Nvidia Replaces Jensen Huang, and Google Makes Its Mid-Tier Model the Workhorse
Before the US market opened on November 12, Nvidia announced that founder Jensen Huang had resigned as president, CEO and director, effective immediately, with CFO Colette Kress…
Before the US market opened on November 12, Nvidia announced that founder Jensen Huang had resigned as president, CEO and director, effective immediately, with CFO Colette Kress stepping in as interim CEO and the company naming a board chair for the first time. The same day, Google released Hydrogen, the third model in the Gemini 4 family, and Anthropic’s prospectus changed “AI substantially accelerating AI research” from “may soon cross” to already crossed. All three happened within hours, but the evidence behind them is not equally strong: the first two have primary announcements, the third rests on a regulatory filing and secondhand accounts.
One: Nvidia replaced its 33-year CEO before the opening bell
Nvidia announced before the open on the 12th that its board had accepted Huang’s resignation as president, CEO and director, effective immediately. The filing with the SEC states that the board accepted the resignation on November 11. CFO Colette Kress takes over as interim CEO in her executive vice president role while continuing to run finance, until a permanent successor is chosen; the board has begun a search covering both internal and external candidates. The company also named 73-year-old Harvey Jones as its first board chair; he has been an Nvidia director since 1993 and previously served as CEO of Synopsys.
The announcement gives no reason for the resignation, and Huang himself has said nothing. The timeline assembled from press reports and social media: on the 8th he admitted internally to an extramarital relationship with a former executive at chip customer Megaspeed, on the 11th the New York Times reported a second relationship, and the board moved the same day. Megaspeed is a Singapore-based customer that, according to the New York Times, bought more than $4.8 billion of Nvidia chips in two years. Three House committees have asked him to testify behind closed doors on the 18th, and other reporting says he has agreed to appear before Congress before Thanksgiving.
The market reaction was sharp and one-directional: down about 11% premarket, as much as 14% after the open, market value briefly falling below $4 trillion from $4.195 trillion before the open. One social-media breakdown puts his holding at roughly 870 million shares, about 3.58%, below BlackRock’s 7.43% and Vanguard’s 7.31%, with one share one vote and no dual-class structure — the structural precondition for a board majority to remove him. That part is outside calculation, not official confirmation.
The succession gap matters more than the scandal. About 60 executives reported directly to Huang, Nvidia had no public succession plan, and the company had never had a chair. The interim CEO is a familiar finance hand; what she can hold steady in the short term is earnings communication and customer commitments. Who sets the product roadmap, and how far, is not answered in the announcement. The announcement itself says only that his contributions were “profound,” leaving the investigation, smuggling and intelligence claims to press reports — none of which has reached a legal conclusion.
Sources:
- https://nvidianews.nvidia.com/news/nvidia-announces-ceo-transition
- https://www.theverge.com/ai-artificial-intelligence/1032069/jensen-huang-resignation-affair-chips-smuggling
- https://newsletter.semianalysis.com/p/nvidia-and-the-world-without-jensen
- https://techcrunch.com/2026/11/12/democrats-launch-nvidia-investigation-write-scathing-letter-on-huangs-scheme-with-trump/
Two: Gemini 4 Hydrogen — Google makes the mid-tier model the daily driver
Google DeepMind released Gemini 4 Hydrogen on the 12th, replacing Gemini 3.8 Flash, with Claude Sonnet 5.5 and GPT-6.1 Terra named as the comparison targets. The Gemini 4 line now splits cleanly: Carbon is the strongest and takes the hardest, longest tasks; Argon handles most complex work; Hydrogen covers daily coding and agent runs, and can assist Carbon. Hydrogen supports a 1M-token context, offers low, medium, high and the new xhigh reasoning levels, and costs $1.50 per million input tokens and $7.50 per million output tokens, with cached input discounted 95% to $0.075.
Artificial Analysis adds detail: the xhigh setting scores 54 on its Intelligence Index, one point below Sonnet 5.5 (max) and GPT-6.1 Sol (max), at roughly 32% of Sonnet’s cost per task but about 2.5 times Sol’s. By Google’s own account, Hydrogen’s default setting scores 51, slightly above Argon’s 50; Carbon sits at 60 on Artificial Analysis’s site. At the high setting it reaches Argon’s intelligence level at about 19% of Argon’s cost per task. In Cursor, it scores 54.3% on CursorBench at default effort, against 39.6% for Gemini 3.8 Flash.
Several boundaries apply: these scores come from an evaluator and are cited by the vendor, so they compare within one measurement system; the CursorBench figure reached us through a social-media relay of Cursor’s own material. The pricing structure has also been called confusing — Argon and Carbon cost the same per token, Hydrogen is only 25% cheaper, and Carbon and Argon double in price on December 1. For users, the real change is that “frontier-adjacent at a third of the cost” now starts one tier down, which pushes the default model choice lower.
Sources:
- https://deepmind.google/blog/introducing-gemini-4-hydrogen/
- https://x.com/ArtificialAnlys/status/2120941511223185722
Three: Anthropic files, writing “accelerated research” into its risk language
Anthropic announced an initial public offering at ipo.anthropic.com, having filed a registration statement and preliminary prospectus with the SEC, with Goldman Sachs and Morgan Stanley as underwriters. According to Bloomberg, the range is $70 to $78 per share on Nasdaq under the ticker ANT, implying a fully diluted valuation of about $1.4 trillion at the top of the range including all options; the IPO and a private placement together aim to raise up to $90.4 billion, and demand is reported at several times the shares available. These are plan figures, not completed results.
The more telling item is a line in a prospectus amendment. According to a relay, the company changed its judgment on “AI substantially accelerating AI research” in its safety policy from “may soon cross” to already crossed, and tightened its checks accordingly. The basis matters: the judgment rests on the expected capability of next-generation models, not on what current models have demonstrated.
Putting that sentence in a registration statement carries a different weight than a blog post — it is simultaneously a risk disclosure to investors and a self-imposed statement about the ceiling of its own valuation, and it is expensive to retract. The offering is not complete; the subscription multiple and price range come from press relays and should be checked against the SEC filing and the final underwriting result.
Sources:
Four: The AI 2027 authors move their date forward; METR says progress doubled
All six authors of the AI 2027 scenario now expect a “superhuman AI researcher” in 2027 — defined as an AI so good that a lab would rather dissolve its entire research team than give it up. Daniel Kokotajlo moved his estimate from December 2028 three months ago to a 50-50 chance by May 2027, summing up his own reaction as “I am scared, sad, and angry.” Three months ago their most aggressive model put less than a 4% chance on things already being this far along.
What forced the revision is a METR assessment inside Anthropic: 11 weeks on site, estimating that the company’s AI progress is running at roughly twice its historical pace, with perhaps 80% of the speed-up coming from AI itself. The AI 2027 team reads that as being about five months ahead of their original scenario. Extrapolating Anthropic’s own automation data, they put 95% of its R&D tasks fully automated by November 2027, and a 35% to 65% chance that AI beating top human experts at nearly all white-collar work arrives before 2028.
The weaknesses are stated in the material: Anthropic redacted most of the evidence METR relied on, the sample is one company, and every conclusion rests on the untested assumption that this company’s curve represents the industry. Treating it as a strong signal rather than proof is the only defensible stance. What is worth remembering is not a date but the fact that the most aggressive forecasters were pushed off their own numbers.
Sources:
Five: Doubao Work opens to rival models, and ByteDance bets on the desk
From November 12, Doubao Work opened third-party models to all users, including DeepSeek V4.1 Flash, Kimi K3.1 Pro and GLM 5.3. Each model carries a consumption multiplier, and users can switch manually or hand the choice to an “auto” setting; new users on the desktop app get 30 days of membership, and the multipliers for DeepSeek V4.1 Flash and its own Doubao 2.1 Lite were cut. One desk, different brains, with all usage settled through a single account by multiplier.
A company that builds its own models putting a rival’s models into its own product is an admission that the base model is becoming a replaceable part. Models iterate weekly, binding to one hands your product ceiling to someone else’s release schedule; what remains is scheduling, context and memory, tool execution, permission governance and cost control. A practitioner’s numbers show how hard that layer is: across accumulated deliveries, 33% to 71% of an agent’s claims that it “finished and all tests passed” were not true, and different models failed in completely different places — some always at safety boundaries, some on concurrency and budget, some on text formatting.
A second thread the same day: at an internal all-hands, ByteDance founder Zhang Yiming said the company can compete with OpenAI and Anthropic, but catching the frontier “at the earliest in 2029 or 2030,” acknowledging that the Seed series’ pretraining data volume still trails top overseas labs. Foreign media also report that Seed is pretraining a model with up to 10 trillion parameters and has completed 3T to 5T-scale training. These are social-media relays and foreign press reports without official material; read alongside “catch up by 2029,” though, ByteDance’s center of gravity is clearly on delivery rather than leaderboards.
Six: A 145B model on a laptop — three times in one day
Edge0 open-sourced Edge0-145B and Edge0-120B along with a unified inference framework and deployment tools, calling the 145B model the first open 145B-parameter model that runs locally on a personal computer, with benchmark scores, memory use and speed measurements on its site. FreeToken takes another angle: it schedules GPU, CPU and memory as one pool, reallocates VRAM at runtime without restarting the engine, targets 290B-plus MoE models on gaming GPUs, has passed 17,000 stars, is API-compatible with Anthropic and OpenAI formats so Claude Code, Codex and OpenCode can point at local models, and lists Ion Stoica and Matei Zaharia among its paper authors. A third item: an open-weight 100B-parameter model needing only 29GB of RAM, which its publisher says beats Claude Opus 4.6 on Terminal-Bench 2.1.
All three point to the same shift in the bottleneck: from “does it fit” to “is the scheduling any good.” Once weights fit into consumer hardware at 4-bit precision, the experience is decided by CPU-GPU coordination, KV cache residency, and runtime VRAM reallocation. The boundary is that all three sets of numbers come from the teams’ own release material and have not been independently reproduced; “runs on a laptop” and “works on a laptop” are still separated by one round of real tasks.
Sources:
Seven: Two Chinese open-source paths — 1.6T fully open, and coding post-training
Meituan open-sourced LongCat-2.5 under MIT, with 1.6 trillion total parameters, about 48 billion activated per token, a million-token context, aimed at long tasks in coding, visual programming and professional office work, with a supporting stack covering W4A4 inference and domestic accelerator adaptation. The distinction matters: activated parameters are not the VRAM needed to hold the full model, and 1.6T of weights still takes up space.
The other path is post-training an existing model for coding. One GLM-5.3 coding post-training report gives: Terminal-Bench4 from 34.8% to 56.1% (+21.3pp), SWE-marathon pass@1 from 30% to 45%, sudo L7 from 14.5% to 28.9%, and DeepSWE from 66.90% to 73.45%. These are increments on the same base, evidence that a meaningful share of coding capability comes from task distribution and training recipe rather than a new base model.
The hardware side comes from an Epoch AI estimate: using satellite imagery and a cooling-equipment model, Huawei’s Horinger data center has about 105,000 Ascend 910C and 8,000 Ascend 910B chips in operation, roughly 81,000 H100-equivalents, about 7% of SpaceX Colossus 2, the largest AI data center in the world. This is an estimate, not a measurement, but it quantifies the combination of “open source is closing fast” and “a single cluster is still a fraction of one.”
Sources:
Eight: Coding agents’ weak spot, admitted officially: browser compatibility
The Chrome team declared Modern Web Guidance stable. Its starting point is concrete: ask an agent to build a tooltip or a modal and it usually pulls in JavaScript to rebuild what the browser already ships; ask it to use newer CSS or HTML and it often skips feature detection and fallbacks, leaving users on older browsers with a broken page.
The evaluation suite quantifies the gap: for features not yet available across all major browsers, unguided models pass checks for feature detection, @supports rules, fallbacks or conditional polyfills only 43% of the time; with Modern Web Guidance enabled that rises to 87%. Every guide carries Baseline status and version data — <dialog closedby>, for instance, is supported only by Chrome 134, Edge 134 and Firefox 141, not Safari, and the guide supplies the matching fallback. Fallbacks are grouped into four patterns: progressive enhancement with no fallback code, @supports detection, a small zero-dependency script, and a conditionally loaded vetted polyfill. Within 48 hours of release the project was the top trending repository on GitHub.
The point is not the score but the fix: agents often write bad code not because they lack capability but because they lack platform knowledge. Making compatibility data a structured input for agents is cheaper than letting them iterate. A comparable case landed the same day — PyTorch says that with AI coding agents, SGLang was integrated with IBM Spyre in under a week with a roughly 1,500-line plugin, delivering about 3.6x the throughput of the reference implementation, matching vLLM on Granite 3.3 8B and Gemma 4 26B A4B, with prefix sharing working out of the box.
Sources:
Nine: Agents got permission to act, and an approval layer with it
Adobe previewed MCP integration for its creative software at MAX, planned for early 2027: third-party agents will be able to make changes directly inside local Photoshop, Illustrator, Premiere and After Effects projects while preserving layers, vectors and edit sequences; Photoshop also showed relighting and enhanced removal that handles leftover shadows and reflections. Figma’s 3D transforms entered open beta the same day for all Full seat users, with non-destructive spatial transforms that keep the original editable.
Replit gave its agent two capabilities closer to business operations: buying supplies for a company through Stripe Link and Visa, and creating and running Google Ads inside Replit across Search and YouTube. The design emphasis is per-transaction approval, with the approval check enforced by code rather than by prompt. On OpenAI’s side, Businesses in ChatGPT began testing merchant identity claims, product information maintenance, engagement data, a Business Agent for eligible merchants to answer customer questions, and Shopify catalog sync. Claude opened whiteboards and sheets in beta to all paid plans, editable by users, colleagues and the model, with model edits marked.
Read together, the common thread is not “stronger agents” but permission boundaries written into the product: per-transaction approval, preserved editable structure, marked model changes. Whether agents reach production depends on whether those constraints are auditable; the material also offers no evidence that payment can be completed inside ChatGPT, so that step remains open.
Sources:
- https://replit.com/blog/replit-agent-ads-commerce
- https://www.figma.com/blog/how-3d-transforms-and-figmas-agent-brought-a-launch-video-to-life/
High-value briefs
- Google smart glasses open for pre-order: orders at glasses.google, shipping December 10, with try-ons at select Warby Parker, Gentle Monster and Google stores, Gemini built in.
- World Labs opens Atlas to everyone: an app and API that generate keyframes, video clips or 3D Gaussian splats.
- Baidu launches “Token Exchange”: Li Yanhong announced it at Baidu World, aggregating Baidu’s own models and Chinese open models (text, image, audio and video) with “trustworthiness” and success-rate data as the pitch; one on-site observer saw little difference from an ordinary API reseller, an unverified comment.
- Claude tightens account checks: a community summary describes two verification paths, Persona (third-party document and selfie check when subscribing) and Portal (Anthropic’s own portal, triggered mid-use), with reports that accounts can still be banned after passing Persona; the same summary says OpenAI began a wave of bans on November 7 and that from November 20 third-party platforms need official approval to offer GPT. Social-media accounts only, not official announcements.
- OpenAI third-quarter revenue reported at $10 billion: up nearly 50% quarter over quarter, enterprise revenue doubling and splitting roughly half with consumer, paid subscribers above 100 million; the material also cites an annualized figure near $50 billion without stating its basis, so it is not quarterly revenue times four.
- OpenAI ships two cybersecurity products: Daybreak Defense Factory and Daybreak Red Team, announced by Greg Brockman, positioned as putting frontier cyber capability in defenders’ hands.
- Service degradation under prefix caching: research finds short bursts of new traffic can push vLLM and SGLang into sustained degradation, halving throughput in some runs with backlog lasting the remaining 70 minutes of the experiment; bounding active sessions’ KV residency by cache pool capacity prevented the failure in every tested run, while limiting new requests alone was weaker.
- Single-GPU proofs for integer model inference: Sable produced a proof for Qwen2.5-32B generating 4096 tokens in 92.5 seconds on one RTX PRO 6000 Blackwell; it proves a well-defined integer computation, not bit-exact agreement with the original floating-point model, and it is not zero-knowledge.
- Image and genome models: MiniMax released H3-Image weights, claiming 4096×4096 images in 10 seconds without listing test hardware; DeepMind released AlphaGenome 2, predicting how DNA sequences and variants affect hundreds of human cell types — a research tool, not a clinical conclusion.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 07:00 | Premarket falls about 11% after the Nvidia announcement | The first board chair was named at the same time; the transition mechanics are more specific than the resignation |
| 08:00 | Someone had Claude Code pull 8-K filings and customer data from SEC EDGAR to assemble the Megaspeed trail | Public filings are searchable enough that investigations no longer belong only to newsrooms |
| 08:00 | A social-media breakdown of the share structure: about 3.58%, one share one vote | Explains how the board could remove him by majority vote |
| 09:00 | Three House committees ask for a closed-door briefing on the 18th | The story moves from corporate governance into a regulatory process |
| 11:00 | Artificial Analysis scores Hydrogen at 54 on xhigh | The mid-tier price-performance claim gets quantified by a third party |
| 13:00 | GLM-5.3 coding post-training gains 21.3pp on Terminal-Bench4 | The same base model still yields large gains from post-training |
| 15:00 | Modern Web Guidance tops GitHub trending within 48 hours | Packaging platform knowledge for agents meets real demand |
| 17:00 | Simon Willison points out Gemini 4’s confusing pricing | Three tiers at similar prices, Hydrogen only 25% cheaper, doubling on December 1 |
| 19:00 | A roundup notes Anthropic at ANT with a $70–78 range, and Gemini Robotics 2.5 fitting parts on a Mercedes line at 96% (proof of concept) | A single social-media roundup; usable as a lead, conclusions await primary material |
Editorial conclusion
The real divide today was not any single story but two things happening at once: the most valuable company in the sector admitted it had no succession plan, and the most aggressive forecasters admitted they had been too conservative. The first shows that governance debt gets settled at the worst possible moment; the second shows the capability curve is still outside prediction. The model-level news — a mid-tier model taking the workhorse role, a 1.6T open release, coding agents patching their platform knowledge — remains within the range of what can be evaluated. What cannot is whose preparation is real when the uncertainty sits in organizations and forecasts themselves.
Sources and method
Reviewed 20 hourly captures and 5 named sources with substantive content (AI HOT morning, HubToday, AI Valley, the Chrome developers blog, the OpenAI blog); the signal pool is rich. The Chrome RSS has been stalled since October 7, so its items were cross-checked between the listing page and the sitemap; claude-blog, cline-blog, google-research and xiaohu-ai were stubs under 500 bytes and were not used. Company benchmarks and social-media relays are attributed as such and were not independently reproduced.
