Nvidia has drawn a line in the silicon. The company has stopped production of H200 chips bound for China and redirected that manufacturing capacity at TSMC toward its next-generation Vera Rubin platform, the Financial Times reported Thursday, citing two people with direct knowledge of the decision. The move lands less than a week after Nvidia disclosed it had received U.S. government licenses to ship "small amounts" of H200 chips to Chinese customers — and confirms what the company's own forward guidance already implied: for Nvidia, China's AI chip market is no longer something worth building for.
The Production Halt: What Actually Happened
The H200 is Nvidia's second-most advanced AI chip, based on the Hopper GPU architecture and manufactured by TSMC on a 4-nanometer process. It is also — critically — the most capable chip that Washington currently permits Nvidia to sell in China under the tightening export control regime that has defined U.S.-China semiconductor policy since 2022.
According to the Reuters report citing the FT, Nvidia has stopped allocating TSMC wafer capacity to H200 chips intended for the Chinese market. Instead, that capacity is being reallocated to the Vera Rubin platform — Nvidia's newest generation AI hardware, announced at CES 2026 in January and already in full production at TSMC on a more advanced 3-nanometer process.
Neither TSMC nor Nvidia has publicly confirmed the decision. TSMC declined to comment; Nvidia did not respond to a request for comment. But the move is consistent with a stark disclosure Nvidia made in last week's record earnings: the company said it was not assuming any data center revenue from China in its forward guidance. At the time, some analysts treated that caveat as conservative posturing. It now looks like the literal truth.
A Deal That Was Born Dead
The backstory here is a slow-motion collapse of what briefly looked like a geopolitical opening. In January 2026, the Trump administration formally approved H200 exports to China, subject to a 25% sales fee payable to the U.S. government — a novel arrangement designed to let Nvidia access the Chinese market while directing revenue back to Washington. Bloomberg reported last week that Nvidia had secured a license to ship a small number of H200 chips to Chinese customers.
The commercial case looked compelling on paper. Chinese technology companies had placed orders for more than 2 million H200 chips for 2026, against an Nvidia stock of roughly 700,000 units. At approximately $27,000 per chip, meeting even a fraction of that demand would have added several billion dollars to Nvidia's already record-breaking revenue stream. ByteDance alone was reportedly prepared to spend around $13.8 billion on Nvidia chips in 2026, up from roughly $11.7 billion in 2025.
But the orders never became shipments. Beijing did not formally approve imports despite Washington's authorization. Chinese officials were reportedly considering a rule requiring domestic chip purchases alongside any H200 imports — a measure designed to protect local semiconductor companies, particularly Huawei. By early January, the Chinese government had asked some technology companies to temporarily pause H200 orders while officials finalized terms. The U.S. Commerce Department confirmed as recently as February 24th that no H200 chips had been sold to Chinese customers.
That last fact is the one that matters most. The deal was approved. The licenses existed. And still: zero chips shipped. Nvidia apparently decided that holding TSMC production capacity for a product with no visible path to revenue was no longer a rational use of its most constrained resource.
Why TSMC Capacity Couldn't Wait
TSMC operates at the frontier of semiconductor physics. Its most advanced nodes — particularly the 3-nanometer process on which Vera Rubin is built — represent years of capital investment and engineering refinement, and they are chronically oversubscribed. Every wafer allocated to one product is a wafer unavailable for another.
Vera Rubin, introduced by Jensen Huang at CES 2026 in January, is already in full production at TSMC. The platform represents a generational leap: it delivers approximately 50 petaflops of inference performance, five times the output of the Blackwell generation, with partner availability scheduled for the second half of 2026. AWS, Google Cloud, Microsoft Azure, and Oracle are among the hyperscalers already committed to Rubin deployments.
The Rubin platform also includes a purpose-built inference variant, Rubin CPX, which packs 30 petaflops of compute at NVFP4 precision and 128GB of GDDR7 memory optimized for million-token context workloads — the kind of long-context processing that software coding assistants, generative video platforms, and next-generation AI agents require at scale. Together, the Vera Rubin NVL144 CPX rack system delivers 8 exaflops of AI compute, 7.5 times the performance of the current-generation GB300 NVL72.
Against that backdrop, the H200 — running on older Hopper architecture on a 4nm node — looks increasingly like yesterday's silicon. The question for Nvidia's manufacturing planning was never really about China's regulatory situation. It was about whether TSMC's most precious capacity should be tied up waiting for Beijing to decide, or redirected toward a platform with a $500 billion committed order book.
Huang told analysts last week that the $500 billion figure for combined Blackwell and Rubin orders was, in his assessment, already too conservative.
China's Countermove: Huawei Ascend and Self-Sufficiency
Beijing has been anticipating this moment for years, and its response is already underway. The Huawei Ascend 910C and 910B chips — manufactured through a combination of domestic Chinese fabs and SMIC's 7nm-equivalent processes — have become the de facto AI accelerator for Chinese technology companies unable or unwilling to wait for Nvidia clearances. Alibaba, Baidu, ByteDance, and Tencent have all significantly expanded their Huawei Ascend procurement over the past 18 months.
Huawei's ambitions in AI silicon are not modest. The company's Ascend NPU roadmap reportedly targets 4 zettaflops of FP4 performance by 2028, a figure that — if achieved — would represent a serious competitive challenge in the mid-tier inference market. Current-generation Ascend chips still lag Nvidia's Blackwell and Rubin architectures in raw throughput and software ecosystem maturity, but the gap is narrowing, and Chinese firms are building workflows specifically tuned for the Ascend stack.
The regulatory dynamic favors Beijing's self-reliance narrative. As The Diplomat reported last month, by tightly regulating the domestic distribution of any H200 imports, Chinese officials could simultaneously allow limited Nvidia purchases for prestige AI applications while ensuring the bulk of procurement stays with Huawei — protecting the local semiconductor ecosystem and its strategic independence objective in a single policy stroke.
Nvidia's China market share in AI accelerators is now projected to fall from 66% to approximately 8%, according to analyst estimates — a staggering reversal from the pre-2022 era when China represented nearly 20% of Nvidia's data center revenue.
What the Numbers Actually Say
Nvidia's decision to walk away from China H200 production would be harder to justify if the company's finances were under pressure. They are not. Nvidia reported record quarterly revenue of $68.1 billion on February 25th, up 73% year-over-year, with data center contributing $62.3 billion — more than 91% of total revenue. Full fiscal year 2026 revenue reached $215.9 billion, up 65%. The company guided $78 billion in revenue for the current quarter, roughly $5 billion ahead of analyst consensus.
The H200 generated zero data center revenue in Q4. That detail appeared in the earnings materials almost as an afterthought, but it captures the commercial reality: the chip that Nvidia is walking away from in China was already contributing nothing. The theoretical upside from Beijing approving imports at some future point cannot compete with the concrete, contracted demand from hyperscalers lining up for Vera Rubin deployments in H2 2026.
What This Means for the AI Hardware Arms Race
The implications of this decision extend beyond Nvidia's quarterly revenue figures. TSMC's 3nm and advanced node capacity is the single most constrained resource in the global AI hardware supply chain. Every major AI chip company — Nvidia, Apple, AMD, Qualcomm, and increasingly Google and Amazon with their custom silicon programs — is competing for access to the same finite wafer starts.
By fully committing TSMC's advanced capacity to Vera Rubin, Nvidia is effectively foreclosing one competitive variable: the question of whether H200 supply to China would constrain Rubin ramp availability for Western hyperscalers. It will not. GTC 2026, scheduled for March 16-19 in San Jose, is now set to be the platform where Nvidia publicly confirms the Rubin transition is complete — and where Jensen Huang is expected to unveil what he has described as a chip announcement that will "surprise the world."
For the geopolitical dimension of AI hardware, the implications are just as significant. The H200 production halt effectively ends the ambiguity about whether the U.S.-China AI chip trade could resume in any meaningful form under the current export control architecture. Nvidia has answered that question with a manufacturing decision rather than a press release: the company is not structuring its supply chain around the possibility of Chinese sales. It is structuring it around Vera Rubin, and Vera Rubin is being built for everyone except China.
That posture — a strategic decoupling encoded in TSMC wafer allocations rather than diplomatic statements — may be the most accurate signal yet of where the global semiconductor industry is headed.
