China’s long-anticipated break with foreign AI accelerators has moved from policy and forecasts into operational reality. In recent weeks, major technology groups have signaled concrete procurement commitments and infrastructure builds that make domestic GPUs and accelerators the backbone of China’s AI stack, a transition amplified by record token consumption, red-hot markets for chipmakers and fresh security controls that enshrine chips as strategic assets. According to reporting by Pandaily, ByteDance is negotiating to buy at least 50,000 inference-focused GPUs from Tianshu Zhixin, while Alibaba has also opened talks with the firm. Those moves capture a broader, fast-moving decentralization of compute away from Western suppliers.
Big-tech buying spurs market convulsions
Procurement by China’s platform giants is no longer rhetorical. ByteDance, whose Doubao assistant reports 368 million monthly active users and an eye-popping 120 trillion daily token calls, has materially reallocated capital: the company increased its AI infrastructure budget by roughly 25% to RMB 200 billion and plans domestic chip purchases exceeding RMB 40 billion. The scale of that demand is reflected in market prices: Hong Kong–listed Tianshu Zhixin jumped nearly 15% on news of ByteDance and reached an all-time high of HK$887.50 days later. Alibaba’s parallel interest underscores how hyperscalers are lining up behind local designs for inference and specialized workloads.
A supplier ecosystem steps up
The procurement push comes amid rapid growth in domestic supply. IDC estimates show that China received about 4 million AI GPUs in 2025 and domestic designs accounted for 41% of deliveries. Huawei’s Ascend family led shipments with roughly 812,000 units; Baidu’s Kunlun Chip and Cambricon each shipped about 116,000 units. Baidu’s chip arm, Kunlun Chip (also marketed as Kunlunxin), is preparing a Hong Kong IPO that aims for a valuation near US$50 billion, more than 15 times the RMB 21 billion valuation it held after a Series D in July 2025, signaling investor willingness to assign high strategic value to onshore compute suppliers.
Nvidia’s China market share collapses
That strategic reorientation has a geopolitical backstory. Nvidia’s H20 line ceased shipments into China in 2025, a shift CEO Jensen Huang publicly acknowledged, telling audiences that market share in China fell from an estimated 95% to effectively zero after shipments stopped. The gap created by halted deliveries ignited demand for alternatives and accelerated enterprise experiments running major models on local silicon. The combination of Western export friction and Chinese industrial policy has produced a self-reinforcing dynamic: as foreign kit becomes harder to source or leaves, domestic designs scale to absorb demand.
Operational scale: from prototype clusters to production fleets
Baidu and others have translated chip-level progress into large-scale infrastructure. Kunlun shipped its third-generation P800 in 2024 and in 2025 constructed what it billed as China’s first 32,000-card, all-domestic AI cluster. That platform has already won centralized procurement orders, Kunlun secured a billion-yuan contract from China Mobile — and serves as proof that local chips can be deployed at hyperscale for training and inference. ByteDance and Alibaba’s moves suggest those clusters will multiply and interconnect with the country’s growing token economy.
Token explosion drives compute intensity
Underlying the procurement race is a surge in usage intensity. EastFrontier reporting earlier this year documented a leap from roughly 100 billion daily token calls in early 2024 to more than 100 trillion by the end of 2025, and token traffic reached about 140 trillion per day by March 2026. The token explosion transforms AI from an R&D problem into a continuous, massive operational load — especially for consumer-facing assistants and agentic services that run always-on pipelines. Firms that can lock in reliable, lower-cost compute sources (local chips, green energy hubs, optimized stack) gain both cost and product advantages.
Security review and industrial policy codify the shift
Regulatory moves have followed the market. In May 2026, Chinese information-security authorities for the first time brought AI chips into the national security review system; nine domestic training and inference chips received the highest I-level rating, effectively clearing them for state procurement and sensitive deployments. That formal recognition reduces procurement friction for domestic suppliers and signals that Beijing sees chip autonomy as central to strategic tech sovereignty. The new rules will also shape how foreign partners and customers engage with Chinese data centers and cloud services.
What this means for competition and risk
The practical consequences are immediate: cloud pricing, product road maps and enterprise procurement strategies are being rewritten around Chinese accelerators. Lower-cost domestic chips combined with aggressive token pricing have enabled a wave of price competition in model serving and made agentic experiences economically viable at scale — a dynamic we explored in our April analysis of China’s token boom here. And this isn’t only theoretical: policymakers and markets have rapidly re-assessed China’s chip self-reliance after domestic suppliers captured roughly 41% of the 2025 AI GPU market, a trend we tracked in April here.
Strategic implications and the road ahead
China’s compute transition closes one chapter of dependence and opens others. Strategically, the emergence of a domestic compute ecosystem reduces Beijing’s vulnerability to export controls and alters the bargaining position of Chinese cloud and AI firms in international markets. For global supply chains, the shift raises the prospect of two parallel compute stacks, one China-centric, one Western, each optimized for its regulatory and commercial environment. That bifurcation will complicate interoperability and standard-setting, while creating export opportunities as China pushes its chips into friendly or nonaligned markets.
Risks remain. Performance parity, long-term supply of advanced packaging and memory, energy constraints, and potential overheating in valuations all present real challenges. Rapid scaling of token-intensive services is also creating substantial power demand and puts a premium on energy and cooling innovations. Finally, tighter security reviews and protectionist rules could lock China’s compute stack into a politically distinct ecosystem, limiting overseas partnerships and increasing friction for multinational cloud providers and customers.
Still, the operational facts are stark: procurement commitments from ByteDance, Alibaba, and Baidu; IDC shipment data showing near-half penetration by domestic chips; a dramatic token surge; and formal security ratings that make local chips procurement-ready. Those elements together mark a tipping point. China’s AI industry is no longer merely promising alternatives to Western silicon; it is deploying them at scale, and that matters for product strategy, investor calculus, and the geopolitical balance of AI capability.