and Baidu. Top Chinese internet firms are projected to invest more than $70 billion on AI infrastructure in 2026. The instruments are guidance funds, state-owned banks, and direct mandates to listed national champions.

The Gulf model is the third and most explicit. There is no equivalent of Microsoft or Tencent on the cap table. There is no $50 trillion debt market to draw from. What the Gulf has is sovereign capital with patient duration, political alignment with the state’s mandate, and energy reserves that translate directly into compute capacity. The wealth fund becomes the principal vehicle because nothing else fits the shape of the problem.

What the three share matters more than what divides them. Each is a structure that lets the state set the terms of the stack. The US sets terms through export controls and procurement. China sets terms through ownership and guidance. The Gulf sets terms through wealth fund mandate and chairmanship. The form is different. The function is the same. One contrast sharpens the picture. As observed after visiting most of the leading Chinese labs in May 2026, Meituan, Xiaomi, and Ant Group, none of them traditionally AI companies, all build their own general-purpose LLMs. The equivalent companies in the US buy services. Ownership mentality has saturated the Chinese corporate layer in a way it has not in the American one.

Now the rest of the world. The EU’s InvestAI initiative targets €200 billion through 2030, with €50 billion public and €150 billion pledged by Blackstone, KKR, EQT, and other private investors who can walk away if returns disappoint. The UK’s Sovereign AI sits inside a private vehicle backed by Accenture, Palantir, Dell, and Nvidia. India’s IndiaAI Mission has over 38,000 GPUs subsidised and targets $200 billion under Mission 2.0. France has €109 billion announced, much of it private. Most of Africa, Latin America, and Southeast Asia have none of these instruments at scale. They will host other people’s stacks.

There is one architectural development that changes this map.

Smartphone penetration in Africa is projected to reach 81 percent of mobile connections by 2030, up from 51 percent in 2023. Asia-Pacific and Latin America already exceed 80 percent. Modern phones can run capable language models on-device. A country that orchestrates inference across its citizens’ phones has a distributed compute layer that does not depend on any single foreign cloud region staying online. The catch is the substrate underneath, where iOS, Android, and the on-device AI runtimes are owned by US firms. A few significant states are already addressing it. China requires the Cyberspace Administration to approve any model running on a foreign device sold in the country, which is why Apple Intelligence ships in China only with Alibaba’s Qwen and a state-mandated filtering layer. Saudi Arabia’s HUMAIN signed with Qualcomm to integrate ALLAM into the AI200 and AI250 silicon rack, building the inference layer at the chip level. The UAE’s Technology Innovation Institute released Falcon-Edge as a 1.58-bit model designed to run on edge devices outside the major OS runtimes. The EU is using the Digital Markets Act review to bring AI assistants under interoperability obligations.

In other words, the substrate is contestable. The states with the means are contesting it. Countries that build this layer over the next five years will have an answer to disruption that does not necessarily require a hundred-billion-dollar wealth fund.

Governments are becoming infrastructure companies because infrastructure has become a precondition for governing at scale. The chair of the AI vehicle can now be the head of state because the AI vehicle is where part of the state's work now gets done. Sovereignty has always required an instrument the state could capitalise at scale. Land, then navies, then reserves, then deterrents. Compute is the instrument this generation of statecraft is built on.

SOURCES

PIF (HUMAIN structure, Aramco minority stake) · Arab News (NIF and HUMAIN $1.2B framework) · House of Saud (PIF 2026 to 2030 strategy, FII Miami) · QIA / Brookfield press releases (Qai $20B JV, December 2025) · Mubadala / AGBI / Bloomberg (MGX structure, $100B AUM target) · Reuters / Salem Radio Network (Temasek and KIA into AI Infrastructure Partnership) · OpenAI press releases (Stargate sites, financing structure) · The Information / WSJ (Stargate equity contributions) · CNBC and Wikipedia (Anthropic-DOD dispute) · Anthropic public statements (February 2026) · OpenAI usage policy update (29 October 2025) · CNN / NDRC (China state venture capital guidance fund, March 2025) · Second Talent (China 2026 AI investment statistics) · Futurum / CreditSights (hyperscaler 2026 capex aggregates) · TechPolicy.Press / European Commission (InvestAI €200B structure) · GSMA Mobile for Development (smartphone penetration projections) · McKinsey (Sovereign AI agenda, December 2025) · Acadlore JORIT (peer-reviewed sovereign LLM bias study, 2025) · Nathan Lambert, Notes from inside China’s AI labs, Interconnects, 7 May 2026 (https://www.interconnects.ai/p/notes-from-inside-chinas-ai-labs) · TechCrunch and SCMP (Apple-Alibaba Qwen partnership, China CAC requirements) · Middle East AI News (HUMAIN-Qualcomm AI200 / AI250 deployment, October 2025) · TII press releases (Falcon-Edge BitNet 1.58-bit model) · ComplianceHub.Wiki and European Commission (DMA two-year review, AI services and cloud, April 2026).

A NOTE ON INDEPENDENCE

All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

In 2050, one in four humans will be African. The continent’s median age today is 19.5. Europe’s is 44. The mismatch between where the population is heading and where the compute lives is the single largest infrastructure arbitrage of the next two decades.

Africa today has 1.55 billion people. By 2050 it will have 2.5 billion, according to the UN Population Division’s 2024 World Population Prospects. That is half of all global population growth between now and mid-century, concentrated on a continent currently served by 223 data centers across 38 countries (under 0.02% of the world total). Per capita compute capacity sits below 1 MW per million people. The Americas have 88.5 MW per million. Europe has 73.9. The five largest African markets combined operate around 400 MW of installed capacity, roughly half of what France alone runs.

Africa is not a country and the data makes that clear in striking ways. South Africa hosts 56 data centers, roughly 41% of the continent’s total alongside Kenya (19) and Nigeria (17). Morocco dominates the upcoming pipeline with 35-60% of planned new power capacity. Ethiopia has 1 facility. Gabon has 1. The Democratic Republic of Congo, projected to add nearly 200 million people by 2050, has effectively no commercial AI infrastructure today. The demographic dispersion is equally wide. Niger grows at 3.7% per year with a fertility rate above 6, doubling its population in under twenty years. Tunisia, Mauritius, and Cabo Verde are below replacement, the same demographic profile as Italy. Egypt is slowing toward 1.57% annual growth thanks to a national family planning campaign. Treating the continent as one market is the first error any infrastructure investor makes.

Three capital channels are racing to position inside this asymmetry, each with different motives.

Gulf sovereign wealth is treating Africa as a post-oil hedge with geopolitical upside. Mubadala, ADQ, PIF, and MGX deployed $66 billion into AI and digital infrastructure in 2025. The Microsoft-G42 geothermal campus at Olkaria, Kenya, anchors East Africa with an initial 100 MW phase scaling toward 1 GW; the first build is explicitly sized for inference and training of Swahili and East African language models, not generic cloud storage. Khazna operates into Egypt through the Benya Group. ADQ’s $35 billion Ras al-Hikma deal sets the template: sovereign capital deployed at a speed Western development finance cannot match, US technology alignment baked in at the chip layer.

China is running the Digital Silk Road playbook with a decade of head-start on physical infrastructure. Huawei equipment runs roughly 70% of African 4G networks. Its Cairo cloud region (the first in Northern Africa) grew 140% in year one and serves 28 African countries. Senegal’s national data center, hosting government data, was built by Huawei on a $79 million Chinese government loan and modeled on Beijing’s data governance architecture. Beijing provides more financing for African ICT than all multilateral agencies and Western democracies combined.

The third channel is African capital. Strive Masiyiwa’s Cassava Technologies committed $720 million to five sovereign AI factories across South Africa, Nigeria, Kenya, Egypt, and Morocco, with 15,000 NVIDIA GPUs running on Cassava’s own 110,000-km fiber backbone. The Johannesburg facility is reportedly 90% pre-booked. Liquid Intelligent raised $605 million in March 2026. Stanlib (Standard Bank’s asset arm) is acquiring a stake in Africa Data Centres. Nxtra (Bharti Airtel) opened a 38 MW hyperscale facility in Lagos. Teraco crossed 200 MW in Johannesburg. African capital is showing up at hyperscale for the first time, and it is using NVIDIA silicon, not Huawei.

Power decides where the campuses land, and the answer differs by country. Kenya’s grid is over 60% renewable, with the Olkaria geothermal field supplying half the country, which is why every East Africa hyperscale project routes there. Ethiopia is essentially all hydro, anchored on the Grand Ethiopian Renaissance Dam. Egypt is building the El-Dabaa nuclear plant with Rosatom and pairing it with gas, positioning the Suez corridor as a Europe-Africa-Asia hub. Morocco combines abundant solar with the planned Morocco-UK power interconnector to Europe. Nigeria has gas in abundance and a grid that delivers four hours of power per day on average, forcing operators to keep tens of thousands of liters of diesel on site. South Africa has Eskom load-shedding and the only mature electricity wheeling regime on the continent, which is why new projects there cluster around Northern Cape solar.

The bear case is real and worth stating plainly. UNCTAD reports African FDI fell 42% in the first half of 2025. Few countries have regulatory frameworks for long-term power purchase agreements. Microsoft-G42 Olkaria has not broken ground 18 months after announcement, with reports suggesting equity stakes between G42, Microsoft, and Kenyan state entities are still being negotiated. Chinese-built data sovereignty is a second layer of risk: the data is physically Kenyan, Egyptian, or Senegalese, but the architecture above it stays in Beijing. Modeling a smooth ramp to 2 GW by 2030 misprices both the political and the electrical risk.

There is one variable that analyses tend to leave out. By 2050 the working-age population (20-64) shrinks by 49 million in Europe, 195 million in China, 22 million in Russia, and 20 million in Japan. Sub-Saharan Africa’s working-age cohort triples to roughly 1.3 billion, almost twice the combined total of all high-income countries. If longevity science delivers anything close to what Sinclair, Kurzweil, and de Grey are projecting for the 2030s, the dependent population in advanced economies grows on top of a working population that is already shrinking. Compute demand in those economies rises, but the workers who build, train, and operate AI systems will increasingly need to come from somewhere else. Africa is the only large region where labor supply and power supply both grow.

The infrastructure decisions being made now (where the cables land, which governments get hyperscale partnerships, which power source anchors which campus) will determine who serves that demand and who collects the rent on it for the next forty years. The question is not whether African compute gets built. It will. The question is which countries cross the threshold from frontier to tier-one in time to capture the demographic wave, and which capital channels are positioned inside those specific markets when they do.

Sources

• UN Population Division, World Population Prospects 2024 Revision (medium-fertility variant)

• UN Economic Commission for Africa, “As Africa’s Population Crosses 1.5 Billion,” July 2025

• European Commission, “Developments and Forecasts of Increasing Demographic Imbalances”

• McKinsey Global Institute, “Dependency and Depopulation,” January 2025

• Institute for Health Metrics and Evaluation (IHME), Global Burden of Disease Study 2021, The Lancet

• World Government Summit 2026, David Sinclair remarks on aging-reversal trials

• African Energy Chamber, State of African Energy 2026 Outlook

• Mordor Intelligence, Africa Data Center Market Outlook, January 2026

• McKinsey, “Africa’s data center demand set to triple by 2030”

• ISS African Futures, “Data centre investments are a gamble for Africa,” 2025

• Microsoft / G42, $1B Olkaria Kenya geothermal campus announcement

• Cassava Technologies, $720M Sovereign AI Cloud announcement

• Billionaires Africa, Liquid Intelligent $605M recapitalization, March 2, 2026

• CNAS, “Countering the Digital Silk Road: Kenya,” 2025

• SWP Berlin, “Sovereign Wealth Funds and Foreign Policy,” February 2026

• A&O Shearman, “Sovereign wealth funds boost Middle East M&A growth,” January 2026

• CFR, China Digital Silk Road tracker

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Europe's AI buildout is running into a hard constraint: power.

Ukraine, for entirely different reasons, is building the answer.

Europe’s data centre sector consumed roughly 96 TWh in 2024, about 3% of continental demand. Installed capacity sits near 18.7 GW today and is heading toward 36 GW by 2030 on 451 Research projections, a compound annual growth rate of roughly 12%.

The binding constraint is power, not capital. Dublin has operated under a de facto grid-connection cap since 2022, with €5.8 billion of data centre projects stranded for lack of electricity. Frankfurt, Amsterdam and London absorb hyperscaler loads equivalent to one-third to 40% of local demand. Grid connection queues across Western Europe now run 7 to 10 years against construction cycles of about 24 months. The Commission’s January 2026 approval of 10 strategic gigafactory sites, matched by roughly €40 billion in public-private funding, sits on top of this bottleneck.

Ukraine is building a different stack, under fire, for reasons that have nothing to do with AI.

Since the full-scale invasion, roughly two-thirds of Ukraine’s pre-war generating capacity has been damaged, destroyed or occupied. The response has been systematic physical decentralisation. DTEK Grids connected 5,144 distributed generation and renewable installations in 2025, twice the 2024 figure. The Solar Energy Association of Ukraine estimates at least 1.5 GW of new solar was added in 2025, enough for around 1.1 million homes. DTEK and Fluence commissioned 200 MW / 400 MWh of grid-forming battery storage across six dispersed sites in September 2025, the largest such portfolio in the country. The government’s Strategy for the Development of Distributed Generation, approved August 2024, targets 27% of consumption from DERs by 2030. By November 2025 the district heating sector already operated 182 cogeneration units and 239 block-modular boilers providing roughly 780 MW combined, with 500 MW of additional decentralised capacity planned for 2026.

That decentralisation profile - thousands of small, dispersed sources feeding autonomous load pockets - matches what modern AI compute wants: distributed dispatch, resilience to single-point failure, and on-site firming capability. Dublin’s problem is a concentrated 100-MW hyperscale load arriving at a constrained node. Ukraine’s evolving topology is the inverse.

Three layers being assembled in parallel: nuclear baseload (Energoatom + AP1000s), distributed firming (DER + BESS + cogeneration), credential layer (Diia.Engine GO Registry).

The baseload layer is being rebuilt in parallel.

Energoatom’s nuclear fleet generates over 55% of national electricity. In July 2025, Westinghouse and Energoatom signed a fuel-assembly partnership at the Ukraine Recovery Conference, advancing the plan to build nine AP1000 reactors: 11,250 MWe of new capacity targeting Khmelnytskyi, South Ukraine, and Chyhyryn. Following the November 2025 corruption scandal, the government appointed a new seven-member supervisory board on 28 January 2026; at its first meeting on 26 February it elected Rumina Velshi (former CEO of the Canadian Nuclear Safety Commission and former Chair of the IAEA Commission on Safety Standards) as Chair, and Patrick Fragman (former President and CEO of Westinghouse Electric) as Vice-Chair. Fragman also chairs the new Strategy and Financial Planning Committee, a direct operational link between Ukraine’s nuclear governance and the technology provider for the AP1000 build.

The credential layer makes the whole thing legible to European buyers. Ukraine launched its Registry of Renewable Energy Facilities on the Diia.Engine platform in July 2025, the same low-code state infrastructure that Germany’s GIZ helped scale into 57 operational government registries and that drew architectural lessons from Estonia’s X-Road. Cross-border circulation of Ukrainian Guarantees of Origin is scheduled to launch in 2026, with full integration into the EU Association of Issuing Bodies by 2027. Under incoming Corporate Sustainability Reporting Directive obligations and the 2026 revision of GHG Protocol Scope 2 guidance, auditable GoO provenance is moving from voluntary to mandatory for European corporates. Hyperscaler contracts increasingly require 24/7 hourly matching, which only a digital registry can verify. The registry is the precondition for procurement-grade clean-compute exports.

Two constraints remain live. Ukraine’s export capacity to the EU is 900 MW as of the ENTSO-E July 2025 revision, and the 2.45 GW cross-border capacity operating in January 2026 is import-weighted during wartime. Ukrenergo’s Recovery Plan targets substantial transfer expansion by 2030, but the lines are not built. And the sector’s political economy - DTEK on distribution, Naftogaz on gas, Energoatom on nuclear - requires parallel governance reform for any compute-export contract to hold. The Fragman appointment suggests Kyiv knows this.

Ukraine has 31 bcm of underground gas storage, the largest in Europe, third globally, currently around 90% full with roughly 10 bcm held on behalf of traders from 21 countries. The country has been operating as Europe’s gas battery for years; the same 21-country counterparty footprint that uses Ukrainian storage as flexibility today is the contractable customer base for clean-compute output tomorrow. For an AI buildout that increasingly demands on-site firming (Cleanview’s February 2026 analysis projects 30% of incremental US data centre capacity from on-site generation in 2026, rising toward 50% - a ratio US-led but increasingly visible across Western European projects), gas storage adjacent to ENTSO-E transfer capacity is a firming reserve that Frankfurt cannot replicate.

For infrastructure investors, the question is whether the EU’s €20 billion InvestAI facility, the EIB’s gigafactory co-financing track, and the World Bank’s roughly $67 billion reconstruction envelope get treated as three separate files or as one. The inputs for a Ukrainian sovereign-compute node are already being assembled by necessity. The financing stacks that would contract the output are still being written.

Sources

· EuroHPC JU — AI Factories and AI Gigafactories

· European Commission — MoU on AI Gigafactories (Dec 2025)

· S&P Global 451 Research — European DC power demand to double by 2030 (Jul 2025)

· ENTSO-E — Go-live of explicit monthly long-term auctions on Ukrainian borders (Dec 2025)

· Ukrenergo — EU import capacity increase to 2.45 GW (Jan 2026)

· IEA — Empowering Ukraine Through a Decentralised Electricity System (Dec 2024)

· CSIS — Striving for Access, Security, and Sustainability (Feb 2026)

· DTEK Grids — 2025 distributed generation connections report

· Fluence Energy / DTEK — 200 MW BESS commissioning (Sept 2025)

· Westinghouse–Energoatom — AP1000 partnership and fuel assembly agreement (Jul 2025)

· Euromaidan Press — Energoatom supervisory board appointments (Jan 2026)

· Ukrainian Energy — Guarantees of Origin registry integration timeline (May 2025)

· CGEP Columbia — Ukraine’s Underused Gas Storage Capacity (May 2025)

· Naftogaz — Underground gas storage business unit

· FP Analytics — Investing in Energy Security: Ukraine (Nov 2025)

· CEPA — A Rebirth in Flame: Ukraine’s Beleaguered Energy System (Dec 2025)

· TechPolicy.Press — Ireland’s data center crisis and EU AI sovereignty (Dec 2025)

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Five US hyperscaler balance sheets now carry capex equal to 2% of US GDP. Sovereigns carry the rest. In a multipolar system with thinner public budgets, the Gulf's public capital stack is building the redundancy needed.

At the IMF–World Bank Spring Meetings in Washington on April 15, Kristalina Georgieva warned that most countries are “less prepared to respond to a major economic downturn” than they were a decade ago. The April Fiscal Monitor backs her: global interest payments have climbed from roughly 2% to nearly 3% of GDP in four years, public debt sits at 94% of world GDP, and primary balances in most large economies are now below the level needed to stabilize that debt. Fiscal consolidation space has narrowed across the G20 in parallel, not sequentially. The policy toolkit most finance ministries brought to Washington is smaller than the one they used in 2020.

The missing buffer has a private substitute. Apollo’s Torsten Slok estimates 2026 hyperscaler capex at $646 billion, which is approximately 2% of US GDP, and larger than the combined 2025 defense budgets of Germany, France, the UK, Japan, Italy and Canada. Pantheon Macroeconomics attributes nearly one-third of Q4 2025 US GDP growth to AI-linked capex plus the wealth effect from tech equity gains. Apollo’s own measurement of real private fixed investment shows corporate capex outside AI at essentially zero growth. A fair question is whether AI capex is stabilizing US growth or crowding out the rest of it; either way, the concentration itself is the fragility, because the growth contribution now depends on the investment plans of five companies.

Sources: Apollo 2026 Outlook (Dec 2025); Apollo “How Much Is $646 Billion?” (Feb 22, 2026); Pantheon Macroeconomics via Fortune (Feb 26, 2026); IMF Fiscal Monitor Press Briefing and WEO (Apr 14–15, 2026).

That concentration has strategic-autonomy consequences the fiscal framing obscures. When Amazon, Microsoft, Google, Meta and Oracle are financing infrastructure at the scale the US Treasury used to fund, the United States has outsourced a core growth engine, and the physical stack of its AI capability, to five CEOs and the private-credit vehicles sitting behind them. I examined that single-provider pattern across cloud, models and launch in Three Dependencies, One Stack: the Anthropic FASCSA designation, the AWS strikes in the UAE and Bahrain, and SpaceX’s role in US military launch. The fiscal architecture now underwriting US growth sits on the same architecture. Monetary sovereignty, growth sovereignty, AI sovereignty and strategic autonomy have collapsed into one operational question: which balance sheets will still be able to build in 2028?

Three layers of the same question: public fiscal space, US private AI capex, and Gulf public capital, measured on the same axis of who can still deploy at scale.

The financing is already moving off the hyperscaler balance sheet

Hyperscaler capex intensity has already passed levels associated with regulated utilities rather than software businesses. Across Amazon, Google, Microsoft, Meta and Oracle, capex now absorbs approximately 60% of operating cash flow. Companies that invest like utilities tend, over time, to trade on utility-like multiples; a re-rating of AI equities toward that baseline is the specific mechanism by which the IMF’s flagged “AI repricing” would move from a tail risk to a valuation event. The bridge between operating cash flow and stated capex plans is already being built outside the hyperscaler balance sheet. Meta’s $30 billion Hyperion data center sits in a special-purpose vehicle, Beignet Investor, with $27 billion of debt from Pimco, BlackRock and Apollo and $3 billion of Blue Owl equity, which is debt that does not appear on Meta’s accounts. Oracle has structured roughly $38 billion of comparable financing for Stargate sites in Texas, Wisconsin and New Mexico; Pimco added a $14 billion commitment to Oracle’s Saline, Michigan campus on April 7. The BIS catalogued the shift in its March Quarterly Review. Morgan Stanley and JP Morgan project $1.5 trillion in new tech-sector debt issuance over the coming years, arriving as global sovereign interest payments reach nearly 3% of GDP. That is a margin pincer: rising capex against rising cost of capital. The Atlantic Council GeoTech Center has argued that financing itself is part of the AI infrastructure stack. The 2026 capex cycle is the direct evidence.

Sources: Apollo 2026 Outlook; BIS Quarterly Review (Mar 16, 2026); Financial Times reporting on SPV structures (Dec 2025); Bloomberg on Pimco-Oracle financing (Apr 7, 2026); Mellon Investments on AI debt issuance (Dec 2025).

The Gulf is deploying public capital on interconnected infrastructure

George Friedman’s recent essay locates the next emerging global industrial power in Latin America, dismissing the Middle East and Africa as “regions, not nations” too burdened by conflict to qualify. The empirical record in the Gulf disagrees. The GCC Interconnection Authority has operated a unified 400 kV super-grid since 2009, spanning roughly 1,200 kilometers of overhead lines and submarine cables across Saudi Arabia, the UAE, Qatar, Kuwait, Bahrain and Oman. GCCIA reports it has prevented more than 1,300 potential outages since commissioning. A $205.5 million expansion will raise UAE–GCCIA transmission capacity from 2,400 MW to 3,500 MW by Q1 2027. A direct Oman–GCCIA interconnection, the Ibri and Baynonah substations, is under construction.

The grid also carries OPGW fiber-optic cables along its rights-of-way, creating a terrestrial data corridor that bypasses Red Sea and Bab-el-Mandeb subsea chokepoints. For infrastructure investors and policymakers pricing subsea cable risk after repeated cuts in the Red Sea and Baltic, fiber on desert-grid towers carries a measurable financial return priced into every cross-border data flow that routes through it. The configuration converts physical redundancy into expected yield.

Sources: GCCIA official network disclosures; GCCIA “Our Plans” — Direct Oman Interconnection and UAE Expansion projects.

On that grid, sovereign capital is being deployed directly. HUMAIN, the PIF subsidiary, has $23 billion of signed agreements with NVIDIA, AMD, AWS and Qualcomm; a $10 billion joint venture with AMD targets 500 MW of capacity; its roadmap commits to 1.9 GW by 2030 and 6.6 GW by 2034 on a $77 billion plan. G42’s Stargate UAE, a Khazna–OpenAI–Oracle–NVIDIA–Cisco–SoftBank consortium inside the 5 GW UAE–US AI Campus in Abu Dhabi, delivers its first 200 MW in Q3 2026. Gulf SWFs deployed approximately $25 billion of new investments in Q1 2026, per Global SWF data reported by Semafor, despite running this pace through an active regional conflict. Combined Gulf sovereign wealth stands at approximately $5 trillion today, with projections to $18 trillion by 2050. The financing model is public equity committed directly into infrastructure, not private debt structured through SPVs, which gives it a different risk profile to the shock channels that could hit the US stack.

Sources: NVIDIA-HUMAIN press release (Nov 19, 2025); Arab News on HUMAIN-AMD JV (May 2025); The National on Stargate UAE Phase 1 (Dec 5, 2025); Global SWF data via Semafor (Apr 13, 2026); SWP Research Paper 2026/RP 03 on Gulf SWFs.

What it means for allocators

The US and Gulf models now function as two sides of the same macro portfolio. The US side delivers scale, model capability, and the world’s leading supply chain, funded through five corporate balance sheets that run at 60% capex-to-cash-flow intensity and increasingly rely on private credit structured outside shareholder disclosure. The Gulf side delivers public equity committed on a regionally interconnected grid, directly tied to energy generation, with fiber routing around maritime chokepoints, in a $5 trillion pool that is projected to triple within a generation. Each side carries the shock the other cannot absorb. A rate shock or AI-equity re-rating compresses hyperscaler multiples and tightens the private-credit channel financing US capex; an energy or maritime shock raises the value of the Gulf’s in-region redundancy. The IMF’s April WEO lists AI-productivity disappointment as one of its three named downside risks for 2026. The corresponding upside scenario, AI capex translating into sustained productivity gains, assumes the capital stack keeps deploying through the downside. That assumption needs a second source.

For infrastructure investors, energy executives and sovereign strategists, the practical question is exposure. Specifically, whether their AI-related positions are concentrated in the single capital stack most exposed to the downside risk the IMF itself named, or whether they have begun pricing access to the public-capital redundancy layer now being commissioned on the GCCIA grid. The capital flows in 2026 will tell us which jurisdictions have built sovereign infrastructure the rest of the system can lean on.

SOURCES

• Sovereign Compute: Three Dependencies, One Stack (Mar 17, 2026)

• Apollo: How Much Is $646 Billion? (Feb 22, 2026)

• Apollo 2026 Outlook — Torsten Slok (Dec 2025)

• Fortune: AI capex + wealth effect = ~1/3 of US GDP growth (Feb 26, 2026)

• IMF Fiscal Monitor Press Briefing, Spring Meetings 2026 (Apr 15, 2026)

• IMF World Economic Outlook, April 2026

• Atlantic Council: Inside the IMF-World Bank Spring Meetings (Apr 17, 2026)

• BIS Quarterly Review: Financing the AI infrastructure boom (Mar 16, 2026)

• GCCIA — Our Network

• GCCIA — Our Plans (UAE expansion & direct Oman interconnection)

• NVIDIA-HUMAIN Partnership Expansion (Nov 19, 2025)

• The National — Stargate UAE Phase 1 Q3 2026 (Dec 5, 2025)

• Semafor — Gulf wealth funds Q1 2026 (Apr 13, 2026)

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

The states building the most data centers are building the 21st century's strategic infrastructure. The states building across all three layers — ground, orbit, and edge — are building the infrastructure that survives.

The sovereign compute orthodoxy runs in a straight line. Compute follows energy. Energy follows geography. Geography follows politics. Every major AI infrastructure strategy on Earth — the US hyperscaler buildout, the Gulf’s sovereign data center investments, India’s Digital India stack — operates within this binding logic. Training a frontier model consumes enough electricity to power a small city. Cooling a hyperscale cluster requires millions of gallons of water annually. Permitting a grid interconnection takes years. The physical layer is real, expensive, and slow, and it anchors the entire AI stack to specific coordinates on a map.

This framing is correct. It is also incomplete. The sovereign compute race as currently understood is a race to concentrate AI infrastructure in fixed, targetable locations — data centers with GPS coordinates, grid substations feeding 500 MW to a single campus, subsea cables landing at known points. The Iran-Gulf conflict demonstrated what this concentration means in practice: three AWS facilities struck, Hormuz shipping down 90–95%, and sovereign AI strategies built on the assumption of physical security suddenly facing the oldest problem in strategic infrastructure. Concentration is capability. Concentration is also vulnerability.

Two emerging vectors suggest that the geographic binding constraints of sovereign compute — while real for training — may not be permanent for inference, which is where AI meets the population. Call the current orthodoxy the geopolitics of the physical: the assumption that AI power is determined by who controls the most land, energy, and water in the most politically stable territory. The two vectors I’m going to describe represent something different — the arbitrage of the digital: the possibility that inference capacity can be distributed across domains where physical geography no longer binds. One points up. The other points out. Neither is a clean solution. Together, they redraw the map.

The first vector is orbital.

In January 2026, Axiom Space launched the first two orbital data center nodes to low Earth orbit. Weeks earlier, NVIDIA-backed Starcloud ran Google’s Gemma large language model on a satellite carrying an H100 GPU — the first time a frontier-class chip operated from orbit. NVIDIA followed at GTC with a full space computing platform: Vera Rubin modules, IGX Thor, and Jetson Orin, with Axiom, Kepler Communications, Planet Labs, and Starcloud as launch partners. Google’s Project Suncatcher envisions constellations of TPU-equipped satellites connected by free-space optical links, with prototype launches planned for early 2027. China’s Zhejiang Laboratory has already deployed a 12-satellite computing constellation carrying an 8-billion-parameter AI model, while Beijing-based BAIST plans a megawatt-scale orbital data center by 2035. Two weeks ago, Syntiant and Novi Space demonstrated real-time AI object detection on a commercial satellite — with model retraining in under 24 hours.

Space sidesteps the terrestrial binding problem almost entirely. Solar power in LEO is continuous and abundant — a solar panel in the right orbit produces up to eight times more energy than on Earth. There are no permitting delays, no water rights conflicts, no grid interconnection queues. Thermal dissipation, though, remains the ultimate engineering frontier. On Earth, data centers use convection and liquid cooling. In vacuum, the option is radiation, and radiating waste heat from a megawatt-scale cluster requires massive deployable radiator arrays. Small-scale inference modules like Jetson Orin handle this with passive radiation. The 5-gigawatt architectures Starcloud envisions will require radiator wings at an industrial scale — heat management as the “orbital water rights” of the next decade. The strategic logic extends further: if compute dispersal is a resilience strategy, the Moon’s south pole, and specifically the Peaks of Eternal Light, where near-continuous solar exposure meets permanently shadowed craters for thermal management, becomes a redundancy node. The Artemis core program’s infrastructure is about depth-of-field for critical systems.

Orbital compute also introduces a survivability profile that terrestrial infrastructure cannot match. Starlink demonstrated this in Ukraine: a distributed constellation of thousands of satellites has no single node to destroy. A 5,000-satellite compute constellation presents a fundamentally different target than three data centers in a desert. Russia’s development of nuclear anti-satellite capabilities — a weapon the US Department of Defense warned could render LEO unusable for an extended period — is the only theoretical counter. But an EMP that disables all satellites in a wide orbital zone would destroy Russian, Chinese, and commercial assets alongside the target, making it a mutual-destruction weapon, not a practical one.

The second vector runs in the opposite direction — not up to orbit, but outward to the edge of the network, to every device in every pocket.

There are 7.4 billion active smartphones in the world today. India alone accounts for 1.2 billion smartphone connections. Mobile neural processing units now deliver throughput approaching data-center GPUs from 2017. Billion-parameter language models run in real time on flagship devices. IDC forecasts that over 370 million GenAI-capable smartphones shipped in 2025 — 30% of all shipments — rising to over 70% by 2029. Deloitte estimates inference workloads now account for roughly two-thirds of all AI compute, up from one-third in 2023. The on-device AI market is projected to grow from $33 billion in 2026 to $157 billion by 2033.

The strategic significance of edge compute is not mainly about latency or privacy. It is about distributing the physical substrate of sovereign AI to a target set so large it becomes indestructible. Iran can strike three data centers. It cannot strike 1.2 billion smartphones. No adversary can. The compute does not live in a building with coordinates — it lives in every pocket, on every wrist, behind every pair of smart glasses. Each device is a micro-sovereign compute node. Individually trivial. Collectively, an AI infrastructure that is physically unconcentrable and therefore physically undefeatable. The parallel to virtual power plants is direct: just as distributed solar and batteries aggregate into grid-scale capacity without centralized generation, distributed NPUs aggregate into national-scale inference capacity without centralized data centers.

India’s position in this framing shifts dramatically. In the current orthodoxy, India’s Intelligence layer — models, training capacity, frontier research — is its weakest dimension. But in a distributed inference world, India’s 1.2 billion connected devices become the largest sovereign compute substrate on Earth. The critical question becomes who controls the model weights on those devices. Apple’s Neural Engine runs Apple Intelligence. Qualcomm’s NPUs run whatever Qualcomm certifies. Google controls Android’s ML runtime across 3.9 billion devices globally. Edge AI may not democratize compute sovereignty — it may transfer it from territorial states to platform companies, making Apple and Google de facto sovereign compute providers for billions of people. This is why nation-state foundation models — India’s domestic efforts, the UAE’s Falcon family, Saudi Arabia’s SDAIA programs — become more strategically important in an edge-first world. Whoever controls the weights controls the intelligence at the edge, regardless of who manufactured the silicon.

Neither vector is a replacement for terrestrial infrastructure.

Training frontier models still requires concentrated, high-power, liquid-cooled GPU clusters. The data center buildout is necessary and serious. But training is not where most AI value will ultimately be captured. Inference is. And inference is splitting into three domains: terrestrial data centers, orbital compute, and edge devices. The sovereign compute race is no longer one-dimensional.

This reframing matters most for geographically constrained states. Singapore, the UAE, Israel suddenly have alternative paths to sovereign AI capability. The Gulf’s centralized data center investments are essential for training. Adding orbital partnerships and edge model ecosystems creates a three-layer resilience architecture where the ground layer can be struck, the orbital layer requires weapons that don’t yet practically exist, and the edge layer cannot be struck at all. A sovereign AI strategy built across all three layers is not just more capable. It is more survivable. And survivability, as the past two months have demonstrated, is not a theoretical concern.

The sovereign compute map most strategists are drawing — color-coded by data center capacity and grid megawatts — is a map of training infrastructure. The inference map has two additional layers that appear on no territorial chart: orbit and the edge. The states that see this early will not just build the most AI infrastructure, but the AI infrastructure that endures.

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sources

NVIDIA, “NVIDIA Launches Space Computing, Rocketing AI Into Orbit,” GTC 2026

Axiom Space, “Orbital Data Centers,” January 2026

CNBC, “Nvidia-backed Starcloud trains first AI model in space,” December 2025

Google Research, “Project Suncatcher: Exploring a space-based, scalable AI infrastructure system design”

China Daily, “Chinese tech firms race to build AI computing capabilities in space,” December 2025

SatNews, “Syntiant and Novi Space Demonstrate Low-Power AI Inference in Orbit,” March 2026

Arms Control Association, “U.S. Warns of New Russian ASAT Program,” March 2024

Atlantic Council, “Russian nuclear anti-satellite weapons,” February 2024

DataReportal, Digital 2026 Global Overview

Deloitte, “Why AI’s next phase will likely demand more computational power, not less,” TMT Predictions 2026

IDC, Worldwide Smartphone Market Forecast, August 2025

Coherent Market Insights, On-Device AI Market, 2026–2033

Future Today Strategy Group, Convergence Outlook 2026

Washington is addressing chip enforcement in two halves: location verification and remote access controls. The third gap — end-use verification at the workload level — remains open in U.S. legislation. G42 built an operational layer for it.

On March 26, the House Foreign Affairs Committee advanced the Chip Security Act, a bill requiring location verification mechanisms embedded directly into every advanced AI chip before export.

The legislation advanced one week after prosecutors unsealed the indictment of Super Micro’s co-founder for conspiring to divert $2.5 billion worth of Nvidia servers to China through Taiwan and Southeast Asian intermediaries. Since November 2025, the Department of Justice has brought at least five separate smuggling cases involving fake audit servers, shell companies in Florida, and routing through Malaysia and Thailand. The Chip Security Act addresses this specific problem: physical diversion of chips from authorized destinations. Ping-based location verification, using network latency triangulation across a dozen geographically dispersed servers, can confirm a chip’s country-level position with roughly 60-mile accuracy. Working prototypes already exist on Nvidia H100 chips. Nvidia has acknowledged its GPUs already transmit the telemetry data needed for this, and implementation could come through a firmware update.

The enforcement question, though, is what happens when verification detects a chip in the wrong country. The bill’s text explicitly prohibits kill switches: it “may not be construed as requiring any chip security mechanisms that may hinder the capability or functionality of a covered integrated circuit product, such as a kill switch or geofencing mechanism.”

The chip keeps running. The exporter is required to report the discrepancy to the Bureau of Industry and Security. After that, enforcement becomes a law enforcement matter — DOJ investigations, sanctions, diplomatic channels. The Senate version directs Commerce and Defense to study whether future mechanisms could “modify the functionality” of illicitly acquired chips, but that assessment is at least a year away from any mandate. For now, detection and reporting is the entire enforcement architecture. Spoofing individual chips through latency manipulation is possible; spoofing thousands of GPUs simultaneously in a data center cluster is substantially harder, though the Center for Cybersecurity Policy has warned that any embedded verification mechanism introduces new attack surfaces that hostile actors could exploit.

Location verification addresses physical diversion. It does not address a second, arguably larger gap: Chinese companies legally accessing advanced chips in third countries. In March, the Wall Street Journal reported that ByteDance is deploying 36,000 Nvidia Blackwell B200 GPUs through Aolani Cloud in Malaysia — a $2.5 billion cluster that will be among the largest commercial AI deployments in Southeast Asia. The chips stay in Malaysia. They are formally owned and operated by a Malaysian entity. Nvidia confirmed the arrangement complies with export controls, noting that “by design, the export rules allow clouds to be built and operated outside controlled countries.” Location verification would confirm these chips are exactly where they are supposed to be. The problem is that ByteDance’s engineers access the compute remotely from China. The House passed the Remote Access Security Act 369–22 in January 2026 to close this “cloud loophole” by extending export controls to cover remote access to controlled chips. It has not yet cleared the Senate, and enforcement — monitoring who logs into every GPU cluster globally — is substantially more complex than verifying where hardware sits.

A third gap remains fully unaddressed by either bill. The Chip Security Act verifies where chips are. The Remote Access Security Act, if passed, would regulate who accesses them remotely. Neither verifies what workloads are running on the chips or whether the entity operating them is the authorized end user in practice, not just on paper. A Malaysian cloud operator that leases capacity to a Chinese AI lab is legal under current rules, and neither bill’s enforcement mechanism reaches the workload layer. This is the structural opening that G42’s compliance architecture occupies. The Regulated Technology Environment handles chip-level compliance — the State Department described it as a “gold-standard framework for managing sensitive technologies” at the first U.S.-UAE AI Acceleration Partnership Working Group meeting on March 26. The April 6 State Department readout went further, describing the RTE as operating under “jointly agreed security and governance requirements”. Digital Embassies, launched at Davos in January, establish government-to-government legal constructs that define who has jurisdictional authority over AI workloads, regardless of where infrastructure is physically located. And crucially, Microsoft serves as the primary infrastructure partner for this framework, reinforcing the UAE’s position as a collaborative node for U.S. technology rather than a sovereign replacement for it.

Greenshield, the operational layer run by Core42, applies cryptographic controls that verify where workloads are processed and ensure they remain within authorized sovereign environments. Together, the three layers create an end-to-end compliance architecture that covers physical location, access control, and workload verification — the full stack that U.S. legislation, taken together, still leaves partially open.

The Vietnam deal shows the model already scaling. In February, G42 signed a $1 billion framework agreement with FPT Corporation and Viet Thai Group to deploy three hyperscale data centers with full sovereign cloud capacity. Vietnam gets AI infrastructure without negotiating chip allocations or export licenses with Washington directly — it buys into a G42-managed ecosystem that has already cleared those compliance hurdles across all three layers.

The Atlantic Council’s William Wechsler argued this month that the Gulf’s interdependence with the West extends far beyond energy, noting the region “will have a disproportionate role in the future of the global infrastructure for artificial intelligence, just as Taiwan today plays an outsize role in semiconductors.” Taiwan’s leverage is its position as the irreplaceable manufacturing node. The UAE is building the analogous position for compliance: the node through which verified sovereign AI access flows to the Global South.

The critical question for policymakers and infrastructure investors: if chip traceability becomes the global standard for hardware access — and bipartisan momentum suggests some version will — then verified compliance infrastructure replaces raw chip supply as the binding constraint on sovereign AI. The Semiconductor Industry Association and the Information Technology Industry Council are right that mandated tracking could push some buyers toward Chinese alternatives. That risk strengthens the case for a trusted intermediary that keeps nations inside the U.S. technology stack while providing sovereign control at the workload level. Five federal smuggling cases, a $2.5 billion legal cloud workaround in Malaysia, and an open gap at the end-use layer all point in the same direction. Whoever controls the compliance architecture that determines where, by whom, and for what purpose GPUs operate holds the next structural advantage in sovereign AI.

Sources

• H.R. 3447 / S. 1705 — Chip Security Act, 119th Congress (congress.gov)

• H.R. 2683 — Remote Access Security Act, passed House 369–22, January 12, 2026 (eeNews Europe)

• House Foreign Affairs Committee press release, March 26, 2026

• U.S. State Department — First U.S.-UAE AI Acceleration Partnership Working Group, April 6, 2026

• CAIS Action Fund — “The Chip Security Act: Separating Fact From Fiction”

• CSIS — “The Architecture of AI Leadership,” February 2026

• Center for Cybersecurity Policy & Law — analysis of CSA cybersecurity risks, July 2025

• Foundation for American Innovation — “The Chip Security Act: A Bipartisan Solution,” July 2025

• Wall Street Journal — ByteDance / Aolani Cloud / Blackwell deployment, March 13, 2026

• Tom’s Hardware — Nvidia confirmation: “export rules allow clouds outside controlled countries,” March 2026

• BISI — “AI Chip Smuggling: The Limits of US Export Controls,” April 6, 2026

• SIA Statement on Chip Security Act, March 2, 2026

• ITI — “The Unintended Consequence of the Chip Security Act,” February 27, 2026

• G42 — Digital Embassies and Greenshield launch, January 20, 2026 (Davos)

• G42 / FPT / Viet Thai Group Framework Agreement, February 9, 2026

• Atlantic Council — William F. Wechsler, “Why the Gulf Should Keep Looking West,” April 2026

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Eighteen months ago, utilities sold electricity and hyperscalers bought it. That boundary is dissolving. The AI buildout is producing vertically integrated compute-energy entities—companies that own the models, the data centers, the storage, and the generation under one roof. Three models of integration are forming: partnership, coalition, and full vertical control. Which one you build on determines which dependencies you carry.

In December, NextEra Energy and Google Cloud expanded their partnership to develop multiple gigawatt-scale data center campuses across the United States, each with co-located power generation purpose-built for the facility. NextEra already has roughly 3.5 GW in operation or contracted with Google. More interesting than the megawatts, though, is what’s happening to the utility itself. NextEra is taking Google’s AI models—TimesFM for time-series forecasting, WeatherNext for weather prediction—and applying them to its own grid operations through an internal initiative called Project REWIRE to run its power grid as a predictive system. Rather than simply adopting generic AI logic, NextEra is using these models (fine-tuned versions of them) on its own operational data and combining them with proprietary grid and optimization models. This allows the company to forecast demand, renewable output, equipment failures, and weather impacts in real time, improving reliability, reducing costs, and shifting from reactive operations to proactive decision-making—while also positioning this capability as a scalable product for other utilities.

The first commercial product is scheduled to launch on the Google Cloud Marketplace by mid-2026. Similar to how Amazon monetized its delivery network that was previously a cost center, NextEra is turning an internal operations system into a commercial software product, allowing it to sell predictive grid intelligence as a service. Petter Skantze, NextEra’s chief risk officer, oversees the program. A utility executive running an AI transformation that sells software on a hyperscaler’s platform: that is an organizational mutation rather than a procurement deal.

In parallel, at CERAWeek in March, NVIDIA and Emerald AI announced that six major US energy companies—AES, Constellation, Invenergy, NextEra, Nscale Energy & Power, and Vistra—would collaborate on flexible AI factories designed to operate as grid assets. NVIDIA’s Vera Rubin DSX reference design now includes DSX Flex, a software layer connecting compute workloads directly to power-grid services. Emerald’s Conductor platform demonstrated the ability to cut AI data center power consumption by 25% during grid stress events without degrading compute performance.

The math behind that 25% matters. Emerald estimates that if AI data centers can flex their power requirements by 25% for roughly 200 hours per year, that unlocks approximately 100 GW of stranded grid capacity across the US—more than the entire American nuclear fleet. The bottleneck, in other words, is not simply generation. It is also the inflexibility of the load. Utilities have been reluctant to approve large interconnections for AI facilities because a data center that draws 500 MW around the clock is a liability to grid stability. A data center that can throttle to 375 MW during peak stress, on software command, is something a utility can work with. That is why the orchestration software—the layer between compute and grid—has become the critical enabler for getting AI facilities connected faster.

Emerald was founded by Varun Sivaram, former CTO of ReNew Power in India and former senior US climate diplomat—a background that straddles exactly the energy-compute-sovereignty intersection. The coalition model accepts that compute and energy will remain under separate ownership and uses a shared software standard to coordinate between them.

It is worth noting what is taking shape here in aggregate.

NVIDIA, Emerald AI, Google, NextEra, and five other major US utilities are assembling a domestic compute-energy integration ecosystem where the software connecting AI workloads to grid services is becoming a proprietary American advantage. REWIRE, DSX Flex, and Emerald’s Conductor are not generic tools—they are trained on US grid data, validated against US interconnection rules, and sold through US cloud platforms. Countries building sovereign AI infrastructure today can import GPUs. They cannot easily import the orchestration layer that makes those GPUs grid-compatible at scale.

Then there is the Musk model.

When SpaceX acquired xAI in February for a combined $1.25 trillion, the headline was orbital data centers. The structure underneath is more telling. Across his companies, Musk now owns every layer of the AI infrastructure stack: compute (Colossus in Memphis, 200,000+ GPUs targeting one million), energy storage (Tesla Megapacks, over $375 million deployed at the Memphis site, with a 50 GWh/year Megafactory in Brookshire, Texas currently ramping up), power generation (a 1.2 GW dedicated plant, plus a joint venture with Solaris Energy carrying 1,140 MW in its orderbook), connectivity (Starlink, 9,000+ satellites), a proprietary data pipeline (X’s real-time feed), and end-user distribution (Grok embedded in Tesla vehicles).

Musk didn’t negotiate a PPA. He didn’t join an interconnection queue. He shipped a power plant across an ocean to Tennessee. That is vertical integration the technology industry hasn’t seen since AT&T owned the phones, the wires, and Bell Labs.

The three models sit on a spectrum of integration—and each carries different dependencies. In the partnership model, the orchestration layer (REWIRE, Emerald’s Conductor, NVIDIA’s DSX Flex) exists because the energy entity and the compute entity are separate organizations that need software to coordinate between them. In the coalition model, that software intermediary serves even more parties. In the vertically integrated model, the orchestration layer disappears entirely—there is no organizational boundary to mediate.

Build on the partnership model and you get speed, but you layer your hyperscaler’s AI deep into your energy operations. Build on the coalition model and you get vendor flexibility, but you depend on a reference architecture you don’t own. The fully integrated model eliminates those trade-offs—but it requires controlling rockets, batteries, generation, data, and distribution simultaneously, a portfolio that exists exactly once on earth.

For anyone making infrastructure or sovereignty decisions around AI, the organizational model matters as much as the technology specification. Signing a PPA with a utility is a different structural commitment than co-developing grid AI tools on a hyperscaler’s marketplace. And neither is the same as owning your own power plant. The question worth asking before your next capital allocation: does your current architecture create the optionality to deepen integration over time, or does it lock you into a permanent intermediary?

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sources

• NextEra Energy / Google Cloud partnership press release, December 8, 2025

• NVIDIA and Emerald AI, CERAWeek announcement, March 23, 2026

• NVIDIA Vera Rubin DSX AI Factory reference design, GTC, March 16, 2026

• Emerald AI Conductor demonstration, EPRI DCFlex Initiative, Phoenix AZ (July 2025)

• Emerald AI $24.5M seed round (Radical Ventures, NVentures, CRV); $42.5M follow-on (October 2025)

• SpaceX–xAI merger, February 2, 2026 (CNBC, Bloomberg, TechCrunch)

• SemiAnalysis, “xAI’s Colossus 2,” September 2025

• NextEra Energy Q4 2025 earnings; January 2026 investor presentation; NEE leadership page (Skantze bio)

Governments around the world now treat compute as strategic infrastructure. The UAE’s Stargate campus is targeting 5 gigawatts of sovereign compute in Abu Dhabi, backed by G42, OpenAI, Oracle, and NVIDIA. Saudi Arabia’s HUMAIN is scaling one of the world’s largest government-owned facilities at Hexagon in Riyadh, a 480-megawatt Tier IV complex. India is scaling from 38,000 to 100,000 public GPUs by December 2026. France, Japan, Canada, and more than 20 other governments have signed sovereign AI partnerships with NVIDIA. But the structural question embedded in all of this capital is: what workload are they building for?

The answer, almost universally, is training — massive centralized clusters designed to produce AI’s foundation models. That assumption is about to collide with a workload shift that changes the entire calculation. Deloitte estimates that inference already accounts for two-thirds of all AI compute in 2026, up from one-third in 2023. McKinsey projects inference will represent over half of all AI compute and 30–40% of total data center demand by 2030. The inference chip market is projected to exceed $50 billion in 2026, a shift that prompted NVIDIA’s $20 billion 'reverse acquihire' of Groq’s inference unit, a move designed to monopolize the low-latency 'token-shuffling' layer that now defines the global AI stack.

Lenovo’s 2026 industry analysis projects that the 80/20 split between training and inference spending is reversing. Training a model is a one-time cost that does not directly generate revenue. Serving it to hundreds of millions of users via inference is a continuous, scaling cost that generates scaling revenue — and it is now the dominant workload in the global AI stack.

This matters because training and inference have fundamentally different infrastructure requirements — and different vulnerability profiles. Training workloads are geographically more flexible: build a massive cluster wherever power is cheap and baseload is guaranteed. Inference workloads follow users. They require low-latency, metro-adjacent infrastructure close to population centers. A piece published last week in Communications of the ACM argued that edge inference is increasingly a sovereignty problem, not a latency problem; regulatory mandates under the EU AI Act, India’s DPDP Act, and Brazil’s LGPD are pushing compute to the edge faster than latency requirements alone would dictate. The implication is significant: India’s 1.4 billion potential revenue-generating inference endpoints (read, people) begin to look more strategically consequential than a multi-gigawatt training campus. The country that hosts the users generates the sustained demand that funds the ecosystem. A country that hosts only the training cluster can technically repurpose it for inference — the same GPUs handle both workloads — but at significantly worse cost-per-token than dedicated inference silicon such as Groq’s LPU or Amazon’s Inferentia, in a facility optimized for remote cheap power rather than metro-adjacent low latency, and without the user base that generates continuous demand. The hardware is not stranded, but the strategic logic behind its location, economics, and demand source is.

The current conflict in the Gulf has made this asymmetry more visible. Any transparent assessment of a sovereign compute strategy requires examining at least three dimensions simultaneously: the political trust architecture (alignment with technology partners and in-country legal and regulatory clarity), the intelligence layer (access to models, data, and talent), and the physical power base (energy, capital, grid capacity, and physical security). The war, is stress-testing the power dimension in real time, so long as Hormuz remains functionally closed. The IEA warned this week that the disruption surpasses the combined oil crises of 1973 and 1979. For Gulf sovereign AI strategies funded by energy export revenue, the disruption cuts both directions: the revenue financing the buildout, and the physical infrastructure that sits in a conflict zone where three commercial data centers have already been hit. An inference-centric world distributes this risk differently. Distributed nodes across multiple geographies are harder to disable than a single mega-cluster — though they require a trust framework, energy supply, and governance architecture in every jurisdiction where they operate. The fragility shifts from concentration to connective tissue.

For infrastructure investors and policy planners, the practical takeaway is clear. The sovereign AI strategies announced over the past 18 months were designed for a training-centric model. The workload mix is shifting toward inference faster than those plans are adapting. Countries with large, digitally active populations and regulatory frameworks mandating local processing — India, Indonesia, Brazil — may find themselves in stronger structural positions than countries that optimized for raw training throughput. The European Data Centre Association’s 2026 report signals that future capacity growth will be constrained by grid readiness, not capital. Sovereign compute strategies that can flex between centralized training and distributed inference, while maintaining trust and security across both modes, will prove more resilient than those locked into a single architecture. The sovereignty question is no longer just “can you train a model?”, but is increasingly: where do the inferences run, who governs them, and what happens when the workload your infrastructure was built for is no longer the workload that matters?

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sources

McKinsey, “The Future of AI Workloads,” February 2026

Deloitte, “Technology, Media and Telecom Predictions 2026: AI Compute”

Lenovo, CIO Playbook / 2026 Industry Analysis

JLL, “2026 Global Data Center Outlook”

European Data Centre Association, 2026 State of European Data Centers

ACM, “Inference at the Edge Is a Sovereignty Problem, Not a Latency Problem,” March 2026

G42/OpenAI/Oracle, Stargate UAE announcement, May 2025; The National, January 2026

Saudi Press Agency, Hexagon Data Center launch, January 2, 2026

India PIB, AI Impact Summit GPU announcements, February 2026

CNBC/NPR, Iran ceasefire developments, March 25, 2026

IEA Director Fatih Birol, remarks March 23, 2026

The instinct is to ask whether all of this slows the AI buildout. The critical question is which layers of the AI stack break under a sustained energy shock—and which accelerate because of it. On the software layer, the answer is unambiguous. The US military struck over 1,000 targets in the first 24 hours of Operation Epic Fury using Palantir’s Maven system integrated with AI for real-time targeting. Palantir stock is up 12% since the campaign began, its US government revenue having grown 55% year-over-year to $1.9 billion in 2025. In parallel, Bloomberg reported this week that Maxence Visseau, founder of investment firm Arkevium, used AI to cut research time by roughly 80%, stress-testing conflict scenarios across asset classes in minutes. When a CFO faces Oman crude above $165 and Brent above $110, compressing margins and spiking input costs, the business case for AI-driven efficiency stops being aspirational.

The hardware layer faces a different calculus. The Middle East AI buildout—one of the fastest-growing in the world through 2025—is obviously stalling. CSIS analysts now frame data centers as legitimate military targets in modern armed conflict. Pure Data Centre Group’s CEO told CNBC his firm would “slow down” in the region. The Atlantic Council’s Tess deBlanc-Knowles notes the Gulf remains attractive for its sovereign capital, energy, and Global South market access—but hyperscalers are running scenario analyses on redirecting next-wave capacity to Northern Europe, India, and Southeast Asia. This is geographic redistribution, though, not aggregate reduction. The CapEx still flows.

The stress-test matrix above maps where specific AI stack components begin to fracture, plotted against oil price and conflict duration. The thresholds are directional, anchored to observable data. Saudi Arabia’s fiscal breakeven sits at $78–94 per barrel depending on whether PIF spending is included (IMF, Bloomberg Economics), meaning Gulf sovereign wealth funds are currently generating windfall surpluses of $49–72 billion annualized. Goldman Sachs estimates European gas at €74/MWh would replicate 2022-level demand destruction. On March 18, TTF spiked to $24/MMBtu—approximately €75/MWh—touching Goldman’s crisis threshold intraday before settling lower, with Brent at $109. TTF has since traded in the €50–67/MWh range, trending toward the crisis line as South Pars damage compounds.

The direction matters more than any single close. Energy constitutes 15–25% of data center operating costs at baseline—but the South Pars escalation changes the timeline. Production infrastructure damage means LNG supply disruption persists beyond any ceasefire, repricing gas-dependent data center markets in Asia and Europe on a horizon measured in quarters, not weeks. Columbia’s Center on Global Energy Policy described March 18 as the moment the war moved “from logistical disruption to meaningful and potentially longer-lasting damage to core energy infrastructure assets.”

Duration remains the critical variable, but the South Pars strike has compressed the escalation curve. At three weeks with only shipping disruption, this was a market shock with geographic reshuffling. With production infrastructure now damaged, some portion of the LNG supply hit is locked in regardless of diplomatic outcomes. Extend the conflict to eight or sixteen weeks and the 1970s secular inflation analog—what Bloomberg’s Simon White calls “Act Three”—begins repricing every energy-linked asset class. Janus Henderson is already flagging that hyperscaler credit spreads, which started widening in Q3 2025, still don’t reflect the historical pattern where large debt-funded CapEx cycles produce meaningful spread widening.

The fertilizer disruption—roughly a third of global supply transits the Gulf—introduces a second-order food inflation channel that most AI infrastructure analyses ignore entirely. Several energy strategists, including Rystad and Financial Sense, frame the current disruption as the front edge of a structural energy supply cycle extending to 2030—which means this matrix is less a war-duration exercise than a preview of the operating environment for the next generation of AI infrastructure buildout.

For sovereign compute, the structural implication is clear. The gap between countries with installed AI infrastructure and those still building is widening in real time. The US benefits from its massive installed base and insulated domestic gas prices—Henry Hub near $3/MMBtu while Asian JKM trades above $20. India and Southeast Asia benefit as redirected buildout destinations. The Gulf states face a paradox: their sovereign capital has never been stronger, but the physical security case for hosting critical compute has weakened. Every government watching this conflict—from Brasília to Jakarta to Nairobi—is drawing the same conclusion: domestic compute capacity is a national security requirement, not a commercial preference. The war doesn’t slow the AI transition. It sorts who controls it.

Sources

• Columbia Center on Global Energy Policy, “Iran Conflict Brief: The High Cost of Attacking Energy Infrastructure,” March 20, 2026

• Gulf Mercantile Exchange (GME), Oman Crude OQD Marker Price, March 19, 2026 ($166.96/bbl)

• Fortune, “Oil prices hit nearly $110 as Iran vows to escalate the war,” March 18, 2026

• MacroVoices Episode #524, March 19, 2026

• EIA Short-Term Energy Outlook, March 10, 2026 (Brent $94/bbl as of March 9)

• CNBC, “How the Iran war could impact hyperscalers’ massive AI buildout in the Middle East,” March 11, 2026

• The Hill, “Iran war casts shadow over Middle East AI investments,” March 17, 2026

• Morgan Stanley, “AI Capex and the Iran War: Market Investing Risks,” March 2026

• Goldman Sachs, LNG supply and European gas price analysis, March 3, 2026

• Bloomberg, “Traders Overwhelmed by Iran News Are Turning to AI for Help,” March 19, 2026

• Janus Henderson, “How to play the AI mega-theme in fixed income,” March 2026

• IMF and Bloomberg Economics, Saudi Arabia fiscal breakeven estimates

• American Gas Association, Natural Gas Market Indicators, March 19, 2026

• CNBC, “Middle East war sends natural gas prices soaring,” March 3, 2026

• The Middle East Insider, “Saudi Arabia Economy March 2026,” March 14, 2026

• Rystad Energy, energy infrastructure supply chain analysis, March 2026

• Financial Sense, “Geopolitical Risk and Energy Infrastructure,” March 2026

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

The data points are specific.

On March 1, Iranian drones struck three AWS data centers in the UAE and Bahrain — the first physical military attack on commercial cloud infrastructure. Two of three Availability Zones in AWS’s UAE region (me-central-1) went offline simultaneously, bypassing the multi-AZ redundancy that enterprises like Emirates NBD and Careem relied on for failover.

On March 3, the Pentagon issued a FASCSA (Federal Acquisition Supply Chain Security Act) designation against Anthropic — the first time this legal mechanism, historically reserved for foreign adversaries, was applied to an American company. Claude, the only AI model cleared for classified military networks, was ordered removed within 180 days.

As of mid-March, Claude is still running the Iran targeting cycle — processing roughly 1,000 potential targets per day, per CBS News — because there is no replacement ready.

Meanwhile, SpaceX handles over 90% of U.S. military satellite launches and its Starlink supports nearly 50 military commands. During the Trump-Musk dispute last summer, Musk floated decommissioning the Dragon capsule — America’s only crewed access to the ISS.

Three different infrastructure layers.

Three different vendors.

The same structural condition: single-provider concentration where the architecture assumed redundancy.

I’ve been developing an analytical framework I call the Trust–Intelligence–Power triangle, which scores sovereign compute strategies across three dimensions: Trust (political stability and alignment with technology partners), Intelligence (access to models, data, and talent), and Power (energy, capital, and physical security). The March events tested all three legs at once. The AWS strikes hit Power. The Anthropic phaseout hit Intelligence. And the dual-use problem — military AI running on commercial cloud, which gave Iran its stated targeting rationale — hit Trust. In each case, the vulnerability was concentration.

The Trust dimension deserves particular attention.

The IRGC said it struck AWS facilities because of their role in supporting military operations. Under the Geneva Conventions, civilian infrastructure that contributes to military action can become a lawful target. When the same cloud region hosts both consumer banking and classified targeting workflows, commercial SLAs effectively become military liabilities — a risk that neither the operator nor the customer originally priced.

Iran’s state media subsequently published a target list naming Amazon, Microsoft, Palantir, and Oracle.

Lloyd’s Joint War Committee has widened the Gulf high-risk zone, maritime war risk premiums have risen fivefold, and analysts project a 15–20% security premium on new Gulf data center capital expenditure for physical hardening alone. Reinsurers are debating whether cloud-based conflict falls under standard war exclusions or requires a new category entirely.

India is the near-term beneficiary.

Submarine cable infrastructure already connects Mumbai to the Gulf. IT load capacity stood at 1.4 GW as of mid-2025, and absorption rates in Mumbai and Chennai are pushing the 2026 forecast closer to 3.2 GW by year-end — ahead of schedule. AWS advised Gulf customers to migrate workloads to alternate regions, and the network path to India’s data center hubs already exists. But the architectural question goes further than geographic diversification. Every sovereign compute strategy built on a single foundation model, a single cloud region, and a single energy supply chain now faces a concrete version of a question that was previously theoretical: how does this architecture perform under pressure? If the world’s most capable military cannot swap AI vendors during an active conflict, the answer matters for Riyadh, New Delhi, and Brussels alike.

The sovereign compute strategies being funded today will operate for a decade or more. For infrastructure investors and policymakers, the question is practical: are the architectures being built today designed for the conditions that March 2026 is making visible?

A note on independence: All opinions shared in this newsletter are my own and do not reflect the views of dmg events, ADIPEC, or any affiliated organizations. This is personal analysis, not institutional positioning.

Sources

CBS News: Internal Pentagon memo orders military commanders to remove Anthropic AI technology from key systems

Mayer Brown: Anthropic Supply Chain Risk Designation Takes Effect

Euronews: Data centres are the new target in modern warfare during Iran war

Fortune: Iranian drone attacks on Amazon data centers signal a new kind of war

Just Security: Iranian Attacks on the Amazon Data Centers: A Legal Analysis

CNBC: Anthropic’s Claude would ‘pollute’ defense supply chain: Pentagon CTO

Lloyd’s List: Gulf war risk premiums topping double-digit millions of dollars per trip

The National Interest: US Space Strategy Can’t Rely on SpaceX Alone

CSIS: If Compute is the New Oil, War in the Gulf Significantly Raises the Stakes

Capacity: The Gulf gamble: Could the war in the Middle East drive a data centre exodus to India?