Bargo
Semiconductors · July 20, 2026

The H200 opening to China produced almost no sales

Washington finally cleared H200 sales into China in July 2026. Beijing had told its firms to stop buying six months earlier.

Bargo · 2026-07-20

Washington spent two years deciding whether to let Nvidia sell H200 chips into China. In July 2026 it finally cleared roughly ten Chinese firms to buy them. On July 14, Commerce Under Secretary Jeffery Kessler testified that "very few" had actually shipped.

The door opened and almost nobody walked through it. That is the story worth understanding, because it changes what the China line in an Nvidia model is actually worth.

What happened, in order

Beijing got there first. In January 2026, Chinese authorities told domestic firms to halt H200 orders, and customs began blocking shipments the following week. When the US eased its rules six months later, the buyers were already under instructions not to buy.

The sequence:

Nvidia (NVDA) trades at $205.70 intraday on July 20, 2026, down 0.82 percent from a prior close of $207.40.

For two years the market treated China access as a free option on Nvidia earnings. The option got exercised and paid almost nothing. That is a cleaner outcome than an open question, and it is worth pricing as settled rather than pending.

There is a second reason the shipments went nowhere. The Financial Times reported in mid July 2026, quoting a source in the supply chain, that Chinese domestic chip suppliers are sold out, including lower grade parts nobody previously wanted. Demand inside China is real. It is being met locally.

Demand outside China is not the problem

Nvidia booked two of its largest non hyperscaler commitments on record in the same stretch.

Foxconn was awarded a roughly $52 billion SpaceX order covering about 13,000 GB300 racks, with deliveries running from late 2026 into the first quarter of 2027. At 72 GPUs per rack that is close to one million GPUs, and about $4 million per rack.

Japan committed separately. The Japanese government is investing up to 1 trillion yen over five fiscal years into Noetra, its national AI company. Noetra's Nvidia build is specified at 13,750 Vera CPUs and 27,500 Rubin GPUs drawing 140 megawatts, with operations starting in 2028.

Put those against the China number. One private space company committed more dollars in a single order than the entire cleared Chinese buyer list was ever likely to spend.

What rental prices say about real H200 demand

The chip China will not buy is the tightest chip in the rental market. As of July 20, 2026, the H200 reads 66.6 on the Bargo Compute Tightness Index, a tight regime, up 10.9 points in 30 days.

GPU rental pricing, July 20 2026
GPU On demand $/hr Spot $/hr Spot discount Tightness 30 day change
H200 $4.28 $3.19 25% 66.6 (tight) +10.9
B200 $7.18 $3.69 49% 51.6 (balanced) +2.9
H100 $3.79 $1.67 56% 38.7 (loose) 0.0

Read the spot discount first. Spot capacity is what a provider rents out cheaply when it has spare machines sitting idle. An H200 renter saves only 25 percent by going spot. An H100 renter saves 56 percent. When that discount narrows, providers have nothing idle to give away.

Prior generation H100 capacity sits loose at 38.7 across 146 listings and has not moved in 30 days. The blended index across all three reads 47.5, balanced, tightening by 2.1 points over 30 days. Live data: H200 rental price and the Compute Tightness Index.

Chinese models are winning volume, not revenue

Export controls were meant to slow China down. On token usage they did the opposite.

Moonshot released Kimi K3 in mid July 2026, a 2.8 trillion parameter model with open weights, free to use. The semiconductor index fell 12.5 percent that week on the assumption that free models destroy compute demand. Within days Moonshot paused new subscriptions because it had run out of GPU capacity. A free model still burns chips, which is the part the selloff got backwards.

Six of the top ten models by token usage on OpenRouter are now Chinese, as of July 2026. CNBC reported that Chinese models have accounted for at least 30 percent of US enterprise token volume every week since February 8, 2026, reaching as high as 46 percent.

Bargo's own token data shows where the money still sits.

Share of weekly inference tokens by provider group, July 19 2026

Open source models carry 53.6 percent of tracked token volume at a blended $0.44 per million tokens. Claude carries 12.4 percent of volume at $10.00 per million. OpenAI carries 7.2 percent at $3.50. Volume and revenue are moving in different directions.

Blended effective price per million tokens

Total volume rose 32 percent in 30 days to 58.5 trillion tokens per week while the blended price fell to $1.97 per million. That is a market getting bigger, not margins collapsing. See Token Demand for the live split.

Does Vera Rubin change the math

Nvidia detailed the Vera Rubin NVL72 on July 18, 2026. HBM4 memory bandwidth of 22 terabytes per second, which is 2.8 times Blackwell. NVLink 6 at 3.6 terabytes per second per GPU, double the prior generation. Nvidia claims 4 times fewer GPUs needed for post training work.

The catch is what it does to the installed base. SemiAnalysis published the arithmetic in "AI Value Capture, The Shift To Model Labs" on May 1, 2026: a Vera Rubin NVL72 has to rent for under $0.70 per petaflop, at roughly a 15 percent five year project return, before a customer is indifferent between Rubin and the existing GB300. Above that, the older chip is the better buy.

Every generation is faster, and every generation squeezes the rental economics of the one before it. That pressure lands on the independent cloud providers first. CoreWeave (CRWV) trades at $76.29 intraday on July 20, roughly 50 percent below its 52 week high of $153.20. Nebius Group (NBIS) trades at $185.60.

A newer competitive question has nothing to do with China at all. On July 20, 2026, The Information reported that Alphabet is in talks with several independent cloud providers to sell them TPUs, and is weighing backstop deals that would guarantee payments on TPU loans. That would put Google silicon directly inside Nvidia's own customer base. Alphabet (GOOGL) trades at $348.69.

What to watch

This is a research note, not investment advice.

Sources

More research at bargo.ai/research.

Get Bargo research in your inbox
One email when we publish. No spam, unsubscribe anytime.
Get Bargo research in your inbox
One email when we publish. No spam.

Related research

AI
AI data center power stocks are still beating estimates

Chip stocks fell into a bear market this month. The companies building the physical plant behind AI kept beating their own guidance.
AI
AI Inference Demand Just Surged 149% — and the GPU Market Is Tightening

Token consumption hit 58.5 trillion per week after GPT-5.6 and DeepSeek V4 launched in July. H200 rental capacity is already tight. The Jevons effect is real.
AI
The 5x That Changes Everything: How NVIDIA Cut DeepSeek V4 Token Costs Without a New Chip

In one month, software optimizations alone made DeepSeek V4 five times cheaper to serve on Blackwell hardware. The implications go far beyond one benchmark.