This morning’s briefing opens with the sell-off in US technology shares. There’s also a Claude-assisted claim to a landmark maths proof, a US threat to tax overseas datacentres, and new releases from ChatGPT, Claude, Copilot and Perplexity.
Nasdaq falls 5.3% as Anthropic and Nvidia shares slide
What happened. The Nasdaq Composite fell 5.3% on Friday 20 November, taking its loss for the week to 10% and wiping out its gains for the year. The S&P 500 fell 3.7% and the Dow Jones Industrial Average fell 1.9%. Both finished more than 10% below their recent highs, according to Ameriprise.
Claude developer Anthropic fell 8.5% to US$88.91 on its second day of trading. The company closed its public offering at US$93.5bn in gross proceeds after its underwriters bought all the additional shares available to them. Anthropic received US$91.6bn before costs and separately closed a US$26bn private share sale to investors including Google, Microsoft and Nvidia.
Nvidia closed 14% lower. The US Securities and Exchange Commission is examining the chipmaker’s accounting practices, Reuters reported, citing two people briefed on the matter.
Why it matters. Anthropic’s US$26bn private placement links the finances of a major AI model developer and the chipmakers and cloud providers building the infrastructure it uses. Ameriprise chief market strategist Anthony Saglimbene says investors are reassessing whether customers can finance the computing power they have committed to buy. He points to elevated borrowing costs and to selling in banking, industrial and materials shares as investors weighed slower AI investment. For organisations buying AI, the practical question is which planned services depend on computing capacity that is still being built.
Index points
The Nasdaq slides as Anthropic goes public
Nasdaq Composite, 2–20 November 2026
People are asking
**Did Friday’s fall reduce the money Anthropic raised?
** Anthropic received US$91.6bn before costs from the offering at US$83 a share. Friday’s US$88.91 close was the price of shares trading on the market.
**What does a market correction mean?
** A correction is a fall of about 10% or more from a recent high. The S&P 500 and Dow crossed that threshold on Friday.
Claude-assisted Riemann proof passes independent computer checks
What happened. On 20 November, Anthropic mathematician Levent Alpöge published a four-page account saying he and Claude had found a proof of the Riemann hypothesis and a broader version of it. The 1859 conjecture concerns how prime numbers are distributed. Alpöge made the argument available as roughly 17 million lines of Lean code, in a language designed to express and check mathematical proofs.
In his 22:05 GMT update, MIT mathematician Andrew Sutherland reported that two independently written programs accepted the full exported proof. His checks also confirmed that the submission used the standard Lean mathematics library’s statement of the conjecture.
Why it matters. Alpöge describes identifying a promising route and developing it with Claude, which he says pulled ahead near the end. Releasing the argument in a form that a computer could check let an independent mathematician test it the same day. The workplace equivalent is a spreadsheet analysis whose totals a colleague can recalculate from the original records. The evidence travels with the result, so someone else can check it before relying on it.
timeline
The Riemann hypothesis from conjecture to computer checks
Milestones in a problem about prime numbers
- 1859Riemann states the conjectureBernhard Riemann links a question about a mathematical function to the distribution of primes.
- 9 August 1900Hilbert lists it among 23 problemsThe mathematician David Hilbert includes the conjecture in his list of major problems for the new century.
- 24 May 2000Clay puts up a US$1m prizeThe Clay Mathematics Institute names the conjecture as one of its seven Millennium Prize Problems.
- 10 August 2026Claude advances a related resultAnthropic reports that Claude improved a bound used in work on the hypothesis.
- 20 November 2026Alpöge publishes his proof claimHe describes his work with Claude and releases the argument as code in the Lean proof language.
- 20 November 2026, 22:05 GMTSutherland reports independent checksThe MIT mathematician reports that two independently written programs accepted the exported proof.
People are asking
**Did Claude do it on its own?
** Alpöge describes a collaboration. He set the direction and worked through the argument with Claude. UCLA mathematician Terence Tao identifies the model as internal to Anthropic.
**What did the checks establish?
** Sutherland reported that two independent proof-checking programs accepted the code and that it addressed the standard statement of the conjecture.
**Has the US$1m prize been awarded?
** The Clay Mathematics Institute still lists the problem as unsolved. Its prize rules require publication in a qualifying outlet, at least two years since publication and general acceptance among mathematicians.
Trump threatens to tax US companies that build AI datacentres abroad
What happened. At a White House briefing on 20 November, US President Donald Trump said he wanted American companies to build datacentres in the US and threatened a tax on companies that build overseas. He said he had yet to discuss the idea with his staff.
The push to build in the US comes amid local opposition. About three-quarters of voters in this month’s midterm elections opposed building a datacentre in their community, according to the AP Voter Poll. In a separate July survey of US adults, 73% backed requiring operators to pay for grid upgrades and 62% backed limits on the number of new datacentres.
Why it matters. The threat brings the location of future computing capacity into US policy. Microsoft told Parliament it was building datacentres in Wales and London and working on sites in northern England as part of its £22bn UK investment. For UK organisations buying cloud or AI services, those projects matter for how much capacity suppliers can offer locally and when.
Domestic construction faces a separate cost question. The US Senate has passed a bill aimed at speeding up energy infrastructure approvals and making datacentres pay for the energy they use. The bill goes to the House of Representatives next.
%
US support is strongest for making datacentres pay for grid upgrades
Support and opposition to four proposed US datacentre policies
People are asking
**What did Trump announce?
** He raised the idea of a tax on US companies building AI datacentres overseas. The proposal is at the press briefing stage, with no rate or start date set.
**What is Microsoft building in the UK?
** Its evidence to Parliament names datacentres in Wales and London, work on sites in northern England and plans to bring more than 20,000 AI chips to the UK.
New releases
ChatGPT. Space is rolling out worldwide to Plus subscribers aged 18 or over, with pages and presentations they can share with other Space users and set to update from connected apps. More
Claude. Claude’s computer-use beta is rolling out to Team and Enterprise, letting it see the screen and work in approved desktop apps once an organisation owner enables it. More
Claude Code. New recurring tasks in Claude Code 2.1.330 can run for up to 30 days instead of seven while the session stays open. More
Microsoft Copilot. GPT-6.2 Sol is rolling out to eligible users in Copilot Chat, Studio, Word, Excel and PowerPoint, while GPT-6.2 Astra is rolling out in Cowork and Studio and reaching Chat and those Office apps through Frontier. More
Perplexity API. The Agent API’s extra-high preset now uses GPT-6.2 Sol when no version is specified, and successful webpage fetches in the Decisions API’s web-grounding mode have fallen from US$0.001 to US$0.0005 each. More
Training your team on ChatGPT, Claude or Microsoft Copilot? See our ChatGPT training, Claude training and Microsoft Copilot training.
Elsewhere in AI
The UK and EU said at their Brussels summit that they intend to launch a digital partnership in early 2027, with AI safety and security among the proposed areas of cooperation. Read
AI evaluation group METR found that the length of software and research tasks AI models could complete half the time doubled about every 3.6 months across its benchmark tests. Read
US Democratic senator John Hickenlooper proposed a Senate committee dedicated to AI and named independent evaluations of AI systems among his priorities when Democrats take the majority in January. Read
US Treasury secretary Scott Bessent said sales of older Nvidia chips to China need no further restrictions after the administration moved to curb Chinese companies’ remote access to newer chips. Read
Cybersecurity company Zscaler announced partnerships with Anthropic and Mistral to help businesses secure AI agents and defend against AI-assisted attacks, including a Mistral integration for European customers. Read
