China Drafted a $295 Billion State AI Grid. The Compute Race Now Runs on Two Different Rails.

Bloomberg surfaced the draft on June 9, and the follow-ups landed across the week: China's National Development and Reform Commission is preparing a five-year plan to spend roughly 2 trillion yuan, about $295 billion, on a nationwide AI data center network. State carriers China Mobile and China Telecom would operate the bulk of the facilities. At least 80% of the underlying technology, AI chips included, must come from domestic suppliers led by Huawei. The financing rails are sovereign debt and ultra-long special government bonds. The grid is supposed to connect by 2028.

The headline is the scoreboard read: $295 billion is about a year of US data center construction spending right now (US construction spend on data centers cleared $50 billion in April alone, per the Census Bureau monthly series). On cash burn the plan does not dwarf the American buildout. On structure it is something the American buildout does not have at all: a single national-scale operator stack, financed off the sovereign balance sheet, with a procurement mandate that excludes the foreign vendors who are doing 80% of the work on the other side.

The Numbers

Two facts about the plan are not in the headline and matter most. First, the operator layer is the state telco duopoly. China Mobile and China Telecom run the fiber backbone they will need to stitch the grid together; putting them in charge of the compute layer on top of that fiber consolidates the network and the compute substrate inside the same national balance sheet. Second, the financing is overwhelmingly fiscal. Ultra-long special government bonds are not commercial paper. They are an instrument the Ministry of Finance issues for strategic infrastructure, with tenors past ten years, often past thirty. Whoever holds them is implicitly underwriting national policy, not a cash-flow forecast.

Two Rails, One Bottleneck

The cleanest way to read this is against the American rail. We laid out the math on Anthropic's $200 billion five-year commitment to Google TPU in May. Anthropic is funding the deal off a $30 billion run-rate revenue line and an equity stake that Google is recycling back into compute spend. Microsoft and OpenAI ran the same mechanic with the Maia inference deal. The American compute rail is private, equity-backed, demand-pull. The lab promises a forward revenue curve, the hyperscaler books that curve as backlog, and the chip supplier scales to a customer-named ramp.

The Chinese compute rail is the photographic negative. State telcos take the operator role hyperscalers play in the US. The Ministry of Finance takes the capital-markets role venture and private credit play in the US. Huawei takes the supplier role TSMC and Nvidia play in the US, with the caveat that the demand is by mandate, not by purchase order. The two rails converge at the same physical bottleneck (power, fab capacity, high-bandwidth memory) but the failure modes are different. A misallocated US buildout means a hyperscaler eats a write-down and the lab renegotiates. A misallocated Chinese buildout means an off-balance-sheet liability sits on a state telco for a decade.

The trade-off cuts both ways. The American rail can reprice quickly when demand softens; when a labs revenue line undershoots, hyperscaler capacity migrates to the next customer on the queue, the way it did after the Anthropic Fable 5 disablement we covered earlier this month. The Chinese rail cannot reprice the same way because the customer is the state. What it can do, that the American rail cannot, is internalize the externalities. Power transmission, substation siting, water permits, fiber rights of way, and chip procurement all run inside a single planning process. That is why the $295B compute number quietly balloons to $735B when grid upgrades are folded in. The Chinese plan is being scoped at the level the US plan was supposed to be scoped at when AI was being framed as a national-security item, and largely is not.

The Huawei HBM Ceiling

The 80% domestic technology floor is the part of the plan that is technically the hardest to deliver, and it is the part that decides whether 2028 is a real target or a slide. Two numbers tell you why. Huawei shipped around 812,000 AI accelerators in 2025, and the processor business is projecting roughly $12 billion in 2026 revenue. Nvidia, by comparison, ran roughly 4 to 6 million data-center-class accelerator shipments in the same window depending on whose count you trust. The Chinese plan needs Huawei to scale Ascend output multiple times over and to do it on a domestic high-bandwidth memory supply chain that has not yet demonstrated the volume.

HBM is the silent constraint. Every modern AI accelerator pairs a logic die with a stack of HBM dies sitting on a silicon interposer; without HBM you do not have an inference chip, you have a digital signal processor. SK Hynix, Samsung, and Micron control the global HBM supply; CXMT in China is the named domestic alternative and is still ramping. The same Huawei Ascend cluster that trained Z.ai's GLM-5.2 in our June 13 piece is the physical demonstration the plan is built on; scaling that demonstration to a national grid is the part of the math that is not yet proven. If domestic HBM does not arrive at volume by 2027, the 80% floor either slips to something softer or the 2028 grid completion does.

What It Does to the Frontier Race

The structural answer is that the compute race stops being a single global queue and becomes two separate queues with different price discovery, different procurement, and different points of failure. That is not a return to a closed border; both rails still touch each other at the chip layer (Nvidia continues to design China-specific parts that fit under whatever export ceiling the US sets at any given week, and Chinese frontier labs still publish open weights into the global ecosystem). It is a return to two markets, where the marginal price of a token in 2028 is going to be set by the cheaper rail at any given moment, not by a single global clearing price.

For US labs the procurement read is unchanged: the equity loop that funded the Anthropic-Google relationship and the Microsoft-OpenAI relationship is still the way the forward compute curve gets pre-financed, and the sovereignty-as-procurement bundle Anthropic just installed in Seoul is the way the labs answer the next jurisdiction that wants to opt out. For Chinese labs the read is that the inference floor at home is a policy variable, not a market price; if the state grid delivers, Chinese token economics are going to undercut anything an American lab can offer inside the wall.

For builders shipping into both markets, the practical implication is a routing question that no major framework handles cleanly yet: which inference call lands on which rail, under what compliance posture, with what pricing reference. Multi-region routing solved the last decade of latency arbitrage. The next decade is going to be multi-rail routing, and the rails are not going to share a billing surface.

Our Take

The plan is a draft. The $295B is a planning number, not a contracted spend, and the NDRC has surfaced numbers like this before that landed at half size by execution. What matters is the institutional commitment to a parallel compute fabric on a separate balance sheet, not the precise figure. Even at half size, the structural fact that Chinese AI compute will be supplied, operated, and financed inside a single sovereign loop is a different category of bet than the American hyperscaler loop, where the lab, the cloud, and the chip vendor are still three different income statements.

The deeper signal is the deadline. 2028 is the same horizon every American frontier deal we have written about this year is targeting: it is when the Anthropic-Google gigawatts start to land at scale, when Microsoft Maia second-gen is supposed to anchor inference, when the Vera Rubin platform reaches steady state on the Nvidia side. Both rails are racing to the same wall, and both are pre-financing capacity that physically does not yet exist. Whoever lands the 2028 grid first sets the inference floor for the back half of the decade.

Three signposts in the next ninety days. First, whether the NDRC plan moves from draft to a formal Two Sessions follow-up with named bond issuance; that is the trigger that converts the $295B number from a target into a budget. Second, whether CXMT or a peer announces a credible HBM3e-class roadmap; without it the 80% mandate has a 2027 cliff. Third, whether the US responds with anything that looks like an industrial-policy match on the operator and grid side, or whether the American buildout continues to run as a pure private-balance-sheet exercise. The first two answers come from Beijing. The third comes from Washington and is the more interesting one.