Nvidia Vera CPU and the 2026 AI Chip War

TL;DR — Nvidia is trying to reopen China through the side door: its new Arm-based Vera CPU faces far fewer export restrictions than its GPUs, and the company is already taking Chinese orders for August delivery. At the same time, Taiwan is weighing tighter controls that would cut AI-chip sales to all Chinese customers to align with Washington. The result is a 2026 where the most advanced silicon on Earth is governed less by engineering and more by geopolitics.

---

What the Vera CPU actually is

When Nvidia unveiled its Vera Rubin platform in March 2026, most of the attention went to "Rubin" — the next-generation GPU line that succeeds Blackwell. But the platform has two names for a reason. Vera is the CPU half: an Arm-based, 88-core processor that Nvidia designed in-house for agentic AI workloads, and which the company says is now in full production.

In a complete Vera Rubin rack, the Vera CPUs feed and coordinate the Rubin GPUs that do the heavy matrix math. On its own, though, a CPU is a far more ordinary device — it schedules work, runs orchestration layers, and handles the general-purpose computing around an AI cluster. That distinction, CPU versus GPU, turns out to be the whole story for China.

This is the same architectural split we've covered before in how specialized AI silicon is reshaping the market, from inference-focused accelerators like Groq's LPU to the data-center buildout driving demand for all of it.

Why Nvidia is selling the CPU, not the GPU

Nvidia's high-end GPUs are the chips Washington cares about, because they are what trains and runs frontier AI models. CPUs sit below the performance thresholds that US export rules target. According to reporting by Reuters on June 12, 2026, Nvidia has begun telling Chinese clients that Vera could be available as soon as August and that they can start placing orders now.

The interest appears real. Reuters reported that one major Chinese cloud provider is planning to order more than 300 servers, each fitted with two Vera processors, for initial testing before any larger rollout. The economics are not trivial: a single Vera chip runs well north of $20,000 before bulk discounts, and a fully configured rack of 256 chips comes to roughly $10 million, per the same reporting carried by Investing.com.

The urgency comes from how far Nvidia has fallen in the market. On the company's late-May earnings call, CEO Jensen Huang said Nvidia's market share in China had "effectively fallen to zero." Huang has separately framed the broader CPU opportunity as a roughly $200 billion market, and when asked on that call whether his forecast included China, he answered, per CNBC: "I would think so."

In other words, Vera is a way back into a market Nvidia has effectively lost — through a product class regulators have mostly left alone. It is a deliberate pivot rather than a coincidence, and it sits alongside China's own enormous push into domestic AI data-center capacity.

The GPU side is frozen

The contrast with Nvidia's GPU business in China is stark. The H200 — its second-most-powerful AI chip — was technically cleared for some Chinese buyers, but shipments have stalled for months.

The US framework around those sales is unusually intricate. On December 8, 2025, the Trump administration announced it would allow H200-class chips to ship to approved Chinese customers under a revenue-sharing arrangement that sends 25% of the proceeds to the US government. The Commerce Department's Bureau of Industry and Security followed with a final rule on January 13, 2026, effective January 15, that shifted license reviews from a presumption of denial to case-by-case evaluation — but only under strict conditions, including independent US lab testing before export and a cap limiting China sales to less than 50% of the same chip's US sales.

Even with that pathway open, the chips have not really moved. Nvidia has also told Chinese clients that H20 inventory is limited and that production of the older part will not resume, while it develops a new B30 chip for China — a Blackwell-based design that uses conventional GDDR7 memory instead of the export-restricted high-bandwidth memory found in its flagship parts, according to Tom's Hardware. A CPU that ships in August is, by comparison, a much cleaner proposition.

Taiwan tightens the other end of the supply chain

While Nvidia works one angle, the manufacturing heart of the industry is moving in the opposite direction. Bloomberg reported on June 9, 2026 that Taiwan is weighing significantly tighter AI-chip export controls aimed at China, in order to align more closely with the United States.

The crucial change is scope. Today, Taiwan's restrictions focus on blacklisted firms such as Huawei and SMIC. The measures under consideration would, for the first time, restrict sales to all customers in China above a certain performance threshold — directionally following the US approach, though the exact threshold remains undecided, per Bloomberg.

A second motivation is enforcement. Taiwan does not currently treat unauthorized AI-chip exports to China as a crime, which is part of why a new legal framework is needed. The Taipei Times reported that Taiwan made its first known detentions of alleged chip smugglers last month, on charges of falsifying documents — a sign of how much hardware may be slipping across the strait through diversion. President William Lai's administration is handling the issue through the Ministry of Economic Affairs and ongoing trade consultations.

The stakes are hard to overstate: Taiwan hosts the vast majority of the world's AI-chip manufacturing and many of the firms that assemble Nvidia processors into finished servers. Tighter rules there ripple straight through the supply chain to server makers and, ultimately, to Chinese buyers.

A timeline of the 2026 chip-control landscape

Why this is really about geopolitics, not chips

Step back and a pattern emerges. The technical question — is this part a CPU or a GPU, does it use HBM or GDDR7, does it clear a performance threshold — has become inseparable from the political one. Each side is optimizing around the other's rules. Nvidia engineers products to fit inside whatever export envelope is permitted, then routes its China strategy through the least-restricted category available. Washington calibrates thresholds, revenue cuts, and testing requirements to slow China's frontier-AI progress without crushing a major US exporter. And Taiwan, caught between its largest trading partner and its most important security partner, is being asked to police the chokepoint it physically controls.

None of this happens in a vacuum. The same export pressure helped push China to build out massive domestic capacity and to back homegrown model developers — the dynamic behind the rise of efficient open models like DeepSeek. Restrictions intended to slow China's AI progress have, in part, accelerated its drive for self-sufficiency, which is one reason Nvidia is so eager to keep some channel open before local rivals fill the gap.

The bottom line

Nvidia's Vera CPU is a genuinely capable piece of hardware, but in mid-2026 its most interesting feature is regulatory: it is the part of the Vera Rubin platform that China can actually buy. By leaning on a product class that export rules largely ignore, Nvidia is trying to reclaim a market where its share has cratered — even as Taiwan moves to tighten the very controls that created the gap. Whether August deliveries happen will depend less on fabs and yields than on how Washington, Taipei, and Beijing read the next move. In the 2026 chip war, the silicon is the easy part; the politics are the product.