Chinese buyers expect to move nearly half their AI accelerator budgets to local chips within a year, and Nvidia is losing budget share in the place it can least control.
A Bloomberg Intelligence survey published July 7 found that executives at Chinese companies expect to allocate 46% of their AI accelerator spending to homegrown chips over the next twelve months, up from roughly 30% today. That is the number to sit with. It is not a slogan from Beijing or an analyst's neat curve on a slide. It is buyers telling Bloomberg where they expect their own chip money to go.
If you own Nvidia, or build around its hardware, that should make you pay attention. China is not just talking about replacing American accelerators. Companies are budgeting for it, even while the local alternatives still carry real performance and software tradeoffs. The shift is being forced by policy, but it is being recorded in procurement plans.
The same Bloomberg Intelligence survey found that 80% of the executives polled said their total infrastructure spending is running over budget this year, with AI buildouts blamed for the pressure. That tells you something uncomfortable about the substitution drive. Switching to domestic chips isn't suddenly cheap or easy - nobody claims that. Washington has restricted what Nvidia can sell, and Beijing wants state-linked buyers to stop depending on it anyway, so companies are spending through the pain regardless.
The performance gap is still real. Tom's Hardware, citing IDC numbers reported in April, said Chinese semiconductor firms delivered 1.65 million AI GPUs in 2025 out of a domestic market of about 4 million units. Nvidia still led with an estimated 2.2 million GPUs shipped, but its share had fallen to 55% from a reported 95% before sanctions. Huawei was the big local name, with about 812,000 AI chips, while Cambricon was listed at 116,000. That is not parity with Nvidia. It is a market being rebuilt under pressure.
Look at the procurement rules and the direction becomes easier to understand. China's Ministry of Industry and Information Technology has added Huawei and Cambricon processors to its approved government procurement list. Nvidia is not on it. For state buyers, that can matter more than a benchmark. A better chip that cannot be bought, approved, imported or defended inside a government project is not really competing on equal terms.
Huawei is trying to turn that opening into volume. Reports this spring said its 950PR chip was already in mass production, with shipments expected to rise in the second half of 2026, and analysts have pointed to Huawei's growing AI chip revenue as Nvidia's H200 shipments into China remain stuck in regulatory limbo. Cambricon has been pulled along by the same demand. The company has gone from an also-ran in the global AI chip conversation to one of the Chinese names buyers now have to consider, whether they wanted to or not.
Software is following the hardware, which is the part Nvidia should worry about most. Z.ai, formerly known as Zhipu AI, released GLM-5 in February 2026, and its GLM models have been adapted for Huawei's Ascend processors. Reuters reported that DeepSeek released its V4 series in April, and Tom's Hardware later reported that the model was optimized for Huawei's Ascend chips. A Huawei-linked research group also said it completed full-parameter post-training of DeepSeek's V4-Pro using more than 1,000 Ascend 910C chips, according to reporting that cited the Shenzhen municipal government.
Do not overread that. Post-training is not the same as proving that Chinese accelerators can replace Nvidia for every frontier model workload. The same reporting noted that Ascend 910C had shown about 60% of an Nvidia H100's inference performance in earlier DeepSeek testing. But every model ported to Ascend makes the next port less strange. Every lab that makes the toolchain work leaves less reason for the next lab to wait for Nvidia.
Nvidia has been warning Washington about exactly this problem. Jensen Huang has argued repeatedly that cutting the company out of China only accelerates Huawei's rise, and The Wire China recently detailed how Nvidia invokes its Chinese rival while lobbying for looser export rules. Frankly, he has a point. Export controls are meant to slow China's AI buildout, but they also give Chinese buyers a reason to tolerate worse hardware today so they are less exposed tomorrow.
None of this means Nvidia is gone from China. The Bloomberg survey still points to roughly 30% of accelerator budgets going its way, and China's data center buildout is large enough that a smaller share can still be real money. Nvidia's software moat is also not some minor detail you can wish away. CUDA, developer habits and proven deployment stacks still matter when you are spending millions on clusters that need to work.
But the direction is no longer just a political forecast. Executives who control budgets say nearly half their AI accelerator spending will go to local suppliers within a year. For investors watching chip stocks swing on every export-control headline, that is the cleaner signal. Nvidia is not being beaten in China because Huawei has already matched it chip for chip. It is being crowded out because two governments have made dependence on Nvidia harder to justify.
Also read: Illinois Becomes the First State to Mandate Annual AI Safety Audits • An AI Agent Ran a Ransomware Attack Almost Entirely by Itself • US Investors Can Finally Buy Into SK Hynix Ahead of Its Nasdaq Debut
