Compute, Code and Robots: Who Is Winning the AI Stack

Opening

Welcome to today's China report. Today we're on technology competition, compute, the AI stack, and where the race actually gets decided. Four reports, all published this spring and summer, and they don't agree. Carnegie says it comes down to megawatts and months. Bruegel says it comes down to software nobody outside the industry has heard of. And two reports from MERICS in Berlin look at Europe's response and at China's robots. Let's take them one at a time.

Carnegie on the race to build compute

First, the Carnegie Endowment for International Peace, published on the eighth of June, by Alasdair Phillips-Robins, Teddy Tawil and Sam Winter-Levy. It's called "The Compute Coalition."

The claim is that the AI race is a construction race. The United States currently hosts about seventy-five percent of the world's advanced AI compute clusters, China about fourteen percent, the European Union under five. And the thing that decides where the next tranche gets built isn't cheap electricity or tax breaks. It's what they call time to power, how many months from decision to a running data centre.

Here's how they get there. They build a financial model of a hypothetical hundred-megawatt data centre, run it over twelve years across ten countries, and compare what different variables do to lifetime returns. A one-year delay in opening costs about five hundred and fifty million dollars. Fifty percent higher electricity prices? Two hundred and twenty-one million. A twenty percent tariff on the servers? A hundred and seventy-two million. Delay is the biggest single lever, by a wide margin. And their time-to-power numbers are brutal for Europe: twenty-two months in the United Arab Emirates, twenty-four in the US, thirty-four in the UK, forty-two in Germany.

That's a genuinely useful finding, and it cuts against a lot of current policy. The report says plainly: don't subsidise electricity, these projects are already profitable, fix the queue.

But look at what the model rests on. It's built from public data, equity research and industry interviews, and the authors say themselves it shows trends, not forecasts, and that they're working around proprietary gaps. The load-bearing assumption is revenue: roughly two hundred million dollars a month from a hundred-megawatt facility at full capacity. Every number I just quoted is a fraction of that assumption. If AI revenue disappoints, the cost of delay shrinks, and so does the urgency the whole report is built on.

And here's the awkward bit. The framing is democracies versus authoritarians. But the model's own fastest builder is the UAE, at twenty-two months. The report then spends real effort modelling war risk in the Gulf, a prolonged conflict putting Emirati returns seven point eight percent below the US, to get to the answer it wants. That's modelling doing political work. What would settle it is boring and public: actual utilisation and revenue per chip-hour from the people running these clusters.

Bruegel on Nvidia's software moat

Second, Bruegel, in Brussels, published the twenty-second of June, by Alicia García-Herrero and Bertin Martens. "Stack battles."

Their argument is that counting chips misses the point. Nvidia holds roughly half the world's installed AI chip stock but as much as two-thirds of the computing capacity. The reason that's durable isn't the silicon, it's CUDA, Nvidia's software layer, with a two-decade head start. Developers write in PyTorch, tuned for CUDA, and moving off means rewriting code, losing optimised libraries and not knowing what performance you'll get. That lock-in is where Nvidia's seventy to eighty percent chip margins come from.

So the mechanism is network effects, and Huawei is attacking exactly that. It open-sourced its CANN software stack in August twenty twenty-five. It ships a plugin that lets ordinary PyTorch code run on Ascend chips, aimed straight at the single biggest switching cost. And DeepSeek's open-weight models are engineered to run on both. If Huawei drains the moat, even only inside China, then America's hardware and software advantages stop reinforcing each other. That's the conclusion.

Now, what's it resting on? Mostly public market estimates, corporate announcements and stated strategy. Which is fine as far as it goes. But this is an argument about network effects with no measurement of the network. There's no count of CANN contributors, no share of Chinese AI workloads actually running on Ascend rather than smuggled or stockpiled Nvidia, no figure for how much of PyTorch works on Ascend today. The two counter-facts they do offer, that Ascend burns two to four times the energy for equivalent output, and that Chinese developers report serious usability problems with CANN, are asserted, not benchmarked.

And there's a tension they don't resolve. They also say Nvidia has largely conceded the Chinese market. If that's right, Huawei's ecosystem is growing because of export controls and subsidy, not because it's winning developers on merit, which tells you much less about whether it travels. The test is simple: does a single serious lab outside China choose to train on Ascend.

MERICS on Europe's fragmented China policy

Third, MERICS, the thirtieth of June, "Fragmented Europe," edited by Bernhard Bartsch and Claudia Wessling, with twenty-four national chapters and an EU chapter from the European Think-tank Network on China.

The claim is in the title. Europe doesn't have a China technology policy, it has about twenty-five of them. Brussels has moved from open collaboration to de-risking, excluding Chinese entities from much of Horizon Europe, tightening investment screening, while member states pull in different directions.

The method is bottom-up: a domestic expert per country, then sorted into groups. France, Germany, Italy and the UK as most affected. Hungary, Ireland, Portugal, Spain as prioritising cooperation. The Baltics and others as barely engaged. The spread is the evidence.

The concrete parts are the good parts. Over eighty percent of Europe's solar inverters come from China, with more than two hundred gigawatts of European solar capacity sitting behind them, Lithuania banned Chinese inverters in twenty twenty-three, most of the EU didn't. That's a countable physical dependency and a countable policy divergence in one sentence. The Nexperia episode is the other: seventy to eighty percent of that company's wafers went to China for processing, Beijing blocked shipments, Volkswagen and Bosch cut production.

Where it's thin: fragmentation is demonstrated more than measured. Twenty-five authors writing to a loose template will produce twenty-five different-sounding chapters, and the evidence base swings wildly, the Germany chapter leans on national statistics and industry associations, others on much less. There's no common indicator applied across all of them. And the report cites its own survey, where nearly eighty percent of China experts expected major Chinese AI progress in twenty twenty-six. That's the network polling itself. Treat it as mood, not measurement.

MERICS on China's robotics bet

Fourth, also MERICS, from the thirtieth of April, "Embodied AI," by Wendy Chang, Rebecca Arcesati and Altynay Junusova.

The headline: China made about twelve thousand eight hundred humanoid robots in twenty twenty-five, roughly ninety percent of the world's total. But the report refuses to be impressed. The kickboxing and dancing robots you've seen are mostly teleoperated or preprogrammed. Deployment is small-scale, single-point trials.

The logic is the electric-vehicle playbook: China already has the batteries, motors and sensors, and Chinese firms hold sixty-three percent of the key companies in the global humanoid component supply chain. Add a trillion yuan of state guidance-fund money over twenty years, and unit costs fall from three to five hundred thousand yuan toward the hundred and sixty thousand that analysts reckon makes them commercially viable. Hardware stops being the constraint. Autonomy software becomes the constraint, and that still runs on Nvidia.

Check the sourcing, though. The twelve thousand eight hundred figure is a state planning agency estimate, the same body that set the target, counting units produced, not sold or working. The sixty-three percent is proprietary bank research you can't reproduce. And "half as efficient as a human" comes from one founder talking down his own product. The claim I'd hold loosest is the one about seventy percent of manufacturing jobs being replaceable; that's a cited outside estimate, not the report's own work.

The report is right that the software gap is the real story. What it doesn't test is whether the Nvidia dependency is eroding, which is the exact question Bruegel spends a whole paper on.

What to watch

So: Carnegie and Bruegel genuinely disagree. One says the race is won in megawatts and permitting queues, the other says it's won in a developer toolchain. They can't both be the decisive variable.

Three things would move me. Whether any lab outside China chooses to train on Huawei silicon. Whether any democracy actually gets under twenty-four months to power. And whether anyone publishes humanoid deployment hours rather than production counts. And notice that Bruegel and both MERICS reports share one premise, that Nvidia is the chokepoint. That's three reports agreeing, but they're reading the same handful of sources. Agreement like that is thinner than it looks.