America Holds 45% of the World's Data Centers. Everyone Else Must Build, Buy, or Ally

This is episode something north of 200 of *The Cognitive Revolution*, and it is the first time Nathan Labenz has put sovereign AI at the center. Governments, he says, are staring at a blunt fact: if the United States and China keep dominating talent, data, and compute, what is everyone else supposed to do? His guest is Anjney Midha, a general partner at Andreessen Horowitz, an early Anthropic investor, and someone who now spends much of his time shuttling between governments and companies. Labenz’s own card is on the table. He thinks states should take the wave seriously. He is also confused by the slogan.

Midha starts with first principles: there is no canonical definition that survives contact with a second region. Satya Nadella had just tweeted Microsoft’s European sovereign cloud; Midha went looking on the website for what they meant by sovereign AI and could not find a sentence. He now uses the question as an icebreaker with policymakers, the official-world twin of “what is your definition of an AI agent.” Ask five people, get five answers.

For a prime minister, the phrase is a way of saying you want control over your destiny. For an engineer, it tends to mean models and services that run on-prem, not on a cloud beholden to someone else’s geography. For a European CIO, it is more concrete: I want to know where my AI workloads live, which governments can send a backdoor request, and which regulations I am signing up for without noticing. For an elected official, it slides from the technical into the cultural. If a large share of the population will talk to models as daily companions, and mission-critical industries will run on them, those systems are cultural infrastructure—trained on someone’s data, post-trained on someone’s values. Sovereignty, at that altitude, is the right to refuse an adversarial country’s value system.

Labenz’s instinct, advising a boutique consultancy whose conservative enterprise clients want everything on-prem, is to push through it. Get legal comfortable with Microsoft’s terms of service. You already trust a hyperscaler with important data; Azure might even be harder to hack than your own shop. Midha answers with a supply-chain red team. An “enterprise,” in his etymology, became a 2000s word once the person who bought the software stopped being the person who used it. Top-down licensing gave way to bottom-up SaaS: swipe a card for Dropbox, Slack, GitHub, and let the CIO chase compliance afterward. Governments, he notes, are simply the largest enterprises. A minister of IT is a CIO with a flag.

Before the cloud, a CIO went to Red Hat for Linux licenses, deployment consulting, and a sentence that said: when this breaks, we are on the hook. That was a bundle of technology plus insurance. SaaS unbundled it. Individuals bought the product; risk landed on the CIO and the chief compliance officer. Hyperscalers then sold the insurance back. Use Azure, and when something goes wrong you can say it is Microsoft’s problem—job security as a SKU. The number Midha likes is the terms of service. Two years ago Azure offered no indemnity on generative-AI services. Today it indemnifies a pile of them. The insurance layer got repriced.

He thinks the industry is in a back-to-the-future moment. Large European buyers who want supply-chain independence are going to firms such as Mistral and asking for the whole package: pick the model, implement it, and stand behind the cyber and copyright risk. His example is CMA CGM, the world’s third-largest shipping company, $70 to $80 billion in revenue, trying to automate ports and cargo while drowning in a weekly flood of 7B, Mixtral, DeepSeek R1, reasoning, non-reasoning. In the United States he keeps running into Palantir in that same role for mission-critical workloads. The sorting question for Labenz’s friend is therefore simple. Are you trying to buy Azure insurance, or the best technology? Right now, if what you want is a frontier open-source model, you often cannot get both from the same place.

The case for leaving Azure is not that on-prem servers are magically safer. It is law. US cloud companies live under the CLOUD Act. If the workloads sit in mission-critical industries outside the United States, and the customers would not be comfortable with a US-government back door, Azure is the wrong bundle. A European company may still want technology, advisory, and indemnity in one contract. It may just want that contract from Mistral rather than from Microsoft. Rebundling is the trend. Jurisdiction is the reason the bundle sometimes has to be local.

The same rebundling is now happening at national scale. Stargate is a fully integrated US offering: chips, data center, models, ChatGPT. Two weeks before the taping, OpenAI internationalized the pattern with the UAE. ChatGPT would be free for every resident, the government would pay, and the workloads would run on OpenAI’s stack rather than on Azure. That, for Midha, is the first clue that hyperscalers are being taken apart and put back together by AI companies. For fifteen years the organizing principle of cloud was the marginal cost of compute and storage. Centralize the workloads, and economies of scale were almost impossible to fight. Most companies, and most countries, chose to buy instead of build.

The old purchasing triad is back: build, buy, or partner. The UAE’s move is a buy. Other governments and some companies are choosing to build—own the chips, run open-source models, handle fine-tuning, post-training, and a local ChatGPT-alike for civilians. Last week, in Midha’s timeline, Nvidia and Mistral announced Mistral Compute: American chips in what will be Europe’s single largest data center, with a cloud on top that treats open-source models as first-class citizens and lets the customer pick the application. He calls the new animal an “open scaler,” as against the closed scalers—Google Cloud, AWS—whose last fifteen years were cheaper centralized compute on proprietary technology.

A year earlier, Midha says, he would have laughed at the idea of Mistral operating tens to hundreds of thousands of Blackwell chips for European workloads. They announced it anyway. The live players rewriting where AI jobs land are no longer only the traditional clouds. Those clouds can keep the CPU workloads. The AI factory layer is up for grabs.

On talent, Labenz and Midha agree about the present. Mistral is one of a tiny number of national champions still near the frontier; Germany does not have an equivalent, and most populous countries do not either. Zuckerberg is a bazooka. You cannot bring a knife. Outside the United States, Europe, and China, private companies cannot fight that fight, which is why states are stepping in with sovereign dollars. G42 in the UAE is, in Midha’s analogy, a state-backed AWS: the capital required to play at the frontier has left the buying power of ordinary firms.

That does not mean the rest of the world is a spectator forever. After Bretton Woods made the dollar the reserve currency, countries that were neither America nor Europe’s trade partner had to decide how to amass reserves. Singapore had fewer than three million people, no oil, and no printing press. Lee Kuan Yew’s bet was to insert the island into the flow of global dollars: Middle Eastern crude arrived in containers, was refined with a value-added tax, and went west, while Singapore ran the lowest corruption and highest ease-of-doing-business scores it could manufacture. Today it has among the world’s highest GDP per capita and a top-ten sovereign wealth fund. Midha’s category for that move is a “hypercenter partner nation”—not big, not deep in talent, but essential to the flow of value. In AI, the equivalent is inserting yourself into the flow of tokens.

The UAE and Saudi Arabia are running a cousin of that play, with a crucial difference: they already have the petro-dollars. They do not need a Singapore moment. They need to convert oil into prompt-flops—buy vast Nvidia infrastructure, tell the world to run workloads there, and, over thirty to fifty years, decouple GDP from the weekly oil price. Midha says that in the Gulf, people watch a barrel the way Americans watch the Fed. Entire investment projects are green-lit or killed on that week’s number. G42’s pitch is to make inference the new number.

Labenz puts the singularity on roughly a five-year clock, so a twenty-year plan sounds long. Midha’s reply is that timelines sort countries. The UAE is more AGI-pilled than Mexico, so it is partnering with OpenAI and trying to partner with everybody, on the theory that it has five years to establish independence. Less urgent governments should still plan twenty years of useful work for a young population: post-training, integration, the job Midha personally brands as a forward-deployed AI solutions engineer. India in 1992 did not leapfrog into auto manufacturing; it invited Suzuki, created Maruti Suzuki, produced most of the country’s cars for a decade, and handed the sector to local talent. Standard Oil’s exploration in Saudi Arabia became Saudi Aramco the same way. Nvidia’s work with Mistral in Europe is, in his telling, the healthy template. If the United States does not bring allies along, they will turn to China.

Labenz’s advice to Mexico, Argentina, Australia, and the rest of the middle tier is brutally efficient: run a massive collection-and-curation project, then walk the corpus to the frontier labs on a silver platter and ask to be in the next training run. They want data; you have data. Midha agrees you cannot compete at the talent frontier *today*, and he agrees you should not try to recreate pre-training in the short term. He does not agree that you should outsource the tokenization itself. “If your culture is not tokenized, then fundamentally you’re reliant on some other country to tokenize it for you.” He calls that the most underdiscussed part of sovereign AI, especially for languages that do not sit on a European root.

There is, he adds, no secret RL recipe. Mistral published. DeepSeek put the method out. The unwritten part is the infrastructure craft of online reinforcement learning, which you only learn by doing. Start the partnerships, the universities, and the clusters now, and there is no fundamental reason Singapore, the EU, or Mexico cannot pre-train and post-train over a long enough arc. Short-term, though, a local champion should not be a pre-training lab. Pre-training gets commoditized. Find an open-source pre-training partner, because weights in hand give you far more room for on-policy updates. The value is last mile: which companies will actually consume the technology, how to productize it, how to do RL on non-verifiable rewards, the fuzzy work that needs a local hand. OpenAI and Anthropic do not have that distribution. Midha is a native Hindi speaker. ChatGPT in Hindi, he says, talks like an American tourist—accent, diction, linguistic choices. Civilians do not want to talk to a tourist, and mission-critical industries do not trust a tourist with the last mile.

Labenz floats the AI 2027 world in which frontier developers pull away, close down, and maybe even hide models. Midha’s empirical bet is lockstep. It took 26 days to go from o1 to DeepSeek R1; China has shown it can fast-follow reasoning within about 60 days. Sitting in June, six-plus months after o1, with o3 Pro out, he does not see a step-function gap to the next R1. Closed and open are moving at six months or less. A year earlier, experts had told Congress the United States was five to six years ahead of China. He thought he was on the wrong planet. DeepSeek had already been releasing strong open-source models for eight months before R1.

Do not analyze DeepSeek as a hyperscaler, he says. It is a hedge fund that came under regulatory pressure—the CCP treating market-making instability as a net negative—pivoted into frontier research, and then got a meeting with Xi Jinping to double down. Inside the current regime its job is to keep a Chinese model at the open-source frontier and adopted worldwide as soft power. The first six months of 2025 stacked R1, a Unitree humanoid, and the animated film *Ne Zha 2* into a soft-power revival after two years of real-estate malaise. The stable equilibrium he sees is division of labor: DeepSeek as the AGI research lab that vacuums talent and open-sources ruthlessly; Alibaba as the enterprise arm that turns an unwieldy mixture-of-experts model into something a business can deploy. That arrangement helps both the party and GDP, so he expects it to last.

Labenz is not Dylan Patel, but he has Deep Research. The snapshot: the United States holds about 45 percent of global data-center infrastructure; the top 25 countries hold 88 percent; the remaining 190-odd countries split the rest, many with no meaningful footprint. Compute is the most centralized layer, storage less so, networks least. His working theory has been that AI will converge to ordinary cloud. Midha splits the answer. Traditional CPU, storage, and networking will keep centralizing, and public cloud will keep being good for most software. AI workloads are a different story.

X-ray a hall and the bill of materials has changed. Ten years ago GPUs were under 10 percent of a data center. Now they are 60 to 70 percent. That is why Jensen Huang will not stop saying “AI factory.” It is not branding; the under-the-hood object is new. Developers also move clouds for models. Midha was an early Anthropic investor and watched Amazon lose hundred-million-dollar-plus contracts to Azure in 2023 because it had no GPT-4 alternative—hence, in his telling, the $8 billion Anthropic bet. Google is stitching Gemini to TPUs. Amazon is stitching Anthropic to Trainium. Microsoft has an eight-year custom-silicon effort that Midha thinks is behind, but Satya wants to control his destiny and does not really have an in-house Anthropic equivalent yet.

If two of the top three models are on non-Nvidia hardware by year-end, that is existential for Jensen. Midha’s credit to him is that he saw it and started funding AI-first clouds: CoreWeave, Nebius, and now Mistral. The end-state he sketches is three or four species: closed scalers (Google, Amazon); open scalers (Mistral Compute as a tightly integrated lab-plus-Nvidia offering); regional clouds (Nebius, CoreWeave, G42, Saudi Arabia’s newly launched Humain); and the rest of the world picking an open stack or a closed one. Eighteen months out, he will not call the traditional clouds the automatic winners of the GPU layer.

Should a country that never hosted data centers—much of Africa, probably Mexico as a net importer of US compute—now put GPUs on its own soil? Midha’s test is the defense-industrial one. Most governments do not build their own jets or tanks; they buy them, and then they insist on strategic autonomy over a few subcomponents. Europe spent seventy years outsourcing core infrastructure and is now trying to buy it back with an $800 billion defense bill after JD Vance told the Munich Security Conference that Americans would not subsidize it anymore. Entertainment for 30 percent of the population can live on a foreign cloud; if that cloud 404s, you live with it. Defense and healthcare should not. If inference is national infrastructure, you decide which slices you cannot afford to have turned off, and you put those locally.

Labenz is a self-described globalist who liked the old McDonald’s peace and hates watching mutual dependence get liquidated in a prisoner’s dilemma. Midha does not disagree about the destination. He disagrees about the precondition. If America wants that integrated future, it has to be a stable technological partner and it has to take the Chinese threat as real. His preferred template is a Marshall Plan for AI: sovereignty over the model layer, partnership on semiconductors. “It’s a huge win for America that Mistral in Europe is going with American chips and not Huawei chips.” The continent, he says, has been courted by the Chinese semiconductor industry like never before in the last eighteen months, and still chose the American stack. Closed, centralized models do not give wealthy countries the independence they want. Open models on US silicon do, and they are how he would like the world to settle into two stable blocs, because China is not pausing.

Labenz wonders whether China even has chips to sell. DeepSeek has said publicly, against the official narrative, that it is GPU-limited; domestic demand could eat every Ascend that exists. Midha’s evidence is the sanction itself. Export controls produced an enormous doubling-down on Huawei. The alternate future was a twenty-year Chinese semiconductor arc still dependent on the United States. That ship sailed under Biden, when Beijing was told it could not count on a stable supply of leading-edge American chips. Three years ago Huawei was not a frontier-chip player. It is now sovereign-backed, motivated, and closing from inference first. He would not be surprised if DeepSeek’s next model is Ascend-compatible on day one. Static snapshots of Chinese capability are, in his view, dangerous. Workloads keep getting more efficient. Decoupling at least the inference half of the ecosystem from US silicon in two to three years is not crazy; frontier training might take three to five.

The threat model, if America’s tech tree stalls and China’s does not, is not Labenz’s sarcastic “will my grandkids speak Chinese.” It is mission-critical workloads running under an authoritarian government whose worst victims, Midha says, are often its own population. He is happy to admire the engineering. He is not happy to accept the civil-liberties trade. On export policy he and Labenz are closer than the China-hawk framing: Midha wants to ship the best technology so the world, including China, adopts it, the way the US never reciprocated the Chinese internet and then lived with ByteDance. Labenz remains unconvinced that denial is worth the tension. Both treat the judgments as perishable.

What the United States gets from the Gulf deals is more concrete. Midha was not in the room for the Trump-administration announcements, but founders in his orbit have been. First, a one-to-one match: infrastructure built there has to be matched with infrastructure built here, which is foreign direct investment and domestic capacity. Second, those halls go up on the American stack, not Huawei’s; Huawei has been courting every Middle Eastern sovereign for two years. Third, allies rather than defectors. Gulf states, by doubling down on America, run the risk of alienating China; for now they seem to think the US is a reliable partner. That was not obvious under Biden, when they were being denied a local AI ecosystem and certainty of chips. Infrastructure is destiny on a twenty-to-thirty-year horizon. Lose an ally in chapter one and you have predetermined chapter later. Three to five years until Huawei can export surplus capacity sounds long until you notice that the first Ascend rack writes the path for the next 500 megawatts and the 500 after that. On a per-FLOP basis, a liquid-cooled Blackwell node in the UAE or Saudi Arabia is, after Midha handicaps the marketing, about 18 to 20 percent cheaper on energy than the same node in an unsubsidized American site, partly because some of those plants are cooled with oil. Europe can still take two years of carbon approvals. The Gulf mostly does not.

Labenz’s remaining questions are personnel, kill switches, and missiles. Do Americans have to fly over to build the halls, and can Washington order them home? He has heard an unverified rumor of a remotely controlled off switch. Aircraft carriers have been dodging missiles; a gigawatt box in the desert cannot. Does the deal silently expand the American defense umbrella? Midha’s frame is dual-use. A classified frontier facility for mission-critical workloads does not get built like a Google hall and does not show up in ordinary filings. Training and inference are already split, and military inference does not need 100,000 Blackwells in one place. The building can look like a department store. If an adversary finds it, the playbook is the host country’s UAV and air-defense playbook, not a special AI treaty. He does not think the physical footprint requires a net-new American expansion.

Weights are the unsolved half. Open source is effective, Midha says, because if your entire national-security apparatus is predicated on the idea that weights do not get stolen, he does not know how to protect you. Chinese citizens have been caught boarding planes from the United States with TPU schematics. There is no paradise inside US frontier labs that prevents exfiltration. If China wants the weights, it is getting the weights. Labenz’s corollary is that nobody should expect the UAE to succeed at a control problem America has not solved.

They never got to Taiwan. Midha wants a part two, because how Taiwanese sovereignty goes over the next two or three years will rewrite the rest of the compute-and-sovereignty script. Labenz files it next to his standing objection to export controls: whatever else they do, they make a big move on Taiwan easier, not harder. Cycle times, they agree, are shrinking. Everything in the conversation should be treated as having a shorter and shorter half-life. See you Monday.