The Unlock That Wasn't: What the H200 Standoff Tells Us About the Chip War's Endgame

By Duncan Galbraith | Economics | Offworld News AI

The United States government spent years trying to keep its most advanced AI chips out of China. Then, last week, it approved ten of China's largest technology companies to buy them. China said no thank you.

This is not a diplomatic story. It is an industrial policy story about what happens when a sanctions regime succeeds too well — and in doing so, changes the economics it was designed to protect.

The Transaction That Didn't Happen

On May 14, 2026, NVIDIA CEO Jensen Huang flew to Beijing aboard Air Force One, invited by President Donald Trump to accompany the administration's delegation for a summit with President Xi Jinping. The occasion had the shape of a breakthrough: the US Commerce Department had already cleared roughly ten Chinese firms — Alibaba, Tencent, ByteDance, JD.com, and distributors including Lenovo and Foxconn — to purchase NVIDIA's H200 AI processors. Each approved customer could buy up to 75,000 chips, either directly from NVIDIA or through authorized intermediaries.

Before the export controls tightened, NVIDIA commanded approximately 95 percent of China's advanced chip market. Jensen Huang has estimated the Chinese AI market would be worth $50 billion this year. China once accounted for 13 percent of NVIDIA's total revenue. The opportunity on the table was real.

Not a single H200 has been delivered.

According to Reuters reporting, Chinese firms pulled back from finalizing purchases after receiving guidance from Beijing. Commerce Secretary Howard Lutnick confirmed the picture at a Senate hearing: "The Chinese central government has not let them, as of yet, buy the chips, because they're trying to keep their investment focused on their own domestic industry."

Beijing's stated concerns include hidden vulnerabilities — the US licensing terms require H200 chips to physically pass through US territory before delivery, for inspection, before Washington collects its 25 percent tariff cut. China's security apparatus is apparently unwilling to accept that condition.

But the security argument, while real, is not the full explanation. To understand why, you need to look at what four years of export controls actually built.

The Efficiency Moat

In October 2022, the Biden administration imposed the first major export controls on advanced AI chips to China. The logic was straightforward: deny China the compute it needs to train frontier models, and you preserve American AI leadership. The assumption embedded in that logic was that hardware scarcity would translate into capability scarcity. More chips equals more AI.

This assumption turned out to be correct in the short term and potentially wrong in the long term.

In January 2025, Chinese startup DeepSeek released an AI model that shook US markets, demonstrating Chinese AI capabilities significantly more advanced than most Western analysts had assumed — achieved on hardware several generations behind what American labs were using. The model had been trained under the constraints the export controls created.

A reporting team from Exponential View that spent a week visiting fourteen Chinese AI labs in 2026, including DeepSeek, MoonshotAI, ByteDance, and Alibaba, documented the mechanism directly. Their finding: Chinese labs are "extracting 4 to 7 times as much intelligence per unit of compute as naive scaling predictions would suggest." The compute constraint, they concluded, has become "capability-generating."

The stock of AI compute in China runs roughly two to three years behind the United States. Chinese open-source models are approximately six to eight months behind the US frontier. That ratio — two to three years of hardware disadvantage producing six to eight months of capability disadvantage — is the efficiency moat the export controls created by accident.

Huawei's Ascend 910C, the current Chinese flagship AI chip, delivers roughly 60 percent of NVIDIA's H100 inference performance on large language model workloads, according to DeepSeek's own research — limited, but functional. The Ascend 950 series, with Huawei's first proprietary HBM memory technology, is scheduled for 2026. The roadmap extends through 2028.

Beijing's calculation about the H200 offer is, in this light, rational. Accepting the chips means accepting supply chain dependency on US hardware, US inspection requirements, and US licensing conditions — all of which can be withdrawn at political will, as NVIDIA experienced when the Commerce Department took a $5.5 billion charge against its quarterly earnings in April 2025 due to sudden H20 restrictions. Why purchase reliability risk from a supplier who has demonstrated the willingness to terminate supply for policy reasons — especially when you have a domestic alternative you are actively developing?

The Economics of Industrial Substitution

This is where the analysis moves from geopolitics into the economics the Brookings Institution made explicit in August 2025: "Starving China's supply of U.S.-designed AI chips will push China to more effectively develop its own AI chip capacity, ultimately weakening America's leadership in AI and related technologies."

The mechanism Brookings identified is standard import-substitution theory, the same framework that explains post-WWII development economics in Brazil, South Korea, and Taiwan: protect a domestic industry from import competition long enough, and it develops the capabilities to compete. The United States has, through its export controls, been running an involuntary import-substitution program for China's domestic semiconductor industry.

The numbers reflect this. China's domestic AI chip market share is projected to reach 50 percent in 2026, up from near zero before the controls. That shift represents the redirection of the investment that would otherwise have flowed to NVIDIA.

By December 2025, the US had partially reversed course, approving NVIDIA to sell the H200 to approved Chinese customers under specific conditions. By January 2026, that policy was formalized. By May 2026, ten firms had been approved and zero chips had been delivered. The policy cycle took three years to complete and arrived at a market that had partly moved on.

NVIDIA's Specific Problem

The commercial damage to NVIDIA is documented in its own SEC filings. The company's April 2025 8-K filing disclosed a $5.5 billion quarterly charge due to restrictions on H20 exports — the downgraded chip NVIDIA had specifically designed to comply with earlier Biden-era rules. The Commerce Department stated the license requirement "will be in effect for the indefinite future." It lasted approximately three months.

That kind of regulatory volatility is not costless to the companies operating within it. NVIDIA designs chips on multi-year cycles, builds supply chains on longer-term assumptions, and prices products against competitive dynamics that the regulatory environment keeps reshaping. Jensen Huang's presence on Air Force One to Beijing — added to the delegation after a personal invitation from Trump, picked up in Alaska — reflects how directly the regulatory regime has linked the company's commercial interests to US foreign policy outcomes.

Before the controls tightened, NVIDIA commanded approximately 95 percent of China's advanced chip market. It now faces a market where domestic alternatives are under active development, software ecosystems are being built around those alternatives, and the regulatory conditions for selling its own products remain unpredictable. Huang has previously warned the US government that its restrictions are costing NVIDIA its China position. The H200 approval was the administration's partial response. China's non-purchase is the market's.

The Structural Question

The H200 standoff is worth analyzing precisely because it cannot be dismissed as a temporary political miscommunication. The conditions that produced it are durable.

The US wants to sell chips that convey strategic advantage. China wants to buy chips without creating strategic dependency. These preferences are not reconcilable on a single transaction — they reflect different theories of what the chip relationship is for. Washington treats advanced semiconductor access as leverage. Beijing now treats it as vulnerability.

The question this raises for the broader AI economy is not whether NVIDIA will eventually sell H200s in China (it may; the summit produced renewed talks and additional tariff relief). The question is whether the episode changes the calculus for future chip generations.

NVIDIA's Blackwell architecture — the current generation beyond the H200 — remains restricted under proposed legislation. The AI OVERWATCH Act would prohibit sales of Blackwell-class chips to entities of concern for two years. If that restriction holds, the cycle repeats: China develops domestic alternatives under constraint, the US partially relents when the commercial cost becomes visible, and by then the market has adapted.

Brookings put the logical endpoint plainly: every cycle of restriction followed by partial relaxation followed by restriction accelerates Chinese domestic capability development while degrading the US competitive position it was designed to protect. The H200 non-sale is not a failure of diplomacy. It is a confirmation of that dynamic running exactly as the analysis predicted.

What NVIDIA is discovering — and what the US semiconductor policy has not yet fully absorbed — is that market positions lost to substitution are harder to recover than market positions lost to competition. A competitor who outcompetes you on price or performance is one you can respond to. A market that has built its supply chain around an alternative ecosystem, for structural reasons, is one you have partially exited.

Huang flew to Beijing. The H200 chips stayed home. The Ascend roadmap runs through 2028.

Sources

• Channel News Asia, "US clears H200 chip sales to 10 China firms as Nvidia CEO looks for breakthrough," May 14, 2026. <https://www.channelnewsasia.com/east-asia/us-nvidia-h200-chip-sales-10-china-firms-6120746>
• Engadget, "US Reportedly Allows 10 Chinese Companies To Buy NVIDIA's Coveted H200 AI Chips," May 2026. <https://www.engadget.com/2172686/us-reportedly-allows-10-chinese-companies-to-buy-nvidia-h200-chips/>
• Brookings Institution, "How overly aggressive bans on AI chip exports to China can backfire," August 15, 2025. <https://www.brookings.edu/articles/how-overly-aggressive-bans-on-ai-chip-exports-to-china-can-backfire/>
• Exponential View, "Inside Chinese AI labs — American AI sanctions created its toughest competition," May 2026. <https://www.exponentialview.co/p/inside-chinese-ai-labs-efficiency-moat>
• Tom's Hardware, "DeepSeek research suggests Huawei's Ascend 910C delivers 60% of Nvidia H100 inference performance," 2025. <https://www.tomshardware.com/tech-industry/artificial-intelligence/deepseek-research-suggests-huaweis-ascend-910c-delivers-60-percent-nvidia-h100-inference-performance>
• Huawei Central, "Huawei reveals 3-year Ascend AI chip roadmap — 950 coming in 2026." <https://www.huaweicentral.com/huawei-reveals-3-year-ascend-ai-chip-roadmap-950-coming-in-2026/>
• NVIDIA SEC Filing (8-K), April 2025. <https://www.sec.gov/ix?doc=/Archives/edgar/data/0001045810/000104581025000082/nvda-20250409.htm>
• ML Strategies, "2026 AI Policy and Semiconductor Outlook," February 4, 2026. <https://www.mlstrategies.com/insights-center/viewpoints/54031/2026-02-04-_026-ai-policy-and-semiconductor-outlook-how-federal>
• KnightLi, "Nvidia H200 China export license approved," May 16, 2026. <https://www.knightli.com/en/2026/05/16/nvidia-h200-china-export-license-approved/>