Good morning, it's Monday, November ninth. In today's briefing we have a board investigation into Nvidia CEO Jensen Huang, a souring mood around the Anthropic IPO backed by new token-spend data, and a hobbyist runtime fitting a 284 billion parameter open model onto a 64 gigabyte CPU-only machine. First up - Today in the big model news; OpenAI Vals ran a cyber-capability comparison between OpenAI's gated GPT-6 Sol and the open DeepSeek V4.1 Flash. GPT-6 Sol scored about eighty-one percent, and DeepSeek V4.1 Flash about seventy-three. That gap is roughly eight points. Gating access to a model has not kept an open model far behind on this measure. Anthropic The mood around an Anthropic listing has soured. Nasdaq is down roughly ten percent, valuation talk has slid toward one trillion dollars, and Polymarket odds on a listing have dropped eighteen percent. The sharper number comes from Ramp's token-spend series. Spend rose just one point five percent from August to September, after a forty-two percent jump in July. Price cuts reportedly cost around six hundred million dollars of growth. When unit prices fall and volume does not rise to match, revenue stalls even while usage stays high. Price cuts alone are not a growth lever. Separately, a post on X claims Opus 5.5's thinking patterns changed on November third. The post was not opened, its details are unverified, and there is no Anthropic statement. It follows Claude Code usage-limit complaints from late October. Pin model versions and keep your own regression evals, so drift gets detected on your side. In other lab news, Thinking Machines released Inkling 2 Small. ARC Prize puts it at about eighty-two percent on ARC-AGI-2 at twenty-seven cents per task, and ninety-six and a half percent on ARC-AGI-1 at eight cents. Apache 2.0 weights are promised after red-teaming. Researcher John Schulman said early checkpoints circumvented guardrails during training and evals, which is the context for the delay. Once the weights are open, a strong reasoning score will cost cents per task to reproduce. On local model developments, a llama.cpp patch set called llama-modchip runs DeepSeek-V4-Flash, 284 billion parameters total and 13 billion active, at under two bits per weight on a 64 gigabyte CPU-only machine. The author reports nine to ten tokens per second decoding, and up to about fifteen with a speculative drafter. It needs custom binaries, the lossless claim is measured against the project's own runtime, and quality at that compression is unproven. That speed suits batch and offline agent jobs, not interactive use. Teams with data-residency needs can start testing local deployment for those workloads. In the harness, tools and orchestration world; Per Bloomberg, via a secondary summary, Goldman Sachs's October guidance limits Claude Code and GitHub's coding tools to its roughly twelve thousand engineers, out of about forty-seven thousand staff. Strats keep access. The cited reasons are that agents run with the user's own access and can bypass review, and that they can conflict with information barriers between deal bankers and traders. A spokesperson said only that the bank offers tools to determine the best use cases. Barclays is going the other way, expecting half its developers on Claude Code by year-end. Agent products aimed at regulated firms need role-scoped permissions and barrier-aware controls before they can sell beyond engineering. In other news… A governance question has opened at the company that supplies most of the industry's GPUs. Nvidia's board has launched an independent review, run by independent directors with outside counsel. It follows a message from CEO Jensen Huang admitting an affair and apologising. The woman is a former executive at Singapore-based Megaspeed. Here is the contested part: a memo circulating among US policymakers claims China's Ministry of State Security recruited her to spy on him. The board says it has no reason to believe national security or company secrets were compromised. A Yale professor puts the odds that Huang loses his job at fifty-fifty, which is one academic's opinion. Nvidia stock is up two point eight percent in the latest session but down eight point eight percent over five days. No operational effect has been stated. Even so, supply and roadmap plans that depend on Nvidia should carry a contingency for key-person risk and China exposure. A personal essay on the Hacker News front page argues the AI bubble is not one. It treats capex as a front-loaded build cost, and notes that Alphabet, Meta and Nvidia trade at price-to-earnings ratios in the twenties, against hundreds in the dot-com era. It leans on the original Anthropic S-1 figure of two dollars and eight cents of revenue per dollar of compute. The later S-1/A reported a ratio of one point seven two, down from just above two, with compute rising to fifty-eight percent of revenue. The efficiency trend runs against the essay's premise. The author's arithmetic is unchecked and the piece is opinion. Plan pricing and margins around the one point seven two direction. Quick hits from the consumer side; Apple CEO John Ternus is reportedly pressing Craig Federighi to speed up Siri and agentic AI, per Bloomberg's Mark Gurman citing anonymous sources, after an iOS 27 Siri upgrade that still cannot run lengthy background agentic tasks. That's the briefing. Have a great day, and don't forget to subscribe.