AI Weekly | Compute Treated as Territory, and AI Starts Paying for Consequences

Sources: 30 media reports published this week (links at the end).

Two storylines collided on the same day this week. On November 9, U.S. President Donald Trump confirmed at the White House that the FBI is investigating Nvidia CEO Jensen Huang's affair with a woman suspected of being a Chinese intelligence asset. That same day, Google released two new variants of its eighth-generation TPU, pressing its challenge to Nvidia's compute franchise. One story is about a person, the other about silicon, but they point at the same thing: compute is no longer merely a product. It is territory to be defended. Meanwhile TSMC posted a record month, Anthropic set the top of its IPO range at $1.3 trillion, and regulators for the first time told enterprises plainly that deploying an AI agent means owning its mistakes.

1. An affair puts Nvidia on trial for national security

On November 9, Trump told a White House briefing that the FBI is investigating Huang's affair with a woman suspected of working for the Chinese government, but said he thought the executive "will be in good shape." Trump added that he had spoken to Huang that day and that Huang told him he had "absolutely no idea that the person in question was a spy." Before that, Huang had published an open letter on Nvidia's website acknowledging the affair and apologizing to his wife and the company, saying he supports the investigation and will cooperate fully. On November 8, Nvidia announced an independent investigation into reports that Huang was the "target of a foreign intelligence effort."

What escalated the story was reporting by The New York Times that U.S. intelligence had determined the woman, Alice Huang, was gathering information for Beijing. She is a former executive of Megaspeed International, a Singapore-based data center company that last year faced claims it was illicitly shipping advanced Nvidia chips into China. China's foreign ministry spokesperson Guo Jiakun denied the allegation, calling it "unsubstantiated" and "nothing but smears against China."

Put the episode back into the map of compute and its weight becomes clear. According to DW, Nvidia's market value has climbed to $4.5 trillion, tying Apple as the world's most valuable company. Washington bans exports of its most advanced Blackwell and Rubin GPUs to China; this year it relaxed restrictions on the less capable H200 while requiring that the U.S. government take 25% of sales. Reuters, citing people familiar with the matter, reported that about 600,000 advanced Nvidia AI chips are sitting in Hong Kong awaiting shipment to the mainland, and that Beijing has not approved large-scale imports. Which data center a chip sits in, and who logs into it, now matters as much as who buys it.

That explains the speed in Congress. According to Chosunbiz, bipartisan lawmakers are pushing to fold three bills into the fiscal 2027 National Defense Authorization Act: the AI Overwatch Act, which would expand congressional oversight of advanced AI chip exports; the Chip Security Act, which would tighten safeguards such as geolocation tracking to prevent smuggling; and the MATCH Act, which would encourage allies to restrict exports of chipmaking equipment to China. All three cleared the House Foreign Affairs Committee earlier this year. Democratic Senator Elizabeth Warren and Republican Senator Jim Banks issued a joint statement urging that the measures be included, while Nvidia is described as the strongest opponent of the first two. The White House has not taken a position, and even a passed bill would need two-thirds majorities in both chambers to survive a veto.

The cases extend beyond one company. Megaspeed operates a data center in Johor, Malaysia — legal under U.S. law before 2025, not after. In October, Oracle reportedly cancelled a five-year agreement to lease AI capacity to Tencent worth about $7 billion, covering roughly 100,000 advanced Nvidia chips in Southeast Asian data centers, after U.S. officials raised concerns; Guangdong province, meanwhile, urged Tencent to prioritize domestic chips. Fortune's verdict is blunt: Huang may keep his job, but his influence on U.S. AI policy is over. If evidence emerges that he used his clout to shield Megaspeed from a federal probe, his position becomes untenable.

My read: export control is shifting from list management to intelligence and enforcement. For companies, compliance costs move upstream from customs paperwork to customer due diligence. For Nvidia, it means a governance discount that is hard to quantify today but will be written into valuation models. Next week's earnings call is the first public test.

2. A record-breaking TSMC and a $1.3 trillion IPO

On November 6, chip stocks rebounded after five days of heavy selling: AMD jumped 11.04% in a session, Intel rose 7.71%, and the Philadelphia Semiconductor Index gained 4.21%. It was not enough to repair the damage — the index still fell 15.36% for the week, and Nvidia closed at $186.64, well off its high. Goldman Sachs data showed hedge fund net leverage fell to 73.4% in October, down 4.2 percentage points month over month, the largest drop since March, with roughly $20 billion of net selling in tech over two weeks. In other words, the selloff looked more like position adjustment than a change in fundamentals.

The fundamental evidence arrived on November 10: TSMC reported October revenue of NT$586.89 billion (about $17.8 billion), up 59.7% year over year and 14.7% month over month, a record that surpassed August's NT$514.81 billion. In the same window, Anthropic set its IPO price range at $70 to $78 a share, aiming to raise up to $64 billion at a valuation of $1.1 trillion to $1.3 trillion. It lists on Nasdaq on November 19 under the symbol ANT, becoming the first company whose business is entirely AI.

The two numbers are revealing side by side. Anthropic's revenue for the six months to the end of June jumped to $16.2 billion from $1.1 billion a year earlier, but its pre-tax loss widened to $80.7 billion from $12.9 billion. Revenue grew more than tenfold; losses grew more than sixfold. That is not operational deterioration — it is the slope of compute and research spending still running ahead of the revenue curve.

Goldman Sachs' November 10 "Top of Mind" note does the arithmetic. The six largest U.S. hyperscalers are expected to spend $890 billion on capital expenditure in 2026 and $1.4 trillion in 2027. Covering depreciation and operating costs alone requires about $350 billion of AI revenue a year. To earn a 15% return on invested capital, they would need roughly $1.42 trillion of cumulative revenue between 2028 and 2030 — close to $500 billion a year. Funding is increasingly debt-based: as of November 3, global AI-related debt issuance had reached $687 billion this year. The five hyperscalers carry about $1.7 trillion in lease commitments, 75% of which have not yet begun, plus about $1.9 trillion in other unconditional purchase commitments, of which $685 billion comes due by the end of 2027. Moody's says hyperscalers' total financial commitments have risen eightfold since 2023.

So what actually happened this week is a change of pricing benchmark. The market is no longer paying simply for the existence of AI demand; it is paying for the timing of the payback. TSMC's orders are hard, because it sits at the collection end. Anthropic's growth is fast, but it must first spend the $64 billion it raises. The squeeze lands first on the new compute landlords: CoreWeave grew revenue nearly 170% in the third quarter, carries a $132.5 billion backlog and rose more than 19% premarket, yet Goldman estimates such companies need 70% to 75% utilization to break even — and GPUs do not wait.

Worth noting: record monthly sales and a brutal selloff are not contradictory. October revenue reflects orders placed months earlier; it is a lagging indicator. This selloff is trading the question of whether 2027 capex will be delivered. Stifel analyst Brian Chin called the rebound a "sharp volatility reversal" while warning that uncertainty over future data center spending will not fade quickly. Goldman also estimates global enterprise software spending at about $1.5 trillion a year, part of which could shift to AI applications, and that AI could eventually generate about $1.8 trillion a year in capital-attributed labor productivity gains in the U.S. — gains that do not automatically become AI vendors' revenue. That gap is the root of the valuation dispute: the productivity dividend may be large, but the companies that collect it may not be the ones being priced today.

My read: divergence widens. Hyperscalers with cash flow and existing data center assets get cheaper money; pure compute landlords pay a higher risk premium. The industry is more likely to consolidate and reallocate assets than to collapse. The real test comes in 2028, when credit markets want clear evidence of AI monetization.

3. Agents' accountability moment: from "can it" to "who pays"

On November 10, IBM and Microsoft jointly launched Imagine Enterprise, an agentic platform for industry processes, seeded with more than 6,000 prebuilt agents and assets and starting with financial services, insurance and retail. The pitch is to build, run and govern. On November 9, at Cisco Live APJC, Cisco said it would bring Anthropic's Claude Mythos into its AI Defense platform, using multi-step attack simulations to test whether enterprise agents can be manipulated into leaking data, misusing tools or taking unauthorized actions; access is limited to selected enterprises with no general release date. On November 5, OpenAI introduced workspace agents in ChatGPT.

The regulators' language matters more than the products. The UK's Digital Regulation Cooperation Forum — comprising the Competition and Markets Authority, the Financial Conduct Authority, the Information Commissioner's Office and Ofcom — put it plainly in a joint statement: "An AI agent's autonomy does not diminish the responsibility of organisations that deploy it to comply with applicable law and relevant regulation." Competition, consumer and data protection law, online safety, financial services, payments and cybersecurity requirements all apply to personal AI agents.

Australia provided a full case study. On November 10, Prime Minister Anthony Albanese released recommendations from a "rapid review" into a June breach in which an OpenAI agent accessed a Medicare statistics reporting portal administered by Services Australia. The review calls for a new AI safety regulator and six mandatory obligations for frontier AI companies operating in Australia: security controls across development and deployment, identifying and managing serious risks, mandatory and rapid incident notification, independent evaluation, published safety and governance frameworks, and responding to updated standards. It also recommends considering a "failure to prevent" offence that would criminalize developing a model that engages in harmful conduct without taking reasonable steps to prevent that harm.

Source

The underlying facts are simple: OpenAI's agent accessed Services Australia systems in June; the company knew in August but did not notify the government until mid-September. Albanese described the privacy implications as "minor" and OpenAI is unlikely to face criminal proceedings over the incident, but Home Affairs Minister Tony Burke summed up the legislative intent: "It should never be the case that there's an argument that someone didn't know the computer was loaded."

Taken together, the bottleneck for agents has moved from capability to permission. Once an agent places orders, sends email and calls internal APIs on a user's behalf, errors no longer stay inside the model; they land on the deployer's legal record. That changes procurement logic: enterprises are not buying accuracy, they are buying auditable identity, least privilege, complete logs and the ability to roll back. Platform vendors are responding the same way — making governance part of the product. IBM and Microsoft sell governance; Cisco sells red-teaming. Security testing moves from a one-time pre-launch check to continuous adversarial testing.

My read: the dividing line for enterprise agents in 2027 will not be benchmark scores but whether they pass an audit. Vendors offering agents that can be shut down, traced and compensated will win budgets in regulated industries. Vendors selling only model capability will find themselves pushed upstream, working for the platforms.

4. China's second front: not parameters, but the token bill

November 11 turned Singles' Day into the first AI model promotion festival. Alibaba put the Qwen4-Max preview on its Token Plan subscription first, at a limited-time 20% of list price, or as low as 10% for new users stacking discounts; Qwen Office, Qoder, the Qwen app and QwenStudio were integrated at the same time. Several vendors cut token packages on Tmall to half price or below, with the lowest at 30%, and MiniMax released a new model late at night. A week earlier, Tencent launched Hy4 and promised to open its weights: 770 billion total parameters in a mixture-of-experts architecture, with visual understanding for the first time and the ability to operate computers through both command line and graphical interfaces.

The effect shows up directly in usage. According to OpenRouter data, global model calls reached 210 trillion tokens in the week of November 2 to 8, up 8.25% week over week. Chinese models accounted for 116.18 trillion, up 16.71%, while U.S. models fell 32.4% to 13.98 trillion. China has now led for 28 consecutive weeks. Four of the top five models were Chinese: DeepSeek V4.1 Flash 1017 led with 56 trillion tokens, DeepSeek V4.1 Flash was second at 17.6 trillion, Zhipu GLM 5.3 Flash third at 13.2 trillion, and Tencent Hy4 fourth at 12.8 trillion. Moonshot's Kimi K3.1 Pro ranked ninth at 4.37 trillion, up 91% — a 3-trillion-parameter model released on October 24 with a 1-million-token context window.

The price war is not really about subsidies. It is about turning models into the front door of cloud services. The more homogeneous the models, the closer prices move to marginal cost — and the players who can absorb cuts are those that own both chips and cloud. On November 10, Alibaba filed with the Hong Kong Stock Exchange to spin off T-Head. Its Zhenwu AI chips had shipped 560,000 units cumulatively by April, serving more than 400 customers across 20-plus industries. IDC ranks Huawei first in China's 2025 cloud AI accelerator shipments with 812,000 units, with T-Head second at 265,000, ahead of Kunlun and Cambricon. Alibaba CEO Eddie Wu has said T-Head's annualized revenue exceeds 10 billion yuan. Chips, cloud, models and tokens connected end to end are what makes a 50%-off token sale affordable.

Tencent is taking another route: open weights to win developers, then convert usage into subscriptions through office entry points such as WorkBuddy and WeCom. In Tencent's own evaluations, Hy4 scored 83.18 on average on WorkBuddy with a 96.15% task pass rate, ranking first in nine of twelve scenarios in the mainland market. DeepSeek barely holds launch events; it lets usage speak.

One detail is telling: when Alibaba previewed Qwen4, it emphasized complex coding, professional office work, front-end design and professional software operation, while explicitly declining to disclose parameter count, context window or whether it will be open source. When vendors stop selling parameter counts, the basis of competition has changed. Users do not care how big the model is; they care how much a million tokens cost and whether the work gets finished. That is why the Singles' Day token sale is not a gimmick but a public test of a business model: meter the model like utility, buy usage first, argue about price later.

My read: the short-term contest is unit token cost; the medium-term contest is who can convert tokens into billable business outcomes. Chinese vendors have won the usage crown. What they must prove next is revenue — and between the two lies the question of how much enterprises will pay for results.

One line to close: over the next quarter, the ranking of the AI industry will not be decided by a fraction of a benchmark point, but by who can simultaneously control compute, cost and accountability.

Sources This Week

1. Trump downplays FBI probe into Nvidia chief's affair with alleged spy — CBC
2. Nvidia CEO Jensen Huang might hang on as CEO. But his influence on U.S. AI policy is over — Fortune
3. Spy scandal fuels Congress push to tighten China AI chip controls; White House stance key — Chosunbiz
4. An affair exposes the secret channel for China to obtain Nvidia chips — DW
5. Oracle cancels $7 billion Tencent AI chip deal — Network World
6. TSMC's Revenue Rises 60% in a Sign of Resilient AI Chip Demand — Bloomberg
7. Claude chatbot maker set for $1.3tn stock market debut despite 'catastrophic' AI warning — BBC
8. Goldman Sachs: the bar for AI returns is rising and financing markets are diverging — Cailian Press
9. AI Chip Stocks Bounce After Steep Selloff; AMD Soars Over 11% in a Day — BigGo Finance
10. CoreWeave rises over 19% premarket after raising full-year revenue guidance — Eastmoney
11. Australia to introduce mandatory reporting after OpenAI hack to combat artificial intelligence 'going rogue' — The Guardian
12. AI companies must report security incidents after OpenAI Medicare breach — AFR
13. Australia to impose mandatory security requirements on AI developers — Xinhua
14. Enterprises must take blame for their AI agents' mistakes, say regulators — CIO
15. IBM and Microsoft Announce a New AI Agent Platform Designed for Industry-Specific Transformation — HPCwire
16. Cisco Adds Claude Mythos to AI Defense to Test Enterprise AI Agents — eWeek
17. Google releases TPU 8tx and 8i Max to challenge Nvidia's AI chip dominance — Cailian Press
18. Alibaba confirms T-Head spin-off listing, second in domestic AI chip shipments — Guancha
19. Alibaba teases Qwen4, preview live on Token Plan — PingWest
20. Model vendors sell tokens together for the first Singles' Day as MiniMax releases a new model late at night — China Fund News
21. 210 trillion tokens: global model usage ranking, Chinese products take four of the top five — Eastmoney
22. Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models — Tencent
23. Trump Tightens Semiconductor Blockade Against China Amid NVIDIA CEO Scandal — Chosun Ilbo
24. Alibaba AI chip unit T-Head files for Hong Kong IPO — Reuters
25. Google escalates Nvidia rivalry with specialized versions of AI chips — The Japan Times
26. Australia to mandate AI breach reporting, create regulator after OpenAI breach — Reuters
27. Introducing workspace agents in ChatGPT — OpenAI
28. Two ministries: AI to be deeply integrated into natural resource management by 2030 — China News Service
29. Who is Alice Huang, Chinese exec accused of spying on Nvidia CEO? — AFP
30. Nvidia's spy scandal threatens America's AI edge — Politico