ANALYSIS: Qualcomm and Nvidia in China — the semiconductor war plays the edges of the rulebook

Introduction: Skirt without violate, the art of the regulatory corridor

The market that resists everything

In late June 2026, a new map of the Sino-American tech war emerged in outline. Qualcomm announced at its investor day on June 25 in New York that it would bring all four of its Dragonfly product lines for data centers into China, including AI accelerators custom-built to stay within the thresholds of American export controls. Nvidia, meanwhile, had been quietly pitching its Vera CPU to Chinese clients since June, with deliveries possible as early as August 2026. This is not a rollback of American restrictions. It is the birth of compliance engineering — a commercial architecture built precisely on the regulatory demarcation line.

That narrow corridor has become the most sophisticated battlefield of the Washington-Beijing rivalry. Not missiles, not direct sanctions, but computational performance thresholds, chip category classifications, and export licenses administered by the Bureau of Industry and Security at the U.S. Department of Commerce. China accounts for 46% of Qualcomm's total revenue in 2025 — essentially from smartphone chips. Abandoning that market is not an option for a company that depends on those revenues to fund its R&D. But capturing it without triggering a regulatory backlash demands a surgical precision that few companies in the world currently master.

Qualcomm and ByteDance — the deal that rewrote the rules of the game

ASICs calibrated below the legal threshold

In May 2026, Bloomberg reported that Qualcomm had secured a deal with ByteDance — TikTok's parent company — to supply custom AI chips for data centers. Application-specific integrated circuits (ASICs) whose performance parameters are deliberately calibrated below the thresholds that trigger American export controls. The model is clear: design compliance from the start, not after the sale. Cristiano Amon, Qualcomm's CEO, told Nikkei Asia on the sidelines of the investor day: "We have versions of all our products that comply with these guidelines." That is not a vague promise. It is an industrial strategy built around a legal boundary as a baseline engineering constraint.

The Dragonfly portfolio includes the Dragonfly C1000 CPU — more than 250 cores on the Oryon architecture, clocked above 5 GHz, with announced efficiency twice that of competing server processors —, High Bandwidth Compute (HBC) technology, the AI300 inference accelerator, and custom silicon offerings. Meta signed on as the first named client for a multi-generation deployment, with production targeted for 2028. Data center revenue targets stand at 5 billion USD for fiscal year 2027, and 15 billion USD by 2029. The Chinese share of that ambition is an explicit pillar, openly acknowledged by management.

The mechanics of the computational threshold

The core of Qualcomm's strategy rests on a precise technical reality: the Bureau of Industry and Security defines thresholds of total computational performance beyond which an export license is mandatory to sell into China. By calibrating its AI accelerators to stay below those thresholds, Qualcomm legally avoids that licensing obligation. The AI250 — the first HBC accelerator, expected in mid-2027 — is not simply a stripped-down product. It is a product designed around a legal boundary as an engineering constraint.

The HBC approach also offers an additional advantage: it does not use HBM (High Bandwidth Memory) — the high-performance memory that Nvidia and AMD require and whose supply from SK Hynix or Samsung is under heightened scrutiny. Qualcomm uses standard memory technologies derived from smartphones, reducing both cost and regulatory exposure across the supply chain. In a Chinese market where HBM is scarce and expensive, that is a concrete commercial advantage — not merely a regulatory compliance artifact.

Nvidia and the Vera CPU — entering through the back door

The GPU blocked, the CPU open — for now

The logic of Nvidia differs from but runs parallel to Qualcomm's. Its most advanced AI GPUs remain under strict control. Sales of the H200 to ten Chinese companies authorized by Washington were blocked for months by Chinese authorities pushing their own domestic chipmakers. But the Vera CPU — an ARM-based processor with 88 cores designed for agentic AI workloads — sits in a different regulatory category. Reuters reported in June that Nvidia was pitching Vera to Chinese clients with a delivery target of August 2026, with roughly 300 pilot units ordered to test software compatibility before larger volumes.

Nvidia CEO Jensen Huang said his forecast for a 200 billion USD CPU market included China. In March 2026, Nvidia had restarted manufacturing of the H200 adapted to export rules, under an unprecedented mechanism: licensed sales with 25% of revenues remitted to the U.S. government. Members of Congress denounced this structure as creating a perverse incentive to approve more sales at the expense of national security — in a letter addressed to Howard Lutnick in January 2026 noting that the H200s had six times more power than previously authorized chips.

The reclassification question

The central question around Vera remains unresolved: is it a general-purpose CPU or an AI accelerator that deserves stricter regulatory classification? If the Bureau of Industry and Security decided to reclassify the category, Nvidia's commercial window in China would close before the first deliveries could materialize. That fundamental uncertainty is what makes the strategy risky, even for a company as well-capitalized as Nvidia. The precedent of the H200 — approved, then blocked within a few months — is a warning the industry has absorbed.

The Chinese data center ecosystem is built on x86 architecture, and the migration to ARM entails significant software costs. Compatibility with existing AI software — trained and optimized for Nvidia's CUDA GPUs — is not guaranteed on an ARM CPU. This is not a guaranteed sale: it is a technological bet as much as a regulatory bet, in a market where Huawei and Cambricon are advancing rapidly on their own architectures.

Beijing's 295 billion — self-sufficiency as doctrine

Huawei, Cambricon, and the national technology wall

China's objective of reducing its dependence on foreign AI technologies to 20% — funded at 295 billion USD over five years — is not wishful thinking. It is the direct consequence of the lesson drawn from American chip restrictions: any technological dependency is a strategic vulnerability. Beijing's response is systemic. Huawei Ascend, Cambricon, and the accelerators of Chinese web giants like Alibaba (Hanguang) and Baidu (Kunlun) benefit from guaranteed state orders, tariff protections, and a software ecosystem being developed at accelerated pace under direct government oversight.

By late 2025, Beijing had banned the use of foreign AI chips in state-funded data centers. For Qualcomm and Nvidia, the still-accessible segment is essentially the private sector — ByteDance, Alibaba Cloud, Tencent — whose calculations can still favor American chips that are cheaper or more capable than current domestic alternatives. But that window is narrowing. According to analysts, Huawei Ascend and Cambricon are expected to reach significantly higher production volumes by 2028, mechanically reducing the share accessible to foreign suppliers.

Howard Lutnick and the next wave of restrictions

U.S. Commerce Secretary Howard Lutnick declared in mid-June 2026 that "the era of the robotics arms race is coming," signaling the imminence of new restrictions on robots manufactured in China. That statement is part of a trend of progressive tightening of the regulatory perimeter — robots, drones, chips, software — which makes the precise compliance strategy of American companies structurally fragile. In June 2026, the Trump administration had already tightened the rules, requiring licenses for Chinese companies headquartered in China even when they operate abroad — a significant expansion of the regulatory scope.

Political pressure in the United States is structurally oriented toward tightening. Members of Congress had sent a letter in January 2026 denouncing the relaxation of controls on the H200 as a threat to national security. That bipartisan pressure creates an unpredictable regulatory environment for companies planning three to five years out. The instability of the rules is itself a form of strategic risk — and it constitutes a comparative advantage for China, whose semiconductor industrial policy is consistent, funded, and not subject to electoral cycles.

Mutual dependence — a decoupling that takes a decade

Why cutting the cord is not a short-term option

That paradox is the defining feature of the entire Sino-American technology relationship in 2026: American companies need the Chinese market to fund the R&D that will keep them ahead technologically, while China still needs American chips and tools to close its current gap. 46% of Qualcomm's revenue comes from China. Cutting that flow abruptly would weaken the company's investment capacity at the precise moment when global technology competition is intensifying. It is a dependency that slows decoupling without stopping it — and creates a commercial gray zone where geopolitics and balance sheets permanently contradict each other.

On the Chinese side, domestic accelerators still have a significant performance gap compared to the best American chips, especially on large language model training workloads. That window of dependence is precisely what Qualcomm and Nvidia are trying to monetize before it closes — what the industry calls the "window of mutual dependence." It is a rational short-term bet. But its shelf life is limited — commercially, by the rise of Chinese national champions, and regulatorily, by increasing pressure in Washington.

The structural irony of the tech war

There is a fundamental irony in this race: American companies indirectly fund, through their sales in China, a technology ecosystem designed to replace them. The revenues generated by ByteDance and other Chinese clients contribute — through taxes and learning effects — to the rise of China's semiconductor industry. But refusing that market means ceding the ground to international competitors who will find ways in. There is no good option — only different trade-offs between national security and commercial competitiveness.

For the West collectively, the central question is not whether Qualcomm should sell chips in China. It is deciding what level of technology transfer is acceptable with a declared strategic adversary, and establishing rules clear and stable enough for companies to plan five years out. The current regulatory instability — relaxations followed by tightenings, licenses granted then blocked — harms American long-term interests as much as it claims to protect them in the short term.

The ByteDance deal — a model inspiring the entire American industry

Qualcomm creates an industrial precedent

The Qualcomm-ByteDance deal of May 2026 is not simply a commercial contract. It is a model — a proof of concept that the entire American semiconductor industry is watching closely. By demonstrating that it is possible to secure significant data center revenues in China without violating export controls, Qualcomm has created an industrial precedent. Other American chipmakers, particularly those specializing in ASICs for AI inference, are now actively examining similar contractual structures. Compliance by design is becoming a differentiating commercial skill, not merely a legal obligation.

That precedent has clear limits, however. The ByteDance deal covers inference chips — chips that run already-trained AI models, not chips that create them. Training chips, far more computationally powerful, remain beyond the reach of the current compliance model. That distinction is fundamental: China can buy chips to deploy AI but remains theoretically constrained in its ability to train new frontier models with American chips. Theoretically — because alternative supply routes remain documented and concerning.

Robotics controls — the next frontier of restrictions

From chips to robots, one same logic of interdependence

Howard Lutnick's declaration about "the era of the robotics arms race" signals an extension of the regulatory perimeter well beyond semiconductors. Robots manufactured in China — whose components often integrate sensors, embedded processors, and potentially dual-use control algorithms — are now in Washington's crosshairs. This logical extension of the export control paradigm responds to a real concern: a country that masters advanced industrial robots and autonomous drones holds a potential military advantage that chip restrictions alone cannot counterbalance.

For American robotics and automation companies, that prospect creates a new zone of uncertainty analogous to the one that semiconductor manufacturers already face. Global industrial robotics supply chains are deeply integrated with Chinese components and subsystems — a reality that new restrictions will need to navigate without causing major economic disruption to American manufacturing industries themselves. The tech war generates its own collateral costs for the American economy, a reality that policymakers cannot ignore indefinitely.

What this means for Qualcomm and Nvidia

For Qualcomm and Nvidia, the extension of controls to robots is not a direct immediate threat — their products are chips, not robots. But it is a clear policy signal: the American regulatory trend is toward systematic tightening, category by category, until the architecture of technological dependence between the United States and China is fundamentally restructured. Each new category subjected to restriction reduces the perimeter of what American companies can market in China without regulatory friction.

That trajectory reinforces the thesis of a long-term decoupling that, even if it takes ten years rather than one, calls into question all revenue models dependent on access to the Chinese market. The 15 billion USD in data center revenues anticipated by Qualcomm by 2029 — of which a significant share is expected from China — is conditional on regulatory decisions that neither the company nor its shareholders control. That is the fundamental systemic risk of a commercial strategy built on the regulatory tolerance of a government that is sending clear signals of progressive tightening.

TSMC and the supply chain — the link that changes everything

Manufacturing in compliance via Taiwan

All of Qualcomm's strategy rests on an element that years of discussion about computational thresholds tend to obscure: the chips must be manufactured somewhere. The vast majority of the ASICs calibrated for ByteDance will pass through TSMC's foundry facilities in Taiwan — a geopolitically explosive reality in its own right. TSMC is subject to American export control rules under the applicable technology control agreements, and must verify that the chips it manufactures for American clients destined for Chinese buyers comply with the regulatory thresholds. That is an additional layer of legal compliance that reinforces the structure — but also an additional dependency on an actor whose own geopolitical security is uncertain in the face of pressure from Beijing.

China has no access to TSMC for the most advanced technology nodes — a direct consequence of American and Dutch controls on EUV lithography equipment. That is one of the few restrictions that genuinely holds: China cannot manufacture 3 or 4 nm chips domestically today, regardless of how much capital is invested. That manufacturing bottleneck is the real Achilles' heel of China's AI ambition — and that is why the Chinese 295 billion includes a massive investment in alternatives to ASML equipment and in the development of domestic foundries like SMIC. The race is not only about chips. It is about the machines that make the chips.

Conclusion: Inside the margins of a rulebook that will move again

A precarious but clear-eyed balance

The strategy that Qualcomm and Nvidia are deploying in 2026 is technically legal, commercially rational, and geopolitically risky. It rests on a regulatory precision that can be called into question by a memo from the Bureau of Industry and Security, a bipartisan congressional decision, or a unilateral restriction by Beijing on foreign chip imports into critical infrastructure. The room to maneuver exists — but it is narrow, contested, and temporary by nature. This is not a stable growth strategy. It is a positioning strategy in a period of technological and geopolitical transition.

What is structurally clear in July 2026 is that the semiconductor war is no longer played out solely in R&D labs or in Washington speeches. It plays out in the technical specifications of chips designed to pass exactly below the legal threshold, in contracts signed between ByteDance and Qualcomm, in decisions by Chinese clients to choose between a compliant American accelerator and a sovereign Huawei accelerator. The outcome of this war will be decided over ten years, not ten months — and no regulatory corridor strategy can replace a long-term industrial vision.

What this signals for the West

The lesson for the West is twofold. First, strategic consistency matters as much as technological superiority in a long-term competition. China has a five-year plan funded at 295 billion. The West holds elections every four years and export rules that change with each administration. Second, companies cannot be the sole bearers of industrial and national security policy. Without a clear and stable collective vision, they optimize short-term revenues — which is their legal mandate — while long-term strategic competition remains without coherent direction.

The semiconductor war will continue. Qualcomm will deliver its first compliant chips in 2027, if the rules do not change. Nvidia will attempt to ship its first Vera CPUs in August 2026, if Beijing does not block the imports. And in the meantime, China will advance on its 295 billion with a regularity that Western democracies struggle to match. The regulatory corridor exists. But it is not a strategy. It is a valuable reprieve — on condition that it is used to build something more durable.

By Maxime Marquette, columnist

Columnist's transparency note

Who I am and what I am not

I am a columnist and analyst, not a semiconductor engineer or a lawyer specializing in export controls. My analysis draws on verified public sources — corporate reports, specialized press coverage, official statements — and not on confidential information or industry contacts. I do not claim to know the precise technical parameters of the chips in question or the undisclosed details of the commercial contracts between the parties. The revenue projections cited are those of the companies themselves and carry significant uncertainty.

My stated biases and methodological limits

I believe that China's technological rise represents a structural challenge to the Western liberal order, and that economic dependence on a declared strategic adversary is a vulnerability that Western democracies collectively underestimate. That bias orients my analytical framing toward caution regarding compliance strategies that maximize short-term access without resolving the underlying strategic contradiction. I acknowledge that the decisions of companies like Qualcomm or Nvidia are rational in their context — and that the responsibility for a coherent industrial policy lies with governments, not with publicly traded companies.

Sources

Primary sources

TrendForce — Qualcomm Reportedly Plans China AI Chip Push With Export-Control-Compliant Custom Chips — June 25, 2026

AI Weekly — Qualcomm Builds Export-Compliant AI Chip for Chinese Data Centers — June 25, 2026

Slicast / Tom's Hardware — Qualcomm plans China-specific data center chips, Dragonfly lineup export-compliant — June 25, 2026

Secondary sources

Yahoo Finance — Nvidia pitches Vera CPU to Chinese clients with August delivery target — June 12, 2026

US News / Politico — US Commerce Secretary Signals Possible Action on Chinese Robots After Review — June 23, 2026

ChosunBiz — Qualcomm supplies millions of AI data center chips to ByteDance — May 27, 2026