A months-long diplomatic and regulatory standoff over Nvidia's H200 GPU broke decisively this week. On Tuesday, March 17, Reuters confirmed that Beijing has granted regulatory approval for Nvidia to sell H200 chips to Chinese customers — the same day CEO Jensen Huang, speaking at GTC 2026 in San Jose, announced that H200 manufacturing was restarting and that the company had received purchase orders from "many" customers in China. The dual unlock — from both Washington and Beijing — ends one of the most closely watched chip-policy standoffs in years, and reopens a market that once accounted for 13% of Nvidia's total revenue.
How the Freeze Happened
The H200's path to China was blocked by a layered tangle of U.S. and Chinese regulatory hurdles that accumulated through 2024 and 2025. The U.S. Commerce Department's October 2023 export control expansion restricted shipments of high-performance AI accelerators to China without an export license, and the H200 — Nvidia's Hopper-architecture flagship — fell squarely in that net. What followed was an extended period in which Nvidia sought U.S. licenses for specific Chinese customers while simultaneously waiting for Beijing to issue its own import approvals.
The dual-approval process proved maddening. Reuters reported in February 2026 that China sales of H200 had been stalled by a prolonged U.S. national security review. By that point, Nvidia had already wound down H200 production to redirect TSMC wafer capacity toward its Blackwell generation — a decision that looked reasonable given the regulatory uncertainty but proved costly when Chinese demand remained strong.
In a late February SEC filing, Nvidia disclosed that the U.S. had granted a license allowing "small amounts of H200 products to specific China-based customers." But U.S. approval alone was insufficient. Beijing's hesitation to authorize imports had been the main barrier to actual shipments, according to Reuters sources. Without both approvals in place simultaneously, no chips could legally move.
The GTC 2026 Announcement
The moment of resolution came at Nvidia's annual GTC developer conference, which drew more than 30,000 attendees from over 190 countries to San Jose's SAP Center. In a Tuesday press conference, Jensen Huang confirmed that Nvidia now has clearance from both governments and is actively restarting H200 production lines. "Our supply chain is getting fired up," Huang told reporters, adding that the company had been licensed for "many customers in China" and had received purchase orders. Bloomberg reported that Huang described the outlook as "different than it was a couple of weeks ago."
CNBC separately confirmed that Huang told them Nvidia has clearance from both the U.S. and China sides. Tom's Hardware noted that Nvidia had largely wound down Hopper-class production in preparation for Blackwell and the upcoming Vera Rubin generation, but confirmed the company plans to reopen H200 orders in 2026 to meet Chinese demand.
The announcement was timed against a backdrop of Nvidia's most ambitious product launch in years. GTC 2026 was dominated by the full production rollout of the Vera Rubin platform — seven co-designed chips spanning GPUs, CPUs, LPUs, networking, and storage — and Huang's projection that Nvidia's AI chip revenue opportunity could reach $1 trillion through 2027, up from a prior $500 billion estimate through 2026. The H200 news arrived as something of a bonus alongside a platform announcement of historic scale.
Why the H200 Specifically
It is worth being precise about what is and is not being approved. The H200 is Nvidia's second-most powerful commercially available AI accelerator — powerful enough to train and run large language models at scale, but not Nvidia's current frontier silicon. The Blackwell generation (B200, GB200) and the incoming Vera Rubin chips remain off-limits to China under U.S. export controls. What Beijing has cleared is a chip that, while still highly capable, sits one full generation behind the state of the art and was already being phased out of production.
That distinction matters commercially and strategically. For Chinese AI companies — Baidu, ByteDance, Tencent, Alibaba and others — access to H200 chips represents a meaningful improvement over the H800 and A800 chips that were the de facto ceiling after earlier export restrictions. The H200 features HBM3e memory with substantially higher bandwidth than its predecessors, making it noticeably better for large-scale inference and some training workloads.
But the H200 is not Blackwell, and it is certainly not Vera Rubin. Chinese AI labs will still operate with hardware that is one to two generations behind what U.S. hyperscalers are deploying. The approval is a strategic détente, not an end to the underlying technology competition — a distinction Huang seemed to acknowledge without dwelling on.
The Groq Dimension
One detail in the Reuters report that received less attention than the H200 headline: Nvidia is also preparing a China-market version of the Groq AI chip. Reuters reported that "sources familiar with the matter" confirmed this adaptation is in progress, though details on specifications, timelines, and whether it would clear U.S. export thresholds were not available.
Nvidia acquired Groq's technology through a $17 billion licensing deal announced in December 2025, integrating the company's Language Processing Unit architecture into the Vera Rubin platform as the Groq 3 LPX — a rack-scale inference accelerator designed to handle the low-latency decode phase of AI token generation. The Groq 3 LPX, co-deployed alongside Vera Rubin NVL72 GPUs, claims up to 35x higher inference throughput per megawatt and 10x more revenue opportunity for trillion-parameter model serving.
A China-specific Groq chip, if it clears regulatory review, would give Chinese AI companies access to Nvidia's inference architecture at a level of performance that could be significant — though it almost certainly would not include the full specifications of the Groq 3 LPX deployed in U.S. data centers. The pattern established with H800 and A800 (scaled-down versions of H100 and A100 designed to stay below export control thresholds) would likely apply again.
The Revenue Math
Analysts are already revisiting Nvidia's revenue models. China generated approximately 13% of Nvidia's total revenue during peak periods before restrictions tightened — a figure that dropped substantially as the regulatory freeze solidified. Industry analysts at SemiWiki noted that reactivating even a fraction of that market adds materially to Nvidia's near-term revenue picture, particularly as H200 production can be restarted against existing infrastructure rather than requiring new tooling.
Chinese demand for AI compute has not diminished during the freeze — it has intensified. Companies unable to secure Nvidia chips at scale pivoted to Chinese alternatives (notably Huawei's Ascend 910B and 910C) and to hoarding whatever Nvidia hardware could be obtained through third-party channels. The return of a legal, direct H200 supply channel will likely trigger a significant initial surge of purchase orders as companies replenish stacks and expand deployments they had deferred.
The critical unknown is how quickly Nvidia can ramp H200 production back up. TSMC capacity that was redirected toward Blackwell and Vera Rubin cannot be instantly reclaimed. The H200 uses a mature 4nm process node and should be less constrained than bleeding-edge processes, but meaningful supply will take quarters to materialize. Nvidia's comments about "restarting" production rather than "resuming full supply" suggest a measured ramp rather than an immediate flood of chips.
Geopolitical Context: A Calculated Thaw
The approval did not emerge in a vacuum. The broader U.S.-China tech relationship has been in a complex negotiation posture throughout early 2026, with the Trump administration weighing the economic cost of export restrictions against national security objectives. Permitting H200 sales — while maintaining the embargo on Blackwell and Vera Rubin — represents a middle-ground position that generates commercial value for a major U.S. technology company without releasing frontier AI capabilities.
Beijing's decision to grant import approval reflects its own calculus. Chinese AI companies are under competitive pressure to close the capability gap with U.S. counterparts, and access to H200 chips accelerates that effort even if it doesn't eliminate the gap. For the Chinese government, facilitating this access while continuing to develop domestic alternatives (Huawei, Biren, Moore Threads) is a rational hedging strategy rather than a capitulation.
Neither government appears to view this as a permanent normalization. The underlying structure of technology competition between the two countries remains intact. What changed this week is one specific data point in an ongoing negotiation — one that happens to be worth billions of dollars to Nvidia and gigawatts of AI compute to China's leading technology companies.
What Comes Next
The immediate practical effect will be purchase order flow. Chinese hyperscalers and AI labs that have been starved of Hopper-class supply will move quickly once a purchase mechanism is established. Nvidia's supply chain ramp will be the throttling factor in 2026, with meaningful volume likely achievable by Q3 or Q4.
Longer term, the H200 China approval does not change the fundamental dynamic: Vera Rubin and whatever succeeds it will remain restricted, meaning the hardware gap between U.S. and Chinese AI infrastructure will persist and likely widen as Nvidia's roadmap accelerates. The H200 represents a floor, not a ceiling, for what Chinese companies can access — and that floor is one the industry set more than 18 months ago.
For Nvidia, the approval is unambiguously positive. It adds a revenue stream that was effectively written off, validates the company's export compliance strategy, and — as Huang noted at GTC 2026 — arrives precisely as the company is also forecasting a $1 trillion chip revenue opportunity through 2027. The H200 China reopening is not the biggest story out of GTC this week. But it may be the most commercially consequential one in the near term.
This article updates our previous coverage: Nvidia Abandons China H200 Production, Pivots TSMC Capacity to Vera Rubin (March 6, 2026).
