Tuesday, November 17, 2026

US Closes the Remote-Compute Loophole: BIS Rule's Hit to AI Chips and Cloud

NVDAMUTSMNBISHK 09903HK 00981
Why today:On 11/16 BIS issued an interim rule requiring licenses for advanced AI chip exports to data centers in Singapore, Malaysia, Thailand, Indonesia and Brazil, and for any data center that may serve China or other D:5 countries. As of the 11/16 close, the top five US heat-list names were all semis; Nasdaq fell 2.49%, Micron 7.62%, SanDisk 8.58%, Nvidia 4.12%, Nebius 7.7%. Iluvatar CoreX (09903) fell 11.13% in HK before the rule was announced.

This piece asks one question: who does the Commerce Department's Nov 16 interim "remote compute" rule actually constrain, and by how much? How much of the impact falls on chipmakers, neoclouds and China's domestic compute names? US prices are as of the Nov 16 (Mon) close ET; HK prices are as of the Nov 16 HK close. Hong Kong's reaction at the Nov 17 open had not happened when this was written (Nov 16, 17:33 ET).

1. One rule, and a selloff bigger than the rule

On Nov 16, the Commerce Department's Bureau of Industry and Security (BIS) issued an interim final rule with two requirements.

First, exports of advanced AI chips and servers to Singapore, Malaysia, Thailand, Indonesia and Brazil now need a license. Second, any data center anywhere needs a license if it provides more than 12 million TPP of compute to companies headquartered in China, Macau or other D:5 countries. TPP is the metric US export controls use to measure chip compute; per TechCrunch, 12 million TPP is roughly 750 Nvidia H100/H200s.

The same day, the Nasdaq fell 2.49%, Nvidia 4.12%, the TSMC ADR 6.34% and Nebius 7.55%. The top five names on Aya's US buzz board were all semiconductors.

The conclusion up front: the revenue this rule directly blocks is small. But one day before Nvidia's earnings, it put three things on the table at once: the "quality" of Nvidia's past revenue, the customer mix of neoclouds, and the pace of China's domestic compute substitution. The day's selloff went far beyond what the rule alone can explain.

2. Background: what is the "remote-compute loophole"?

Since 2022, US chip controls on China have governed where hardware goes: high-end GPUs can't be sold to China. But AI compute can be rented. A Chinese company doesn't need to import chips; it can rent a full GPU cluster in a Singapore or Malaysia data center and train models over the network. That is what the rule text calls the "cloud loophole."

The direct trigger was the Megaspeed case. Per reporting carried by Yahoo News, the Singapore company and its Malaysian affiliate Speedmatrix bought about $4.8B of Nvidia hardware over three years. US investigators allege they used shell companies to obtain about 240,000 chips banned for China. Megaspeed denies wrongdoing. The case also led to the controversy around Jensen Huang's resignation as Nvidia CEO. Per the same report, Treasury Secretary Bessent pressed Commerce over why Megaspeed had not been sanctioned, saying "Commerce has the pen."

The main terms are below (as described by TheNextWeb, TechCrunch, ai-tldr and others; the Federal Register text itself was not checked directly):

TermContent
Data-center licenseAny data center worldwide providing >12M TPP to companies headquartered in China, Macau or D:5 countries needs a license; indirect access through intermediaries also counts
Five-country export licenseAdvanced chip exports to Indonesia, Malaysia, Singapore, Thailand and Brazil need a license, on the grounds of "higher-than-normal risk of diversion"
ExemptionsBuyers headquartered in the US (or with a US parent) that meet reporting requirements; recommender systems for e-commerce and social media
TimelineEffective on Federal Register publication; 90-day window for existing data centers; 60-day comment period after publication
OtherFormally rescinds the never-enforced Biden-era AI Diffusion Rule; countries that tolerate evasion may face import restrictions or Entity List designations

D:5 is the US arms-embargo country group. TechCrunch describes it as 22 countries, including China, Russia, Iran and North Korea.

3. Question one: how big is Nvidia's direct exposure?

View: On the reported numbers, the revenue the rule could block is at most about 7% of Nvidia's last fiscal year, and very likely far less. The hard part to price is a different question: how much past revenue reached China through Southeast Asia.

Evidence. Nvidia's FY2026 10-K (year ended Jan 2026) splits revenue by customer headquarters: US $149.6B, Taiwan $42.3B, China (incl. HK) $19.7B, Other $4.3B, for a total of $215.9B. China's share is 9.11%. But Nvidia's Q3 guidance states it "is not assuming any Data Center compute revenue from China," so China revenue going to zero is largely priced in.

Nvidia FY2026 revenue by customer HQ: Southeast Asia hides in 'Other' and 'Taiwa

Note that Singapore and Malaysia are not broken out. Customers headquartered there sit inside "Other," which is only $4.3B.

Reasoning. Many US customers used to invoice centrally through Singapore. In its 2025 reporting, Nvidia said that over 99% of controlled data-center revenue billed to Singapore came from orders by US-based customers. That bucket falls squarely within the rule's US-HQ exemption.

The real exposure sits in two gray zones. One is the $4.3B "Other" bucket. The other is sales routed through Taiwanese ODMs (contract assemblers): Nvidia estimates 76% of data-center revenue from Taiwan-HQ customers goes to US and European end customers, which leaves about 24% going elsewhere.

Adding the two gives roughly $14.5B, or 6.7% of FY2026 revenue. That is a generous ceiling. It applies 24% to all Taiwan revenue rather than only data-center revenue, and it also counts Japan, Korea and the Middle East, which the rule doesn't touch.

For comparison, Nvidia shed about $162B of market value on Nov 16 alone, more than that entire revenue ceiling. "Lost sales" alone can't explain the selloff.

The uncertainty is about compliance. Per the FT (via Bloomberg, Nov 3), Nvidia had verbally told cloud partners worldwide to keep Chinese customers at no more than 49% on new contracts. Ming-Chi Kuo has argued that the share of Nvidia chips ending up with Chinese buyers "may be higher than official statements." If investigators conclude that some Southeast Asian customers were diversion channels, the question shifts from "how much less will be sold" to "how much past revenue shouldn't have counted." Nvidia doesn't disclose this, and this piece cannot quantify it.

4. Question two: compliance cost or order risk for neoclouds?

View: For neoclouds (companies that rent out GPU compute) such as Nebius and CoreWeave, the direct hit is customer screening, not chip access. This matters more for Nebius than for CoreWeave. But Nebius's recent plunge was mostly not about this rule.

Evidence 1: the rule covers every data center, including those in the US. The data-center clause ignores location and looks only at customer HQ and the 12M TPP threshold. At roughly 16,000 TPP per chip, 750 H100-class chips is a small cluster today. Any neocloud renting meaningful capacity to a China-HQ customer now needs a license first, and must check that customer's ownership and control.

Evidence 2: Nebius doesn't qualify for the US-HQ exemption. Per yfinance, Nebius is headquartered at Schiphol, Netherlands. In March 2026 it made Singapore its APAC regional headquarters and said it had "won our first customers" in Asia. So any future chip deployments in Singapore or the other four countries can't use the US-HQ exemption and will need their own licenses. Nebius doesn't disclose APAC revenue or capacity, so this exposure can't be sized. Its disclosed core capacity is in Finland (455MW across three sites), New Jersey (up to 300MW) and Pennsylvania (1.2GW planned). CoreWeave is a US company and can use the exemption route under the five-country clause.

Evidence 3: most of Nebius's drop came before the rule. From Nov 11 to Nov 16, Nebius fell 30.75% in total. Of that, Nov 12 alone saw a 16.84% drop after it guided 2027 capex at $40–45B (see Aya's Nov 15 Nvidia earnings preview). Only the Nov 16 move of −7.55% coincided with the rule. Per yfinance, short interest is about 19.8% of float and the stock is about 64% below its 52-week high. With positioning like that, any bad news gets amplified.

Since Nov 3: US compute names fell while HK domestic-chip names rose, then gave

Look at the direction: US compute names and HK domestic-chip names had already split apart two weeks before the rule. The rule didn't start this divergence; it is the latest event in it.

Reasoning. For neoclouds, order risk is concentrated in two customer types: China-HQ AI companies renting capacity through overseas nodes, and intermediaries with Chinese ownership. Nvidia's early-November 49% cap was already squeezing this business; the rule turns a commercial constraint into a legal one.

The FT report also carries a contrary signal: cloud operators say it is hard to find enough non-Chinese customers outside the US. That suggests Southeast Asian data centers could end up with idle capacity, not just a different set of customers.

Institutional research points to the same structural weakness. Stratechery (2026-11-04) notes that neoclouds constantly need new sites, permits and outside financing, so tighter licensing raises the bar for newer entrants. Epoch AI (2026-04-10) notes that capital already sunk into operating assets is less sensitive, while projects not yet built are the most exposed to policy and financing conditions. By that logic, planned Southeast Asian projects take the hardest hit, not capacity already running in the US or Europe.

5. Question three: the domestic substitution case, and Hong Kong's timing mismatch

View: The rule logically strengthens demand for Chinese domestic AI chips. But Iluvatar CoreX's sharp drop on Nov 16 came before the rule was published, so it isn't a reaction to the rule. The real test for Hong Kong is the Nov 17 open, which hadn't happened at the time of writing.

Evidence: the timeline. Iluvatar CoreX closed −11.13% at HK$68.30 on Nov 16, and SMIC closed −5.03% at HK$46.08. The HK close fell in the early hours of Nov 16 ET, and reports on the rule came out later that US day, so both moves came before the rule. The same day, HK AI-application names MiniMax and Zhipu fell 18% and 13%, and AI hardware broadly weakened. The better explanation is broad de-risking across the HK AI sector.

The run-up before that is more telling. From Nov 3 to Nov 12, while the Nvidia leadership controversy unfolded and US chip stocks kept falling, Iluvatar CoreX rose from HK$56.35 to HK$78.70, up 39.66%. On Nov 16 it was still 21.21% above its Nov 3 level. In other words, the "tighter US controls help domestic chips" trade was already partly priced in before the rule.

Fundamentals: per East Money, Iluvatar CoreX's 1H26 revenue was 946M (labeled HKD by East Money), up 191.6% YoY, with net income of 106M, a turn to profit. It still lost about 1.0B for full-year 2025, and its trailing P/E is negative. SMIC's 9M26 revenue was 58.60B and net income 7.44B (East Money basis), up 27.2% and 115.7% YoY, with a 25.5% gross margin and a trailing P/E of about 74x.

Reasoning: the bottleneck is supply, not demand. The rule cuts Chinese firms' ability to rent Nvidia compute offshore. Some of that demand comes back onshore, and part of it goes to domestic GPUs. TrendForce (2026-08-10) describes this mechanism: the less reliable foreign supply is, the more policy procurement and ecosystem investment tilt toward local vendors.

But ChinAI (2025-01-13) points to the other side: the same controls also squeeze domestic chips' upstream inputs, such as HBM (high-bandwidth memory) and advanced-node capacity. Demand may be pushed onshore, but whether capacity can absorb it is a separate question. For SMIC, the rule mainly adds visibility on advanced-node orders. For a designer like Iluvatar CoreX, the key is securing enough wafers and HBM.

One data point shows how urgent demand is. Per the FT (via Bloomberg), a B300 selling for about $600,000 internationally fetched more than $2.54M in China (the report doesn't say whether that is per chip or per system). A gap that size shows China's compute shortage is real. It also means the arbitrage incentive is large, which will make enforcement hard.

6. What the market is betting on

At the Nov 16 close of $156.14 and yfinance's FY2028 EPS consensus of $15.85 (as of Nov 15), Nvidia trades at about 9.85x forward earnings. From Nov 3 to Nov 16, the stock fell 24.46%.

At that multiple, the market is clearly not just pricing "a few billion dollars less sold into Southeast Asia." It is discounting three harder-to-quantify risks: execution during the leadership vacuum; possible revenue restatement and fines from the Megaspeed case; and whether controls expand from "who the customer is" to "what the compute is used for."

Is that pricing reasonable? On the rule alone, direct revenue exposure is capped and cushioned by the US-HQ exemption, so the rule by itself doesn't justify such a low multiple. But if the Megaspeed investigation finds more customers in violation, today's discount may not be excessive. Neither outcome can be ruled out yet.

Memory stocks show how messy attribution is. Micron (−7.62%) and SanDisk (−8.58%) both fell more than Nvidia, even though memory has the weakest link to a remote-compute rule. That suggests the day's selloff was mostly collective de-risking across the AI chain ahead of Nvidia's earnings, with the rule as one push among several. Aya's buzz data shows Nvidia's bull/bear split at 54%/46%, slightly more bullish than last issue, so retail sentiment was not panicked.

7. What to watch next

DateEventWhat to look for
Nov 16, 20:30 ET (HK open Nov 17)HK market openWhether Iluvatar CoreX and SMIC react to the rule as substitution winners; if they keep tracking US declines, HK is trading AI-chain risk appetite as a whole
Nov 17, after closeNvidia Q3 earningsWhether it quantifies Southeast Asia or five-country exposure; any new risk disclosures on the rule or Megaspeed; Q4 guidance vs the ~$124.9B consensus
Nov 18–19Tencent, Alibaba resultsCommentary on overseas cloud nodes and compute buying: can Chinese clouds keep training AI overseas under the rule?
From Federal Register publication60-day comment period, 90-day transitionIf published in mid-to-late November, comments would close around mid-January 2027 and the transition would end around mid-February (our estimate); watch for industry lobbying to loosen the threshold
Dec 10Commerce Secretary Lutnick testifies before the Senate Banking CommitteeWhether Megaspeed is added to the Entity List, and whether enforcement reaches past sales

Two other signals are worth watching. One is the response from Malaysia and Singapore; Joshua Kurlantzick of the Council on Foreign Relations said, "For Malaysia, the controls could be disastrous." The other is whether Nebius discloses its APAC customer mix on its next earnings call.

8. Conclusion

This rule closes the "rent compute" loophole, not the "sell chips" loophole. On the geographic breakdown in Nvidia's 10-K, revenue that could be affected is at most about 6.7% of FY2026 revenue, and actual exposure is smaller once the US-HQ exemption is taken out. The Nov 16 drop across US compute names mixed together Nvidia's upcoming earnings, the leadership controversy and separate pressure on memory stocks, so it can't all be pinned on the rule. Nebius carries higher compliance costs than CoreWeave, but its recent plunge came mainly from its capex guidance. The domestic-substitution case holds, but it is partly priced in, and the bottleneck is capacity, not demand. Iluvatar CoreX's Nov 16 drop came before the rule; Hong Kong's real reaction will show at the Nov 17 open.

Appendix: Data and Sources

Table A: Key financial and valuation data (with formulas)

MetricValueFormula/inputsSource & date
Nvidia FY26 total revenue$215.94B—10-K, FYE 2026-01-25
China (incl. HK) revenue$19.68B, 9.11% share19,677/215,938Same
"Other" revenue$4.30B, 1.99% share4,299/215,938Same
Taiwan revenue × 24%~$10.16B42,345×(1−76%), using all Taiwan revenue as a proxy (generous ceiling)Same; 76% is a data-center figure
Gray-zone revenue ceiling~$14.46B, 6.7% share4,299+10,163; 14,462/215,938Our estimate
Nvidia market value lost on Nov 16~$162B37,703/(1−4.12%)−37,703 ($100M units)Tencent Securities quote, 2026-11-16
Nvidia forward P/E9.85x156.14/15.85Price Nov 16; FY28 EPS from yfinance Nov 15
Nvidia Nov 3→Nov 16−24.46%156.14/206.69−1yfinance
Nebius Nov 11→Nov 16−30.75%107.13/154.69−1yfinance
Nebius market cap / short interest$31.38B / 19.8%—yfinance, 2026-11-16
Iluvatar CoreX Nov 3→Nov 12+39.66%78.70/56.35−1East Money K-line
Iluvatar CoreX Nov 3→Nov 16+21.21%68.30/56.35−1Same
Iluvatar CoreX 1H26Revenue 946M (+191.6%), net income 106M—East Money financials (labeled HKD)
SMIC 9M26Revenue 58.60B (+27.2%), net income 7.44B (+115.7%), GM 25.5%—East Money financials (labeled HKD)
Nvidia Q2 FY27Revenue $96.2B, data center $89.0B; Q3 guide $108B ±2%, excluding China—Nvidia press release, 2026-08-26

Table B: Event timeline

DateEventSource
2026-03-10Nebius announces APAC expansion, with regional HQ in SingaporeNebius press release
2026-11-03FT/Bloomberg: Nvidia verbally tells cloud partners to keep Chinese customers at ≤49%Dealroom citing FT/Bloomberg
2026-11-12Nebius reports Q3 and 2027 capex guidance; stock falls 16.84%Aya Nov 15 report
2026-11-16BIS issues remote-compute interim rule; US semis sell offTechCrunch, Yahoo News, TheNextWeb, Bloomberg
2026-11-17Nvidia Q3 earnings (after close ET)Company schedule
2026-12-10Lutnick testifies before the Senate Banking CommitteeYahoo News

Table C: Aya buzz data (Nov 16 close)

TickerBuzz rank / scoreChangeBull/Bear
NVDAUS #1 / 82−4.3%54% / 46%
MUUS #2 / 81−7.5%57% / 43%
TSMUS #5 / 66−6.2%65% / 35%
NBISUS #8 / 64−7.7%52% / 48%
09903HK #11 / 70−10.9%No data source
00981HK #12 / 68−5.0%No data source

Note: the buzz board's price changes are program-calculated and differ slightly from the quote snapshot cited in the text (e.g., NVDA −4.12%); the text uses the quote snapshot throughout.

Market views

  • Joshua Kurlantzick (via TheNextWeb): "For Malaysia, the controls could be disastrous."
  • Rule text (via TechCrunch): the restrictions will undermine the ability of "companies headquartered in countries of concern" to use advanced US semiconductors to train models.
  • Ming-Chi Kuo (@mingchikuo, cited in Aya's Nov 15 report): the share of Nvidia chips reaching Chinese buyers "may be higher than official statements."
  • Live X search was not authorized for this issue, so no KOL posts were added.

Sources

For informational and research discussion only; not investment advice.

<!-- AYA_PODCAST -->

🎙️ Podcast script (Chinese, ~20 min, 泓观阳略): read it here

AI-generated by Aya Research from public data and disclosures, for information and research purposes only. Not investment advice.