The Financial Times reported this morning that Beijing has allowed ByteDance and Tencent to each receive about 10,000 NVIDIA H200 chips at their mainland facilities in recent weeks, with similar approvals said to be close for a handful of other Chinese technology groups. It is the first mainland delivery of consequence since the US cleared H200 exports under conditions, and the timeline has spent the day describing it as China reopening.
Provenance first, because it decides how much weight the rest can carry. The FT story is paywalled and every figure below reached us through aggregators quoting it. The ~10,000-per-company number is a reported delivered quantity, not a disclosure by NVIDIA, ByteDance or Tencent, none of whom have confirmed it. Nothing here is in a filing.
What is in a filing is the thing that makes 20,000 chips interesting at all. In its Q1 FY2027 CFO commentary NVIDIA said two things about China, both quoted in our own model:
"No shipments of Data Center Hopper products to China occurred during the quarter, compared with $4.6 billion in the first quarter of fiscal year 2026."
"Revenue is expected to be $91.0 billion, plus or minus 2%. We are not assuming any Data Center compute revenue from China in our outlook."
The company zeroed China out of its own numbers. So any licensed volume is upside that neither the guidance nor any consensus built on it contains — which is exactly why the headline travels. The table and chart above are the reason it should not travel very far.
Mainland or Hong Kong? Both, and the difference matters
Two of the most specific accounts circulating today contradict each other: one says the chips are in mainland facilities, another says geofencing keeps them stationed in Hong Kong. That is a load-bearing disagreement — revenue recognised against a mainland customer is China revenue in a way that hardware parked in Hong Kong may not be.
On the reporting available, both are describing different halves of the same policy. Small batches have entered the mainland; Beijing is separately steering the bulk of the larger licensed allowances offshore, with Hong Kong the preferred destination, to protect the domestic chip ecosystem. Not a contradiction — a managed thaw with a small mainland door and a larger offshore one.
We flag it rather than resolve it, because neither half is sourced to a document we can read.
What 20,000 chips are worth
This is where the piece either works or does not, and it turns on a number NVIDIA does not publish: the price of an H200.
We do not hold it, and we are not going to invent it. What we can do is bound it. China-specific quotes through the past year have clustered near $27,000 a unit, with 8-GPU modules quoted around 1.4–1.5 million RMB, so a range of $25,000 to $40,000 covers the reported evidence with room at the top. Twenty thousand units is then:
| Unit price | Revenue | Share of the $91.0B guide |
|---|---|---|
| $25,000 | $0.50B | 0.55% |
| $40,000 | $0.80B | 0.88% |
Both ends are ours, and a reader who rejects the range should reject the dollar figures with it. So here is the same finding stated in a form that survives any price you prefer:
NVIDIA guided to $91.0 billion plus or minus 2%. That tolerance is $1.82 billion. For 20,000 H200s to fill it, an H200 would have to sell for $91,000 each.
Nobody thinks an H200 costs $91,000. Which means this shipment — the event being read as China reopening — is smaller than the rounding NVIDIA itself put on a single quarter's guidance. It is also 11% to 17% of the $4.6B of China Hopper revenue the company ran in the year-earlier quarter, so it does not restore the old China business either. It restores about a seventh of one quarter of it.
What our model already priced
Our NVIDIA forward model carries a fourth case beside bear, base and bull, called China returns, and it exists for precisely this event: the company excluded China from its own outlook, so the model isolates what happens if that is licensed back rather than burying it in the base case.
| Case | Fair value per share | Terminal-year revenue |
|---|---|---|
| Bear | $63.15 | $380.9B |
| Base | $245.07 | $925.9B |
| China returns | $350.31 | ~$1.12T |
| Bull | $625.45 | $1.68T |
The gap between the China case and the base case is $105.24 a share. On 24.391B diluted shares that is $2.57 trillion of equity value — and about $194 billion a year of extra revenue by the terminal year.
Now put the two together, because this is the whole article: the model's China case is worth roughly $194 billion of revenue a year. The event is worth $0.5–0.8 billion, once. That is around a third of one percent of the thing the tape is being invited to price.
The model's own note on that case says why the mismatch is structural rather than a mistake:
"Read it as a step, not a slope: the model applies the tilt as compounding growth, so it understates the first year and only lands the size of the prize near the end of the horizon."
That warning cuts both ways, and the honest reading is the one that respects both directions. Twenty thousand units is not the prize. It is evidence that the channel can open at all — and a compounding case is the right shape for a channel, if a channel is what this is. Nothing yet published shows that. Roughly ten Chinese firms are reported to hold US licences with per-company ceilings said to run to 75,000–100,000 units, and Chinese regulators still gate individual orders case by case. A licence ceiling is not a shipment, and one batch is not a slope.
So the model does not change today. A case built on a compounding reopening should not be marked to a single reported batch, and export-control stories reverse often enough that doing so would be the more expensive error. What changes is that the case now has a first data point instead of none.
Hong Kong is a power problem before it is a policy one
There is a physical constraint under the geofencing story that is worth doing the arithmetic on, because it bounds how fast the offshore half can matter regardless of what Beijing decides.
NVIDIA is reported to hold roughly 500,000 H200s in inventory, largely built for Chinese customers. At eight GPUs to an HGX server, that is 62,500 servers; at roughly 10 kW a server, 625 MW of IT load. Hong Kong's entire installed data-centre base is cited in secondary reporting at around 580 MW.
The inventory built for this market would, if deployed there, need more power than the territory currently has installed — before cooling overheads, and before anything else in Hong Kong keeps running. Today's 20,000 units are about 25 MW, which fits comfortably. The next order of magnitude does not.
That is the real reason to treat "managed thaw" as a description rather than a slogan. The mainland door is small by policy; the offshore door is small by physics.
The Baidu anomaly this might explain
One reason to care about a small shipment is that it is a demand signal from customers we cannot otherwise see.
Our Baidu Q2 coverage flagged a discrepancy the release never explained: GPU Cloud revenue grew 283% year over year, accelerating from 184% the quarter before, while the AI Cloud Infra segment it sits inside fell 17% sequentially, from RMB 8.8B to RMB 7.3B. A triple-digit growth line inside a shrinking segment is either a mix effect or a constraint.
If Chinese hyperscalers have been supply-constrained rather than demand-constrained, an H200 channel is the missing variable. We are not asserting that — Baidu is not among the companies named in today's report, and one quarter's segment mix has several possible explanations. It is a hypothesis the next Baidu release can settle, and it is the reason this piece carries a second ticker.
What to watch
- NVIDIA's next outlook, expected around 26 August. The single cleanest checkpoint: whether the "we are not assuming any Data Center compute revenue from China" language survives the quarter. If it goes, the China case stops being hypothetical. If it stays after a reported mainland delivery, that tells you what the company thinks this batch is.
- A second batch. One shipment is a batch; a second, to different customers, is a channel. The distinction is worth more than the volume of either.
- Whether any per-company licence ceiling is actually approached. The reported 75,000–100,000-unit ceilings are the difference between a rounding error and a real line. At 20,000 delivered against a 75,000 ceiling, the interesting number is the gap, not the delivery.
- Hong Kong data-centre capacity announcements. The offshore half of this policy cannot scale past the territory's power. New capacity, or a relaxation of the mainland gate, is what would let the licensed volume move.
- Baidu's next AI Cloud Infra line. If the segment grows sequentially again while GPU Cloud stays triple-digit, the supply-constraint reading gains a point. If it falls again, it was mix.
The 20,000-unit figure — about 10,000 each to ByteDance and Tencent, at mainland facilities, in recent weeks — is from the Financial Times (Zijing Wu, 19 August 2026), reaching us through aggregator posts quoting the paywalled story; neither company nor NVIDIA has confirmed it, and it is not in any filing. The reported US licence ceilings of 75,000–100,000 units per company, the roughly ten licensed Chinese firms, the Beijing steer toward Hong Kong, the ~500,000-unit H200 inventory and Hong Kong's ~580 MW installed data-centre base all come through the same secondary reporting and are stated as claims, not disclosures. H200 unit pricing is not disclosed by NVIDIA; the $25,000–$40,000 range is our reconstruction from China-specific quotes clustering near $27,000 and 8-GPU modules quoted at 1.4–1.5M RMB, and every dollar figure derived from it is a bound rather than an estimate. The 8-GPU server and ~10 kW figures are standard HGX configuration, not a company statement. The $91.0B outlook and its ±2% tolerance, the absence of Data Center Hopper shipments to China in the quarter, and the $4.6B year-earlier comparison are quoted from NVIDIA's Q1 FY2027 CFO commentary (CIK 0001045810, 20 May 2026) as carried in our forward model. Fair values of $63.15, $245.07, $350.31 and $625.45, terminal revenues, 24.391B diluted shares and the "step, not a slope" note are that model's own output and assumptions as of its 2026-08-17 basis — they are ours, not anyone's published target. Baidu's GPU Cloud +283% and AI Cloud Infra −17% sequential figures are from our Q2 coverage. The 26 August NVIDIA date is the estimated one on our earnings calendar. No price or market capitalisation is quoted here.
