---
title: "Nvidia board probes Jensen Huang; Anthropic buys $3.4B in gas generators"
date: 2026-11-09
summary: "Nvidia's board opened an independent probe into Jensen Huang and a reported foreign intelligence effort. Anthropic took over about $3.4B of gas generators, and SpaceX filed a $24B Starbase gas plant plus super-compute plan."
cover: "https://morningtokens.com/covers/2026-11-09.jpg"
url: "https://morningtokens.com/issues/2026-11-09/"
---

Pour a slow cup for this one: the biggest AI story of the weekend isn't a model at all. Nvidia's board has opened an independent investigation into whether Jensen Huang was the target of a Chinese intelligence effort, nine days before earnings.

Around it, the week's real theme keeps hardening: labs are buying their own power plants (Anthropic's $3.4B in gas generators, SpaceX's $24B Starbase plant), while governments split between "pause" and "build faster".

## Top news

### 1. Chips · Nvidia's board opens an independent probe into Jensen Huang and a foreign intelligence effort

On Sunday 8 November Nvidia's board said a committee of independent directors, with outside counsel, will investigate reporting that founder and CEO Jensen Huang "has been the target of a foreign intelligence effort". In an email to staff the same day, Huang admitted he had been unfaithful to his wife and said he supports the review.

- The board says it has "no reason to believe that either national security or NVIDIA secrets were compromised", but the inquiry "will examine all available information".
  - It also says it "has confidence in Jensen and his leadership team".
- The trigger is a New York Times report from late October that U.S. intelligence concluded a Singapore-based data-company executive, Alice Huang of Megaspeed, was gathering information for Beijing and had a "close relationship" with Huang (ABC, AP).
  - China calls the allegations "baseless".
- Huang's email: "I was unfaithful to my wife… I take full responsibility."
  - On the security question he writes that he supports the investigation and "will cooperate fully".
- Timing: Nvidia reports Q3 results on 17 November, and House committee chairs have asked Huang for a closed-door briefing on 18 November.
  - Trump told reporters Huang is "an honorable guy".

**Why it matters:** Nvidia sits at the centre of U.S. export-control policy and of every lab's compute plan. A governance question at the top of that company now overlaps with an earnings print and a congressional briefing in the same fortnight, and with the export-licence story in item 4.

Sources: [NVIDIA board statement](https://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors) · [CNBC: Huang's email and the board statement](https://www.cnbc.com/2026/11/08/nvidia-jensen-huang-email-board-of-directors-statement.html) · [AP](https://apnews.com/article/nvidia-jensen-huang-investigation-chinese-spy-68839e7ca6ba2b121529f7e475e501d6) · [The Guardian](https://www.theguardian.com/us-news/2026/nov/08/nvidia-jensen-huang-investigation)

### 2. Power · Anthropic takes over $3.4B of gas generators, 1.7 GW, from a Texas supplier

The Financial Times reported on Sunday that Anthropic agreed in late September to pay about $3.4B for commitments covering 1.7 GW of heavy-duty gas-fired generators. It is described as the largest known single purchase of power-generation equipment by an AI lab.

- The seller is OnePWR Solutions of San Antonio, which had ordered Caterpillar generators about two years ago but needed financing to complete the purchase (Investing.com, citing the FT).
- Anthropic first planned to be a OnePWR customer, then took over the equipment orders instead, plus the engineering and construction capability to install them.
  - It did not buy OnePWR and no staff moved over.
- A Piper Sandler deal sheet lists a $3.4B sale of OnePWR's power-generation platform to an unnamed AI lab in September (TheEnergyMag).
  - Where the generators will go is not public, and neither company commented.

**Why it matters:** This lands in the same week as the post-midterm data-centre backlash and Anthropic's IPO push. Owning generation gear is a way around multi-year grid interconnection queues, and it also puts gas turbines on Anthropic's balance sheet right when voters are punishing data-centre projects.

Sources: [CNBC (FT report)](https://www.cnbc.com/2026/11/08/anthropic-generator-deal-ai-data-centers.html) · [Investing.com](https://www.investing.com/news/stock-market-news/anthropic-reportedly-strikes-34-billion-deal-for-gasfired-power-at-data-centres-5010208) · [TheEnergyMag](https://www.theenergymag.com/news/market-news/anthropic-gas-generation-acquisition-ai-data-center)

### 3. Compute · SpaceX files a $24B "gas-fired power plant and super-compute facility" at Starbase

A one-page notice posted on 6 November by the City of Starbase, Texas, says SpaceX proposes to "construct, own, and operate a gas-fired power plant and super-compute facility" on about 444 acres east of Brownsville, with initial phases estimated at about $24B. The city will consider tax breaks at a public hearing on 9 December.

- The commission will weigh a Chapter 312 property-tax abatement and a Chapter 380 economic-development agreement. Nothing has been approved (The 956 Report).
- The notice gives no megawatts, chip counts or dates.
  - A separate pending air filing, reported by The 956 Report on 3 November, lists 35 temporary gas turbines (about 748 MW) in a first phase, later replaced by eight larger GE Vernova 7E units.
  - Proposed NOx cap: 248.86 tons a year, just under the 250-ton major-source review line.
- A SpaceX job posting says "Starbase is standing up on-site power generation for the Starbase AI supercomputer" (FourWeekMBA).
- Context: incoming Texas governor Gina Hinojosa has pledged to halt new data-centre approvals, so the state may become a gatekeeper.

**Why it matters:** With item 2, this is the pattern of the month: AI compute is now sited where you can burn gas on-site, and the emissions and tax details are where the fights will happen. The 9 December hearing is the first public test.

Sources: [The 956 Report](https://www.the956report.com/post/spacex-wants-a-tax-break-for-a-24-billion-gas-plant-and-super-compute-facility-in-starbase) · [FourWeekMBA](https://fourweekmba.com/ai-starbase-notice-puts-24b-on-spacex-gas-plant-supercomputer/)

### 4. Export controls · Two Chinese firms cleared to buy Nvidia H200s have empty offices

A New York Times investigation, written up by Tom's Hardware on 8 November, found that the listed addresses of two Chinese companies approved to buy Nvidia H200 chips, New Matrix and Maginfra, led to empty offices. Their executives are often linked to firms the U.S. has sanctioned for supplying China's military.

- The approved list also includes well-known names such as Baidu, BYD, Tesla Shanghai, Alibaba, Tencent, ByteDance and ZTE.
- At both offices, receptionists and people working nearby said they had never heard of the companies or seen anyone use the space.
- Former NSC China director Liza Tobin: "I have a ton of concerns that neither the Commerce Department or Nvidia is doing the due diligence needed to ensure that the licensing terms are being followed."

**Why it matters:** The licence regime is the policy compromise that lets H200s flow to China. If approvals can go to shell addresses, expect pressure for end-use checks, and it adds to the questions Nvidia faces before the 18 November briefing.

Sources: [Tom's Hardware (8 Nov)](https://www.tomshardware.com/tech-industry/artificial-intelligence/two-chinese-companies-cleared-to-buy-banned-nvidia-ai-chips-have-empty-offices-and-unknown-addresses-investigation-reveals-businesses-have-ties-to-sanctioned-chinese-military-suppliers) · [The Verge](https://www.theverge.com/ai-artificial-intelligence/1029483/nvidia-ai-chips-china-blacklist-government)

### 5. Policy · The Justice Department sets up a "Super Intelligence Task Force"

On 5 November the DOJ announced the Super Intelligence Task Force (DOJ-SITF), led by Associate Attorney General Stanley Woodward and Associate Deputy Attorney General Stephen Weiss. It will coordinate cases about the misuse of advanced AI by foreign state actors, criminal organisations and fraudsters.

- Attorney General Todd Blanche said the unit will work with "SI Force Chair Jay Clayton", the new White House body.
- Weiss said the department will use law-enforcement tools to "deter improper use of SI, harden regular reporting on SI risk and response, and ensure companies exercise vigorous internal controls so that SI technology behaves as intended".
- It absorbs the department's existing AI task force, and a published memorandum sets out how cases are coordinated.
  - Mind and Machine Weekly reports prosecutors may need to notify the unit before bringing AI-related charges.

**Why it matters:** The phrase "ensure companies exercise vigorous internal controls" is the interesting part: it points toward compliance expectations for AI developers and deployers, not just prosecuting misuse.

Sources: [DOJ press release (5 Nov)](https://www.justice.gov/opa/pr/department-justice-launches-super-intelligence-task-force) · [Nextgov/FCW](https://www.nextgov.com/artificial-intelligence/2026/11/department-justice-forms-ai-task-force/417074/)

### 6. Infra · Google takes over Crusoe's troubled Wyoming data-centre site and plans $17B there

Crusoe confirmed on 6 November that it sold its unfinished southeast Wyoming campus to Google, which plans to invest $17B developing it. The purchase price was not disclosed.

- Crusoe had pitched the site to Google as a customer and failed; Google had raised concerns about the project's cost and timeline (TechCrunch, Bloomberg).
- The project was announced with Tallgrass Energy with ambitions to scale to 10 GW.
  - Crusoe will still build another Google campus near Amarillo, Texas.
- The $17B is Google's planned build-out spend, not the price paid.

**Why it matters:** Hyperscalers are absorbing projects that independent developers can't deliver on time. For builders, the practical signal is consolidation: fewer, bigger owners of AI capacity.

Sources: [Crusoe newsroom](https://www.crusoe.ai/resources/newsroom/crusoe-wyoming-data-center-campus-update) · [TechCrunch (6 Nov)](https://techcrunch.com/2026/11/06/crusoe-sells-troubled-wyoming-data-center-project-to-google/)

### 7. Security · Palo Alto Networks is in talks to buy agent-identity startup Clutch for about $500M

Palo Alto Networks is negotiating to buy Clutch Security, an Israeli startup that secures the credentials and access AI agents and other non-human identities use, for about $500M in cash plus milestone payments (The Information via Reuters, TechCrunch). Neither company has confirmed it.

- Clutch was founded in late 2023, has raised about $28.5M and was last valued at $85M in January 2025.
  - It is said to be approaching $10M in annual recurring revenue (Kill The AI summary of reports).
- That is a roughly 50x revenue multiple for the category, if the deal closes.

**Why it matters:** Yesterday's lead (Claude Code permission files leaking secrets) showed agents are now a secrets-handling surface. The big security vendors are paying up to own "which agent has which key", and this will shape what your agent platform has to integrate with.

Sources: [TechCrunch (5 Nov)](https://techcrunch.com/2026/11/05/palo-alto-networks-in-talks-to-buy-ai-agent-security-startup-clutch-for-500m/) · [Ctech](https://www.calcalistech.com/ctechnews/article/qx69w4bra)

### 8. Media · An anti-Waymo op-ed campaign ran on AI-written pieces, none disclosed

Semafor reported on 8 November that a campaign placing local-newspaper op-eds against Waymo (robotaxis cost jobs, endanger riders, block emergency vehicles) relied on substantially AI-written text. The detector Pangram scored six pieces at 68% to 95% AI-written, and none disclosed it.

- Authors and the PR firms involved said they reviewed, edited and approved the pieces.
- Semafor has tracked AI writing on opinion pages since August; this is a case of it being used for coordinated issue advocacy.

**Why it matters:** Astroturf used to be expensive. If a small PR shop can flood local opinion pages with fluent, signed, "human-approved" pieces, editors will need disclosure rules and detection as standard intake, and policy fights about AI itself will be argued with AI text.

Sources: [Semafor (8 Nov)](https://www.semafor.com/article/11/08/2026/anti-waymo-campaign-relied-on-op-eds-written-with-ai)

### 9. Policy · The UK says it won't slow AI data-centre construction

UK Growth Minister Dan Tomlinson said Britain has no plans to slow AI data-centre construction despite the U.S. backlash, treating the build-out as essential infrastructure, with a Compute Roadmap target of at least 6 GW of AI-capable capacity by 2030 (Engadget).

- Projects seeking government support must invest in local communities and create local jobs.
- It contrasts directly with the U.S. post-midterm mood: Bernie Sanders and newly elected Democrats renewed calls for a national data-centre moratorium.
  - Rep. Yassamin Ansari said a pause on advanced AI should be the new majority's first priority.

**Why it matters:** Capacity may start to follow politics. If U.S. siting gets harder, the UK is openly positioning itself as the place that says yes, with community-benefit strings attached.

Sources: [Engadget](https://www.engadget.com/2310304/uk-has-no-plans-to-slow-down-ai-data-center-construction/) · [IBTimes](https://www.ibtimes.com/uk-expand-data-centers-ai-opposition-3809536)

### 10. Weird & wonderful · A homelabber ran 2.8-trillion-parameter Kimi K3 on a CPU with Optane memory

Tom's Hardware reports a homelabber ran Moonshot's 2.8T-parameter Kimi K3 with no GPU, on a 40-core third-gen Xeon with 2 TB of Intel Optane PMem 200, at about 0.85 tokens/second with 1-bit quantisation and 0.75 tokens/second at 3-bit.

- Optane modules run so hot they can "burn out" neighbouring DRAM, so the build uses blower fans for active DIMM cooling.
- Optane is discontinued, which is exactly why used modules are cheap per terabyte.

**Why it matters:** Under one token a second isn't usable for chat, but it shows that "frontier-sized open weights at home" is a memory-capacity problem more than a GPU problem. Useful if you need offline, batch-style runs of a giant open model.

Sources: [Tom's Hardware](https://www.tomshardware.com/tech-industry/artificial-intelligence/homelabber-uses-second-gen-optane-memory-sticks-to-run-a-2-8-trillion-parameter-ai-model-on-cpu-alone-warns-that-optanes-extreme-heat-can-burn-out-nearby-memory-dram-used-blower-fans-to-prevent-melting-ram)

## GitHub repos

- **[guillaumemeyer/watermarks-remover](https://github.com/guillaumemeyer/watermarks-remover)** (~26.8k★, +860 this week, MIT, Python) — a privacy-first app that strips AI watermarks from content you own.
  - Timely now that xAI and Anthropic watermark text by default.
- **[dsx-ai-factory/nvke](https://github.com/dsx-ai-factory/nvke)** (Apache-2.0, Go, created August, +345 stars today) — "NVIDIA Kubernetes Engine", a management plane for GPU Kubernetes on bare metal.
- **[meta/astryx](https://github.com/meta/astryx)** (~17.4k★, +875 today, MIT, TypeScript) — Meta's open-source, "agent ready" design system, meant to be driven by coding agents as well as people.

## Research papers

- **The Last Mile to the Real World: Agent-Native Primitives** (2611.07461) — agents fail between reaching a resource and applying it; proposes reusable primitives that each carry a resident LLM. [HF](https://huggingface.co/papers/2611.07461)
  - Terminal-Bench 4.0 cited: GPT-5.6 Sol with Codex 37.3%, Claude Opus 5 with Claude Code 51.8%.
- **Beneath the Surface: What We Overlook About Agent Evolution** (2611.07624) — all sixteen self-evolved agents gained new tasks while losing ones they used to solve; most reliability gains also raised deployment cost. [HF](https://huggingface.co/papers/2611.07624)
- **Sketchpad: certified exact arithmetic with a small VLM** (2611.04434) — Qwen3.5-0.8B with a LoRA runs arithmetic on a rendered sheet through a three-cell view, with exactness proved in Lean. [HF](https://huggingface.co/papers/2611.04434)
- **SAE Interprets MoE** (2611.08089) — trains a sparse autoencoder on an MoE layer's output and shows experts specialise in fine-grained, human-interpretable features. [HF](https://huggingface.co/papers/2611.08089)
- **Prefill-to-Autoregressive Cross-Model Transfer** (2611.06976) — PLatour, an open framework that bridges a big model's prompt representations into a small model's generation. [HF](https://huggingface.co/papers/2611.06976)

## AI workflows

**1. [A family budget where the agent decides once and a rule handles it forever](https://x.com/DanyloVorvul/status/2119040306381222329)**
- **Replaces:** several hours a month of hand-categorising bank transactions and pulling a budget together, the "fourth step" every personal-finance book leaves manual.
  - Scripts poll five bank accounts daily, drop duplicates and load everything into Actual (open-source budgeting app). Rules classify anything seen before.
  - Anything new goes to an agent (Hermesa, built on the open-source Hermes Agent), which compares it with the allowed categories and labelled examples, checks payee history via Actual's CLI, and searches the web and a map when unsure.
  - If it still can't tell, it writes its findings to an Obsidian note and the payment lands in the weekly review.
  - Every confirmed answer becomes a rule: 100 hand-labelled transactions to start, 194 rules today (187 set a category across 21 categories).
  - A weekly agent checks the plan and messages on Telegram only when something is over budget or needs review; a monthly agent drafts next month's budget for a chat edit.
- **What to connect:** a budgeting app with rules and a CLI or API (Actual), bank sync or exports, an agent with web and map search, a notes store for unresolved cases, and a messaging channel for exceptions.
  - Cost: the software is free and open source; the model is either a self-hosted open model (Qwen on your own GPU server) or a paid hosted API.
  - Starter prompt in the post: pick 100 varied transactions for me to label, label the rest with my logic, list anything you're unsure about instead of guessing.
- **Who it's for:** households and sole traders who want categorised spending without a monthly bookkeeping session, and anyone building an agent that should get cheaper and more consistent over time.

**2. [Twelve single-job bots run one person's household, with a yes before anything real](https://x.com/aswincode/status/2118747986155208971)**
- **Replaces:** personal-admin juggling across insurance, paperwork, travel, money, content and a side app.
  - This week the Household Manager worked through home-insurance renewal options, drafted the questions to the insurer, read the replies and drafted the next round, while the owner made the call.
  - A Staff Engineer bot opened PRs fixing issues it found in PostHog analytics and Sentry crashes, working through Claude Code CLI on an always-on Mac (real repo, tests, iOS Simulator) and never merging itself.
  - Other bots: Chief of Staff; a bot-builder that creates every new bot from a plain-words request; an Editor that keeps outgoing drafts in one voice and learns from what was actually sent.
  - Also: CMO, Growth Lead, Researcher, Ideas Scout, Travel Agent (can't book without a yes), Accountant, Life Coach.
  - Bots hand requests to each other instead of guessing.
  - What broke: bots filled the reminders list with their own follow-ups (fixed with a rule that only the owner adds reminders), and overnight pings (new routines now weekday waking hours, at most hourly).
- **What to connect:** X, Gmail, Google Calendar, a money app, a notes/reminders app over MCP, GitHub, Vercel, Sentry, Cloudflare, RevenueCat, Convex, PostHog.
  - Secrets in the vault; a spare machine with Claude Code for the coding bot; shared rules (ask before buying, sending, posting, booking or deleting).
  - Start with a few bots and the tools you actually use.
- **Who it's for:** individuals and solo founders who want a small bot team for life admin and a side product, with drafts by default and every consequential action approved.

**3. [Deleting 1,410 safety checks on purpose to see which ones the tests miss](https://x.com/tom_thomasson/status/2118745575113851100)**
- **Replaces:** trusting a growing agent-written test suite on faith, and being the copy-paste layer between two coding agents.
  - Agents removed 1,410 safety checks from the code one at a time and re-ran the tests: 1,297 were caught, 113 slipped through, 91 of those were real gaps and each now has a test pinning it.
  - A supervisor passes work between Claude Code and Codex, which landed 12 packages in 4 days (suite at 6,329 tests).
  - The same idea was used on AI-written market reports: 28 planted errors (wrong prices, bad maths, invented figures), all 28 caught by an AI reviewer.
  - Honest scoreboard included: consulting calls booked 0 of a goal of 3.
- **What to connect:** Claude Code and Codex with a supervisor that routes work between them, a test runner the agents can call, and a reviewer agent for generated reports.
  - A separate model (ChatGPT here) audits live pipeline runs against the governance rules.
  - No cost stated.
- **Who it's for:** solo developers and small teams shipping agent-written code or reports who want proof that their checks actually catch failures.

## Quick hits

- **Apple trims internal AI budgets** — Bloomberg reports some Apple engineers' daily AI-tool cap fell from $300 to $200 in October, with cuts to $100 for some teams under consideration but not confirmed. [TechCrunch](https://techcrunch.com/2026/11/04/apple-is-reportedly-cutting-budgets-for-its-internal-ai-usage/)
- **Qoder's stealth model Qoduck is free to try** — Alibaba's Qoder IDE is offering a daily no-credit allowance of an unnamed model, Qoduck, to individual Qoder International users since 7 November, even at maximum thinking.
  - Speculation that it is Qwen 4 is unconfirmed, and there are no official benchmarks. [Qoder docs](http://docs.qoder.com/events/qoduck-free)
- **"Ban RSI" goes mainstream in Congress chatter** — On 6 November Rep. Greg Casar quoted New York Assemblymember Alex Bores's call for Congress to ban recursive self-improvement, meaning AI fully automating AI research, "this year if possible".
  - It's an advocacy post, not a new bill, but it followed last week's METR report on AI-accelerated development at Anthropic. [The Verge on the METR report](https://www.theverge.com/ai-artificial-intelligence/1029126/anthropic-metr-report-accelerating-ai-development)
- **Musk promises overnight Grok Bot fixes** — [@elonmusk](https://x.com/elonmusk/status/2119617455102333206) said he did a round-table review with the Grok Bot engineering team, fixes are being implemented overnight and users "should notice a dramatic improvement by tomorrow".
- **Theo on the $200 Claude plan** — [@theo](https://x.com/theo/status/2119632946663653740) says he ran 8 to 25 Opus 5.5 threads most of Saturday and used under 20% of a $200 Claude subscription, a counterpoint to last week's grumbling about Opus 5.5 quality.
  - Same feed: [he'll ban bots](https://x.com/theo/status/2119567959052231138) that spam "review" comments on every T3 Code PR.

## Go deeper

- [Anthropic, "Teaching Claude why"](https://www.anthropic.com/research/teaching-claude-why) is resurfacing on X this weekend; it is from May, but still the clearest write-up of why training on reasons beats training on demonstrations for alignment.
- [The 956 Report's line-by-line read of the Starbase notice](https://www.the956report.com/post/spacex-wants-a-tax-break-for-a-24-billion-gas-plant-and-super-compute-facility-in-starbase) is a model of how to read a local filing.
