Nvidia orders from ByteDance, Alibaba and Tencent
$35B+
More than 3 million RTX PRO 5500 cards at about $11,400 each, Reuters, Oct. 20, 2026
Hopper shipments to China
<1%
Share of Nvidia data-center revenue, quarter ended July 26, 2026
H200 chips parked in Hong Kong
600,000
Awaiting mainland import approval, Reuters, Nov. 5, 2026
Nvidia H200 write-down
$0.4B
Excess inventory and purchase obligations, first half of fiscal 2027
Non-Chinese quota on cloud rentals
51%
Minimum share of GPU services Nvidia now requires cloud partners to give non-Chinese customers, Nov. 9, 2026
Remote Access Security Act
369–22
House passage in January 2026; still awaiting a Senate vote
The scramble
Chinese labs are buying access, not chips, before Washington closes the remote-access door
Reuters reported on Nov. 12 that Chinese AI firms are reserving Nvidia-powered computing capacity at data centres outside China — Malaysia, Singapore, Europe — because they expect Washington to tighten restrictions on remote access to controlled processors. The vehicle is the Remote Access Security Act, which would give Commerce authority over cloud-based access to restricted technology. Passed 369–22 by the House in January 2026, the bill still awaits a Senate vote, and that is the deadline the buying is running against.
- The trigger is personal: NVIDIA's board statement of Nov. 10 named former acting US Attorney Joon H. Kim to run an independent investigation of the controversy surrounding CEO Jensen Huang, and the company opened its own probe on Nov. 8.
- The route is rental, not purchase. Overseas data centre operators are being asked to reserve GPU capacity, which sidesteps chip import controls entirely — and Nvidia's own response has been to instruct cloud partners that at least 51% of services must go to non-Chinese customers, cutting off the very channel the rush is using.
- The counterparty is murky. Reuters identified Megaspeed, a Singapore-based cloud operator, as having helped Chinese customers reach controlled processors; Megaspeed denies illegally diverting chips, and Alibaba's unit Apex Enterprise Solutions financed a Malaysian affiliate, Speedmatrix, linked to it.
- The gate is shut at home. Beijing has allowed some H200s in but has withheld approval for large-scale imports, leaving roughly 600,000 chips parked in Hong Kong; the government also proposed its own curbs on AI models, agents, chip-design tools and clusters, turning the chip race into a two-way veto.
- The clock is short. NVIDIA reports fiscal third-quarter results on Nov. 17, five days after the scramble became public.
| Date | Event | Why it matters |
|---|---|---|
| Jan. 13, 2026 | Commerce moves H200 and MI325X licences to case-by-case review | The first crack in the ban |
| Jan. 28, 2026 | China clears a first batch of H200 imports for ByteDance, Alibaba and Tencent | Demand was real and Beijing knew it |
| May 14, 2026 | US clears H200 sales to about 10 Chinese firms; Lenovo and Foxconn approved as distributors | Licences exist, deliveries do not |
| May 31, 2026 | Commerce closes the loophole for shipments to Chinese firms abroad | Overseas access becomes the next target |
| Nov. 5, 2026 | About 600,000 H200 chips sit in Hong Kong awaiting mainland import approval | Demand exists, the door is shut |
| Nov. 8–10, 2026 | Nvidia opens a probe of its CEO; independent directors hire a former acting US Attorney | Political risk repriced into policy risk |
| Nov. 12, 2026 | Reuters: Chinese AI firms race to reserve overseas Nvidia capacity | The hoarding becomes public |
| Nov. 17, 2026 | NVIDIA reports fiscal Q3 2027 | First print after the scramble |
The order book
A $35 billion order Nvidia says it cannot fill
The clearest evidence of the hoarding is a purchase order, not a lease. ByteDance, Alibaba and Tencent have ordered more than 3 million NVIDIA RTX PRO 5500 workstation cards at about $11,400 each, and total orders from Chinese customers have passed 6 million units, a combined sticker value above $35 billion. These are export-compliant parts, aimed mainly at inference rather than training, which needs less compute. Nvidia has told customers it cannot guarantee filling the orders, with partners receiving as little as 5% to 10% of what they asked for.
China's RTX PRO 5500 order book against what Nvidia can build
Units in millions. Chinese customers have ordered more than 6 million cards; Nvidia has indicated it can produce no more than 3 million globally in 2027.
Unit: million units
Total China orders
All Chinese customers, reported Oct. 20, 2026
6
ByteDance, Alibaba and Tencent
More than 3 million cards, worth over $35 billion
3
Nvidia 2027 global production
Upper limit indicated to customers
3
Supply, not demand, is the binding constraint, and it gets worse in 2027. Early samples arrive in December and commercial deliveries begin in the first calendar quarter of 2027, landing after the Nov. 17 print at a rationed fill rate. From 2027 Nvidia expects to devote much of its foundry capacity at TSMC to the Rubin platform rather than to the workstation card, and it has discontinued the RTX 6000D whose price in China has doubled.
Nvidia's book
The China line is too small to carry the story, and it has already been written down
The reason the scramble cannot produce a great quarter sits in NVIDIA's own filings. Hopper shipments to China were less than 1% of data-center revenue in the quarter ended July 26, 2026, and the company took a $0.4 billion charge in the first half of fiscal 2027 for H200 excess inventory and purchase obligations caused by diminished demand. Washington cleared the licences; Beijing never let the chips move.
| Disclosure | Figure | Period |
|---|---|---|
| Data-center revenue | $89.0 billion | Q2 FY2027, quarter ended Jul. 26, 2026 |
| Graphics revenue | $7.9 billion | Q2 FY2027 |
| Hopper shipments to China | Less than 1% of data-center revenue | Q2 FY2027 |
| H200 excess inventory and purchase-obligation charge | $0.4 billion | First half of FY2027 |
| Inventory and excess purchase-obligation provision | $985 million | Q2 FY2027 |
| Provision release on previously reserved inventory | $177 million | Q2 FY2027 |
| Tariff on H200s imported into the US that Nvidia cannot pass on | 25% | Q2 FY2027 |
Restrictions have reduced NVIDIA's China AI data-centre chip share from a near monopoly to effectively nothing.
Scarcity rents
Three prices show where the value actually sits
The hoarding is visible in prices rather than in filings, because most of the transactions are rentals and diversions. The spread between the international list price and the Chinese market price is the rent Chinese buyers pay for access they cannot get through normal channels, and it is the number that makes a crackdown politically inevitable.
| Item | International price | Price in China | Premium | As of |
|---|---|---|---|---|
| Nvidia B300 rack | About $600,000 | About $2.8 million | 4.7x | Nov. 9, 2026 |
| H200 eight-card server | Not disclosed | More than ¥6 million (about $900,000) | About 3x vs. a year earlier | End of October 2026 |
| Blackwell-based server | Not disclosed | More than ¥20 million | Not disclosed | End of October 2026 |
| RTX PRO 5500 card | About $11,400 on order | Not disclosed | Not disclosed | Oct. 20, 2026 |
The 4.7x premium on a B300 rack does two jobs at once. It transfers value from Chinese AI labs to whoever controls access to controlled silicon, and it hands Washington the evidence that the existing rules are leaking. A scarcity premium that large argues for tightening the rules, not loosening them, which is exactly the dynamic now running in both directions.
Who gets paid
A hoard of already-built chips pays the memory cycle, not the chip vendor
| Layer | Company | Link to the scramble | Latest reported figure |
|---|---|---|---|
| Upstream memory | SK hynix | HBM stacks on every H200 and Blackwell board | Q3 2026 revenue ₩104.6T and operating profit ₩83.3T, up 328% and 632% year on year |
| Upstream memory | Micron | Second source of HBM for the same boards | Fiscal 2026 revenue $133.2B, with an 86.8% gross margin in the September quarter |
| Upstream foundry | TSMC | Builds the chips and reallocates capacity away from the China card | Q3 2026 revenue NT$1.49T, up 51% year on year |
| Downstream build | Super Micro Computer | Builds the servers that reach China through diversions | Fiscal Q1 2027 revenue $19.9B, up 297.5%, on a 12.7% gross margin |
| Downstream cloud | Oracle | Leases 100,000 GPUs in Southeast Asia to a Chinese buyer | Lease disclosed Nov. 9, 2026 |
| Downstream buyer | Tencent | Leased capacity abroad and joined the card orders | More than 3 million RTX PRO 5500 cards, about $11,400 each |
| Downstream buyer | Alibaba | Same card orders, plus financing into Malaysian capacity | More than 3 million cards; Apex Enterprise Solutions funded Speedmatrix |
| Private buyer | ByteDance | Largest single buyer of the compliant cards | More than 3 million cards of a China order book above 6 million |
This is the part of the story that gets inverted. Because the scramble is mostly about securing compute that already exists — roughly 600,000 H200s parked in Hong Kong, 100,000 GPUs leased through Oracle in Southeast Asia, Blackwell clusters of more than 8,000 GPUs being considered in Europe by Moonshot AI — it re-routes built silicon instead of adding demand for new silicon. Memory suppliers are paid at the build, not by the buyer, so the China rental wave barely touches their order books.
The exception is the compliant card channel, and NVIDIA is explicitly prioritising it over the China order book by moving foundry capacity to Rubin. Its own explanation is revealing: Chinese producers of workstation chips and cards have seen unprecedented growth since the curbs were first imposed in 2022. China still buys mix, not volume, and the part of the mix that survives is the part Nvidia can ration.
The substitution trade
The only reliable winner of a panic is the domestic supplier
Export controls do two things at once: they starve a buyer and they hand that buyer an excuse. Chinese accelerator and memory programmes are now being funded on the assumption that foreign silicon can be cut off without notice. Cambricon Technologies reported third-quarter 2026 revenue of RMB 3.53 billion, up 104% year on year, with a 38% operating margin, and it is one of two Chinese chip designers testing CXMT's homegrown HBM3E.
- CXMT began small-scale HBM3E production and aims to ramp next year, with overall yields near 25% on the previous HBM3 generation — yields explain why it has to start small.
- Alibaba's T-Head and Cambricon Technologies are both testing CXMT's HBM3E for their accelerators, with commercial products possible as early as 2027.
- The demand pull is one-sided: export restrictions make it hard for Chinese chipmakers to buy HBM, so domestic HBM has a captive market even at lower performance.
- The incumbents keep the near term: SK hynix posted Q3 2026 operating profit of ₩83.3 trillion, up 632% year on year, and the HBM stack it sells is unaffected by who buys the accelerator.
- The complaint is the same either way: NVIDIA shipped Hopper into China at less than 1% of its data-center revenue, so the China channel is a rounding error for the memory complex and everything for Beijing.
Horizons
What moves first, and what the next two years decide
- Nov. 17: NVIDIA reports fiscal Q3 2027. Watch the China language in the filing and whether a mainland import approval for the 600,000-chip Hong Kong stockpile lets the company release inventory reserves.
- December: the first RTX PRO 5500 samples reach the three buyers, the first hard evidence of the 5%–10% fill rate.
- Senate timing on the Remote Access Security Act, which would extend export controls from shipping chips to running someone else's chips.
- Beijing's import decision on the Hong Kong stockpile, and whether mainland prices for H200 and Blackwell servers hold above three times last year's level.
- 2027 foundry allocation: Rubin takes most of NVIDIA's capacity at TSMC, capping how much compliant card capacity exists for the buyers now queuing for it.
- 2027 CXMT ramp: HBM3E moving from trials at T-Head and Cambricon Technologies into commercial boards would reset Chinese accelerator cost curves.
- 2027 demand air pocket: once the leases and the Hong Kong chips are in hand, Chinese labs' incremental compute purchases fall, and overseas AI server orders fall with them.
- Policy: a closed remote-access channel would leave the 100,000 GPUs Tencent has leased in Southeast Asia, and similar arrangements, exposed to enforcement.
Synthesis
Read the scramble as a policy signal, not a demand signal
The most useful thing about this week's story is what it reveals about where the constraint sits. Chip supply is not scarce — NVIDIA booked $96.2 billion of revenue in a quarter and produced more RTX PRO 5500 cards than Chinese buyers could obtain. Scarcity sits in permission: licences, US inspections, tariffs, import approvals and the right to run a GPU someone else owns. Treat the China channel as a policy option, not a demand signal, and treat 2027's air pocket as a Chinese capex event rather than a semiconductor event.
The investable consequences follow the same split. The businesses that gain most are those whose revenue does not depend on export-control geography — Micron, SK hynix and TSMC on the build, and Cambricon Technologies on the domestic substitution. The businesses that lose are those whose growth was built on the grey channel, Super Micro Computer most of all, plus the cloud operators whose leases become unusable if Washington moves to remote-access controls.
- If Beijing approves the Hong Kong stockpile in a single shipment, NVIDIA books a one-time catch-up and a reserve release, which would make the small-segment argument right in the wrong direction.
- If the Senate kills the Remote Access Security Act, the overseas rental channel stays open and Chinese labs' 2027 compute spend normalises instead of air-pocketing.
- If CXMT's HBM3E reaches commercial yields in 2027, domestic substitution graduates from policy-driven to cost-driven and the memory premium compresses.
Where the scramble actually lands on listed stocks
- The China order book lands in a $7.9 billion Graphics segment, not the $89.0 billion data-center line.
- Hopper shipments to China ran under 1% of data-center revenue in Q2 FY2027, so no China order can swing a quarter.
- Near term (Nov. 17 print): a mainland import approval for the 600,000-chip Hong Kong stockpile would release inventory reserves.
- Over one to three years, Rubin takes most of the foundry capacity and caps the compliant card fill rate at 5%–10% of orders.
- Fiscal 2026 closed with $133.2 billion of revenue and an 86.8% gross margin in the September quarter.
- HBM is paid for at build time, so a China rental scramble that re-routes existing GPUs leaves the order book unchanged.
- Over one to three years, CXMT's HBM3E ramp into Chinese accelerators is the one real threat to the premium.
- Near term: China's pull-forward is immaterial to HBM demand because its Hopper share of Nvidia's data-center line is under 1%.
- Q3 2026 revenue of NT$1.49 trillion was up 51% year on year as AI accelerators absorbed advanced-node capacity.
- Moving Nvidia's 2027 capacity to Rubin raises the value of each wafer and shrinks the low-end card business China actually orders.
- Near term, the RTX PRO 5500 queue and Rubin priority cap how much China-linked card capacity can be built in 2027.
- Over one to three years, a closed remote-access channel removes demand for low-end accelerator dies, which are the smallest part of the mix.
- Prosecutors say roughly $2.5 billion of its servers bought in 2024–2025 were diverted toward China.
- Fiscal Q1 2027 revenue rose 297.5% to $19.9 billion, but a 12.7% gross margin shows the volume is not converting into profit.
- Over one to three years, a closed remote-access channel removes the grey channel that carried China demand.
- Near term: the company's exposure to diversion cases is now documented in a public prosecution, which is a repricing event for its China-linked order book.
- Q3 2026 revenue of ₩104.6 trillion and operating profit of ₩83.3 trillion were records, whatever Beijing decides.
- China barely touches the HBM order book: Hopper shipments to China ran under 1% of Nvidia's data-center revenue.
- Over one to three years, CXMT's HBM3E entering T-Head and Cambricon boards puts the first crack in the HBM premium.
- Near term, memory demand is set by the AI build cycle, and Chinese export rules do not slow that build.
- Leasing 100,000 GPUs in Southeast Asia locks scarce training capacity before the Remote Access Security Act is voted on.
- Near term, access costs roughly 4.7 times the international rack price, with a B300 rack reaching $2.8 million in China.
- Over one to three years, those leases become stranded assets if remote access is regulated.
- Its own order for more than 3 million RTX PRO 5500 cards is the cheapest way to hold capacity without a chip shipment.
- Alibaba, ByteDance and Tencent account for more than 3 million of the 6 million-plus cards Chinese customers ordered.
- Its unit Apex Enterprise Solutions financed Malaysian GPU capacity linked to Megaspeed — the structure Washington is legislating against.
- Over one to three years, T-Head's domestic accelerators make the overseas lease optional rather than critical.
- Near term: the July 2025 Alibaba AI chip unit listing plan in Hong Kong gives it a domestic route to the same demand.
- Q3 2026 revenue of RMB 3.53 billion was up 104% year on year, with a 38% operating margin and RMB 1.40 billion of net profit.
- It is one of two Chinese chip designers testing CXMT's HBM3E, which removes the last foreign dependency from its board.
- Near term, every month of export scare accelerates design-in of domestic accelerators.
- Over one to three years, commercial CXMT-backed parts in 2027 would let it price against foreign silicon instead of against shortages.
