China wants to spend $295 billion building AI out of its own chips. The hard part is a stack of memory.

The most expensive sentence in China's reported plan to spend roughly $295 billion building artificial intelligence on its own silicon is not about money at all. It is about a component most people have never heard of, made by a handful of companies, in volumes that no amount of state capital can conjure on command. The plan, as Bloomberg described it this week, is a five-year national buildout of interconnected computing hubs, drafted by the National Development and Reform Commission, operated largely by state carriers, and stocked — for at least 80 percent of the critical technology — with domestic chips rather than Nvidia's. It is an enormous, coherent industrial ambition. And it runs, the way these things always do, into a chokepoint somewhere far upstream that the headline number cannot touch.

Start with what is actually being proposed, because the figure has already been flattened into a slogan. Reports describe the plan as early-stage — 2 trillion yuan, around $295 billion, spread across the next five years, with no fixed allocations or milestones yet locked in. State firms including China Mobile and China Telecom would own and run the bulk of the facilities and, crucially, wire them together into something closer to a single national machine than a scatter of competing private clouds. Add the grid connections and the power, and people briefed on the discussions told Bloomberg the all-in number could drift toward 5 trillion yuan. Treat all of it as reported intent, not signed budget. Beijing has announced computing-infrastructure pushes before; this one is distinguished less by its size than by a single hard requirement written into it.

The 80 percent rule

That requirement is the 80 percent. The blueprint, as described, would have these new data centers source at least four-fifths of their core technology — AI accelerators above all — from domestic suppliers, with Huawei the obvious anchor. The practical effect is to design Nvidia and AMD out of the country's flagship AI buildout almost entirely. For anyone who has watched the last three years of export controls, this is less a shock than a formalization. In May, Nvidia's own chief executive said the quiet part plainly: speaking about the Chinese market, Jensen Huang said the company had 'really largely conceded' it to Huawei. Nvidia's share of China's AI-accelerator market, once dominant, has by several accounts collapsed toward zero. A government rule mandating domestic silicon is, in that light, mostly ratifying a market that Washington's controls and Beijing's response had already split in two.

And Huawei has spent those three years building to meet exactly this moment. The company is reported to be on track for something like $12 billion in AI-chip revenue this year, up from roughly $7.5 billion last year, on orders already booked from Alibaba, ByteDance and Tencent — the same firms that, not long ago, were buying Nvidia by the container. Huawei has laid out a roadmap of new parts, the Ascend 950, 960 and 970, to arrive over the next three years, and it has a system-level answer to Nvidia's rack-scale machines: the CloudMatrix 384, which lashes 384 Ascend 910C processors together into a single AI 'supernode.' On paper, a domestic stack exists, from the accelerator to the cluster. The question the $295 billion cannot answer is whether the factories behind that paper can actually produce it at national scale.

This is the part of the story that the money tends to obscure, and it is the part worth slowing down for. A chip is not one thing. It is a logic die, etched at the leading edge of lithography; a stack of high-bandwidth memory bolted to it; and an advanced package that fuses the two into something a server can use. China's progress is wildly uneven across those three layers, and a buildout is only as fast as its slowest one.

The logic is closer than you think

Take the logic die first, because it is the layer where China has surprised the skeptics. SMIC, the country's leading foundry, makes Huawei's Ascend processors on a 7-nanometer-class process without access to the extreme-ultraviolet lithography machines that TSMC and Samsung use for the same node. It does this the hard way, with older deep-ultraviolet tools and multiple patterning, and for a long time the yields were punishing — a large AI die is exactly the kind of part that multiplies every defect into a thrown-away wafer. But the numbers have moved. The Ascend 910C is reported to have climbed to roughly 40 percent yield, enough to make the line profitable for the first time, with 60 percent — the rough industry comfort line — as the stated goal. Analysts who track the fabs closely now argue that SMIC's advanced-node capacity, on the order of 50,000 wafer starts a month, is sufficient to feed more than a million Ascend units a year. The logic, in other words, is no longer the wall.

China procured enough high-bandwidth memory for 1.6 million accelerators. Its domestic supplier can make enough, next year, for perhaps 300,000.

The memory is. Every modern AI accelerator depends on high-bandwidth memory — HBM — the stacked DRAM that sits beside the logic and feeds it data fast enough to keep the matrix multiplications from starving. It is one of the most difficult mass-produced objects on earth, and it is made, at the leading edge, by three companies: SK Hynix, Samsung and Micron, none of them Chinese, all of them constrained by the same export rules that pushed Nvidia out. Here the arithmetic turns brutal. By the accounting of SemiAnalysis, which tracks this supply chain part by part, China stockpiled roughly 13 million HBM stacks — enough, in principle, to build out around 1.6 million Ascend 910C packages. But that was a one-time hoard, bought ahead of the controls. The domestic replacement, CXMT, is expected to produce only about 2 million HBM stacks next year. That is enough memory for somewhere between 250,000 and 300,000 accelerators.

Hold those two numbers next to each other, because their gap is the whole story. The foundry can supply logic for more than a million Ascend chips a year. The country can supply domestic memory for fewer than a third of a million. A data center cannot run on a die without memory any more than an engine can run without fuel, and you cannot close that gap by appropriating another trillion yuan. HBM capacity is built one painstaking fab at a time, gated by tools and process knowledge that money accelerates only slowly. This is the single point of failure the $295 billion headline walks straight past: the bottleneck is not the chip, and it is not even the dollars. It is a stack of memory.

Brute force has a power bill

There is a second cost hiding in the design, and it is one Huawei has been candid about in its own way. Because each Ascend chip is less capable and less efficient than the Nvidia part it replaces, China's strategy at the system level is to use more of them — to win on scale what it cannot yet win on the transistor. The CloudMatrix 384 is the emblem of this: where Nvidia's comparable rack fuses 72 chips, Huawei's uses 384, and it reportedly draws substantially more power to deliver competitive aggregate performance. That is a defensible engineering choice when electricity is plentiful and politically subsidized, which in China it often is. But it means the same compute target requires more dies, more memory stacks, and more grid — which loops straight back to the HBM ceiling and to the buried 5-trillion-yuan power number. The brute-force path makes the memory chokepoint tighter, not looser, because every unit of compute now wants more chips behind it.

None of this means the plan fails. It means the plan has a speed limit that is set physically, not fiscally. A useful way to read the $295 billion is as a demand signal: a state guarantee that whatever Huawei and SMIC and CXMT can produce, there will be a buyer, a site and a grid connection waiting. That is not nothing. Guaranteed demand is precisely what lets a supplier justify the next fab, and a domestic HBM industry that can today serve 300,000 accelerators might, with that guarantee, serve a million in a few years. The buildout is better understood as a forcing function for the memory and packaging industries than as a thing that gets built on schedule. As one closely-read China-technology analyst argued this week, capital cannot simply buy its way past a chip constraint; the constraint has to be manufactured away, link by link.

The mirror image of America's bottleneck

What makes this moment genuinely interesting is the symmetry. On the other side of the controls, the United States has its own buildout running into its own physical wall — not memory but power, the transformers and interconnection queues that have turned the American AI story into a five-year wait for grid hardware. Two superpowers, two enormous piles of capital, each colliding with a different unglamorous chokepoint that no amount of announcement can dissolve. China can make the logic and is short the memory. America has the memory and the chips and is short the megawatts. The export-control regime that severed the two systems guaranteed that each would have to rediscover, the hard way, that compute is a physical good with a brutally concentrated supply chain — and that the last link in that chain, wherever it sits, is the one that sets the pace.

So watch CXMT, not the press conference. The interesting number over the next eighteen months is not 2 trillion yuan or 5 trillion yuan. It is how many HBM stacks come off a Chinese line a year from now, and at what yield. That figure, made in a few fabs by a few thousand people, will decide how much of this $295 billion actually turns into running silicon — and how much remains, for a while longer, a very well-funded intention. The money was always going to be available. The memory is the thing that has to be made.