Edge AI Daily Morning Brief (November 15): NVIDIA's Jensen Huang Resigns as CEO; Google Releases Gemini 4 Hydrogen
On November 12, NVIDIA announced that Jensen Huang has resigned from his roles as President, CEO, and board member, effective immediately; CFO Colette Kress has taken over as interim CEO, and the board has begun the search for a successor. Signatures on the petition website have exceeded 115,000. Google released its execution-type model Gemini 4 Hydrogen the same day, supporting four thinking levels and a 1 million Token context, with API pricing at $1.50 per million input Tokens and $7.50 per million output Tokens. This morning brief also covers the FAA clearing Starship's 15th flight, Microsoft 365 being down for more than 24 hours, and other news.
Body · AI translation
Silicon Valley Frontier:
1. Nvidia asks Jensen Huang to return; what is really on trial is the board of directors
1. On November 12, Nvidia announced that Jensen Huang had resigned as president, CEO and director with immediate effect; Executive Vice President and CFO Colette Kress would serve as interim CEO while retaining her original position; the board launched an internal and external search for a successor and appointed Harvey Jones as its first chairman. As of the afternoon of November 14, signatures on the petition website exceeded 115,000, including 5,965 current employees, 1,108 former employees and 63,063 shareholders; tech executives such as Musk were also among the supporters, demanding that the board let Huang return and disclose more information while an independent investigation clarifies the facts.
2. The article argues that the core of the dispute is not "whether Jensen Huang should return," but the tension among founder dependence, board procedure and supply-chain trust: Huang has led the company since 1993 and is deeply tied to chip design, cloud computing, AI models, developers and policy relationships; his departure pushes the question to "whether Nvidia can continue to make key decisions without him." A successor must understand chips, systems, the software ecosystem, cloud services and the policy environment at the same time; an interim CEO can stabilize financial communication and operating order, but can hardly fully replace the founder's technical appeal and industry-relations function.
3. On the day of the announcement, Nvidia closed down 4.34%, having fallen as much as about 14.35% intraday before gradually recovering; on the same day the Philadelphia Semiconductor Index fell 0.54%, AMD and Broadcom rose more than 5%, and retail investors made about $111 million in net purchases in the first 30 minutes of trading, a record for the same period; the company said operations, customer commitments and the product roadmap would not be interrupted, and it still planned to release earnings on November 17. The article points out that what AI infrastructure customers really worry about is roadmap and supply stability; management uncertainty raises partners' waiting costs, making this executive change a test of the "continuity premium," but continuity cannot become a reason to bypass governance.
II. Huang's departure: Asia first confirms whether chips can still be delivered
1. On November 12, 2026, Nvidia announced that founder Jensen Huang had immediately resigned as president, CEO and director; Executive Vice President and CFO Colette Kress would serve as interim CEO (while retaining the CFO role); Harvey Jones became the first chairman of the board; and the board simultaneously launched a search for a permanent CEO (internal and external candidates). Previously, on November 8, the board stated that it had launched an independent investigation into reports that "Jensen Huang was targeted by foreign intelligence activities," and Huang promised to cooperate.
2. Asian governments' concerns focused on "whether projects can run without the individual": South Korea's science ministry said Nvidia reaffirmed the supply of 260,000 advanced GPUs and the advancement of joint AI projects, and the government's AI strategy committee held an emergency meeting; Japan regards Nvidia as an important player in AI semiconductors and stressed stable supply; Taipei said the overseas headquarters project "is still progressing."
3. Market analysis suggests AMD and Broadcom may benefit from customers diversifying suppliers, but their ability to take over depends on software compatibility and delivery capability; the article concludes that Nvidia has entered a "de-personalization" testing period, with successor selection, the national security investigation, cross-border chip compliance and performance of Asian projects proceeding in parallel, and the key indicators are whether the next batch of chips can be delivered on schedule and whether the next projects can continue to be fulfilled.
1. (Event) On November 12, 2026, Nvidia announced that co-founder, president and CEO Jensen Huang had resigned with immediate effect; Chief Financial Officer Colette Kress would serve as interim CEO, and Harvey C. Jones would become the company's first board chairman; Huang had led the company for about 33 years, and the market views this as no ordinary management rotation.
2. (Model and data) Nvidia has changed from a GPU supplier into a "capital + computing power" ecosystem organizer: according to PitchBook, it participated in 105 financing rounds this year and was often the largest investor, investing $2 billion in CoreWeave in January and $2 billion in Nebius in March, and on August 19, 2026, announced up to $105 billion in conditional credit guarantees for OpenAI's Ohio data center; the trend is that AI infrastructure competition has become a combined competition of capital, equipment, data centers and customer contracts.
3. (Impact and views) The shock is concentrated on neocloud and computing infrastructure companies, and the financing market is expected to diverge: companies with long-term contracts, stable customers and high GPU utilization can still obtain funding, while those relying on future demand or a single strategic investor will face tighter terms; the deeper risk is that the market loses a unified anchor for expectations, and the industry shifts from "who can get equipment" to "who can turn equipment into profit," with capital no longer flowing in merely because "Nvidia is in it."
IV. The governance questions behind Nvidia's succession rumors
1. According to a November 13 report by the UK's Financial Times, Nvidia's board has approached Arm Holdings CEO Rene Haas regarding CEO succession; Nvidia's board responded that it had not "formally approached" any candidate and that speculation about the relevant candidates was "premature," meaning the contact reported does not equal a confirmed candidate or a completed appointment.
2. Haas worked at Nvidia for about seven years, rising to vice president and general manager of the computing products business; he joined Arm in 2013, became Arm CEO in 2022 and led its Nasdaq listing in 2023; he combines identification with Nvidia's culture and experience running a listed company, but Arm and Nvidia have different business models and there are industry relationships between the two sides, so formal contact must be handled prudently at the board, legal and conflict-of-interest levels.
3. The core test for succession lies in translating Jensen Huang's personal judgment into organizational capability and completing the shift from "Jensen Huang's Nvidia" to "Nvidia's Nvidia," which requires establishing a clearer delegation system, auditable risk management and a stable executive pipeline; at the same time, the next CEO must also cope with the policy uncertainty and regulatory communication pressure brought by U.S. export restrictions on advanced AI chips.
V. Google Swaps Brains: Hydrogen Strikes Back at the Flagship
1. On November 12, 2026, Google released Gemini 4 Hydrogen, officially positioned as an execution-oriented model focused on everyday coding, agentic workflows and high-frequency applications; it supports four thinking levels—low, medium, high and xhigh (default high)—turning the reasoning budget into an adjustable cost dial. API pricing is $1.50 per million tokens for input and $7.50 for output, covering a 1 million-token context window.
2. Google officially says that on the Artificial Analysis Intelligence Index, Hydrogen scores 51 at the default high level, above its sibling Argon's 50, and reaches 54 at the xhigh level, but still below Carbon's 60; Google also says that at the default high level it completes tasks about 2.5 times faster than Gemini 3.8 Flash and scores 57% on Terminal-Bench 4.1 (compared with 18.2%). All the above figures come from Google's release materials, not independent blind-test conclusions.
3. Google uses an Orchestrator to dynamically switch models within the same session, with Carbon responsible for system design, complex judgment and task decomposition, and Hydrogen responsible for code modification, parallel work and verification; Google says that on Terminal-Bench 4.0 this combination achieves a success rate close to using Carbon alone, while consuming an average of 24% less quota and being 19% faster. The article argues that the focus of competition may shift from parameter scale and top leaderboard scores to output per task and workflow-level metrics, and may trigger a more intense price war.
VI. FAA Clears Starship's 15th Flight, SpaceX Must Pass the Regulatory Loop
1. The U.S. Federal Aviation Administration (FAA) has issued SpaceX a launch license for Starship's 15th flight, approving launch from its Starbase site in Boca Chica, Texas. The license assessment covers matters including public safety, overflight of densely populated areas, payload and national security, and sets flight and safety boundaries for the specific mission.
2. The core change in this license is that it covers the "launch—reentry—recovery" loop: SpaceX plans to have the upper stage return from orbit to the Texas launch site and be caught by the launch tower rather than splash down in the sea, and the return trajectory passes over land including northern Mexico, so the scope of the safety assessment expands accordingly. SpaceX President Gwynne Shotwell said the license is expected to involve return to the launch site and tower capture of both rocket stages.
3. Regulation is becoming the "second technology roadmap" for Starship commercialization: the FAA previously discussed up to 25 orbital launches per year and corresponding landing arrangements around Boca Chica's environmental assessment (a project-level assessment, not a term of the 15th flight license). If safety reviews of the flight path, reentry method and recovery maneuvers take a long time, SpaceX's overall iteration pace will still be locked by the regulatory process.
VII. Microsoft Denies Connection to Nationwide Flight Grounding
1. Event and official characterization: on November 13 local time, the U.S. Federal Aviation Administration (FAA) issued a nationwide flight grounding order for "special circumstances and safety reasons," suspending departures from airports because of communication problems at multiple air route traffic control centers; Microsoft responded that there is no indication the grounding order is related to any of its services, and U.S. Homeland Security Secretary Malin also said that the Microsoft service issue and the FAA communications outage are two independent events, with no sign of malicious attack so far.
2. The article's core judgment: at present the causal chain remains only at "temporal overlap" and has not entered "technical linkage"; for the aviation system, suspending departures in the face of communications anomalies is a safety-redundancy measure, not a determination of the source of the fault, and "using Microsoft products" does not equal depending on a particular Microsoft cloud service or having critical safety controls hosted by it. The real risk is that attribution outpaces investigation—unverified inferences spread more easily in an information vacuum, potentially amplifying market panic, interfering with the real investigation and obscuring the communications equipment and emergency procedures that genuinely need improvement.
3. Deeper implications: the incident exposes a new risk in the digital infrastructure era—"systems are mutually visible, yet the boundaries of failure are increasingly blurred"; in the future, critical systems must answer not only "is there a cloud service" but "which functions can still run when the cloud service is abnormal"; the key to resilience lies in independent communication links, verifiable degraded modes, clear supplier boundaries and regular cross-system drills, and the attribution conclusion of a single incident cannot replace an architecture audit of the entire industry.
VIII. Microsoft Restores Service, but the Single-Point Risk of Cloud Office Is Closer
1. On November 14, 2026, Microsoft announced the full restoration of its Microsoft 365 flagship services: previously, Outlook, Teams, Copilot and others had suffered a global outage lasting more than 24 hours caused by a network capacity upgrade issue, affecting more than 70 countries and regions, with U.S. Outlook failure reports peaking at over 130,000; Microsoft said the cause was multiple servers going offline due to excessive load, not a cyberattack, and that this was the second large-scale outage of the year.
2. The outage exposed the risk of cloud service centralization: Microsoft 365 reduces enterprise deployment costs through a unified identity, data and operations system, yet it concentrates risk onto a small number of critical services; underlying failures propagate layer by layer through identity authentication, message delivery, meeting scheduling and AI invocation, ultimately manifesting as multiple products being unavailable at the same time—the more successful a single platform is, the more easily it becomes a shared dependency for everyone.
3. Service restoration does not equal business restoration: after front-end responses recover, backlogged emails, calendars and notifications still need retry delivery, and enterprises must also verify whether data and permissions are fully closed-loop; the article warns that enterprises should not hand critical work to a single platform, and should retain email archiving, offline copies, and backup paths such as phone and SMS that can be regularly drilled and switched to.
IX. Apple Brings Siri AI Down to Older Apple TVs
1. Apple released the new Apple TV 4K in October 2026, listing Siri AI as a core selling point; it subsequently confirmed that tvOS 27.2.1 can bring Siri AI support to second-generation and later older Apple TV 4K models, but requires pairing with a nearby device that has Siri AI enabled (such as an iPhone); first-generation models are not supported.
2. A shift in product tiering logic: from "feature exclusivity" to "experience tiering," with basic AI capabilities spreading downward through backward compatibility, while hardware synergy (such as the new model forming a four-speaker system with HomePod mini 2 and built-in far-field microphones) concentrates upward; the value of new devices is no longer proven by exclusive Siri AI.
3. For Apple (a large-cap tech stock), the move aims to expand AI reach and ecosystem stickiness, changing the reason to upgrade from "must upgrade" to "worth upgrading"; competition in living-room AI shifts its focus from single-point performance to device synergy among phones, speakers and TVs, and the value of older hardware is extended because it sits within the device network.
1. The Dyania Health clinical AI system Synapsis AI, introduced on NVIDIA's blog, is positioned as a clinical reasoning engine rather than a general-purpose chatbot; it links patient medical records across document types and time points, outputting clinical signals traceable to the original records, with physicians retaining verification and final decision authority—that is, a "clinical co-pilot" positioning.
2. In deployment, the system runs inside the hospital's own firewall, relying on in-house NVIDIA accelerated computing infrastructure, with patient records never leaving the hospital, and is trained with NVIDIA NeMo and optimized with TensorRT and TensorRT-LLM; NVIDIA thus extends from chip supply to the computing foundation inside hospitals, but still faces the healthcare industry's long procurement, validation and compliance cycles.
3. The Cleveland Clinic case (7,812 aortic valve replacement patients from 2005 to 2021, 88 clinical questions per person, medical records spanning up to 16 years) shows that AI can accelerate cohort screening and risk modeling, but only proves research efficiency gains, not improved patient clinical outcomes; WHO guidance on the ethics and governance of AI in healthcare requires human oversight, transparency and explainability, and accountability, and AI's value still needs long-term real-world validation.
XI. AWS and NVIDIA Redefine the Cost of Coding Agents
1. A shift in cost logic: the real expense of coding agents is not a single model call, but the number of call rounds, historical context replay and failure retries needed to complete a task; billing trends are moving from "token-based pricing" to "task-based pricing," and enterprise evaluation should shift from input/output prices to average call rounds, failure retry rates and final pass rates.
2. The AWS and NVIDIA NeMo Switchyard solution: requests pass through a load balancer, LiteLLM gateway and Switchyard plugin, and are ultimately sent by Amazon Bedrock to backend models; its Stage routing selects the model tier round by round based on tool activity and results in the full conversation history, differing from static routing that fixes model selection at task start.
3. Routing value and boundaries: the goal should be to optimize "cost per completed task" rather than per-round call price—a low-cost model that frequently reworks may not be cheaper; the solution uses claude-opus-5-5 and nemotron-super-3-120b for a single-model baseline versus multi-model routing comparison, but the original text does not give an extrapolable savings ratio, and it must be distinguished from Intelligent Prompt Routing within the same Bedrock family, with each routing decision logged to control governance and audit costs.
1. The U.S. Department of Transportation launched SMART, an AI-assisted air traffic control tool that consolidates about 200 data streams—including weather, routes, traffic flow and controller staffing—into a single platform, aiming to shift aviation scheduling from "handle problems after they arise" to "predict congestion in advance"; however, the FAA made clear that it cannot replace controllers or take over aircraft, that its suggestions must be reviewed by staff, that local managers may accept or reject them, and that ultimate safety responsibility remains with humans.
2. The FAA has long faced a controller shortage: in fiscal year 2024 it hired 1,811 controllers (against a target of 1,800), while about 1,400 left during the same period, and it hired 5,778 in total over the past five years; in 2025 the FAA lowered its target for active controllers from 14,633 the previous year to 12,563, and the article warns that the risk lies in converting AI efficiency into lower staffing, training and backup configurations, thereby compressing safety redundancy.
3. Air traffic control modernization remains constrained by aging infrastructure: in July 2025 the U.S. Congress appropriated $12.5 billion to support the first phase, and the FAA estimates the second phase requires about $10.2 billion more; Boeing's 2025 forecast projects that the global aviation industry will need about 2.37 million new personnel from 2025 to 2044 (including 660,000 pilots), indicating that AI cannot replace human labor demand—the key is to preserve human veto power, make infrastructure able to withstand failure, and reserve verifiable backup plans.
Open-source trends:
XIII. AWS Turns OpenFold3 into Usable Biocomputing Infrastructure
1. AWS integrated OpenFold3-preview (version 0.4.1, of3-p2-155k checkpoint) into Amazon SageMaker JumpStart. The model was developed by Columbia University's AlQuraishi Lab and the OpenFold Consortium, with the goal of reproducing DeepMind's AlphaFold3 at the residue level, and is released under the Apache 2.0 license supporting academic and commercial use; inputs are protein, DNA and RNA sequences and SMILES ligands, and outputs are mmCIF structures and confidence metrics such as pLDDT, PAE and PDE.
2. Data boundary design is key to this deployment: AWS uses network isolation, and users can generate MSAs on their own infrastructure and pass them to the endpoint with the request, so sensitive sequences need not leave the enterprise account. The article therefore concludes that competition in cloud AI services is shifting from "whose model is stronger" to "who can let customers connect data with confidence," with data residency, task orchestration, cost control and verifiable results becoming core capabilities.
3. The JumpStart deployment supports asynchronous inference and multi-GPU instances (ml.g6.2xlarge, ml.g6.4xlarge, ml.g6.12xlarge, with g6.4xlarge as default), and provides a scale-to-zero mechanism to change the cost structure, letting compute scale in and out with tasks. However, OpenFold3-preview remains a research preview, and its confidence metrics are model judgments rather than experimental facts; it is positioned as a high-value upstream tool to shorten the time from sequence to structural hypothesis, reducing trial-and-error costs without replacing experimental validation.