AI News, Nov 6: The AI Trade Gets Repriced
The Big Story: the AI trade gets repriced
Thursday’s selling landed exactly where the data-center buildout is financed. Nvidia lost 9.3%, AMD 17.3%, Micron 17.5% and Broadcom 10.4%, with Microsoft down 3.6% and Tesla 10.8%. Overnight it crossed to Asia, where the KOSPI fell 8.7% at one point and the Korea Exchange imposed a circuit breaker during morning trading.
Then Anthropic. Investors are now discussing an IPO valuation of roughly $1.2 trillion to $1.3 trillion, down from the $1.4 trillion to $1.5 trillion range communicated to bankers, and from figures as high as $2 trillion in early October. A third of the company’s notional worth is gone in five weeks, with no change to the product.
Neither move is about the technology. Three days after voters backed restrictions on nearly every data-center measure on the ballot, the market began pricing political risk into the one input the buildout cannot substitute. Eaton CEO Paulo Ruiz gave the gap its number this week: announced data-center projects worldwide “amount to 366 GW”, against an installed capacity of 50 GW. Nearly all of it still needs a permit, a substation and a county that wants it.
The Bank of England puts three years of AI at 0.5% of UK productivity
The Bank’s November Monetary Policy Report found “only evidence of a modest AI-driven rise in UK productivity growth”: 0.1 to 0.2 percentage points of extra annual growth since early 2024, or about 0.5% over three years. Just under a fifth of UK GDP growth since the start of 2026 came from AI-related activity, and most of that is upstream development rather than anyone using the output. A Gartner survey quoted alongside it found 3% of organizations had hit their AI targets while 38% were over budget.
Visa’s new Agentic Commerce Benchmark ran 22 models from 12 labs through shopping tasks the same week: 57.2% of runs passed and 42.8% failed. In 54% of the failures the model completed a purchase anyway that didn’t meet the stated requirements. Two measurements, one week, same shape: the spending is real and the return is still mostly a forecast.
Today’s Top Stories
METR says Anthropic’s research is accelerating, and most of the proof was redacted
Eleven weeks, 2.5 full-time staff, about 15 interviews, four unreleased models. METR’s finding: Anthropic’s internal capability metric accelerated 1.9x since October 2025, from 15 to 29 points a year, 95% CI 1.5 to 2.8x. It guesses perhaps 80% of that is AI improving AI rather than extra compute. Across all frontier developers the figure is 1.4x. Anthropic’s public August risk report told readers acceleration was “less than a factor of 2”; METR calculates that the redacted internal figures imply 1.97, and says it is “plausible” the models available by July 15 had crossed Anthropic’s own early-warning threshold. Anthropic cut “a majority of the evidence we found most informative” under unilateral redaction rights, then removed compute verification from the project’s scope midway through. The report drew 39 points on Hacker News.
Protesters broke into an Amsterdam data center site
More than 100 activists marched arm in arm through the fence of a Microsoft AI data center under construction in Amsterdam on Friday, climbed the cranes and danced while police stood by. Placards read “Screw your datacentre.” The site belongs to Pure Data Centres, which asked occupiers to “leave safely and engage with us in dialogue.” Microsoft said it respects peaceful expression but does not condone unlawful action. The Stop Dirty Data Centres coalition has actions in 16 countries across Friday and Saturday. Yesterday this was a press release; today it is a trespass case.
Claude Code was writing users’ email addresses into the User-Agent header
A bug report filed Thursday night says Claude Code disclosed the user’s account email to third parties in the User-Agent header during web searches, and the reporter now receives “dozens of scam emails per day.” It became the window’s biggest Hacker News thread by a factor of three, at more than 1,300 points. Anthropic staff replied in-thread: a built-in mod called cc-plugin-privacy ships on by default in v2.1.317 and prompts before any web call containing your address, and the account email leaves the model’s context entirely in next week’s release. The same complaint was filed as an issue on July 25 and is still open.
An audit found defects in 30% of a benchmark everyone cites
Stanford’s E. Kelly Buchanan, Crystal Zhou, Scott Linderman and Christopher Ré audited Terminal-Bench 2.0, a coding benchmark that had extensive human and model-assisted review plus red-teaming before release. They found issues in roughly 30% of it. Repairs raised accuracy by over 10 percentage points and moved leaderboard positions by up to three ranks, though the overall ranking held at a Spearman rho of 0.958. The clock is a hidden variable too: extending the time limit closed a 20-point harness gap on GPT 5.4 and shrank a 30-point gap on GPT 5.4 mini to 5. They call the failure mode benchmark maintenance debt, and it applies to every leaderboard in a funding deck.
Thinking Machines ships its biggest model, closed
Inkling 2 Small is a mixture-of-experts model with 594B total parameters and 17B active, scoring 31 on the Artificial Analysis Intelligence Index, the highest of any Western model due to publish weights. It launched proprietary, available only through the company’s own Tinker platform, with an Apache 2.0 weights release promised for an unconfirmed later date. Per task on that index it costs $1.18 against $7.77 for Gemini 3.8 Flash. A frontier-adjacent score at a seventh of the price is the news; the licence is a press release until the files appear.
Quick Hits
- Policy: Democrats are planning a flood of AI bills next year, with Schumer prioritizing consumer energy costs and Rep. Lori Trahan drafting independent safety evaluation and liability legislation; no bill text yet.
- Chips: Southeast Asian cloud providers must now enroll in Nvidia’s Fleet Intelligence tracking system, which logs individual accelerator locations, to qualify as Nvidia Cloud Partners.
- Oversight: Three House Republican chairs want Jensen Huang to brief them on Nov. 18, the day after Nvidia’s earnings, about his contacts with a Megaspeed executive.
- Infrastructure: Google committed $17 billion to its first Wyoming data center, air-cooled, with a 702 MW AEP fuel-cell project and carbon storage at a Tallgrass sequestration hub.
- Earnings: IREN said over $4 billion of annualized revenue slipped out of 2026 because Microsoft asked it to reconfigure network infrastructure for 200 MW of frontier-model capacity.
- M&A: Dropbox made its largest acquisition to date, paying $350 million for Tigris Data to power a storage platform for AI agent builders.
- Consumer: Anthropic began testing a Personal Finance hub that reads bank accounts through Plaid, read-only, US beta, with those conversations excluded from training.
- Publishing: Meta paused the commercial terms that would have let it use business customers’ websites, after injunctions in Norway and Brazil and a UK CMA investigation.
- Courts: North Carolina parents are suing Google in California state court over months of messages their 13-year-old exchanged with Gemini about self-harm on a school-issued Chromebook.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
