Edge AI Daily Morning Brief: Nvidia's Jensen Huang Resigns, Anthropic's $90.4 Billion IPO and Google…
Nvidia announced that Jensen Huang is immediately stepping down as President, CEO and board director, with CFO Colette Kress serving as interim CEO; on the day of the news, the stock fell about 14% in early trading and its market value dropped below $4 trillion. Anthropic's IPO attracted about six times subscription, with the public offering raising up to $64.4 billion and accompanied by a $26 billion private placement, for a combined total of possibly $90.4 billion. Google released Gemini 4 Hydrogen, which is 2.5 times faster than Gemini 3.8 Flash, with API pricing at $1.50 per million input tokens and $7.50 for output.
Body · AI translation
Silicon Valley Frontier:
1. Jensen Huang Steps Down, NVIDIA Governance Faces a Major Test
1. Event: On November 12, 2026, NVIDIA announced that Jensen Huang would immediately resign as President, Chief Executive Officer, and member of the Board of Directors; current Executive Vice President and Chief Financial Officer Colette Kress will serve as interim CEO while continuing as CFO, and the Board has launched a search for a permanent CEO including both internal and external candidates; Harvey C. Jones was appointed as the company's first Chairman of the Board.
2. Market reaction: After the announcement, NVIDIA's stock price fell significantly, interpreted as a repricing of governance risk and key-person risk rather than evidence of weakening demand for AI accelerators; this risk will transmit layer by layer through customer procurement, supplier inventory building, capital expenditure, financing, and the software ecosystem.
3. Core judgment: Jensen Huang combined multiple roles—interpreter of the product roadmap, coordinator of customer relationships, ecosystem salesperson, and industry narrator—and his departure touches on expectation management; the moat lies not only in transistors but also in trust in the CUDA ecosystem; the market will subsequently focus on four validation points: the product roadmap, customer commitments, management stability, and the succession process, testing whether NVIDIA can convert personal credibility into institutional credibility.
2. NVIDIA Falls Below $4 Trillion, Who Takes the Baton After Jensen Huang
1. On November 12, 2026, NVIDIA announced that Jensen Huang would resign as President, CEO, and director (effective immediately); Executive Vice President and CFO Colette Kress will serve as interim CEO while continuing as CFO, the Board has launched a formal CEO search (internal and external candidates), and Harvey C. Jones was appointed as the first Chairman of the Board.
2. On the day of the news, NVIDIA's stock fell about 14% at one point in early trading, and its market capitalization fell below $4 trillion (evaporating as much as about $601 billion), before the closing decline narrowed to 4.34% and market capitalization returned above $4 trillion; intraday panic and closing recovery coexisted, showing that the market is torn between "key-person risk" and "business system inertia," rather than denying its technical capabilities.
3. Fundamentally, NVIDIA's fiscal 2026 revenue was $215.938 billion (up 65% year over year), data center revenue was $193.737 billion (up 68% year over year), and the Blackwell architecture accounted for most of data center revenue; the leadership change does not alter existing demand, but the valuation logic must add a new variable: "whether the successor can maintain the product roadmap, supply chain, and customer coordination"; AI sector valuations may shift from "growth certainty" to "growth plus risk discount."
3. Microsoft Outage: Cloud Efficiency Collides with Concentration Risk
1. From November 12 to 13, 2026, Microsoft 365 suffered a major global outage, affecting Outlook, Teams, SharePoint, and Copilot, among others, with the core being Exchange Online, which carries enterprise email and calendars; data from outage-tracking websites showed that reports came from 74 countries and regions within about one hour, with U.S. Outlook problem reports peaking at about 130,000. Microsoft said the failure was not caused by a cyberattack or AI agents, but by an accidental server shutdown combined with a network configuration change that triggered high load; after pausing deployments, shifting traffic, and restarting servers, services took nearly a day to recover in stages.
2. The market has not immediately priced in negative effects from the incident: Microsoft closed at $523.88 on November 13, up 0.17%, and investors still view it as an operational and reputational test rather than an event that immediately rewrites earnings expectations; but this reflects customer stickiness and trust reserves, not that concentration risk has been resolved, and long-term costs may manifest as higher redundancy investment, stricter service terms, and a lower concentration premium.
3. The incident has pushed regulatory focus from data security and privacy to system "resilience"; Japan's GSS government system shared by more than 20 departments, as well as Australian court and police systems, were affected, exposing the risk of core government affairs being tied to a single overseas cloud platform; industry views hold that true resilience lies in critical operations still having viable independent paths when the primary platform is unavailable (independent communication channels, offline copies, vendor switching, and recovery drills), rather than merely adding one more supplier at the procurement level.
4. Anthropic's $90.4 Billion IPO: Capital Begins to Judge the AI Story
1. Background and data: The IPO of Claude developer Anthropic attracted about six times subscription orders; the prospectus shows a public offering raising up to $64.4 billion, plus a $26 billion private placement; if both are completed at the upper limit, the total would reach $90.4 billion, potentially making it one of the largest stock issuance deals, though the subscription multiple, offering size, and final allocation may still change.
2. Structure and implications: The $90.4 billion is a combination of "public offering + private placement"; the former addresses market pricing and liquidity, while the latter is closer to a long-term exchange of resources; large-model companies in the scaling stage still need continuous investment in computing power, chips, data centers, and inference services, combining the attributes of software companies and infrastructure operators, meaning the market is accepting a more capital-intensive growth model.
3. Impact and views: Under oversubscription, underwriting banks may prioritize existing shareholders, and new buyers may receive less than expected; after listing, pricing power will shift from a few private investors to the public market, and the company will be tested by quarterly financial reports and cash flow; the core of competitiveness will shift from a model-capability race to capital conversion efficiency (whether computing power investment can be converted into paid usage and sustained cash flow).
5. Anthropic and SpaceX Push AI Financing Toward Compute Contracts
1. Anthropic announced in February 2026 that it had completed a $30 billion Series G funding round at a post-money valuation of $380 billion (more than double the $183 billion Series F valuation), led by GIC and Coatue, with participation from D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, MGX and others; the funds are directed toward infrastructure expansion, frontier research and enterprise-grade products. The article argues that the implication is a bet placed in advance on the future demand curve, and that AI competition is shifting from the number of models to capital intensity.
2. In an S-1 filing submitted on May 20, 2026, SpaceX disclosed that it had signed a cloud services agreement with Anthropic under which Anthropic will pay $1.25 billion per month from May 2026 to May 2029, for a potential total of nearly $45 billion, covering more than 300 megawatts of compute capacity at Colossus 1 in Memphis, with either party able to terminate on 90 days' notice; the article notes that "contract value" does not equal "realized revenue," but it marks the point at which compute begins to take on financial attributes similar to power purchase agreements, and computing capacity may become a service that can be sold separately.
3. The article's view: AI valuations must ultimately return to the balance sheet, and investors should look at three indicators: the actual activation speed and utilization rate of compute contracts, whether customer revenue can cover inference costs while leaving gross margin, and whether infrastructure assets can migrate across customers; Anthropic (the demand side), SpaceX (the infrastructure side) and Tesla (the application and physical asset side) form a synergistic but concentrated "narrative supply chain," and a slowdown in any link will transmit back through the chain; only companies that can line up financing, capacity, utilization and cash flow in a straight line are the true "picks-and-shovels sellers."
VI. OpenAI Raises $15 Billion, $1.25 Trillion Valuation Faces a Stress Test
1. Financing side: According to CNBC, OpenAI's latest private funding round totaled about $15 billion at a pre-money valuation of $1.25 trillion; the market had previously reported that it planned to raise at least $30 billion at a valuation of about $1.4 trillion, with the target gradually shrinking to about $15 billion, showing that capital is still willing to bet but is becoming more cautious about financing size and deal terms.
2. Operating side: According to The Wall Street Journal, OpenAI's third-quarter revenue was about $10 billion, up nearly 50% quarter over quarter, while its operating loss widened to about $10.9 billion (about $10.1 billion in the second quarter); the adjusted operating margin improved from about -150% to about -109%, with revenue accelerating but not translating into profit, presenting a pattern of "the more it sells, the more it loses."
3. Industry and valuation: Anthropic had third-quarter revenue of about $17 billion and an operating loss of about $567 million and is preparing for an IPO, forming a contrast; the valuation logic for AI companies is shifting from "model capability" to "cash flow capability," and the market is shifting from "give money as long as there is growth" to "growth must explain costs"; the key to OpenAI's next stage lies in lowering inference costs and narrowing losses.
VII. Google Fights Back Inward with Hydrogen
1. Product positioning: Google released Gemini 4 Hydrogen, focused on high-speed execution and agentic workflows. According to Google, it scores 51 points on the default high setting of the Artificial Analysis Intelligence Index (higher than the flagship Argon's 50 points) and 54 points on the xhigh setting, is 2.5 times faster than the previous generation Gemini 3.8 Flash, improves its Terminal-Bench 4.1 score from 18.2% to 57%, and reduces hallucinations by more than fivefold, shaking the traditional tiering of "flagship is strongest, lightweight is a compromise."
2. Cost and orchestration: Hydrogen's API pricing is $1.50 per million input tokens and $7.50 per million output tokens, covers a 1 million token context, and is integrated into Google AI Studio, the Gemini API, the enterprise edition, Search and the paid Antigravity subscription; together with Antigravity's Orchestrator, Carbon handles planning while Hydrogen handles execution and testing, with multiple models dynamically handing off to one another. Google says that on Terminal-Bench 4.0, Orchestrator's task success rate is close to Carbon's while saving an average of 24% of quota and speeding up by 19%, with competition shifting from "model rankings" to "efficiency per task."
3. Industry impact and risks: Through Hydrogen, Google is turning the model layer into a "distributed workforce"—the flagship model handles difficult problems as a fallback, while the fast, low-cost model handles massive high-frequency calls; the pressure on OpenAI and Anthropic comes mainly from Google's existing distribution network rather than from a single leaderboard score. But the faster the speed, the faster errors propagate; for long workflows, stability, failure recovery, rework rates and the number of human interventions deserve more developer attention than the highest leaderboard score.
VIII. Anthropic Makes Its Flagship Model the Default
1. Anthropic has made Claude Opus 5.5 the default model for the Pro, Max, Ultra and Team plans, positioning it as the everyday workhorse for research, analysis, writing and coding, with the flagship model shifting from a "tool for key moments" to the entry point of daily workflows; its API pricing is $4 per million input tokens and $20 per million output tokens, below the $5 and $25 of Opus 5, and the company says that under typical workloads total running costs are 40% lower and output speed is more than 30% faster, with cache reads at $0.20 per million tokens.
2. The model matrix is now layered by the cost of task errors: Haiku 5.5 ($0.10 input, $0.50 output) takes on high-frequency, narrow-scope, easily validated tasks; Sonnet 5.5 ($2, $10) handles documents and spreadsheets with clear boundaries; Opus 5.5 takes on complex knowledge work; and Fable 5.2 ($8, $40) targets open-ended, long-running technical tasks. The logic of model selection has shifted from "capability rankings" to "task risk."
3. Opus 5.5's thinking feature is always on, with the effort parameter controlling thinking depth (medium by default), and model management has shifted from "whether to switch models" to dynamically allocating compute budgets; but vendor benchmark conclusions cannot replace on-the-ground business validation, and scenarios such as finance, law, healthcare and security still require human review of sources and assumptions, while the cost accumulation and path dependence brought by defaulting also warrant caution.
IX. SpaceX Bets $74 Billion, AI Compute Enters the Heavy-Asset Era
1. SpaceX plans to build two supercomputing facilities in Cochise County, Arizona, and near its Starbase headquarters in Texas, with a total planned investment of about $74 billion (about $50 billion and $24 billion respectively). The job figures are projections: Texas has already created more than 3,000 construction jobs and expects to employ 500 people permanently, while Arizona is expected to have about 5,500 construction jobs and 1,000 permanent positions; none of these have all materialized.
2. SpaceX intends to build, own and operate a gas-fired power plant for the Texas project itself, incorporating energy supply and computing capacity into the same infrastructure plan, marking a shift from "buying servers" to "arranging energy supply" through vertical integration; the cost is heavier capital expenditure, longer cycles, and responsibility for the regulation and operation of energy facilities.
3. The article's view: the AI industry is moving from a software startup phase into a heavy-asset infrastructure race, with scarce resources shifting toward power and construction capacity; the key to judging whether the $74 billion investment succeeds lies in whether project approvals and construction can proceed as planned, whether the power system can operate stably under cost and regulatory constraints, and who ultimately uses the compute and at what price it is monetized—synergy does not equal a commercial closed loop.
X. OpenAI and Other Giants Sued Over Their Own Doomsday Prophecies
1. On November 12, 2026, four nonprofit organizations—Public Citizen, the Irreplaceable Education Fund, the Heat Initiative and the Alliance for Secure AI—sued OpenAI, Google, Meta, Anthropic and SpaceX's AI entity in San Francisco County Superior Court, asking the court under California's Unfair Competition Law to freeze frontier model research and deployment, a rare judicial attempt.
2. Litigation strategy: rather than first arguing the merits of the technology, they turn executives' earlier public risk statements (Anthropic CEO Amodei saying there is about a one-in-four chance of a bad AI outcome, Musk warning in 2023 that AI could destroy civilization) into evidence of "advancing despite knowing the risks," turning safety commitments from trust assets into a double-edged sword in litigation.
3. Impact and outlook: as of the report, no court freeze order had been issued, and in the short term the case is more likely to be a test of institutional boundaries than an immediate halt; combined with regulatory pressure from Florida and eight other states that previously sued OpenAI and others, frontier AI competition may shift from capability and compute to safety proofs, audit records and verifiable commitments.
XI. Nvidia Changes Leadership, A-Share Supply Chain Looks First at Resilience
1. Background: on November 12, Nvidia officially announced that co-founder, president and CEO Jensen Huang has resigned with immediate effect, appointing executive vice president and chief financial officer Colette Kress as interim CEO and Harvey C. Jones as the first chair of the board, and launching a search for a successor; the company said operations, customer commitments and the product roadmap are expected to continue, but the founder transition still creates a window for observing governance and execution.
2. Supply chain response (November 13): Runjian said its token factory centers on token output, with diverse compute resources that can be dynamically allocated through leasing; Lingyi iTech said its subsidiary Liminda continues to deliver liquid-cooling and power products in volume to leading North American compute customers and their supply chains, and exhibited the Inner Manifold manifold and a new-generation zero-pressure-drop quick connector; Shengyi Technology clarified that it has not received any notice of order cancellations from any end customer, that production and operations are normal and that orders are full.
3. Core view: The article argues that the medium- to long-term divergence in the A-share supply chain depends on business structure—compute service providers with diversified procurement, adaptation to multiple server types and dynamic scheduling capabilities are more resilient to supply shocks, while manufacturers that enter system-level segments such as liquid cooling and power supplies and continue to pass certification are more likely to gain upgrade opportunities; investors need to confirm supply-chain relationships through company announcements, periodic reports, customer certification and actual deliveries, and guard against three misconceptions: equating "responding that business is normal" with performance growth, equating "participating in the supply chain" with orders coming from Nvidia, and equating "the liquid-cooling trend" with all liquid-cooling companies benefiting.
Open-source trends:
12. Rust 1.100.0 pushes safety into the supply chain
1. Rust 1.100.0 was released on November 12, 2026, stabilizing the never type ("!", the never-return type) after 10 years and 5 failed attempts, and changing the fallback type for all versions from "()" to "!"; it also stabilized extern "custom" and the core API for custom allocators (the Allocator trait), enabling standard types such as Vec and Box to use custom allocators.
2. Lints are becoming Rust's "second type system": duplicate #[repr(...)] attributes trigger checks, drop_in_place-related lints have broader coverage, and repr(align) fields inside repr(packed) are now denied by default, with finer diagnostic boundaries to reduce false positives; Cargo introduces its own lint system for the first time (such as cargo::unused_dependencies), allowing warnings to be controlled centrally in [lints.cargo].
3. Cargo is moving from "resolving dependencies" to "governing dependencies": a minimum publish age dependency cooldown period has been added, setting a time threshold for dependency resolution but not replacing security audits; cargo install now uses the lockfile inside the .crate archive by default, making --locked a no-op; the layout of intermediate build artifact directories has changed, and tools should use Cargo JSON messages to locate artifacts rather than relying on internal directory conventions.
13. What Nvidia finds hardest to blockade is not chips
1. Former OpenAI researcher Will DePue (Unemployment Labs) published a report in November 2026 stating that six open-weight models—DeepSeek V4, Moonshot Kimi K3, MiniMax, Ant Group and others—retain hardware fingerprints consistent with Nvidia Blackwell (148 stream processors for B200/B300, 152 for GB200/GB300), which Hopper and AMD chips cannot reproduce; however, this evidence only suggests that the training process may have had contact with such chips and cannot by itself prove illegal export or violations.
2. Controls target the physical flow of chips, but compute can reach its destination via third-country data centers, cloud services and rented compute; Epoch AI estimates that about 2 million advanced chips entered China via hidden destinations in 2024-2025, with Malaysia importing about 275,000 Nvidia accelerators, corresponding to a transshipment scale of about $3 billion (a trade-data inference, not a judicial finding). In January 2026, the U.S. Bureau of Industry and Security changed H200, MI325X and others to case-by-case review, with additional conditions such as customer screening and third-party testing.
3. The core of the dispute is shifting from hardware blockade to "compute verifiability," and regulators may require retention of hardware logs, cloud instance locations, chip serial numbers and task timestamps; for Nvidia, the risk lies not only in sales but in whether it can build a compute-tracking system convincing to regulators, cloud customers and investors—its moat is shifting from "having the most powerful GPUs" to "knowing who has used the GPUs."