Huang's First Week Out: A Petition, 260,000 GPUs, and Anthropic's Compute Dilemma
The first full week after Jensen Huang's departure saw Nvidia's vacuum filled from several directions at once: a petition demanding his reinstatement passed 115,000 signatures,…
The first full week after Jensen Huang’s departure saw Nvidia’s vacuum filled from several directions at once: a petition demanding his reinstatement passed 115,000 signatures, South Korea said Nvidia had reaffirmed a supply of 260,000 advanced GPUs, and the governments of Taiwan and Japan settled on language of watchful waiting. On the same day, Anthropic conceded ahead of a record IPO that compute is its “deepest business dilemma” and put the scale of its commitments in front of investors.
A second thread ran through products and engineering: Sierra turned how personal agents connect to enterprise services into a public protocol, Cloudflare and the FBI both named AI-orchestrated penetration platforms, and Shopify published how it lets offensive agents attack production under supervision. On the research side, several results pointed in the opposite direction from two popular assumptions — that multi-agent delegation helps, and that memory should be retrieved by semantic similarity.
A note on evidence first: this capture contains no Nvidia corporate announcement, no succession plan, and no investigative finding. Every detail of the management upheaval comes from media accounts. Anthropic’s financial figures are reporting, not filings. Some research conclusions exist only as secondhand summaries.
Theme 1: In the first week out, nobody else can sign for Jensen Huang
Huang resigned as Nvidia’s chief executive on November 12, 2026, effective immediately, and left the board as well. According to The Verge and AI Valley, the acknowledged and apologized-for issue was extramarital relationships; the harder part is a national security question involving Megaspeed, a Singapore cloud company and Nvidia partner that has been investigated over export-restricted chips reaching China.
The petition became the week’s most concentrated public statement. As of Saturday afternoon it had more than 115,000 signatures, per TechCrunch: 5,965 current employees, 1,108 former employees, and 63,063 shareholders, with Elon Musk, Replit’s Amjad Masad, and Epic’s Tim Sweeney among the signatories. It asks for three things: reinstatement until the facts are settled, a fair and independent investigation, and transparency about how the decision was made.
Asian governments were more measured than Silicon Valley. South Korea’s science ministry said Nvidia had reaffirmed it will supply 260,000 advanced GPUs to the country and continue joint AI projects; Taiwan’s economics minister Kung Ming-hsin said the matter is unlikely to significantly change global AI industry growth; Japan’s industry ministry said it is watching closely. What they share is that none of them is treating the event as a supply chain interruption.
The Verge’s framing is longer-range: Huang founded what became the first $5 trillion company, roughly 6% of the S&P 500, and the pivotal bet he drove was the shift from graphics cards to CUDA. The piece argues that the era of the founder-CEO who defines strategy through personal performance is exiting alongside him. That is commentary, not a verifiable conclusion.
The harder thing to quantify is the ecosystem effect. Nathan Lambert argues that Nvidia’s funding for open models will likely dry up with the founder’s departure, because the company’s first priority now is stability and shipping to customers. He splits the problem into three layers: Nvidia’s internal support mechanisms, the regulatory risk that follows, and an American absence at that layer.
AI Valley put it more briefly: no trouble on Tuesday, much less certainty about the next twelve months. That holds because Nvidia’s orders, capacity, and customer relationships are already contracted, while whether the next CEO sustains the same pace of bets is an open question.
The market’s odds are low. Polymarket data from early on November 14 showed roughly 31% of participants expecting Huang back at Nvidia before the end of the first quarter next year. A separate claim circulating on social media — that Nvidia had engineered scarcity by funding AI startups to buy its GPUs — rests on a single media report and is an allegation, not a finding.
The boundary to keep: a petition is employees and shareholders speaking, not a governance process. Whether the board has begun a succession process, and how interim management is composed, is not covered by this material.
Sources:
- https://techcrunch.com/2026/11/14/elon-musk-other-tech-execs-sign-petition-calling-on-nvidia-to-reinstate-jensen-huang/
- https://techcrunch.com/2026/11/14/asian-governments-react-to-nvidia-ceo-jensen-huangs-exit/
- https://www.interconnects.ai/p/open-source-ai-without-jensen
Theme 2: Anthropic calls “how much compute to buy” its deepest business dilemma
Speaking at a JPMorgan investor event, Dario Amodei conceded that “there’s a huge uncertainty in what’s going to happen next year.” The mechanism he described is procurement lead time: compute has to be locked in one to two years ahead, and buying too little constrains customers and revenue while buying too much overextends the balance sheet. Reports say he also outlined an approach built on gross margins and a cash buffer, without giving complete figures.
The scale cited by The Decoder is that Anthropic’s total compute commitments approach $901 billion. HubToday’s summary flags an easy misreading: a commitment is not money already spent, nor is it current debt. It is a purchasing framework stretching over years.
Disclosing that number before a record IPO is itself a pricing decision. The reporting reads the listing as a buffer between revenue growth and a long-term cost mismatch — in other words, the company needs public-market money to pay for compute that has not yet produced revenue.
The risk in this kind of disclosure is an opaque maturity structure. Public reporting does not say how the commitments are spread across years, whether there are minimum purchase volumes, or whether they can be scaled back with demand. Treating them as a fixed cost curve overstates near-term pressure and understates long-term flexibility.
Sources:
Theme 3: Same model, 34 percentage points apart between API and web
Anthropic published a safety evaluation methodology whose central finding is that scores depend heavily on deployment configuration. On a multi-turn suicide and self-harm evaluation, Fable 5.1 scored 60% (±14%) when tested through an API configuration and 94% (±7%) on the web product using a new system prompt. One model, a 34-point gap.
The difference comes not from the weights but from the surrounding conditions — system prompt, interface, and scoring rules. The team therefore argues for measuring the rate of appropriate responses per scenario and reporting all three configuration items; otherwise cross-product comparisons mean little.
Two operational recommendations follow: use at least 100 cases for complex multi-turn evaluations and at least 1,000 for simple single-turn ones. With fewer samples, an interval like ±14% swallows any improvement.
The practical value for readers is that it demotes statements like “this model is safer” to “this model is safer under this configuration.” A vendor’s published score, a third party’s score, and the score you get inside your own product were never the same measurement.
Source:
Theme 4: Sierra’s Poppy 1.0 gives personal agents an enterprise entry file
Sierra co-founders Bret Taylor and Clay Bavor released Poppy 1.0, a personal agent protocol. An enterprise publishes a discovery file under its own domain announcing its website, API, or enterprise agent entry points, along with identity descriptions and login authorization. A personal agent reads that file and knows how to connect.
The deployment matters more than the protocol. According to Sierra, stores on the Shopify platform have already published Poppy files automatically, and Meta Muse has enabled the protocol. Commerce connects through UCP, which carries the actual shopping flow; Poppy handles only discovery and authorization.
The split is deliberate: standardize who may connect and as what identity, and leave how a transaction completes to an existing protocol. For enterprises, integration cost moves from writing a private connector to publishing a file.
The limits are equally clear. This is the protocol publisher’s own account; real coverage among partners and the authorization experience have not been independently verified. Automatically publishing files for platform merchants also raises the question of whether those merchants know.
Source:
Theme 5: Attack and defense are both using the same agent capabilities
A Cloudflare blog post says that over three days it observed more than 50 framework families attempting to scan its global customers, describing a trend of LLM-driven tools chained into scanning followed by follow-on access. Its defensive argument is not complicated: identity isolation, permission constraints, and microsegmentation to shrink the blast radius after a breach. Note that the excerpt supports the “more than 50 framework families” figure but does not show the “30x request growth” statistic mentioned in the summary.
One platform named is ARTEX, an AI-orchestrated penetration testing framework. Social media accounts say the FBI has issued an advisory warning organizations about it, and that a university student previously used it to breach a Korean bank’s systems. That portion rests on a single social media retelling; neither the institutions nor the incident details are independently confirmed.
On the defensive side a reusable design has appeared. Shopify’s public Cyber Chaos Monkey gives an offensive agent source code, infrastructure knowledge, and temporary credentials for a test store, then has it validate vulnerabilities in production while a second, independent judge agent monitors behavior and terminates the task on anomalies. The company says it has run tens of thousands of tasks and found dozens of issues with real impact, with no downtime or incidents to date. Those results are the company’s own account.
A hardware example shows the risk is not confined to the web. According to a disclosure retelling, GCSA says it obtained the highest OS privilege on a robot from a leading vendor; the vendor confirmed the issue by email on November 13 and began remediation and scoping of affected products. The brand and the mechanism have not been made public, and starting a fix does not mean the risk is gone.
Read together, the common thread is not that AI is more dangerous, but that both sides use the same capabilities: reading code, calling tools, executing continuously. The defensive assumption therefore shifts from holding a boundary to assuming the boundary has already been crossed, and limiting consequences through isolation and review.
Sources:
Theme 6: Three routes to the same capability at lower cost — training precision, kernel scheduling, context compression
Thinking Machines Lab released Inkling 2 Small under Apache 2.0: 594 billion parameters, with 4-bit forward compute from pretraining through final reinforcement learning. The material claims an 8% reduction in pretraining time and 20% in RL time, though backward computation was not kept at 4 bits.
The deployment floor needs stating: running NVFP4 requires at least 500GB of VRAM, and MXFP8 at least 800GB. “Small” in the name does not mean a consumer single-GPU deployment. Sebastian Raschka’s observation is that it uses attention residuals, similar to Kimi K3’s approach, but without the linear-attention hybrid.
The fifth version of the CUDA-L2 paper takes another route: an LLM plus reinforcement learning, with CUDA execution speed as the reward, automatically optimizing 1,000 half-precision matrix multiplication configurations. Against cuBLASLt’s autotuned baseline it reports an average 11.4% gain offline and 15.9% in an online mode that simulates real-time inference intervals.
Stanford’s Memorilla compresses context instead. It encodes a document collection into a fixed 16 memory tokens fed to a frozen model, reaching 84.33% on FactKG with a frozen Qwen3-8B against 80.43% for a full-context baseline, with roughly 108x fewer document tokens. It still trails the baseline on benchmarks such as PubMedQA; code, training recipe, and checkpoints are released under Apache 2.0.
All three move cost away from “a bigger model” toward “a cheaper way to execute”: lower precision in training, better kernel scheduling in inference, compressed representations for context. But every figure is bounded by a specific task and configuration, and CUDA-L2’s gains especially should not be read as a general speedup ratio.
Source:
Theme 7: Three counterexamples to the default assumptions about multi-agent systems and memory retrieval
The first concerns delegation. The Lost in Delegation study splits tasks by whether they fall before or after a model’s training cutoff and compares delegation schemes. Its finding is that delegation is not consistently better than not delegating; when the orchestrator lacks post-cutoff knowledge, accuracy can fall markedly, and the same model using tools directly does better. Common failures include distorted queries, inefficient tool use, and information lost between agents.
The second concerns “expert correction.” A two-stage revision study fixes the question, the evidence, and the initial answer, then compares neutral reflection, anonymous disagreement, and domain-expert disagreement. Across 1,989 medical, scientific, and legal judgments, none of the five models reached the selective-correction threshold needed for a net accuracy gain under expert disagreement, with accuracy falling by as much as 16.1 percentage points. Adding an expert label alone moved accuracy between a 5.2-point drop and a 2.4-point gain.
The third concerns memory. MemCredit starts from the observation that most methods pick the memory most similar to the current state, but similarity is not usefulness. It measures each memory’s effect on the action through controlled interventions, then trains a small estimator to decide which memories fit the token budget. With a 500-token memory budget on ALFWorld, success reaches 86.57% against 65.67% for BM25 retrieval, using 152.1 memory tokens per decision — 74.5% fewer — with the agent policy itself unchanged.
The three point the same way: more agents, more memory, and more context are not cost-free correct answers. What determines the result is whether you can obtain a reliable feedback signal, and whether you are willing to pay only for information that changes the decision.
The first two come from paper summaries and the third from the author’s own social media summary of a paper. None has been independently reproduced.
Source:
Theme 8: Model selection shifts from “strongest” to “cost per task”
Anthropic made Opus 5.5 the default model for its Pro, Max, Ultra, and Team plans, emphasizing research, analysis, writing, and document work. The arithmetic in the official practical guide is worth noting: Sonnet 5.5 at a high thinking level already costs nearly as much per task as Opus 5.5, while Opus 5.5 is more likely to deliver a finished result in one pass; the pricier Fable 5.2 is reserved for multi-agent projects that run for hours.
The guide’s test figures are the company’s own account: 16 of 18 passing on a zero-tolerance financial fabrication test, and M&A Excel modeling plus an executive deck done in 63 minutes at half the previous generation’s cost. The usage advice is concrete: ask for the finished file in the first message, delete “think hard” from the prompt, stay at the default Medium level, and upload chart images directly.
The companion “Building always-on agents” guide is about architecture rather than models: an event inbox for external triggers, timers for proactive behavior, memory for continuity. The real value is not in the parts — queues plus scheduled jobs could assemble these long ago — but in assigning clear responsibilities to each component and stating that always-on is an experience, not a model left idling.
The throughput numbers push the same way. The Grok Bot team went from first line of code to launch in seven weeks; engineer Lauren Tan said in an interview that she merges about 4,000 pull requests a month, and xiaohu.ai’s write-up puts it at 11,000 PRs a month with roughly 0.5% reverted. The more important detail in the same interview is the boundary: multiple personal bots under one account share a single cloud computer, with login state and files in common, so “independent identity” is currently implemented in layers.
A fixed pattern for mixed delegation has already appeared in individual practice. One developer describes having Claude plan and build the frontend, handing chunk-sized backend and data work to a local Codex, having Codex run at most three review rounds, and restricting the dispatch script to the working tree and temporary directories — installing dependencies, network tests, and commits stay with the human.
These numbers come from interviews and vendor guides, with no independent audit. But they explain why the selection conversation is shifting: from which model is strongest to which model delivers in one pass on this task, and at what price.
Source:
High-value briefs
- OpenAI seeks $15 billion: a report relayed by AI Base puts the round at $15 billion at a $1.25 trillion pre-money valuation, expected to close within weeks; it also cites roughly $10 billion in third-quarter revenue against an operating loss of about $10.9 billion, and ties the raise to IPO preparations expected in 2027. Size, valuation, and timing should be read as reporting.
- Four nonprofits ask a court to halt frontier model development: the suit was filed in San Francisco Superior Court on November 12, 2026, naming OpenAI, Anthropic, Google, Meta, and SpaceXAI, and seeking an emergency injunction freezing next-generation training and deployment. Filing a complaint is not a court ruling.
- Ernie 5.5 Flash preview, hands-on: a tester reports decode speed of 267 tokens per second, spatial understanding and modeling that stands out among Flash-class models, and long-context recall tested to roughly 256K, suited to fast prototyping and data cleaning; agentic coding and agent capabilities were still being tested.
- Early Gemini 4 Hydrogen results: an outside summary says it performs more strongly on a find-and-fix test across two real repositories and 105 planted bugs, but no complete evaluation methodology was published. A Google weekly list circulating the same day names two eighth-generation TPUs, Gemini 4 Carbon, Gemini Robotics 2.5, AlphaGenome 2, and smart glasses preorders, with no accompanying benchmark data.
- Gödel’s notebooks read with AI assistance: a social media account says Claude processed 1,009 pages in two days and decoded 663 potential theorems, 400 of which passed formal verification; the project team had previously advanced at about one page a week for nine years. Single secondhand source, no formal paper seen.
- Automatically generated evaluations still need auditing: a Google-related discussion says 29 agents in 11.5 hours turned 13 public data tables into 50 causal inference tasks; independent reviewers found wording cues and template shortcuts, and the team converted those findings into follow-up tests.
- How risk attention is measured: 87% of surveyed market participants cited AI risk in a New York Fed survey, up from 50% in the spring; a separate retelling puts AI infrastructure financing debt above $680 billion as of October. Survey opinion, financing definitions, and official risk assessments should be read separately.
- Open-source coding agents competing for the same ground: GitHub trending shows Univer adding 495 stars in a day, Qwen Code offering automatic memory, skills, sub-agents, and MCP, and Tencent open-sourcing LoopForge, which uses Docker sandboxes to split a request across planning, implementation, review, and testing roles before opening a PR. Capabilities are self-described; no side-by-side testing was done.
- Browser capabilities packaged as a skill: the Chrome team turned modern browser capabilities, compatibility, and recommended patterns into a skill that coding agents can consult — native dialog for modals, Popover for overlays — so models stop falling back to old approaches.
- Courses going public: lectures from Stanford’s CS329Z “Engineering AI Agents” and CS329A “Self-Improving AI Agents” are on YouTube, including a comparison of DSPy and LangGraph.
- Secure coding enters the classroom: Semgrep’s co-founder and CEO gave a 47-page lecture at Stanford CS146S titled “Secure coding in an agent-first world,” built around making exploitation more expensive; the speaker’s view is that high-profile exploits will get worse before they get better.
- The “Gemini is now an agent” claim: in a Gemini at Work recap, Google Cloud made “Gemini is now an agent” the announced theme, alongside scenarios such as connecting Google Wallet for personal finance insights, without describing the boundaries of that agent capability or how it handles failure.
- Two local tools: ClamDown keeps a MacBook awake with the lid closed and manages the fan curve for long agent runs; Yogurt does meeting recording, transcription, and summarization entirely on the local machine across 99 languages.
- Starship recovery: SpaceX caught its returning ship with the launch tower’s arms for the first time, and many AI practitioners passed it along. A separate claim that the flight carried an Optimus robot that returned with it has a single social media source and is unconfirmed.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 02:00 | 31% on Polymarket expect Huang back before the end of Q1; a Google model called Krypton circulates as a Fable/Astra peer | Market odds and a product roster both show the story still moving |
| 04:00 | A developer relays Anthropic’s official guidance: Opus 5.5 as the daily driver, Fable 5.2 only for hours-long multi-agent work | A vendor starts tiering models by cost rather than capability |
| 07:00 | TechPowerUp reports Nvidia allegedly funded AI startups to buy its GPUs, engineering scarcity | A single-media allegation that amplifies once layered onto the leadership story |
| 08:00 | Cloudflare names ARTEX-class AI-orchestrated penetration platforms; social accounts say the FBI issued an advisory | Attack tooling moving from experiment to institutional threat lists |
| 09:00 | GitHub daily chart: Univer +495 stars, Qwen Code and Tencent’s LoopForge on the same list | Office foundations, terminal agents, and delivery pipelines all agentified at once |
| 11:00 | Grok Bot interview: seven weeks to launch, ~4,000 PRs merged monthly, multiple bots sharing one cloud computer | Throughput figures and identity-isolation limits in the same conversation |
| 12:00 | A founder argues SaaS should become an MCP server and hand inference cost back to customers | A reverse answer to the margin problem of AI wrappers |
| 16:00 | Claude assists in decoding Gödel’s notebooks; 400 theorems pass formal verification | A case where math’s verifiability produces acceleration |
| 18:00 | A researcher’s OpenClaw review: the web UI is better than expected and the mobile app is usable | An open-source agent client catching up with commercial products on experience |
Editorial conclusion
The real story of the day was not a model release but the way the same set of facts was pushed in different directions by different parties: employees and shareholders tried to change a governance decision with a petition, Asian governments held to supply commitments to preserve certainty, Anthropic priced compute risk before listing, and the security and engineering side accepted a plainer premise — a model’s score, an agent’s output, and a memory’s hit rate all depend on how they are deployed and verified, not on their names. The argument around Huang will continue, but the verifiable part currently stops at who signed, who issued a statement, and who reaffirmed an order.
Sources and method
Reviewed four content-bearing named sources and 20 hourly captures in the 2026-11-14 (PT) archive, about 195KB in total; the remaining named sources had no new posts that day. The signal pool is rich, covering eight main themes and 12 briefs. Main limits: no corporate announcement on the Nvidia governance event, Anthropic’s finances are reporting, and some research conclusions are secondhand.
