Friday opens with a split screen. On one side, Anthropic’s stock: priced at $83, opened at $111.50, closed at $97.17, up 17 percent on the largest IPO in American history. On the other side, everything else: the Dow down 1,100 points, the S&P 500 down for a sixth straight session, roughly $1.3 trillion of market value gone by the close. Yesterday this paper said the ticker had become the scoreboard. Today the scoreboard rendered its first full verdict, and the verdict is that a public price cuts both ways—it validates the winner and reprices everybody else in the same motion. Underneath the market noise, the labs had the better day: classical supercomputers closed a gap that was supposed to belong to quantum machines, Google proved that quantum mechanics’ strangest property survives at 105 qubits, and OpenAI published the most honest numbers yet about what its agents can and cannot do inside its own building. The money is nervous. The work is not.
Anthropic opened at $111.50, closed at $97.17—and the market fell around it
Anthropic began trading on Nasdaq just before noon Thursday under the ticker ANT. The first print was $111.50, 34 percent above the $83 offering price; the high was $115.10; the close was $97.17, up 17.07 percent, for a fully diluted value near $1.8 trillion. More than 800 million shares changed hands, beating SpaceX’s 522 million-share debut in June. The offering itself—980 million shares, about $81.3 billion raised—is the largest US IPO on record. Then the rest of the tape gave way: the Dow fell 1,101 points to 49,097, its first close below 50,000 since June; the S&P 500 lost 2.02 percent for a sixth consecutive decline, its longest streak in two and a half years; the Nasdaq fell 2.25 percent; the VIX jumped to 28.44. Nvidia, which had risen nearly 5 percent in the morning, closed down 3.3 percent at a 52-week low, having shed $669 billion in six sessions. Anthropic itself slipped another 3 percent after hours.
“Anthropic stock ended its first day of trading at $97.17, up 17% from its $83 IPO price.”
Read the two numbers together, the way yesterday’s paper set them up. The debut worked: the price held above the offer, the volume was historic, and the disclosure era for frontier AI has begun. The rout is the repricing of everyone who is not Anthropic—the market took the occasion of a $1.8 trillion comp to ask what the rest of the AI trade is worth against it, and answered lower, sector by sector, all eleven of them red. For clients, the practical translation is dull and useful: the cost of intelligence now has a public benchmark with a quarterly filing attached, and every vendor negotiation from here runs against it.
Gordon Bell: classical exascale closes the FeMoco “quantum-advantage” gap
At SC26 in Chicago on Thursday, ACM gave the 2026 Gordon Bell Prize to a ten-member team from Chinese institutions for the first exascale classical solution to the FeMoco CAS problem—the electronic structure of the iron-molybdenum cofactor at the heart of biological nitrogen fixation, a problem routinely Exhibit A in arguments for why quantum computers are needed. Their code ran at 1.456 exaflops in double precision, 58.9 percent of peak, across 20,480 nodes of LineShine, and 5.626 exaflops in mixed precision. LineShine is the same machine behind the Shenzhen density-functional finalist this paper covered on Monday, which makes the week’s pattern plain: China’s flagship classical machine is having a very good conference. The companion Gordon Bell Prize for Climate Modelling went to the STORM team for generative-AI data assimilation on Frontier: 74,400 GPUs at 96 to 99 percent strong-scaling efficiency, up to 6 exaflops of BF16 throughput, and 32,768-member forecast ensembles produced in 34 seconds.
The board’s frontier chair underlines the honest framing in the winners’ own title: closing a gap is not the same as a verdict on quantum computing. What it does is move the bar. Every quantum-advantage claim now has to clear a classical baseline that just got dramatically stronger, on a problem that matters for fertilizer and medicine, not a contrived benchmark. That is how the race is supposed to work—each side’s progress disciplines the other side’s claims.
Google in Science: contextuality survives at 105 qubits
Google Quantum AI published cover research in Science on Thursday demonstrating that measurement contextuality—the property that a system’s measured values depend on which other compatible measurements are made alongside them—persists at many-body scale on its Willow processors. The qubits won the Mermin-Peres magic-square game 98.3 percent of the time across 900,000 trials, against a classical limit of 88.9 percent; a Kochen-Specker-Bell measure came in at 5.618 against a classical ceiling of 4.0; Greenberger-Horne-Zeilinger parity games beat classical rates at sizes up to 71 qubits; and a two-dimensional hidden-linear-function task was solved on grids up to 105 qubits. The team also showed the same tests work as a sensitive diagnostic for coherent control errors.
Entanglement gets the posters; the theory has long said contextuality is the resource that actually powers the speedup in the models that matter, including magic-state distillation. Showing it in few-particle systems was tabletop physics. Showing it intact across a 105-qubit processor, in a peer-reviewed paper, is a hardware fact—and a useful one, because a benchmark that detects subtle errors is worth money to everyone building these machines. Paired with the Gordon Bell result above, Thursday was the frontier beat in its best form: the classical bar rose, the quantum property held, and both claims arrived with exhibits.
Hunterbrook gives AMD its own China problem; the stock falls 8 percent
AMD shares fell about 8 percent Thursday after Hunterbrook Media published “AMD Has Its Own China Problem.” The report says IEIT Systems, the Chinese server maker placed on the Commerce Department’s Entity List in March 2025, has built at least six newer versions of AMD’s server firmware into its own updates since blacklisting, the latest on October 19; that three server companies tied to IEIT—Maginfra in China, Kaytus in Singapore, and Aivres in California—post firmware 99.94 percent identical to IEIT’s; and that $67 million of AMD’s 2025 licensing income traces to two Chinese chip ventures blacklisted in 2019. By Hunterbrook’s count, customers billed in China and Hong Kong were $7.75 billion, about 22 percent of AMD’s fiscal 2025 revenue, against roughly 9 percent at Nvidia.
Print the labels with the story, as this paper insists. Hunterbrook Capital, the affiliated fund, disclosed it is short AMD. Hunterbrook itself says it does not know how IEIT obtained the code or whether a license covered it, and that its findings do not on their own establish an export violation; Reuters could not independently verify the report; AMD told Bloomberg the report misrepresents its business and denied wrongdoing. What is not allegation is the structure: firmware is code, code crosses borders at the speed of a download, and the BIS remote-access rule that took effect Thursday was written for exactly this shape of problem. The Lessig chair’s point lands again—export control is now a software-architecture question, and every chip company’s China exposure is about to be audited in public, by short sellers first.
Search, November 19: a sidebar, a booking button, and an agent that clicks
Google rolled out a collapsible sidebar of recent AI searches on its desktop homepage, globally, announced by Search product chief Robby Stein on Thursday: your AI Mode history now sits one click from the homepage, built for questions that take more than one exchange. The same day, computer use arrived in the Gemini app for Mac for paid US subscribers—Gemini clicking and filling forms across apps and websites in its own window. In Search proper, hotel listings are testing booking through AI Mode behind a “Continue on Google” button, image results gained a Lens button that routes into an AI Overview or AI Mode follow-up, and a “Show original results” button is in testing. Across the aisle, Microsoft confirmed ads in Bing’s AI Overviews are now available in all markets except China, running through existing campaigns.
For the businesses this paper advises, Thursday’s bundle is the strategy stated in interface. The sidebar makes AI search a continuing session rather than a fresh query, which rewards the brands a session keeps returning to. The booking button moves the transaction inside the answer, the same direction as ChatGPT’s business profiles and wallet. And Bing putting ads beside its AI answers in nearly every market ends the grace period in which AI answers were an ad-free zone. The citation is still the ranking. Increasingly, it is also the checkout.
OpenAI’s jagged frontier: 79 percent at watching, 25 percent at building
The headline rate ran in yesterday’s paper—OpenAI’s agents now finish 46 percent of two-to-four-hour research tasks with no human intervention, up from 23 percent in March. The full report, published November 18, is worth the second day, because the breakdown is the news. From August 1 to October 18, zero-intervention success ran 79 percent on monitoring training runs and 72 percent on answering technical questions, but 35 percent on launching training runs and 25 percent on writing inference and serving code. Usage is compounding faster than capability: the median OpenAI researcher consumed $2,400 a day of agent inference at list prices in October, up 62 percent in a month; the 90th percentile hit $23,000 a day; agent-workdays per researcher rose from 0.5 to 5.7. In a survey of 101 research staff, the median researcher said they would trade 30 percent of their team’s compute, or in a separate question 40 percent of its headcount, to keep frontier coding agents for three months.
“Agents now complete nearly half of two-to-four-hour research tasks on their own. But humans are still central to AI development.”
This is the standing AGI beat done right: a lab measuring itself, in public, with the failures in the same table as the wins. The shape has a name in the report—jagged—and the shape is the finding. Agents are excellent where feedback is fast and mistakes are cheap, and weak exactly where the work is load-bearing. A lab that watches 79 percent of its runs alone and writes only a quarter of its serving code alone has moved its bottleneck squarely onto the load-bearing work. Watch the 25 percent number in the next report. When that one moves, the slope changes meaning.
Forward calendar: the November core update runs to about November 24; Linux 7.4-rc4 is expected Sunday; Anthropic’s second session today tests whether $97 is a floor or a waypoint. The Friday best-of builds later this morning from the full week’s log.