A house newsletter · daily · researched and written by the Kurzia AI
Monday 9 November 2026 · Issue 11 · 3 min read
Goldman pulls Claude Code from its traders; DeepSeek's agent app burns 92T tokens a day; Nvidia's board investigates Huang
The day's one thing
thestateofai.com · yesterday
Goldman Sachs
Takeaway: roughly 47,000 staff
Goldman Sachs has restricted Claude Code and GitHub's coding tools to its engineers, cutting off the traders and bankers who had been using them to monitor client flows and model market moves. Bloomberg reports the change came in internal guidance updated in October, after a review covering both productivity and risk.
This is the first big firm to pull agent access back by role rather than widen it, and the reasoning travels: code an engineer writes with an agent still passes review and deployment gates, while a trader's script does not — and unreviewed code plus information barriers is a compliance problem, not a productivity one. The sensible stance is Goldman's rule without its manner: name who owns any agent-written tool before it touches production data, and record what people were using it for before you take it away. Developers make up about 12,000 of the bank's roughly 47,000 staff, and CIO Marco Argenti was still praising trader-built AI tools in July. Goldman and Morgan Stanley are lead underwriters on Anthropic's IPO, with the listing expected on Nasdaq before the end of November.
How organisations are changing
DeepSeek Harness — the open-source agent app hit 1.54 million daily users on November 4, under three months after launch, and is consuming about 92 trillion tokens a day — roughly 60 million per user, figures the company released to Bloomberg. That is the working scale reference for token-bill planning: at that consumption, a flat per-seat allowance in your organisation breaks on the first heavy user, so meter, cap, and watch the outliers. techmeme.com
US professional services — added 70,000 jobs in October, the most since July 2022, with consulting, legal and accounting — the AI-exposed professions — all growing, the Wall Street Journal reports. One month proves nothing, but it is the hard number to hand the board when displacement comes up: the sectors most exposed to agents are still hiring, and the job-loss story has not yet reached the payroll data. techmeme.com
Monaco — kept Devin after evaluating code-review products but built its own agent workspaces — background agents need somewhere to run the product and reproduce a failure, and that depends on how your own cloud is set up. Mihail Eric, who leads AI there and teaches Stanford's CS146S, states the rule of thumb as buy to learn and build to keep; the course now has students contribute to open-source projects with maintainers reviewing their patches, a hint at what your next graduate hires will expect to be judged on. latent.space
Vendors, pricing and risk
OpenAI status — an incident for elevated errors with image generation was posted on OpenAI's status page on November 8. The page notes availability is reported in aggregate across all tiers, models and error types, and individual customer availability may vary. Decision-relevant detail: if image generation is in your stack, log your own error rates — the vendor's aggregate average is not evidence for an SLA claim or a contract conversation. status.openai.com
CAISSI — officials at the Commerce Department's AI safety centre were asked to informally report to a White House office, the Associated Press reports, raising concerns about the division's independence. If your compliance plans lean on federal AI-safety guidance to settle agent-risk questions, expect that guidance to arrive contested — the referee is itself now a political story. techmeme.com
News the board will ask about
Apple — is said to be pressing for more urgency in its AI efforts: Bloomberg's Mark Gurman, citing sources, reports hardware chief John Ternus is pushing software chief Craig Federighi beyond Siri tweaks toward agentic and other advanced capabilities. Unconfirmed and unnamed-sourced, but the internal-pressure framing suggests Apple's agent move, when it comes, will be judged against what Google and OpenAI already ship. techmeme.com
Anthropic — has bought commitments for 1.7 gigawatts of gas-fired generators from equipment supplier OnePWR for $3.4 billion, the Financial Times reports, citing sources — described as the largest such purchase by an AI lab, bypassing traditional suppliers to secure scarce equipment. Your agent vendor's bottleneck is now turbines and grid power, not just chips; capacity, not demand, is what to plan procurement and contract terms around. techmeme.com
Kristian Polso — a developer with about ten years building production ML for clients has written that he will no longer use LLMs in professional coding, citing cognitive-impact studies, unmaintainable output, and the CLOUD Act's reach over EU-stored data. Your senior engineers are reading this argument; the compliance half — US vendors' EU-region promises do not block US federal law enforcement access — is the part worth answering before your EU teams raise it. polso.info
The big AI stories today
Nvidia — its board has begun an independent investigation into reporting that CEO Jensen Huang was the target of a foreign intelligence effort, saying it has no reason to believe national security or company secrets were compromised; House committee chairs have asked him to a closed-door briefing on November 18. Governance risk at the top of the company your hardware roadmap depends on. nvidianews.nvidia.com
Google — has bought out Crusoe's southeast Wyoming data centre site and will build the project itself, a day after pledging $17 billion on its first Wyoming data centre; Bloomberg reported Crusoe was pushed aside after failing to sign customers. Capacity keeps consolidating into the hyperscalers — one more reason your cloud choice is also an AI-capacity choice. crusoe.ai
One takeaway
Decide who owns agent-written code before an incident does it for you. Goldman's rollback is the pattern to copy and improve on: coding agents by role, not by enthusiasm, with a named engineer owning anything an agent writes before it reaches production data. thestateofai.com