For the better part of a decade, the global AI economy has run on a simple assumption: serious compute means Nvidia silicon, and Nvidia silicon means the CUDA software stack that surrounds it. China is now drafting a plan to break that assumption at national scale — not by out-engineering a single chip, but by building an entire alternative computing fabric and mandating that the country use it.
According to a report from TechTimes citing Bloomberg and China’s National Development and Reform Commission (NDRC), Beijing is preparing a five-year programme worth roughly 2 trillion yuan — about $295 billion — to wire thousands of data centres into a single national computing grid built on at least 80% domestic technology. The practical effect is blunt: Nvidia and AMD, the two companies that have defined accelerated computing, would be written out of one of the largest infrastructure procurements in history. Markets reacted accordingly, with Nvidia falling around 2.4% and AMD around 4% on the news. This is a feature about what that mandate actually changes — and what it doesn’t.
The mandate
The headline number is enormous, but the structure matters more than the size. A roughly $295 billion, five-year commitment to interconnect data centres nationally is significant on its own. Pairing it with an 80% domestic-technology floor turns it into something different: an industrial policy that manufactures demand. Where Western hyperscalers buy the best chip available on a global market, Chinese operators building into this grid would be steered — by procurement rules and security review — toward home-grown alternatives.
That creates what economists call a captive market, and captive markets are rocket fuel for domestic champions. Huawei, with its Ascend line, and a cluster of peers including Cambricon, Biren and Moore Threads no longer have to win purely on merit against the best Nvidia can ship. They have to be good enough, available, and compliant. Guaranteed offtake at this scale changes the maths of chip design entirely: it underwrites the capital expenditure, justifies the fabrication risk, and gives engineering teams the one thing a frontier semiconductor programme needs most — time, backed by revenue, to iterate.
It is also a hedge against Washington. Years of escalating US export controls have made Nvidia’s most capable parts either unavailable or politically radioactive inside China. Beijing’s response is to stop treating access to foreign silicon as a variable it can control and start treating domestic capacity as a problem it can solve with money and mandates. The grid plan is the clearest statement yet of that strategy.
How close is the hardware
The reflexive Western assumption is that Chinese accelerators are generations behind. That framing is increasingly out of date. Huawei shipped roughly 812,000 Ascend chips in 2025 and projects around $12 billion in AI-processor revenue for 2026 — an increase of roughly 60% — according to figures reported by TechTimes. Those are not the numbers of a science project. They are the numbers of a product being deployed across Chinese hyperscalers at meaningful scale.
On a single-chip basis, the best domestic parts still trail Nvidia’s flagship Blackwell-class accelerators on raw performance and, critically, on manufacturing yield at advanced nodes. But the industry has learned a lesson that reshapes the comparison: at frontier scale, the relevant unit is no longer the chip — it is the rack, the pod, and the cluster. If a single domestic accelerator delivers, say, 60-70% of a competitor’s performance, you can often close the gap by wiring more of them together with fast interconnects and disciplined system design. That brute-force, rack-scale approach is power-hungry and capital-intensive, but for a state willing to subsidise both electricity and silicon, it is a viable path to competitive aggregate compute.
Add to this the security-review dimension. Domestic chips clearing China’s own security and procurement vetting gain a structural advantage no benchmark captures: they are permitted. In a system where compliance is a gate, a slightly slower chip that passes review beats a faster one that doesn’t. The hardware gap, in other words, is real but shrinking — and it is increasingly the least interesting part of the story.
The software moat
Here is where the picture gets honest. Nvidia’s true durability was never just the transistors; it was CUDA, the programming layer that thousands of AI researchers, framework maintainers and infrastructure engineers have built their careers and codebases on top of. Roughly a decade of developer lock-in means that the entire scaffolding of modern AI — the kernels, the libraries, the optimisations, the Stack Overflow answers — assumes CUDA underneath. That ecosystem is the moat, and it is far harder to replicate than a chip.
China’s answer is CANN, Huawei’s compute architecture, paired with the MindSpore framework. On paper they are credible. In practice, teams migrating from Nvidia report real friction: porting overhead, missing or immature operators, debugging tools that lag, and performance that has to be hand-tuned rather than inherited. The most telling data point in the TechTimes report comes from RAND, which noted that iFlytek — one of China’s leading AI firms — reported roughly a three-month model-development delay after switching from Nvidia to Huawei’s Ascend 910B.
Sit with that number. A three-month delay is not a verdict that the hardware failed; the 910B is a serious part. It is a measurement of ecosystem immaturity — of every small rough edge in the software stack compounding into lost quarters. For a single company that is an inconvenience. Multiplied across an entire national buildout, it is a strategic tax on velocity. The hardware deficit is closing fast; the software and tooling deficit is the one that will actually determine whether Beijing’s grid runs as smoothly as the spreadsheet promises.
The optimistic case for China is that captive demand fixes this too. Force enough of the country’s best engineers onto CANN and MindSpore, and the ecosystem matures because it has to. Every ported model, every contributed kernel, every internal tool makes the next migration cheaper. Moats can be drained — it just takes years of concentrated, unglamorous engineering. The mandate buys exactly that.
The power problem
The constraint that no procurement rule can legislate away is electricity. Rack-scale clusters of domestic accelerators, deployed to compensate for single-chip gaps, are extraordinarily power-dense. Wiring thousands of data centres into a national grid does not just move data — it concentrates demand for energy at a scale that strains generation, transmission and cooling all at once. AI’s electricity appetite is rising faster than almost any forecaster predicted two years ago, and a buildout of this ambition turns power from a line item into the hard ceiling on how much compute can actually be switched on.
This is, arguably, why the plan is framed as a grid at all. Centralising data-centre planning lets the state co-locate compute with generation — hydro in the south-west, wind and solar in the north, coal where it must — and balance load across regions in a way a fragmented market cannot. It is a recognition that at this scale, AI infrastructure is energy infrastructure, and the two have to be planned together or neither works.
For India, building its own data-centre ambitions, the lesson is pointed and timely. The instinct is to focus the policy conversation on chips and sovereignty. But the binding constraint on an Indian AI buildout will be the same one China is engineering around: reliable, affordable, scalable power. A few takeaways worth internalising:
- Plan compute and power together. Treating data-centre policy and energy policy as separate files guarantees stranded capacity and grid stress. China’s grid framing is the lesson, even if its chip mandate isn’t replicable.
- Don’t fetishise the chip. Access to silicon matters, but ecosystem maturity and power availability decide whether that silicon does useful work. India should weigh developer tooling and skills as heavily as hardware sourcing.
- Co-locate with generation. Renewable-rich regions and predictable baseload should anchor cluster siting, not legacy real-estate or connectivity assumptions.
- Mandates cut both ways. Captive demand can build a domestic industry; it can also entrench slower technology and a three-month tax on every project. Pick the trade-off deliberately, not by default.
The bottom line: Beijing’s $295 billion grid plan is the most serious attempt yet to build frontier AI infrastructure outside the Nvidia–CUDA order. The hardware is closer than the West’s comfort narrative admits. But the two things that will decide whether China’s compute wall stands are the ones that money buys slowly, not instantly: a software ecosystem deep enough to rival CUDA, and a grid strong enough to keep the lights on. The chips were never going to be the hard part.
