Edge AI Daily Morning Brief: Nvidia Under Investigation, Anthropic's $895.4 Billion Compute Commitment and…

Silicon Valley Frontier:

1. On the day Nvidia was investigated, Google launched two TPUs.

1. On November 9, 2026, Nvidia closed at $178.67, down 4.27% for the day; on the same day, Alphabet (GOOG) rose 0.73% to $286.01. Google that day launched its eighth-generation TPU products, TPU 8tx (focused on security and confidential inference) and TPU 8i Max (targeting ultra-large context windows, with approximately 768MB SRAM, 17.2TB/s bandwidth, and 384GB HBM), marking four TPU systems released within a year and aimed directly at the gaps in Nvidia's moat in confidential computing and large memory bandwidth.

2. Nvidia's board launched an independent investigation through a committee of independent directors and hired outside counsel, concerning allegations that Jensen Huang became a target of foreign intelligence; in a letter to employees on November 8, Huang admitted to infidelity toward his wife and apologized, pledged to cooperate with the review, and withdrew from multiple events including a CEO forum hosted by Anthropic. For Nvidia, whose core asset is its founder's personal charisma, this governance shock may be harder to repair than quarterly performance fluctuations.

3. Nvidia's data center revenue for fiscal 2026 was $193.7 billion, accounting for about 90% of total revenue, and its AI accelerator share remains about 80%; according to Silicon Analysts forecasts, as AMD and in-house chips scale up, its share may fall to about 75% by 2026. The biggest structural threat comes from hyperscalers' in-house chips (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA), and Broadcom's AI ASIC business revenue already exceeded $20 billion in fiscal 2025. However, Google TPU is still constrained by the CUDA ecosystem (more than 5 million developers), the fact that the two new products will not ship until 2027, and its "produce and use in-house" model.

2. Anthropic's $895.4 billion computing power bet

1. Valuation and market: Anthropic plans to complete its Nasdaq IPO before Thanksgiving on November 26, with the valuation range for potential buyers lowered from the $2 trillion discussed in early October to $1.2 trillion to $1.3 trillion (some as low as $1.1 trillion); the trigger was the Nasdaq 100 index's single-day plunge of 6.4% on November 5, wiping out about $2.3 trillion in market value (affected by concerns over a slowdown in AI capital expenditure after the Democrats' landslide victory in the midterm elections).

2. Financial fundamentals: The revised S-1 shows revenue of $33.24 billion for the nine months ended September 30 (up about 1,200% year over year), with $16.96 billion in the third quarter alone; but net loss reached $104 billion (versus $25.1 billion a year earlier, mainly due to fair value changes in convertible bonds and derivative liabilities), and cumulative purchase and lease commitments are about $895.4 billion, the vast majority directed to computing power and infrastructure, constituting rigid cash obligations.

3. Product competition: On November 9, it released Claude Fable 5.2 and Mythos 5.2, focusing on vertical scenarios such as agentic coding (49.1%), software engineering (69.9%), and scientific research (74.6%); but its raw knowledge work Elo score of 1,785 still trails OpenAI GPT-6 Astra's 1,800, and the "leading in verticals, second in general-purpose" pattern compresses the margin for error in converting its computing power investment into cash flow and valuation premium.

3. ChatGPT Enterprise lands on AWS Marketplace, weakening Microsoft's distribution advantage

1. Core event and channel logic: On November 9, 2026, OpenAI put ChatGPT Enterprise on AWS Marketplace. The model and enterprise-tier levels are unchanged; what has really changed is the payment channel—enterprises can pay directly using their existing AWS bills and count eligible spending toward AWS multi-year upfront consumption commitments, compressing the procurement cycle from months to weeks.

2. Market data and trends: According to Omdia research, enterprise software sales through hyperscale cloud marketplaces (AWS, Azure, Google Cloud) will grow from $30 billion in 2024 to $163 billion in 2030, a five-year compound growth rate of 29.1%; among this, AI-related market spending will reach nearly $24.4 billion by 2030, with a compound growth rate of 37%, and the total value of existing upfront cloud commitments is about $470 billion.

3. Competitive landscape and impact: This move is seen as a structural blow to Microsoft's distribution advantage. Microsoft, with M365's high enterprise penetration and deep Copilot integration, has long held a quasi-monopoly as the "enterprise AI procurement gateway"; after OpenAI lands on the AWS marketplace, customers centered on AWS have another option that is equally smooth or even financially better, and its advantage shifts from "the only gateway" to "one of the gateways"; Google Cloud is expected to follow, and buyers' bargaining power rises.

4. Agents account for 80% of enterprise AI usage; the software market's judgment was wrong

1. Reversal of sentiment: A Menlo Ventures report on November 3, 2026, shows that enterprise generative AI deployment spending rose from $33 billion in 2025 to $176.8 billion in September 2026, a 5.4-fold increase in one year (versus 3.9-fold the previous year) and accelerating; 55% of this was net new budget rather than diverted software or labor budgets, directly refuting the zero-sum assumption of the SaaSpocalypse (AI devours SaaS) at the start of the year.

2. Structural inflection point: agents have become the main consumers of AI; on OpenRouter, agents generate 80% of tokens, about 7 times human usage, and traffic has grown 37-fold since surpassing humans in February 2026; the software pricing unit has shifted from "seats" to "tokens." Spending is concentrating at the top: Anthropic, OpenAI, and Google together account for 69% of enterprise AI deployment spending ($122.7 billion, up from 50% the previous year), and Anthropic's LLM API share rose from 40% to 55%.

3. Budgets and winners: 55% of new AI spending is net new, 22% comes from software budgets, of which 13 percentage points go to incumbents and only 9 percentage points (about $12.6 billion) flow to AI-native vendors; application spending is about $7.7 billion, with coding ($34 billion) and assistants/customer service ($25 billion) together accounting for 76%. ServiceNow accounts for 91% of AI spending in IT and Workday for 68% in HR; Salesforce rose about 23% in a single day after beating earnings expectations, and the SaaSpocalypse trade is unraveling.

V. Nvidia's Seoul AI Day full-scale push: multiple major partnerships with the South Korean government and giants such as Samsung, LG, and Hyundai

1. Nvidia held AI Day 2026 at COEX in Seoul from November 9 to 10, with about 2,000 developers, startups, and researchers expected to attend, focusing on three main tracks: agentic AI, physical AI, and AI infrastructure; it reached cooperation with South Korea's Ministry of Science and ICT (MSIT) and Samsung, LG, Hyundai Motor, SK Hynix, Kakao Mobility, and others, upgrading its positioning toward South Korea from a mere chip buyer to a sovereign AI showcase market.

2. Sovereign AI implementation: Nvidia and MSIT are expanding cooperation in seven areas including AI semiconductors, data centers, and talent development; the R200 from domestic NPU vendor Rebellions will connect to Nvidia's platform via NVLink Fusion (deployment not expected until 2028), and four South Korean foundation models (SK Telecom, LG, Upstage, Motif) use Dynamo optimized on B200 GPUs, claiming performance improvements of up to 2 times.

3. Deep industry penetration and hidden concerns: Samsung and SK Hynix use CUDA-X and PhysicsNeMo to accelerate semiconductor process simulation, LG uses Isaac GR00T to drive CLOiD robots, and Kakao Mobility advances L4 autonomous driving; but AMD held its Advancing AI World Tour in Seoul on November 4 (about 1,300 pre-registrations), South Korean companies pursue a multi-vendor strategy, and sovereign AI, while improving efficiency, in practice deepens dependence on Nvidia's ecosystem.

1. On November 5, 2026, U.S. stocks fell broadly, with the Nasdaq Composite down 5.52% to about 24,315 points, and the S&P 500 posting its largest single-day drop since April 2025; Nvidia plunged about 9.3% that day, even though its fundamentals remained strong with quarterly revenue of $96.2 billion (up 106% year over year) and a gross margin of 75%. The article argues that this is not weakening AI demand but a market "reckoning" over the return on AI investment, with the pricing focus shifting from "can it make money" to "when will it earn it back."

2. The core contradiction is that capital expenditure growth far outpaces revenue growth: the four major cloud providers Microsoft, Amazon, Alphabet, and Meta will spend a combined total of about $725 billion in capital expenditure in 2026 (up 77% year over year), and Goldman Sachs expects the four to invest a cumulative $5.3 trillion in AI infrastructure from fiscal 2025 to fiscal 2030; meanwhile AWS has annualized revenue of about $142 billion (up 24% year over year) and a backlog of $244 billion, and its monetization pace is seen as difficult to support such an aggressive expansion pace, with return on capital facing dilution.

3. Competitive fissures and market divergence: cloud providers are backing alternative suppliers (Meta shifting some orders to AMD, Alphabet developing its own TPUs, and Amazon Trainium gradually replacing some Nvidia GPUs), eroding Nvidia's pricing power over the long term; Nvidia gave guidance of about 70% revenue growth for fiscal 2028 (as recorded by JPMorgan), still above Wall Street's average of about 63% but failing to fully convince; the market pays only for "monetization"—Meta's stock fell after it raised its capital expenditure guidance, while Alphabet rose on strong cloud growth.

VII. AI developers face a 2027 rate shock after a $1 trillion debt binge

1. AI-related debt is expected to climb to $1.2 trillion by 2027 (expected to be about $750 billion by the end of 2026); in 2026, 33% of the capital expenditure of the five major hyperscalers was financed by bonds (26% in 2025), with investment-grade bond issuance of about $250 billion in 2026, rising to $400 billion to $440 billion in 2027.

2. The credit market has split into two tiers: Microsoft's credit default swap spreads are almost on par with Apple's, while Google and Amazon are only 10 to 15 basis points wider and can still refinance at manageable cost; CoreWeave's five-year CDS once spiked to 1,289 basis points, and the yield on its bonds maturing in 2032 topped 16%; about $18 billion of Oracle's Project Jupiter loan fell to 89 to 91 cents on the dollar, already showing signs of distress.

3. BofA expects AI-related credit issuance to grow 31% to $685 billion in 2026-2027 and to fall 23% to $531 billion in 2028, with the peak possibly close at hand; BofA expects the Fed to raise rates by 25 basis points each in December 2026 and March 2027, with a terminal rate range of 4.25% to 4.50%; Goldman Sachs estimates that a potential pause on data center construction could affect 45% of planned U.S. capacity.

VIII. $11 billion is just the down payment: OpenAI's valuation surges to $1.25 trillion

1. Financing event and valuation: On November 7, 2026, OpenAI completed an initial financing of about $11 billion at a valuation of $1.25 trillion, with the total for this round reaching up to $20 billion; this is up about 46.7% from the $852 billion valuation when it completed a $122 billion financing in March 2026, but below the original September target of $1.4 trillion, when it planned to raise at least $30 billion before scaling back due to a public-market selloff in the AI sector.

2. Revenue and losses diverge: OpenAI's second-quarter 2026 revenue was $6.7 billion, up 18% from $5.7 billion in the first quarter, with annualized revenue approaching $70 billion, but its operating loss widened from $9.3 billion in the first quarter to $12.3 billion in the second quarter, growing faster than revenue; the $1.25 trillion valuation corresponds to a price-to-sales ratio of about 18 times, and revenue growth struggles to keep pace with compute and data center capital expenditure.

3. Investor structure and risks: Investors in this round include Coatue, T. Rowe Price, SoftBank, Sequoia, Founders Fund, a16z and a UAE consortium, while Amazon, Nvidia and SoftBank are exchanging capital for cloud service and compute tie-ups; CEO Altman said there will be no IPO in 2026, private-market valuations lack public-trading validation, industry capital is concentrating in the leaders, and latecomers and public-market investors are under pressure.

IX. Nvidia investigation weighs on Asian markets: market roundup

1. Nvidia's board announced on November 8 that it had launched an independent investigation, with an independent director committee and outside legal counsel reviewing reports that founder and CEO Jensen Huang became a target of foreign intelligence operations; the board said it had no reason to believe national security or company secrets were compromised, did not suspend him and reiterated confidence in his leadership, while Huang publicly apologized the same day over an extramarital affair.

2. U.S. stocks were sold off across the board on Thursday, November 5, amid concerns over tighter AI regulation after the midterm elections and slowing data center construction: the Nasdaq fell 5.52% to 24,315.00 points, officially entering correction territory, the S&P 500 fell 4.36%, and the Dow fell 3.44%, with all three indexes posting their worst single day since April 2025; Nvidia fell 9.26% that day to $181.60, wiping out about $450 billion in market value, before rebounding 2.78% the next day (November 6).

3. Asian supply chains were indiscriminately sold off on November 6: SK Hynix fell 11.7%, Samsung Electronics nearly 10%, TSMC 8.5%, and the Taiwan Weighted Index more than 8%; because Asian suppliers already account for about 90% of Nvidia's production costs (about 65% a year earlier), TSMC's foundry and CoWoS packaging, SK Hynix's and Samsung's HBM memory, and Foxconn's and Quanta's server assembly all depend heavily on Nvidia demand.

X. Meta distills its flagship model into everyone's pocket: the AI race shifts from leaderboard chasing to getting things done

1. Product rollout: In an official announcement on November 6, Meta opened a specialized model distilled from its flagship Muse Wonder to all Muse and Meta AI users, raising its personal superintelligence index from 38% to 47%; earlier, on October 12, Meta had just released Muse Wonder and Wonder Full, whose training compute was about 10 times that of the previous-generation Muse Spark and whose context reached 1 million tokens.

2. Two major capabilities: Active Memory raised the success rate of tasks requiring correct use of key context from 76% to 90% and reduced the error rate of forgetting key information by 2.4 times; Active Focus improved Muse's ability to complete routine tasks when encountering unexpected questions by 67%, marking a shift in AI assistants from "who is smarter" to "who is more reliable."

3. Interaction and commercialization: Zuckerberg announced on Threads the same day (November 6) that Muse would launch real-time voice conversation and avatars, with a latency of about 870 milliseconds from finishing speech to the first audio-video response; Meta differentiates itself from OpenAI's paywall and Anthropic's enterprise focus with a "distill the flagship down to the largest user base" approach, and plans to monetize by taking commissions on shopping transactions users complete through AI agents.

Open-source trend:

XII. Four major South Korean open-source models enter the factory together: Nvidia lays deployment plumbing into the heartland of sovereign AI

1. Event: On November 8, 2026, Nvidia published Dynamo-optimized deployment recipes on its official blog for open-source models from four South Korean organizations, covering SK Telecom's A.X K2 (688 billion parameters), LG AI Research's K-EXAONE 2.0 (750 billion parameters), Upstage Solar Open 2, and Motif Technologies' Motif-3 (314 billion parameters). The recipes are optimized for B200 GPUs, and Nvidia says they work out of the box and deliver performance gains of up to 2x.

2. Analysis (shift of control point): The article notes that for open-source models, "open weights do not equal usable." Multi-node deployment involves high engineering barriers in parallelism, routing, and KV-cache management; Dynamo 1.0 previously achieved up to a 7x increase in request serving throughput on the Blackwell architecture in SemiAnalysis's InferenceX benchmark. Because Nvidia defines and preconfigures the deployment recipes, the control point shifts from the one-time training stage to continuous inference infrastructure, creating a dependency in which "model sovereignty does not equal ecosystem sovereignty": training rights can be self-held, but deployment rights may rest with others.

3. Trend (change in the logic of model competition): The myth of parameter scale is fading—the 750-billion-parameter K-EXAONE 2.0 scored only 31 on the AAII benchmark, ranking last among the four, while the 314-billion-parameter Motif-3 ranked 10th globally with 47 points, making it the highest-ranked model from outside the US and China; all four models have passed the second round of screening under the South Korean government's "National AI Model Project" and have been recognized by Epoch AI as notable AI models. The article argues that there is no clear winner at the infrastructure layer, and startups can quickly launch using Nvidia's framework, but differentiation will converge on the models themselves.

XIII. Apple open-sources HALO, tactile sensing becomes the focus of A-share mapping

1. Event and validation: On November 4 local time, Apple's research team published the paper "Perceptive Humanoid Motion from Simulation" on arXiv, open-sourcing its self-developed humanoid robot autonomous control system HALO (including data generation, model training, and deployment code), deployed on Unitree's G1 robot. In 6 real-world environments that had not been pre-scanned or mapped, it completed 360 single-task tests with an average success rate of 91%; however, in a single scenario, the average success rate for 3 types of combined tasks was only 70%, and the most complex "bend down to pass an obstacle—walk toward the target—push the box" succeeded only 5 times out of 10 (50%). The paper identifies the shortcoming: the system relies mainly on visual perception and struggles to directly perceive contact conditions; when objects leave the field of view or are occluded, performance on tasks involving contact declines, and adding tactile perception is the future direction.

2. Market mapping logic: What HALO validates is the feasibility of "zero-shot transfer at the semantic layer" (a control system trained on simulation data can transfer to real environments that have not been mapped), which has more of a "narrative reinforcement" effect on the A-share humanoid robot sector rather than directly validating tactile solutions or bringing orders. A June 2025 research report by CICC estimates that the market space for humanoid robot force sensors is expected to reach 7.9 billion yuan by 2030, tactile sensors 500 million yuan, and the overall sensor market 11.9 billion yuan (based on shipments of 350,000 units); the penetration rate of six-axis force/torque sensors may exceed 30%, with 2 or 4 units installed per robot. On price, six-axis force sensors are expected to drop from about 20,000 yuan each in 2024 to 10,000 yuan, 8,000 yuan, and 5,000 yuan in 2025 to 2027; tactile sensors are expected to drop from about 400 yuan in 2024 to 200 yuan by 2030.

3. Industry chain and pace: Relevant A-share companies are mainly positioned in the midstream—Hanwei Electronics (300007) has established a product matrix of "tactile—balance—force control—olfaction," and its flexible electronic skin is being supplied in small batches to some robot body manufacturers; Keli Sensing (603662) has completed development of six-axis force/torque sensor series for wrists, ankles, and other areas and has sent samples to multiple customers; Fulaixin Materials' second-generation tactile sensor pilot line has been installed, and the product has been installed at some downstream customers; others include Audiowell and Riying Electronics (electronic skin is in sample validation). Cailian Press VIP data shows that in the dexterous hand industry chain, planetary roller screws and tactile sensors account for 43.32% and 25.27% of value, respectively. IDC data shows that global humanoid robot shipments in 2025 will be about 18,000 units, up about 508% year on year, with a small base and both elasticity and uncertainty coexisting. The trading logic in the article is summarized as follows: short term depends on narrative, medium term depends on sample-to-mass-production conversion, and long term depends on the speed at which the cost curve declines; the core variables are cost reduction and the mass-production pace of leading robot body manufacturers (such as Tesla Optimus and UBTECH WalkerX).