When "Sovereign AI" Means Buying NVIDIA in Bulk

I have spent the last few weeks watching the sovereign AI story turn into something different from what the press releases promise. At first the numbers look enormous. For instance, NVIDIA booked over $30 billion in sovereign AI revenue last fiscal year, roughly triple the year before. Meanwhile, McKinsey sizes the category at $500 to $600 billion by 2030. Yet when I pull on the threads, most of what gets called "sovereign AI" is governments buying the same chips Microsoft buys.

So here is my read on where this goes.

Stargate is leakier than the headlines

The flagship project announced at the White House in January 2025 promised $500 billion and 10 gigawatts of capacity. Yet by February 2026, The Information reported the Stargate entity had not even staffed up. The Texas site is real and shipping, but the international rollout is falling apart faster than anyone wants to admit.

Stargate UK was paused in April 2026 after OpenAI looked at British industrial electricity prices and walked. Then the Norway project transferred to Microsoft after OpenAI pulled out. Meanwhile, Argentina is structured so that Sur Energy carries the construction risk while OpenAI signs on as a future customer. Only Stargate UAE is moving ahead of schedule.

Beyond the field offices, OpenAI now sits on roughly $1.4 trillion in announced infrastructure commitments against $20 billion in annualized revenue. Oracle's order backlog ballooned to $523 billion, and OpenAI represents over 57% of it. Moody's has called Stargate "effectively one of the world's largest project financings." That is not a compliment.

The Gulf bought its way to the front

Here is the move that shaped 2025 more than anyone expected. Trump rescinded the AI Diffusion Rule on May 13, two days before it was scheduled to take effect. So he replaced country tiers with bilateral deals. Then he flew to the Gulf and came back with roughly $2 trillion in pledges.

Saudi Arabia's HUMAIN, run by Tareq Amin, assembled an extraordinary deal book in six months. First, NVIDIA committed up to 600,000 GPUs over three years. Then AMD added $10 billion. After that, AWS, Qualcomm, Google, and Oracle layered in tens of billions more. HUMAIN even put $3 billion into xAI's Series E. That stake then converted to equity in the SpaceX-xAI merger at $1.25 trillion.

The UAE went bigger. Stargate UAE sits inside a 5 gigawatt campus. Fully built, it could host more B200 chips than every other announced AI project combined. UAE gets 500,000 NVIDIA chips a year through 2027. G42 takes 20%, with US firms operating in UAE getting the rest.

So what did the Gulf trade for this access? Capital, mostly, plus cheap energy and convenient geopolitics. Behind it all sits MGX, the Abu Dhabi vehicle chaired by Tahnoun bin Zayed. It now holds stakes in OpenAI, xAI, Anthropic, Databricks, and Mistral. Beyond those, the fund deploys $10 billion annually toward $100 billion in assets.

Europe is quietly conceding it cannot scale alone

Brussels announced €200 billion through InvestAI and €20 billion for AI Gigafactories. The expressions of interest came in fast, 76 of them across 16 member states, demanding over 3 million GPUs. Yet the formal call slipped. Meanwhile, CEPS analysts have warned that existing EU AI Factories risk becoming "cathedrals in the desert."

Then came the tell. In April 2026, Cohere and Aleph Alpha merged at roughly $20 billion. Schwarz Group, the German company that owns Lidl, wrote the check. Beyond that, the combined entity runs on STACKIT, Schwarz's sovereign cloud, under a new Canada-Germany Sovereign Technology Alliance.

Read that again. Europe's flagship sovereign AI play is now a transatlantic Canadian-German construct backed by a supermarket conglomerate. Mistral remains the only credible European model lab, valued at €11.7 billion. That is roughly 36 times smaller than OpenAI's last secondary. ASML taking 11% of Mistral was strategically sharp, since it ties Europe's only frontier lab to its only critical-tech monopoly. Even so, the gap is enormous.

Asia is running three different playbooks

India has installed over 38,000 GPUs under its IndiaAI Mission, with subsidies as deep as 100% for foundation model training. So Sarvam was selected for a 120 billion parameter sovereign model. Meanwhile Krutrim, the rival, has lost ground so visibly that its founder pledged Ola Electric shares as data center collateral.

Japan is running a defense-tinged play. For example, Sakana AI raised $135 million at a $2.65 billion valuation with In-Q-Tel, the CIA's venture arm, participating. Plus Rapidus has consumed ¥1.7 trillion in subsidies and is targeting 2nm mass production by 2027.

China, however, is the most interesting. Last December, SMIC confirmed 5 nanometer production using older equipment. Then Huawei's CloudMatrix 384 reportedly delivers twice the dense compute of NVIDIA's GB200 NVL72. The catch is that it burns four times the power. Looking ahead, Bernstein expects Huawei to capture 50% of China's AI chip market in 2026 with NVIDIA collapsing to 8%.

Power, not silicon, is the real bottleneck

Beyond all the chip drama, the binding constraint by 2027 will be electricity and water. For example, Ireland's data centers already consume 22% of national power, up from 5% in 2015. Meanwhile Singapore has held a moratorium since 2019. As for Germany, industrial electricity at roughly €0.15 per kilowatt hour is why Stargate UK collapsed.

By contrast, Gulf electricity sits at $0.05 to $0.06 per kilowatt hour. Plus the Barakah nuclear complex is running, and Saudi Arabia's first reactor is queued up. So Texas, Louisiana, the Midwest, and parts of Patagonia win on this map. Yet Ireland, Germany, Singapore, the Netherlands, and parts of California lose.

Water adds a second layer. Saudi data centers consumed 15 billion liters in 2024, with projections hitting 426 billion liters a year by 2030. Beyond that, Khalifa University researchers warn of 50x desalination expansion raising Gulf seawater by 0.6 degrees, with marine damage to follow.

So what does sovereignty really mean now?

Here is where I land. Real sovereignty needs four things at once: compute, capital, data, and energy. So far, only one country has all four. Then there is China, with an autarky that does not connect to Western markets but is closing the gap on its own terms. Everyone else is buying differentiated sovereignty wrappers around imported compute. Meanwhile, Gulf capital writes the largest checks while Brussels writes the most regulations.

The "sovereign AI" label, in plain terms, is mostly a framing device. In practice, it sells NVIDIA chips to governments at full price. So the premium accrues to the chipmaker, not to domestic value capture. Beyond models, sovereign cloud has become a resilience purchase. Think insurance for European banks worried about DORA compliance, not a true independence play.

What changes this picture by 2027 is not another export-control round. It is whether Anthropic's compute-efficient scaling re-rates the gigawatt thesis. Or whether Oracle's concentration risk forces a credit event. Or whether Huawei and DeepSeek-style efficiency pull the Global South east. Watch those three. The rest is press release theater.