A pile of Nvidia chips roughly 600,000 units tall is sitting in Hong Kong, a few hours’ drive from the customers who want it, and it is going nowhere. The Nvidia chips in that pile are H200s, the most advanced Nvidia chips the company is allowed to sell abroad under current US rules, and according to a Reuters report published Thursday, Beijing has simply not approved large-scale imports. Customs clearance is stalled. The crates stay crated.
That single fact reframes the AI trade narrative of the past two years. Washington loosened its stance first: the Bureau of Industry and Security now reviews H200 export applications case by case, with conditions. Chinese tech companies want the silicon to train and run their large language models. Their own government, for now, is declining the delivery.
Reuters had the story first, and the numbers have been circulating through supply-chain briefings ever since. What follows is the full shape of the stalemate: how the pile formed, why neither government is in hurry to move it, and what it costs the company whose name is on the boxes.
This is not a story about a shipment delay. It is a story about a stockpile of Nvidia chips with no legal path forward, a Beijing strategy to cut reliance on US technology, and Nvidia sitting in the middle holding inventory that earns nothing while it waits.
Table of contents
The Pile Up Nobody Could See Before
The headline number is what makes this story new. Reuters, citing two people familiar with the matter, put the backlog at roughly 600,000 H200 processors waiting in Hong Kong — a scale of stalled trade that had not been reported before. The Nvidia chips are physically close to their buyers and legally frozen at the border.
Hong Kong is doing exactly what Hong Kong does: warehousing. Nothing is confiscated, nothing is seized. The freight arrives, gets wrapped, and sits in bonded facilities across Hong Kong’s vast warehousing belt while Chinese import approvals never come. Local logistics firms reportedly sublet space for the growing inventory, and the handling invoices alone tell a story about how long everyone expects this to last. Chips cleared to leave one side of the border cannot enter the other.
Precision matters here, because the numbers around this pile will get quoted for months. The 600,000 figure describes H200 units awaiting import approval — not units seized, not units denied outright, and not a permanent ban. Beijing has withheld approval for large-scale imports; there is no published timeline for a decision and no published criteria.
The reporting also notes that as of July only a very small number of H200 chips had actually reached China. Six hundred thousand is therefore not the size of a failed shipment — it is the size of a demand pipeline for Nvidia chips that built up ahead of approvals that keep not arriving. Chinese companies ordered, Nvidia cleared US-side review, and the goods parked.
Why park them in Hong Kong at all? Because the US-side door opened. Once the Bureau of Industry and Security moved to case-by-case licensing for the H200, Nvidia could legally ship. The company positioned inventory at the last stop before the mainland, ready for the day Chinese customs says yes. That day has not come.
| What we know | Where it comes from | How solid it is |
|---|---|---|
| Roughly 600,000 Nvidia chips (H200 units) are held in Hong Kong awaiting Chinese import approval | Reuters, citing two people familiar with the matter | Newly reported scale; figures are approximate and may move |
| Beijing has not approved large-scale imports; customs clearance is halted | Reuters, corroborated by follow-up coverage | Solid — consistent across outlets |
| Only a very small number of H200s had reached China as of July | Research findings cited in follow-up coverage | Directional; exact counts are not public |
| US export review for the H200 is case-by-case, with conditions on volume | Federal Register notice, January 2026 | Solid — published rule text |
Why the H200 Is the Chip Chinese Labs Keep Asking For
The H200 sits near the top of Nvidia’s AI lineup, and unlike the Blackwell parts, it is not banned from export in principle. Chinese internet companies want Nvidia chips of this class for one mundane reason: memory bandwidth. Large language models live and die by how fast data moves between chip and memory, and the H200’s 141 gigabytes of HBM3e memory at 4.8 terabytes per second, per Nvidia’s own spec sheet, is the difference between a model that fits and one that doesn’t.
For Chinese labs the calculus is practical, not political. Frontier training runs are won by clusters, and a cluster of H200s with fat memory pipes absorbs the long-context workloads that choke smaller-memory parts. Homegrown accelerators can serve inference; the hardest training jobs still gravitate toward Nvidia silicon.
Chinese buyers have been here before. Nvidia built the H20 — a China-specific chip deliberately de-rated to stay inside US export limits — precisely to keep serving this market lawfully. Washington tightened the screws again in April 2025 by requiring export licenses for it, and the episode ended in a negotiated arrangement in which a slice of China revenue flowed back to the US Treasury.
The lesson Chinese buyers drew was not “wait for the next de-rated part.” It was that Nvidia chips are policy instruments that either government can rewrite mid-flight. Yet orders for Nvidia chips kept coming, because the alternative — building model architectures around less mature domestic accelerators — still costs engineering months.
So the pile in Hong Kong is the physical residue of that tension: buyers hedging for access, Nvidia hedging for approvals, and two governments refusing to blink first.
Washington's Half-Open Door for Nvidia Chips
What makes the Hong Kong pile bizarre is that the US government is not the blocker this time. In January 2026 the Bureau of Industry and Security published a Revision to License Review Policy for Advanced Computing Commodities — the rule that moved H200 exports to China from presumption-of-denial to case-by-case review.
Case-by-case, however, comes with strings, and the strings are where the real policy lives. Exporters must certify that sales to China will not reduce the semiconductor capacity available to US customers. In practice that condition could cap shipments of Nvidia chips at roughly 50% of equivalent US end-use volumes — Washington keeping its own companies first in line.
The design is deliberate. A hard ban would have handed China a propaganda issue and Nvidia a clean writ to write off the market. A free-for-all would have flooded a strategic competitor with frontier compute. Case-by-case licensing with a volume ceiling keeps every applicant for Nvidia chips dependent on a regulator’s mood — and keeps every approval revocable.
Nvidia leaned into the opening. It sought Chinese import approvals, positioned inventory in Hong Kong, and told investors it was prepared to ship the Nvidia chips overnight. That is why the goods are already at the border rather than still in the supply chain: the US-side process is functioning. Applications clear. Boxes fly.
Beijing now holds the only veto that matters. And every month the veto sits unexercised, the pile grows, which is its own form of leverage in the broader trade negotiation — Nvidia chips as a bargaining chip, literally.
Why Beijing Is Saying No to Chips It Wants
Beijing’s refusal looks irrational only if you assume the goal is maximum compute at minimum cost. The goal, as the reporting frames it, is control. Approving large-scale imports would help Chinese companies build AI faster this year — and would re-attach them to an American vendor that Washington can regulate at will. Every approved pallet of Nvidia chips is another cycle of dependence renewed.
The strategy is a slow decoupling with Chinese characteristics: let domestic suppliers grow behind a wall that is lowered just enough to keep everyone guessing. Import approvals become a dial Beijing can turn — a trickle here to relieve pressure, a freeze there to punish or to save as a concession in trade talks.
Consider what the freeze buys. Chinese buyers keep ordering Nvidia chips in anticipation of approval, which means their procurement data, roadmaps and model plans stay organized around Nvidia’s release train — information both governments can read. Meanwhile the freeze gives state-backed labs a standing justification to spend on domestic alternatives, and gives domestic vendors their most persuasive sales argument: watch two governments fail to agree while Nvidia chips gather dust in a warehouse.
There is also a simpler reading: bureaucratic paralysis. A decision to approve frontier Nvidia chips is a strategic admission that Chinese silicon is not yet sufficient. No ministry wants to sign that. Freezing the paperwork costs Beijing nothing visible while the pile, conveniently, sits outside the mainland.
And the timing is not accidental. With Washington visibly hungry for Chinese purchases of American technology in any trade détente, an unapproved pile is a loaded bargaining table. Approvals can be doled out — priced in something else.
The Domestic Champion Play
Every month of the freeze is a subsidy in disguise for China’s homegrown chip makers, and nobody embodies that better than Huawei. Its Ascend accelerators are already the default fallback for Chinese model shops that cannot get approved Nvidia chips, and coverage this September pointed to Huawei planning a new AI chip for a Q1 2027 launch — a direct shot at the next training generation while the H200 question stays stuck.
The market has noticed. By Nvidia’s own framing earlier this year, Nvidia itself largely conceded China’s AI chip market to Huawei — a remarkable sentence from the company that designed the category. Huawei’s chips carry a political advantage no spec sheet can beat: they cannot be embargoed by a foreign government.
Cambricon, Moore Threads and a parade of provincial startups round out the field, sustained by procurement rules, subsidies and — increasingly — real demand, because a freeze on imports is the best sales pitch domestic suppliers could ask for. The software gap remains: CUDA’s moat is a decade deep, and Chinese labs pay a tax in engineer-hours to port every model.
That tax shrinks with every generation. Models built from scratch on domestic stacks cost nothing to port. The freeze therefore compounds: short-term pain for Chinese labs, long-term entrenchment for domestic vendors. The pile of Nvidia chips in Hong Kong is, from this angle, not a warehouse at all — it is a countdown timer.
| Home team | Where it stands | Why the freeze helps |
|---|---|---|
| Huawei Ascend 910C | In production; the default domestic training fallback | Labs forced to port models stay in the Huawei toolchain |
| Huawei Ascend 960 | Reported for a Q1 2027 launch | Success would land while Nvidia’s China access is still frozen |
| Cambricon / Moore Threads | Inference-heavy niches, expanding into training | Procurement rules favor domestic parts outright |
| Provincial cloud programs | State-funded clusters built on domestic silicon | Guaranteed customers regardless of import policy |
What the Pile Costs Nvidia
Nvidia chips stored in Hong Kong do not generate revenue until they ship, and there is no reported timeline for when it might. Each month of the freeze converts cutting-edge Nvidia chips into depreciating inventory: the H200 is already a generation behind Blackwell, and the clock does not care that the boxes are shrink-wrapped.
There is no disclosed impairment on the Hong Kong stockpile — these are saleable goods awaiting a permit, not chips banned from existence — but the accounting understates the strategic cost. The China question now shapes Nvidia’s guidance the way weather shapes farming: not a line item, an assumption everything else hangs on. Sell-side analysts who cover the company now model China as an option struck at a discount, and management has stopped attaching a timeline to it, which tells you how open that assumption remains.
The pile is also a lesson in industrial policy physics: market access can be lost at either end of a trade route, and sometimes at both simultaneously. Export approval from Washington turned out to be only half the equation; the receiving government’s import approval is the other half, and it is Beijing’s half. No amount of lobbying in Santa Clara files a Chinese customs declaration.
And even the half that Washington controls is rationed. The 50% capacity condition means that if Beijing ever relents, the volume of Nvidia chips that Nvidia can send may still be capped by how much it ships to American customers — a ceiling written into the rulebook so the US rebuilds its own supply first, always.
Add the collateral damage: a frozen China channel tightens global demand assumptions, feeds the broader AI chip supply squeeze, and pushes Nvidia to lean harder on sovereign AI deals and US data-center buildouts to keep the growth line intact. The company is not waiting for China to save the story; it is making the story without China.
Three Ways the Pile Moves
Beijing approves a managed slice — say, a quota of Nvidia chips sized to relieve pressure without rewiring the strategy. Chinese giants get just enough H200 capacity for their hardest jobs, domestic champions keep the rest, and the pile drains slowly enough that approvals remain an annual negotiation. This is the path of least institutional resistance, and the 50% US cap conveniently makes a trickle the maximum legal flow anyway. Quotas would also be quotable: Beijing could announce a number, keep it well below demand, and let both governments claim victory while the reservation list grows.
Nothing moves until the next trade accord. The pile becomes permanent scenery, Nvidia books more write-downs on aging stock, and Chinese labs finish migrating their roadmaps to domestic stacks. This scenario is Beijing’s cheapest option and Huawei’s favorite — every month of inertia rewrites one more procurement contract in Chinese. For the Nvidia chips parked in Hong Kong, none of the three paths comes free.
A deal — rare earths easing, tariff truces, something else — unlocks approvals in a single dramatic stroke, exactly as past détente moments produced sudden chip sales. If so, the pile of Nvidia chips in Hong Kong converts from sunk cost to deliverable inventory almost overnight, and Nvidia ships against demand deferred for quarters.
Which path wins tells you more about where Washington and Beijing think the AI race is decided than any keynote will. The chips themselves are already placed; only the paperwork is missing.
The Lesson Written in Shrink-Wrap
Strip away the geopolitics and the Hong Kong pile leaves every AI company in the room with the same takeaway: the most expensive Nvidia chips in computing are now a document problem. Nvidia chips get manufactured, licensed, insured and flown halfway around the world — and then sit in a bonded warehouse for months because one signature in one ministry is not forthcoming.
This is how supply chains get rerouted under permanent policy risk. Buyers in every regulated market — the Gulf, India, Southeast Asia — are learning to price political time alongside shipping time, and contract lawyers have quietly rewritten force-majeure clauses for a world where the bottleneck is a ministry window, not a port crane. Nvidia’s answer for its chips so far has been to spread exposure: sovereign AI factories, government-backed clusters, China weighing whether to let ByteDance and Alibaba buy newer Nvidia chips as a template for how market access might be rationed and re-imported piecemeal.
If you want the early signal, ignore the summits and watch the customs forms. A first approved declaration is the tell; steady weekly approvals are the trend. The pile itself is a barometer — when it drains without fanfare, the thaw is real. When it grows while diplomacy headlines glow, the freeze is policy, not paperwork.
For now, 600,000 of the world’s most coveted Nvidia chips do what nothing in this industry has done in decades: they wait. The companies that made them, the companies that bought them, and the governments that could free them are all playing a longer game than the inventory can comfortably age in.
References
Nvidia’s H200 chips pile up in Hong Kong as Beijing withholds approval — Crypto Briefing
Revision to License Review Policy for Advanced Computing Commodities — Federal Register
Bureau of Industry and Security — Official Site
NVIDIA H200 — Official Product Page
Huawei Ascend (chip) — Wikipedia
Artificial intelligence industry in China — Wikipedia
Ministry of Commerce of the People’s Republic of China — Wikipedia
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
