---
title: "About 600,000 Nvidia H200 Chips Are Reported to Be Sitting in Hong Kong Warehouses"
date: 2026-11-19
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-h200-hongkong-stockpile
language: en
---

# About 600,000 Nvidia H200 Chips Are Reported to Be Sitting in Hong Kong Warehouses

> Reuters reports roughly 600,000 H200 accelerators waiting in Hong Kong as import approvals trickle out — a computing bottleneck measured in permits, not fabrication capacity.

They are boxed, inventoried and within a short drive of the mainland border. They are also not where the companies that want them need them.

About 600,000 Nvidia H200 chips are stored in Hong Kong, waiting for permission to move into mainland China, according to a Reuters report in November 2026. The demand is real. The bottleneck is a licence.

## Key takeaways

- Roughly 600,000 H200 accelerators are held in Hong Kong warehouses — far more than previously reported volumes — according to Reuters sources.

- Several large Chinese internet companies have each received about 10,000 H200s in recent months, but Beijing has not approved large-volume imports.

- An eight-card H200 server was quoted at more than RMB 6 million (≈ US$840,000 / HK$6.6M) in mainland China by late October — about three times its price a year earlier.

- The H200 is not Nvidia's newest chip, but its combination of capability and software ecosystem still makes it useful for training large models (大模型), which is why the permit pace matters.

## A queue, not a shortage

The image that fits is a customs queue, not a factory fire. The chips exist. They have been made, shipped to Hong Kong and parked. What is missing is a decision to let them cross into the mainland in bulk.

Reuters, citing two people familiar with the matter, said Beijing has withheld approval for large-scale imports even as Chinese technology companies ask for more computing power (算力). Three people said some chips have been allowed in over recent months, and that several major internet firms have each obtained about ten thousand. Smaller quantities have moved through intermediaries and indirect routes; the sources did not say how many.

Neither the Hong Kong government nor China's industry and planning ministries responded to requests for comment, and Nvidia and the Chinese companies reported to have received the chips did not respond either. That silence is normal for a story built on anonymous sourcing, and it is also a reason for caution.

## Why the H200 still matters

It would be easy to dismiss the whole episode as a contest over an older chip. The H200 is not Nvidia's latest AI processor. But "not newest" is not the same as "not useful."

Training a large model depends on more than raw arithmetic. It depends on memory bandwidth, interconnect and a software stack that engineers already know. Chinese alternatives have closed some of that gap, but the H200 retains capabilities that domestic products have yet to match, particularly when paired with Nvidia's tooling. For a team trying to keep a training run on schedule, access to a known platform is worth a great deal.

The restriction story has two layers. Export controls adopted from 2022 onward have limited what Nvidia can sell into China, and a policy shift in 2025 moved certain chips to a case-by-case review rather than a blanket denial. Import licensing on the Chinese side is a separate gate. A chip can be legal to sell and still not be allowed in quickly. The warehouse stockpile is what happens when those two gates are out of sync.

## The price signal

Scarcity shows up in quotes. One person told Reuters that by the end of October, a server carrying eight H200s was offered in mainland China for more than RMB 6 million (≈ US$840,000 / HK$6.6M), roughly triple the price a year earlier. Servers built with Nvidia's newer Blackwell chips, which are also under export restrictions, were quoted above RMB 20 million (≈ US$2.8M / HK$22M) each.

Those are unofficial-market numbers from a single source, not audited transaction prices. They are still useful as direction: when authorised supply is slow, buyers turn to intermediaries, and the premium compounds.

Earlier reporting said officials had told several Chinese firms they might be allowed limited H200 purchases for training, with a total possibly below 200,000 chips and a preference for domestic processors in inference. That figure sits awkwardly beside a 600,000-chip stockpile, which is exactly the point — the gap between what is physically present and what is permitted is the story.

## Honest limitations

The stockpile number comes from Reuters' anonymous sources. I have not seen warehouse records, procurement notices or contracts, and no other outlet has independently reproduced the 600,000 total. Treat it as a reported figure, not a verified inventory.

The price quotes are similarly single-sourced. They describe offers, not necessarily completed sales, and they will move with policy. The distinction between an export rule and an import licence is drawn from public reporting; I have not quoted the current texts of either regime, and both can change.

This article deliberately stays with the industrial and commercial angle. It does not adjudicate the trade and technology dispute behind the controls, name political figures, or speculate about future policy. Those questions need different sources and a different kind of argument.

The piece also does not assess whether domestic accelerators could substitute for the H200. That is a technical comparison requiring benchmarks and supply data I have not gathered here.

## What readers can do now

- **If you buy compute, separate delivery from availability.** A chip that is legal to sell may still be stuck at a border. Build permit timing into procurement plans, not just price and lead time.

- **If you are comparing clouds, ask where the accelerator physically sits.** Remote access raises its own compliance questions, and a contract that ignores jurisdiction is a risk you inherit.

- **If you are reading the next headline about AI compute, watch the licence count.** In this episode the binding constraint is not fab capacity or model design. It is a queue of approvals.

The 600,000 chips in Hong Kong are a strange kind of bottleneck: the hardware is finished, the buyers are waiting, and the deciding variable is a stamp. For anyone planning an AI project that depends on accelerated computing, that is a more practical thing to track than the next benchmark release.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-h200-hongkong-stockpile
Free to quote with attribution and a link to the original.
