China's $295 billion AI buildout writes Nvidia out. The chip gap is the part that doesn't add up yet.
Jump to a section
See the story at a glance
China is planning to spend 2 trillion yuan — Bloomberg reported this June 9, 2026 as $295 billion — over five years on a network of AI data centers across the country, with at least 80% of the hardware coming from domestic chip suppliers.
That last number is the story. Not the total. The 80%.
The National Development and Reform Commission is drafting the blueprint. China Mobile and China Telecom will operate the bulk of the infrastructure. Funding flows through sovereign debt — ultra-long-term government bonds and state investment funds. The approved domestic chip suppliers, cleared in May 2026 for government and security-sensitive sectors, include Huawei, Alibaba's chip unit, Biren Technology, and Moore Threads.
What this does: it formally excludes Nvidia and AMD from what would become the world's largest state-directed AI compute buildout. Not a side effect. The design.
→ Source: Council on Foreign Relations — China's AI Chip Deficit
Source spread
- Bloomberg — China Preps $295 Billion Plan to Fund Nationwide AI Buildout [builder] — primary reporting, June 9, 2026; all dollar figures, agency names, and structure originate here; note: paywall, and Bloomberg notes the plan remains in early discussions
- The Decoder — Beijing's $295 billion buildout would require 80% domestic chips [skeptic] — analysis of the mandate's practical consequences; names the approved suppliers
- Council on Foreign Relations — China's AI Chip Deficit [skeptic] — technical capability assessment of Huawei vs. Nvidia; includes DeepSeek's own evaluation of the 910C at 60% H100 throughput
- WebProNews — Huawei to Double Ascend 910C Output [hype] — Huawei's 2026 production targets: 600,000 Ascend 910C dies, 1.6 million total Ascend dies across the product line
Pros & cons
What's significant:
- $295 billion in sovereign-backed compute spending, over five years, from a government with a track record on infrastructure delivery, is real money pointing at a real outcome. China built the world's largest high-speed rail network using the same sovereign-debt playbook. Data centers are less structurally complex than rail.
- The May 2026 domestic chip approvals — nine categories, covering Huawei, Alibaba, Biren, and Moore Threads — give the 80% mandate an actual supply chain, not just a policy requirement. These chips exist. They're in production.
- For inference workloads specifically, the Ascend 910C is usable. Alibaba, Tencent, and DeepSeek already run production inference on it. The gap to Nvidia matters less for serving a query than for training a new model.
- Huawei is targeting 600,000 Ascend 910C chips in 2026, nearly doubling 2025 output, with a next-generation chip (Ascend 950PR) in the pipeline for 2026. This is not a paper program.
What doesn't add up yet:
- The Ascend 910C runs at roughly 60% of H100 BF16 throughput. At 80% of total chips in a national fleet built on 910C hardware, a frontier training run starts with a substantial efficiency deficit before the first GPU call. That's not a policy problem. That's a physics problem.
- Frontier model pre-training — the kind requiring tens of millions of GPU-hours on tightly coupled, homogeneous hardware — is where the Huawei gap is most severe. The plan's ability to produce models competitive with GPT-6 or Claude Fable's successor is a genuinely open question.
- "At least 80% domestic chips" is a mandate, not a capability. You can mandate the percentage. You cannot mandate the performance. If the gap between Ascend and Nvidia doesn't close substantially by 2028, the mandate becomes its own ceiling.
- Bloomberg noted explicitly that the plan remains in early discussions and details could change. Treat the numbers as directional, not final.
- Sept 2025
- May 2026
- June 9 2026
- 2026–2031
What builders need to know
- If you're building for Chinese markets: The 80% domestic chip mandate is your future infrastructure. Start testing your workloads on Ascend-compatible frameworks (MindSpore, CANN) before you need to. The toolchain differences from CUDA are substantial, and the migration cost is lower now than when you're under a deadline.
- DeepSeek is already doing this at scale. DeepSeek's production inference runs on Huawei 910C chips. If you're benchmarking Chinese open-weight models and find that performance numbers from Chinese labs don't always translate to your Nvidia cluster — this is why. They're optimized for different hardware.
- Nvidia's China exit is structural. The H20 (Nvidia's export-control-compliant China chip) is already restricted. The $295B plan formalizes a trajectory that was already underway with or without a NDRC blueprint. Don't plan around a policy reversal that brings Nvidia back to China at scale.
- Watch Biren Technology and Moore Threads. Huawei Ascend gets all the attention, but both Biren (BR100) and Moore Threads are on the approved list and are producing inference-optimized silicon with different architectural tradeoffs. If you're evaluating deployment on this stack, you'll be choosing among suppliers, not just between "Huawei or nothing."
- This doesn't change your roadmap today. The plan is in early drafting. The data centers aren't built. The chip gap exists. The dollar figure is real but the execution timeline is five years out. Don't let the headline distort near-term decisions unless your product explicitly targets Chinese government or security-sector deployment.
Further reading
- Bloomberg — China Preps $295 Billion Plan to Fund Nationwide AI Buildout — primary reporting, June 9, 2026
- The Decoder — Beijing's $295 billion AI buildout would require 80% domestic chips — chip mandate analysis and supplier list
- Council on Foreign Relations — China's AI Chip Deficit — technical capability assessment of the Huawei vs. Nvidia gap
- WebProNews — Huawei doubles Ascend 910C output to 600,000 in 2026 — production volume context
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Mythos wrote AMD kernels in three days. The CUDA moat is narrowing — not gone.
Anthropic's IPO roadshow revealed that Claude Mythos Preview ported Claude's prompt processing to AMD MI355X GPUs in roughly three days, versus the six to twelve months that traditional engineering takes. The claim is narrow — one model, one task — but the direction it points matters for every builder thinking about GPU lock-in long-term.
AI adds 8% to GDP and flatlines knowledge-worker wages. Anthropic's economists are saying this out loud.
Anthropic's economics team published a model in September showing three futures for US jobs and wages through 2030. In the most likely scenario, GDP grows 8% above baseline and average wages rise — but knowledge-worker wages stay flat, and capital captures a larger share of output than any time in modern history. Worth understanding before the transition is baked in.
Anthropic's IPO doesn't change what Claude can do. It changes who Claude has to answer to.
Anthropic's public S-1 is out. The equity structure and PBC charter got thorough coverage here four days ago. What hasn't been addressed: what going public specifically means for builders running production workloads on Claude API — pricing stability, feature velocity, safety commitments, and whether the post-IPO incentive environment changes your platform bet.