Key points
- Anthropic — On November 19, 2026, Anthropic made its Nasdaq debut, with valuations reported from roughly $1.3 trillion all the way to $2 trillion, making it the most closely watched listing in the AI industry.
- MAI Code 2 — Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price, and against Opus 5.5 reaches 96% of the performance at only 5% of the price.
- Google — Google DeepMind announced that Genie 4 is out and available to eligible Google AI Ultra subscribers.
Here is Robin’s Daily AI Hotpot — an unfiltered roundup of the day in AI.
Anthropic Lists on Nasdaq: The $1.3 Trillion to $2 Trillion Gap Is a Matter of Basis
Capital markets first: a leading model company started trading on Nasdaq on November 19, and the same pricing was turned into very different market caps.
On November 19, 2026, Anthropic made its Nasdaq debut, with valuations reported from roughly $1.3 trillion all the way to $2 trillion, making it the most closely watched listing in the AI industry. The difference mainly comes down to the counting basis: the $2 trillion in headlines may be on a fully diluted basis that includes common shares, unexercised stock options and restricted stock, while outstanding shares at the then-$98 share price put the market cap at about $1.28 trillion. Reuters reported on November 18, 2026 that the IPO priced at $83 per share and raised about $81.3 billion, corresponding to a fully diluted valuation of roughly $1.5 trillion and about $1.1 trillion undiluted — which is exactly where the $1.3 trillion to $2 trillion spread in coverage comes from. The company simultaneously carries $900.9 billion in long-term spending commitments, $451.6 billion of which is expected to land before the end of 2030, which will visibly lift demand for AI infrastructure suppliers; its third-quarter operating loss was $567 million, so new investors are taking on the risk of heavy spending against profitability as well. In community discussion, several commenters argued that $2 trillion is too high for any software company, AI included, describing the pricing as “insane,” while others read it as investors cashing out early; one commenter clarified that fully diluted valuation and the market cap Nasdaq displays on outstanding shares are not the same basis, and another asked for an explanation of how revenue of close to $17 million a quarter supports $2 trillion; the reporting offers no listing filing or complete financial data to check these claims.
The word to remember is “basis”: different reports on the same day, options and restricted stock counted or not, and the same company lands hundreds of billions of dollars apart. From here on, whenever you hear a market cap for any AI company, first ask which method it uses; what is genuinely hanging in the air is revenue — the spending commitments run past 2030, so this pricing amounts to buying a very long execution period in advance.
The AI Selloff Pushes the S&P and the Dow Into Correction Territory, and the Nasdaq Gives Back the Year
The selling spread after the first trading day: the broader market entered a correction while the new company itself was still above its issue price.
An AI selloff led lower by Anthropic pushed the broader market into correction territory: the S&P 500 and the Dow Jones Industrial Average entered a correction and the Nasdaq gave back all of its 2026 gains. Anthropic’s record IPO opened at $111.50 on November 19, about 34% above the $83 issue price, then fell about 15% to $95.50, still above the issue price; at the time the Nasdaq and the S&P 500 were each down 1.5%. In an earlier tech selloff on November 5, retail investors sold $265 million of stock, the largest single-day selling since 2020, when the market’s worry was regulation of AI data center construction. The selling was not confined to the United States either: the Toronto composite index fell and global equities moved lower in step, and the strength of Anthropic’s first day did not dispel the concern — investors are watching what it will cost to build AI data centers.
What this round actually registers is index levels and financing costs, not any single company’s first-day pop — entering a correction and record retail selling since 2020 show the market starting to price compute spending; when you run into talk of an “AI bubble,” first check which range the index sits in and whether data center financing conditions have changed.
SoftBank Falls More Than 16% in a Single Day as the Same AI Selloff Reaches Asia
The selling did not stop on Wall Street; one day later it pressed down on Asian tech stocks.
According to CNBC, SoftBank shares fell more than 16% in a single trading day as the same AI selloff bore down on Asian technology stocks. After Anthropic’s Wall Street listing, investor concern over whether AI valuations are sustainable did not fade and kept weighing on Asian tech; holders of SoftBank and other AI-linked Asian tech stocks saw clear paper losses, and doubts about the durability of the AI boom outweighed the strength of Anthropic’s first trading day.
The same selloff gets priced at different magnitudes in different markets: in the United States you look at indices and correction ranges, in Asia at the share price of the heaviest-weighted names like SoftBank. Treat it as a brand-new event and you will misread it; whether it transmits into financing is the thing still worth watching.
The U.S. Senate Passes a Permitting Reform Bill That Could Speed Up Data Center Power Timelines
Over to Washington: a bill covering energy and data center permitting has just cleared the Senate.
The U.S. Senate passed an omnibus bill aimed at simplifying project permitting nationwide, covering the development of energy and AI-related projects. The Bipartisan American Affordability and Jobs Act is part of a long-running federal permitting simplification effort intended to streamline federal approvals for energy and infrastructure projects, and after Senate passage it has been sent to the House for consideration. The reporting gives no vote tally, no specific provisions, no implementation timetable, and does not say whether it has taken effect. If it does become law, the permitting changes could speed approvals for power transmission lines, energy projects and AI data centers, directly affecting developers’ schedules and project timelines.
It is still several steps from taking effect: the bill can still change in the House, the reporting offers neither a vote count nor a timetable, and what it actually presses on is permitting schedules for transmission and data centers.
Microsoft Ships Seven In-House Models at Ignite, With MAI Code 2 Selling on Code Price-Performance
Microsoft put seven of its own models on the table at Ignite in one go, and the coding one is being priced against rivals.
At Ignite 2026 Microsoft launched seven in-house AI models in one pass, covering coding, image generation and editing, transcription and other tasks, with the code-focused MAI Code 2 positioned on price-performance. Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol’s coding quality at about 40% of the price, and against Opus 5.5 reaches 96% of the performance at only 5% of the price. The company also cites Artificial Analysis rankings and price comparisons to show its positioning, saying MAI Image 2.7 ranks first for image editing and that MAI Image 2.7 Flash targets high-throughput generation. All of these numbers come from the official launch framing; the materials carry no benchmark definitions, no test methodology, no latency or stability data, and no complete specifications, open-weight status or licensing terms for the seven models. In the community, some say MAI Code 2’s real coding experience is comparable to Opus or Sol, while others question how bold the comparison is and ask why it is not yet in the Artificial Analysis index; one view holds that Microsoft should ship open-weight or open-source versions if it wants share, and other commenters criticized this Ignite as lacking innovation, saying they have no incentive to adopt Microsoft’s AI stack.
Comparisons of the “this much quality at a fraction of the price” kind only hold on the same benchmark and the same workload, and these figures currently rest on Microsoft’s own framing alone, with no benchmark definitions and no independent verification; what is missing is exactly the measurement on a shared workload.
Microsoft Calls Copilot “the New OS for Work” and Moves Office Inside It
In Microsoft’s vision the hierarchy between apps and assistant has been inverted — speak first, then let Office do the work.
In the Ignite keynote, Microsoft CEO Satya Nadella laid out a new “frontier ecosystem” vision and announced that Word, Excel and PowerPoint are now generally available inside Copilot, which he called “the new OS for work.” This is an inversion of the hierarchy: apps like Word used to be the star with an assistant helping alongside, and Microsoft now imagines the agent leading while apps become plug-in components hanging off it, backed by a layer of organizational knowledge that pulls context from email, meetings, calendars, chats and files. Users can create documents, spreadsheets and presentations directly in Copilot, and the generated files can still be edited further in the respective Office apps.
With the entry point changed, Office goes from protagonist to a component that gets called, and the organizational knowledge layer behind it pulls context from email, meetings, calendars and files; what is worth watching next is which data sources that layer connects to, where the permission boundaries are drawn, and whether enterprise and personal accounts get tiered authorization.
Google DeepMind Opens Genie 4 to Some AI Ultra Subscribers
Progress on the games side too: a world model that generates scenes from your words in real time is starting to roll out to subscribers.
Google DeepMind announced that Genie 4 is out and available to eligible Google AI Ultra subscribers. The rollout is staged by design: first a limited set of trusted testers, then eligible AI Ultra subscribers, with plans to open it to early partners through the Gemini API and AI Studio. It builds on the Gemini family, with Gemini Omni responsible for making the simulation’s appearance and behavior close to the real world and other Gemini models providing the complex reasoning that keeps a simulation coherent and continuous. Subscribers can drive scenes with natural language, run interactive world simulations in real time, and get ambient sound, music and dialogue generated alongside. Eligibility and regional availability are vague, however: some users found that workspace accounts could not subscribe and that their personal account page still showed Genie 3 as the Ultra offering, which limits its impact for now. There is also no evidence yet that it works as a programmable game engine. Discussion includes admiration for DeepMind’s creative investment and a view of it as a future game engine in the vein of Roblox or GTA, alongside worries about weak profitability and negative social consequences, and questions about whether scene detail stays consistent as the camera rotates.
The staged rollout says it is not yet a general product: trusted testers first, then subscribers, then API partners. Worth watching is whether scene detail holds up as the camera turns, when it genuinely opens to developers, and whether world models get sold as programmable engines — there is no evidence for that layer yet.
Google Launches Gemini 4 Neon: Higher Overall Score, Weaker Than Carbon on Research-Grade Reasoning
Another name joins the model lineup, and the official materials do not even give specs, pricing or availability.
Google launched Gemini 4 Neon, positioning it as the more capable variant in the Gemini 4 family. The materials contain no official technical specs, pricing, availability scope or verified benchmark results, so its real advantage over Carbon is still unclear. Based on Artificial Analysis data, the community sees Neon scoring higher than Carbon overall but lower on research-scale reasoning evaluations such as CritPt, which fits the division of labor between the two: Neon is more of a generalist, Carbon more oriented to long-horizon, complex research tasks. One user reported Neon hitting 127 tokens/s on OpenRouter versus 46 tokens/s for Carbon, an unverified community observation. In Artificial Analysis’s Xhigh test, Neon’s intelligence index was 61, above the comparable-model median of 43, but that evaluation run generated 283 million tokens against a median of 65.1 million, so production selection still has to price token cost and response load against real workloads. For developers choosing between the two, Neon can be a candidate for general benchmark priority, but not a full replacement for Carbon on long-horizon or research-grade reasoning tasks; leading a general ranking does not mean it has the edge in sustained long-running work or research-style reasoning. Some commenters are weary of the growing pile of model names and question whether scale can be inferred from a name; others feel the latest Gemini is comparable to or slightly better than Opus 5.5 on some benchmarks while still behind it in overall intelligence experience.
Leading a general ranking does not mean being stronger on a specialty: to use it for long-horizon or research-grade reasoning, look first at the gap on evaluations like CritPt; that pair of numbers — “intelligence index 61, but one evaluation run generating 283 million tokens” — is also a reminder that token consumption, meaning the metered usage unit you pay for, has to be calculated against real workloads.
Anthropic Employee Claims a Complete Proof of the Riemann Hypothesis, With Roughly 15 Million Lines of Lean Formalization
The tech world also picked up a math story today, and for now it is only a claim.
Levent Alpöge, a mathematician employed at Anthropic, published a PDF claiming a complete proof of the Riemann hypothesis and saying the method generalizes to the generalized Riemann hypothesis; discussion says the accompanying Lean formalization is about 1.3GB and roughly 15 million lines, an abnormally large artifact whose machine verification burden is equally abnormal. The claim is under intense scrutiny, and the available materials contain no independent confirmation; what is being debated is only whether the problem is encoded correctly in the formalization and whether the dependencies hold — a large artifact does not by itself prove the proof correct. The Riemann hypothesis concerns the distribution of the zeros of the Riemann zeta function and is one of the most famous unsolved problems in mathematics; Lean is a formal proof assistant that can check a proof at the code level, but experts still have to confirm whether the conjecture and related definitions are encoded correctly and whether the premises the proof depends on hold. If independently verified, it would be a mathematical breakthrough and a demonstration that AI-assisted formal mathematics can scale. Commenters are moved both by Alpöge’s credentials and by the sheer size of the Lean artifact, reading it as a sign of AI progress; others worry that the handful of mathematicians employed at AI companies will gain a decisive advantage from it. The judgment of the mathematics community as a whole is not yet clear, and until experts confirm it, this can only be treated as a claim awaiting verification.
An independent check sits between claim and proof: first confirm that the formal statement in Lean is the Riemann hypothesis itself and that no extra assumptions hide in the dependencies; until the mathematics community reaches a conclusion, this can only be broadcast as an unverified claim — size alone does not prove a proof correct.
Today’s items circle the same thing: once capability takes the stage, numbers need a stated basis, claims need verification, and errors need someone to own them. The ledger and the lab are all asking the same question — who actually pays this bill.
🎧 This episode is also available as a podcast: listen to Daily AI Stew · 2026-11-20.
