THE ADVANCE: PUT THE AGENTS TO WORK, AND THEY SCORE 12%
The most useful number of the cycle did not come from a lab. It came from a benchmark that pushed labs to do real work. On November 2, SemiAnalysis released GDPx, 206 agentic tasks across 11 domains chosen for economic return — chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering — built with 14 specialized data partners, with tasks running up to eight hours in realistic environments. SemiAnalysis’s own line is the headline: “GDPx is a true frontier-level benchmark: even the best model scores just 12%.”
The best model is Anthropic’s Claude Opus 5.5 at 12.0%. Google’s Gemini 4 Carbon is second at 10.9%, OpenAI’s GPT-6 Astra third at 9.6%, and Meta’s Muse Wonder Full fourth at 9.0%. In finance and legal, SemiAnalysis says that, under binary grading, essentially every model scores exactly 0. Two caveats SemiAnalysis itself states, and both matter. First, Opus 5.5 and GPT-6 Astra are, in its words, “sandbagged” by their safety classifiers in cyber, biology, and alignment — exclude those three domains and the ranking shifts. Second, this is SemiAnalysis’s own product, run by SemiAnalysis, with partner-built tasks that are largely private; no outside group has re-run it. Cost figures come via secondary reporting of the leaderboard: Opus 5.5 at $51.41 per task, Carbon at $16.29 per task — Carbon is nearly as good on this test at roughly a third of the per-task cost, which may be the more durable finding. (SemiAnalysis post — primary source; leaderboard cost details via Unite.AI — secondary.)
Anthropic, meanwhile, filed the document that details the frontier costs. On November 2, it filed its first S-1 amendment (S-1/A), adding third-quarter results to the October 26 public S-1. The amendment, as summarized from SEC EDGAR by the tracker we read, reports: Q3 revenue of $16.96 billion, up from $11.56 billion in Q2 and $4.73 billion in Q1; nine-month revenue of $33.2 billion; a Q3 operating loss of $567 million and a net loss of $23.3 billion, mostly a $21.9 billion non-cash change in the fair value of convertible notes, derivatives, and SAFEs; and about $895.4 billion in committed spending, substantially all on compute and infrastructure. It also adds a new risk factor: internal forecasts show future models may soon meet Anthropic’s own Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI” — the company’s filing-language way of saying recursive self-improvement is getting close by its own threshold.
The amendment does not include a price. The price range and share count are still blank. Reported expectations, not filing facts: pricing around November 19–20 in some reports, a roadshow that cannot lawfully start before November 10 (15 days after the public S-1), and institutional investors pushing for a valuation of about $1.5 trillion to $1.6 trillion, down from the roughly $2 trillion discussed in September. (S-1/A figures via a secondary tracker summarizing EDGAR — flagged; the filing itself is the primary document, and the tracker separates filed facts from reported ones.)
Anthropic’s other release was smaller, stranger, and aimed at last week’s question — who decides, and how would you know? On November 2–3 it released FlashProve, an open-source research prototype for “provable inference”: a lab publishes a cryptographic fingerprint — a commitment — of one exact set of model weights, and each response can carry a proof that it came from that model, without revealing the weights. The target failure is named plainly in the reporting: a provider secretly swapping in a cheaper model behind the API you paid for.
The numbers, from Anthropic’s own report via secondary coverage: tested on a 4-bit version of Meta’s Llama 4 Maverick (about 400 billion parameters), proofs of about 19 MB per 1,024-token request, generated at about 110 tokens per second on a single Nvidia H100, at 70 to 95 times the GPU time of serving the same model normally. And the limits, also plain: it is a research prototype, the code is described as not maintained, and no lab — including Anthropic — has plans to deploy it. A proof that you got the model you paid for, at 70-plus times the cost of the answer, that nobody is deploying: that is where verification actually stands. (Anthropic research post via secondary reporting — flagged.)
One deadline lands today, unresolved as we write. Moonshot promised that the full weights of Kimi K3.1 Pro would be released by November 4 — its 2.8-trillion-parameter (104 billion active) flagship, 1-million-token context, $2 per million input tokens and $10 output through the API, with cyber-related data deliberately removed for the open release, cutting its own internal exploit-benchmark score from 90.5% to 81.0%. As of the November 3 check, no K3.1 repository had appeared on Moonshot’s Hugging Face organization and no license had been published, and a Beijing-morning November 4 check reported the same. The day is not over. We score this in Prediction Watch below as pending, not missed — yet. (Moonshot post — primary source; Hugging Face checks via secondary trackers — flagged.)
CHIPS & IRON
Five days after Intel CEO Lip-Bu Tan told his earnings call that Intel was evaluating “unique and highly differentiated opportunities to accelerate our Ohio build-out” and hoped “to share more on this in the near future,” a name finally attached itself to the sentence — from Seoul, not Santa Clara. On November 3, the Seoul Economic Daily reported that SK hynix is considering producing DRAM at Intel’s New Albany, Ohio plant, with options including leasing part of the plant or forming a joint venture with Intel and major U.S. cloud companies, according to industry officials. The paper notes a fact that makes the idea more than a rumor: in 2023, the U.S. Foreign-Trade Zones Board approved production activity at the New Albany site whose finished products included “electronic memory circuits” — memory was already permitted there. It would be SK hynix’s first front-end DRAM fab in the United States, alongside its Indiana HBM packaging plant targeted for 2029 mass production and a New York NAND plant pursued through its Solidigm subsidiary. The hurdles are also in the report, and they are real: DRAM lines need different processes and equipment from Intel’s logic lines, so someone has to pay for a redesign; SK hynix says its DRAM and HBM output is effectively sold out through 2029, so Ohio production would start after that; and an SK hynix official said there was nothing specific the company could confirm about cooperation with Intel or DRAM production in the United States. Nothing is signed.
A secondary summary of the report notes Reuters first reported Intel–SK hynix talks in September — so this is a two-month-old conversation getting more specific, not a deal. For New Albany, though, it is the first concrete candidate answer to the question the earnings call left open: who else might pay to fill a fab Intel delayed to 2030 and cannot fill alone? A memory tenant brings its own demand, customers, and money. Watch for a signed agreement that says who pays for the equipment. Until then: reported, plausible, unconfirmed by either company. (Seoul Economic Daily — primary reporting; SK hynix non-confirmation is in the same report.)
Nvidia, meanwhile, is reportedly writing export control into its sales contracts because the law has not. On November 3, the Financial Times reported — via secondary summaries of the FT report — that Nvidia is telling cloud providers who rent out its chips that at least 51% of their services under new contracts must go to customers outside China, capping Chinese customers at 49%. The reported details: it applies to future contracts only, not existing leases; it was passed verbally through sales and regional teams, not in writing; and it is the first such limit Nvidia has imposed. Nvidia neither confirmed nor denied the report, saying in response that data centers and clouds built on American chips are the foundation American model and application makers need to compete worldwide. The context is the loophole everyone knows: U.S. rules bar selling the most advanced chips to Chinese companies, but Chinese developers have rented the same compute from data centers outside China. A bill to bring that remote access under export controls has stalled. So the chip seller is doing privately, by contract, what Congress has not done by statute — with a percentage it chose itself, communicated verbally, that no regulator published. Flag it twice: the rule is reported, not confirmed, and if it is real, its enforcement is exactly as visible as a verbal instruction. (FT via secondary reporting — flagged; Nvidia’s non-denial is in the same reporting.)
OpenAI’s infrastructure news was about the grid, not the chips. Its Effingham County, Georgia page for Project Camellia says OpenAI is contracting with Georgia Power for 3.2 gigawatts delivered in phases between 2028 and 2032, will pay the full cost of the infrastructure and electric service required, will use a closed-loop water system, will provide $80 million in community benefits over the life of the project plus up to $71 million in Codex credits for Georgia students, and — the new part — that the project will be “one of the first data centers designed to proactively reduce its power consumption before residential customers are impacted during periods of high demand.” Secondary reporting on the companion Grid Protectors announcement puts a number on that flexibility: shedding up to 1,000 megawatts during regional emergencies at the Georgia campus.
THE DEBATE: THE ELECTION AFTER THE ELECTION
The midterms were yesterday, November 3. As of this morning, early reporting says Democrats regained control of the Senate, with the House still being counted in some reports — treat the House as uncalled here, because the sources we read conflict on whether it is projected or decided. The subject doesn't need a final count. AI, and specifically data centers—their electricity, their water, their tax breaks—was a campaign issue this cycle in a way it never has been before, and both parties’ strategists are now on record saying so before the results were even in.
The clearest statement came the day before the vote, from inside the industry. On November 2, OpenAI chief global affairs officer Chris Lehane made public a memo titled “2028: The AI Election”, arguing AI will be the number-one issue of the 2028 presidential race and laying out five questions candidates must answer — safety and human control, taxes, children and jobs, affordability, and the race with China. His framing is a “compelling third path” between letting AI advance “without speed limits” and keeping control “in the hands of a self-appointed elite.”
The polling he cites — his own and polling he says he has seen, so label it as an interested party’s numbers: 65% of voters said AI would be important to their 2026 choices; by early October, one in five ads aired by Democratic candidates or groups mentioned data centers; 66% of voters support continued AI development, including 53% who want stronger safeguards and oversight, while 24% favor a pause. Secondary reporting of the memo adds the warning underneath it: a pledge to halt data center construction could become, in Lehane’s phrase via Politico’s full publication, the 2028 version of the 2002 Iraq war vote. And OpenAI’s own sequencing tells you this is a campaign, not a memo: reporting by Axios, carried secondarily, says the company launches its Washington push for mandatory “speed limits” on the most powerful systems in the lame-duck session that opens November 9 — the same week Anthropic’s roadshow can earliest begin. One lab goes to Congress, the other to investors, on the same Monday.
The other side’s polling points the same direction, which is why this is a debate about who owns the issue rather than whether it exists. A Guidant Polling and Strategy survey of 21 swing districts, reported November 2, found voters trust Democrats over Republicans on AI regulation 43% to 36%, independents 37% to 26%; 80% said they would be more likely to back a candidate supporting “common-sense” AI regulation, and 66% said they would be less likely to back one who opposes guardrails. A poll commissioned by one side, reported by an outlet sympathetic to the other — the cross-pressure is itself the evidence: AI safety is now a battleground-district issue with numbers attached, argued over by professionals who expect it to move votes.
Add the week’s third document and the triangle closes. Anthropic’s S-1/A — a securities filing, not a campaign text — chose this week to tell public investors, in risk-factor language, that its own models may soon cross its own recursive-self-improvement threshold, while a senator’s October 27 letter to Dario Amodei demanding answers on existential risk, bioweapons, and a near-$2 trillion valuation had a reply deadline of November 3, election day itself, per secondary coverage of the letter. No public reply has been reported as we write. The debate is no longer labs versus critics. It is labs versus labs versus Congress, each writing in a different genre — memo, filing, letter — about the same question Lehane put first: how do you make sure AI stays safe, and people stay in control, and who gets to say what that means?
PREDICTION WATCH
Dario Amodei (Sept 12): agent swarms able to “take over the entire internet” within 6 to 12 months. Bearing news: GDPx is the hardest evidence yet against the timeline — the best agent system in the world completes 12.0% of realistic professional tasks, and 0% in finance and legal under binary grading. Swarms that cannot finish a finance task won't take over the internet by March. Check-in: March 2027.
Sam Altman (to Fortune, mid-September): no OpenAI IPO in 2026. Bearing news: still holds, and the sequencing sharpened — OpenAI’s public move this week was a political campaign for the lame-duck session, not a filing. Anthropic has amended but not priced. Check-in: December 2027.
“Intelligence explosion” paper authors (Sept 28): months-long AI research projects automated by mid-2028; one developer commanding a workforce equivalent to “millions” of top researchers. Bearing news: the strongest support yet, and it came from a securities filing — Anthropic’s S-1/A says internal forecasts show future models may soon meet its own Responsible Scaling Policy definition of dramatically accelerated, AI-attributable progress. A risk factor is not a demonstration, and GDPx’s RSI domain is one of 11 in a benchmark the best model scores 12% on overall. Both facts stand. Check-in: June 2028.
Elon Musk (Sept 25): xAI’s Colossus 2 roughly doubles to 1.2M+ Nvidia chips by year-end (his own post — unverified). Bearing news: no verified chip count seen this cycle. Check-in: January 2027.
Rep. Ro Khanna (letter dated Sept 30): “The theft of an advanced model weight by a hostile non-state actor could endanger all of humanity.” Bearing news: no theft reported. The access question moved by contract instead — Nvidia’s reported 51% rule aims at Chinese firms renting advanced compute abroad, the rental equivalent of the weights leaving. Check-in: as events warrant.
Jay Clayton / Super Intelligence Force (Oct 4, via WSJ): the Force has 120 days — until about February 1, 2027 — to deliver its report on AI risks and opportunities. Bearing news: no public move seen this cycle; the week’s governance documents were a campaign memo, a securities filing, and a reported verbal sales rule. Check-in: February 2027.
Industry insiders via Axios (Oct 9): a major AI incident within 6 to 12 months — a cyberattack on financial services, internet, power, or water. Bearing news: no such incident this cycle. Check-in: April 2027.
Jensen Huang (repeatedly, on export controls): U.S. curbs are ceding China’s $50 billion AI chip market to Huawei. Bearing news: his own company is now reported to be extending the curbs itself — the 51% rule would cap the rental workaround his customers’ customers use, which either undercuts his argument or proves his point about demand leaking around every wall, depending on whether Chinese rental demand moves to Huawei-powered domestic clouds instead. The next test remains the IDC half-year numbers. Check-in: January 2027.
Jensen Huang (BG2 podcast, Oct 15): annual AI infrastructure spending reaches $3 trillion to $4 trillion by late this decade, and Nvidia’s content per gigawatt keeps rising. Bearing news: Anthropic alone now files $895.4 billion in compute and infrastructure commitments — one customer, albeit over years, at a scale that makes his range harder to dismiss. Check-in: November 17, 2026, on Nvidia’s Q3 results; full check-in 2030.
Sources via Reuters (Oct 19–20): Anthropic’s public debut lands in mid-November, after the midterms, and OpenAI lists in 2027. Bearing news: the first half is now “late November” in the telling. The S-1/A landed November 2 with the price range still blank; reported pricing dates cluster around November 19–20, the roadshow cannot start before November 10, and reported valuation talk has cooled to $1.5–$1.6 trillion from about $2 trillion. Still scorable as on-track for a November trade — barely, and only if the next amendment carries a range. Check-in: December 2026.
Demis Hassabis (to the Wall Street Journal, October 2026): AGI within two to three years, by his demanding full-cognitive-range definition. Bearing news: GDPx cuts against the short end of his window on the evidence that matters — deployment-grade work. A benchmark where finance and legal score zero for every model isn't what two years to full cognitive range looks like, unless the gains are about to be very discontinuous. Check-in: October 2027.
Moonshot AI (Oct 23): the full Kimi K3.1 Pro weights will be released by November 4, 2026. Bearing news: pending on deadline day. Checks on November 3 and on the morning of November 4 (Beijing time) found no K3.1 repository on Moonshot’s Hugging Face organization and no published license. If the weights post today U.S. time, this scores as kept; if not, next edition scores it missed. Check-in: November 5, 2026.
Jack Clark, Import AI (Oct 26): ProgramBench will be saturated by the end of 2027. Bearing news: no new ProgramBench read. The adjacent evidence is GDPx’s design premise — SemiAnalysis built it precisely because saturating older benchmarks stopped meaning the work was done, and its finance example buries the model in roughly 4,000 files instead of spoon-feeding the prompt. Check-in: January 2028.
Andy Jassy, Amazon CEO (Oct 29): AWS “can grow north of 40% in each of 2027 and in 2028” and be “a trillion-dollar revenue business in the next five to 10 years.” Bearing news: no new AWS figures this cycle. Check-ins: January 2028 on 2027 growth; the trillion-dollar leg by 2031–2036.
Lip-Bu Tan / Intel (Q3 call, Oct 29): the server CPU market (TAM) expands from $40 billion to $200 billion over the next five years. Bearing news: the Ohio-specific sequel to his call was about memory, not CPUs — the reported SK hynix DRAM discussions would fill New Albany with a tenant whose product is the other half of the AI server. It neither confirms nor dents the CPU forecast; it suggests Intel’s path to filling Ohio may run through partnerships rather than its own product lines. Check-in: 2031, with interim reads each earnings season.
Evercore and Bank of America, via Fortune (Oct 30–31): AI-related capital expenditure reaches $1 trillion in 2027. Bearing news: Anthropic’s $895.4 billion in filed compute commitments is a multi-year commitment total, not 2027 capex — do not add them together — but as a single company’s promise it shows the order of magnitude the banks are extrapolating from. Check-in: February 2027, on full-year 2026 results and 2027 guidance.
NEW — Chris Lehane, OpenAI chief global affairs officer (memo made public Nov 2): AI will be the number-one issue of the 2028 presidential election, and the candidate who articulates a “compelling third path” on it will have “an opportunity to define a new contract for this new era.” This is a prediction from an interested party selling the agenda it predicts — which does not make it wrong. The midterm evidence he cites (one in five Democratic ads mentioning data centers by early October) is checkable in principle and was directionally confirmed by both parties treating data centers as a live issue on November 3. Scored on the 2028 cycle, not before. (Memo via secondary reporting of the Politico-published text — flagged.) Check-in: November 2028.
NEW — Bob Peck, New York Angels board member (CNBC, reported Nov 4): Anthropic’s IPO could price near $1.5 trillion, at the low end of the reported $1.5 trillion to $2 trillion range, and become a yardstick for hyperscalers valued at $2 trillion to $4 trillion. Against it, in the same reporting: University of Florida IPO scholar Jay Ritter counts only 14 offerings out of 4,110 priced between 1980 and 2024 above 40 times trailing sales — and $1.5 trillion is about 43 times Anthropic’s trailing sales, which would put this listing in that rare cohort on day one. A named investor’s dated, falsifiable pricing call, made before the price range is even filed. (COINOTAG reporting the CNBC appearance and the amended filing — secondary.) Check-in: pricing night, expected late November 2026.
ONE NUMBER
$895.4 billion. The committed spending — substantially all of it compute and infrastructure — that Anthropic’s November 2 S-1 amendment lists, per the EDGAR-summarizing tracker we read (secondary reporting of a primary filing, labeled as such). One company. Not annual spending, not a forecast by an analyst: commitments, in a securities filing, by a company that took in $33.2 billion in the first nine months of the year and lost $23.3 billion in the third quarter alone, mostly on paper. Last edition’s number was $828 billion for five companies’ 2026 capital spending. This edition is larger, and it belongs to one company that hasn't yet sold a single public share. Whatever you think the advance is worth, that is what it costs to stay in it.
WORTH YOUR TIME
SemiAnalysis’s GDPx announcement — read it before you read anyone’s summary of a model release this month, including ours. It is a benchmark vendor selling a benchmark, and it says so by its structure; it is also the most concrete attempt yet to define what “the agents are coming for professional work” would actually have to mean: 206 tasks, real environments with thousands of files, eight-hour limits, grades that do not spoon-feed. Its worked example — rebuilding a real private-equity deal, a fund buying about 62% of a snack manufacturer for $880 million plus an earnout, across roughly 4,000 anonymized files, and asking the model to do an associate’s actual task inside it — explains in one page why every model scores zero in finance, and why a high score on a test that hands the model its context in the prompt is not the same claim. Twenty minutes. When the next lab announces a number in the high 80s, you will know which question to ask. (SemiAnalysis — primary source for its own benchmark.)
Enjoyed this roundup?
Subscribe for a clear, signal-over-noise take on the biggest developments in AI, technology, and the shifting power dynamics behind them.
klaatu barada nikto: the safe phrase to prevent world domination.
