← April 26, 2026
tech power

DeepSeek Builds the World's Best Open-Source AI on Chinese Chips. The US Just Made It Official.

DeepSeek Builds the World's Best Open-Source AI on Chinese Chips. The US Just Made It Official.
Reuters

What happened

DeepSeek released a preview of V4, its most capable model yet, built specifically to run on Huawei's Ascend chips rather than Nvidia hardware. The model features a 1.6 trillion parameter architecture, a 1M token context window, and prices that undercut every major US AI model. Commerce Secretary Howard Lutnick confirmed the same week that China has purchased zero Nvidia H200 chips since Washington lifted export restrictions four months ago, because the Chinese government is actively blocking its companies from buying American hardware and steering them toward domestic alternatives. The US government simultaneously escalated accusations that DeepSeek and other Chinese AI firms stole intellectual property from US AI models to achieve their performance gains.

The US export control strategy assumed China would need American chips to build world-class AI. China decided it would rather build world-class chips than buy American ones, and DeepSeek V4 is proof the strategy worked. Washington has handed Beijing a forcing function that American market pressure never would have created.

Prediction Markets

Prices as of 2026-04-26 — the analysis was written against these odds

The Hidden Bet

1

Export controls slow China's AI development by denying access to cutting-edge hardware.

DeepSeek V4 launched on Huawei chips and undercut GPT-5 pricing while matching frontier performance. The controls may have accelerated China's domestic chip ecosystem by eliminating the easier path of just buying from Nvidia.

2

The IP theft allegations, if true, explain DeepSeek's capabilities.

DeepSeek's key innovations — mixture-of-experts architecture, efficient training at low cost — are documented in their public research papers. Accusing them of theft may be deflection from the more uncomfortable truth that China solved efficiency problems American labs deprioritized.

3

Nvidia's loss of the China market is a temporary setback that can be reversed through diplomacy.

Chinese companies are now optimizing their AI systems specifically for Huawei chips. Every month that passes deepens that ecosystem lock-in. The window for Nvidia to recapture this market may already be closed regardless of what export policy does.

The Real Disagreement

The core tension: did export controls protect American AI leadership, or create the competitive threat they were designed to prevent? The case for controls is that China's military would otherwise have direct access to chips enabling autonomous weapons and surveillance systems at scale. The case against is that the controls forced China to build its own chip stack and AI ecosystem, which is now cheaper and in some ways more efficient than the American alternative. Both concerns are legitimate. The controls cannot simultaneously protect US national security and preserve US commercial dominance in AI — those goals are in direct conflict. The national security case is stronger, but the commercial cost is now real and measurable.

What No One Is Saying

Washington is now accusing DeepSeek of IP theft precisely because the alternative explanation — that Chinese engineers solved hard problems Americans hadn't prioritized — is more threatening than a crime. A theft narrative implies a fixable vulnerability. A genuine innovation narrative implies a structural shift in where AI progress happens.

Who Pays

Nvidia shareholders and US AI hardware companies

Ongoing, worsening over 12-18 months as Huawei Ascend ecosystem matures.

Exclusion from China's $50 billion annual AI infrastructure market. DeepSeek's architecture shows that efficient models can be built on non-Nvidia hardware, threatening Nvidia's pricing power in other markets too.

US AI labs competing on frontier model costs

Immediate: price pressure is already visible in enterprise AI procurement.

DeepSeek V4 at $0.14/million tokens undercuts every major US model. Labs that built business models on premium pricing face direct commoditization from an open-source competitor they cannot counter with export policy.

US national security planners

3-5 years: Huawei Ascend ecosystem is still maturing, but the trajectory is set.

A China with a fully domestic AI chip stack and frontier AI capabilities is more dangerous than a China dependent on American hardware. The controls created the independence they were meant to prevent.

Scenarios

Huawei catches up, Nvidia locked out permanently

Huawei's Ascend chips reach performance parity with Nvidia H100 within 18 months. Chinese AI companies, already optimized for Ascend, have no incentive to switch. Nvidia's China revenue goes to zero.

Signal DeepSeek or another Chinese lab releases a model that outperforms GPT-5 on standard benchmarks while running exclusively on Huawei hardware.

IP theft case leads to sanctions, decoupling accelerates

US government files formal sanctions against DeepSeek citing evidence of systematic model distillation from OpenAI and other US labs. China retaliates across tech trade. Semiconductor supply chain splits definitively.

Signal Commerce Department adds DeepSeek to the Entity List or files criminal referrals against specific individuals.

US relaxes controls as commercial pressure mounts

Nvidia and other US chip firms lobby successfully to ease export restrictions, arguing the controls cost more in commercial revenue than they prevent in military capability. Limited H200 sales to Chinese cloud providers resume.

Signal Lutnick revises the 'delicate balance' framing; Commerce initiates review of H200 export policy.

What Would Change This

If Huawei's chips are revealed to have fundamental architectural limitations that prevent scaling to the next generation of AI models, the current Chinese trajectory stalls. DeepSeek V4's efficiency advantage matters less if the hardware ceiling is near. Detailed independent benchmarks of Ascend versus Nvidia at scale would answer this directly.

Sources

Reuters — DeepSeek V4 launches on Huawei Ascend chips, outperforms other open-source models on agent tasks, priced at $0.14/million tokens — cheaper than all major US competitors.
CNBC — V4 intensifies AI race pressure on US firms; DeepSeek's R1 already disrupted markets in 2025, V4 escalates with 1M token context window and Huawei integration.
The Star (via AP) — US Commerce Secretary Lutnick confirms China has bought zero H200 chips despite export clearance; Chinese government is forcing domestic companies to use Huawei hardware instead.
Tom's Hardware — US government alleges DeepSeek and other Chinese AI firms stole intellectual property from US AI models; timing of V4 launch alongside these accusations is notable.
MoneyCheck — Nvidia excluded from China's $50B annual AI infrastructure sector; DeepSeek V4 demonstrates that Huawei Ascend can now support frontier-level AI workloads.

Related