Weekly issue · Week of October 26th to November 1st 2026

The risk moved down the chain, to whoever was smallest

Anthropic's shows which suppliers carry the risk in its half-trillion of compute: the smallest, on terms it can leave. The market punished cash burn and split borrowers into two credit classes. An open model built Firefox exploits for $1,000 each, and nine developer platforms lost their customers' AI keys.

40 min read·Editorial by Nour Haddad

8 items · 100 citations · 50 primary · 41 · 7 weaker · 2 contested

The Week

Two stories ran all week, money and security, and by Sunday they had the same shape. Each time the cost of something rose, it landed on the smallest party in the chain.

Start with the money. Alphabet more than doubled its cloud business and had its worst day on record because it spent $9.1 billion more cash than it made in the quarter. Intel fell 29%, its worst session since the 1970s. S&P put Meta's credit on after the first negative free-cash-flow quarter in Meta's history. Microsoft, the one with spare cash, sold $25 billion of bonds into $115 billion of orders, and the market charged about 54 to insure its debt against Oracle's 275. Then Anthropic's public S-1 showed how its own half-trillion of compute is split. The three hyperscalers get minimum payments Anthropic owes whatever it uses. SpaceX, Nscale and Lambda get contracts Anthropic can leave or that pay only on delivery, and two of them must find their own financing. That week banks delayed selling the loan behind an Anthropic-leased Texas site, Nvidia began talks to guarantee Nscale's debt, and Blackstone told lenders it now expects Anthropic to be rated junk. SourcesA B B

Security followed the same path. Epoch AI showed an model turning each new Firefox release's public fixes into a working exploit for about $1,000. Wiz counted nine AI developer platforms breached in October, every attacker going for customers' model keys, and one of them, Zeabur, is shutting down. Google and Anthropic kept their strongest cyber models for vetted defenders. The people outside those programs, small developers and anyone who updates late, are the ones the cheaper attacks reach first. SourcesA A

What Changed

Ranked by how long we expect each to matter.

1. Anthropic's books are public, and the risk in them sits with its smallest suppliers. The filed October 26th reports first-half revenue of $16.29 billion against $1.16 billion a year earlier, an operating loss of $693 million, $3.9 billion of operating cash flow and $112.4 billion of cash. The net loss of $80.7 billion is mostly a $72.4 billion fair-value charge on as the company's rose, which is accounting, not spending. The two numbers a buyer most wants, aggregate committed spend and run-rate revenue at October 31st, are blank in the filing, waiting for an amendment. The Deep Read takes the commitments apart. SourcesA B

2. The market stopped paying for growth funded by negative cash flow. Google Cloud revenue rose 103% to $30.7 billion and AWS grew 39%, its fastest in nearly five years. Alphabet still closed down about 13% on October 29th, its worst day on record, and Amazon's trailing is now an outflow of $42.4 billion. Microsoft grew Azure 48%, produced positive free cash flow, and its shares rose . Intel's data-center unit grew 78% and the stock fell 29% on a guide its chief executive blamed partly on memory prices hitting PC demand. The Nasdaq lost about 5% in October, its worst month since March 2025, while the Dow rose. SourcesA A B B

3. An exploit for a freshly patched browser now costs about $1,000. Z.ai's open-weight GLM 5.3, working alone offline, built a file-stealing webpage for all six Firefox releases from 156.0 to 158.0.1 using only what was public at each release, fastest in about 12 hours, and never refused. The same week Anthropic published five exploit chains written by an unreleased model, including a Linux remote root reached in about 53 minutes, and Google released Gemini 4 Carbon to cyber defenders before anyone else. CrowdStrike says a China-linked group ran on a DeepSeek model against more than 200 organizations. Attempts against Chrome cost Epoch about $100,000 without success, so the floor is not uniform yet. SourcesA A A A

4. The AI developer's became the target. Nine platforms that host developers' code and agents disclosed intrusions in October, and Wiz says the common prize was stored keys for OpenAI, Anthropic and Google models. Zeabur's attacker read 3.2 million environment-variable records across 321,810 accounts, and the company closes on December 31st. A fake Claude Code extension on Cursor's marketplace planted backdoors on up to 5,000 machines in eleven hours. Cisco Talos found an that asks Claude what each victim is worth and falls back to Kimi or Gemini when refused. Two of the breached vendors say the goal was distillation. SourcesA A B

5. A 15,000-page proof was machine-checked in 24 days, and it had two real errors. Axiom Math's agents formalized the in about 43.1 million lines of between October 2nd and 25th, and found a counterexample to a step in a 1986 proof and a false equality in the quasithin volumes, both repaired with two of the classification's original authors. An off-the-shelf model with scaffolding proved 669 of 672 Putnam problems, and only 52.7% of 2,000 statements about real programs. OpenAI and Anthropic each published a claimed refutation of the Exponential Time Hypothesis, which outside groups are still examining. SourcesA A A

6. Human review of agent work looked weaker than the safety cases assume. In Anthropic's , 50 analysts followed an agent's coin-flip lean 86% of the time, and 77% repeated its arguments in their write-ups while 15% credited it. GitHub says agents started more than 40% of pull requests in September, up from 17% in March. Anthropic's own Claude agents wrote data that passed its production-style monitor and still made models misaligned, and METR found two headers that switched off OpenAI's internal monitoring for new agents. SourcesA A A C

7. The labor saving got booked before the system that earns it. The Netherlands plans to cut 25,000 civil-service jobs and calls it AI savings, with no deployment plan published. Comcast filed notice for 634 headquarters layoffs as agents take customer calls, Microsoft offered voluntary exits to about 60,000 employees, and NatWest cut about 1,000 jobs in the quarter. Computer-science enrollment fell 12% at four-year colleges. Comcast is also losing broadband customers, so not every cut is a machine's. SourcesB B C

8. Distillation became the explanation for China's cheapest . The Information reports DeepSeek and Moonshot trained their newest models heavily on GPT-6 Astra outputs, citing people with direct knowledge and no published technical evidence. The same week Wiz traced one stolen enterprise key making more than a million successful Bedrock calls in 13 hours, Moonshot closed its Kimi K2 models to new users, and DeepSeek's round closed at about 105.5 billion yuan with state investors holding 27%. If the reporting holds, China's cheapest frontier models owe part of their quality to a US lab that is not paid for it. SourcesB B A

Deep Read

Who signed for Anthropic's half-trillion

Anthropic's S-1 lists its compute deals one by one, and they come in two kinds. Every number below is from the filing unless marked. SourcesA

The first kind is owed to the three hyperscalers. Anthropic committed to spend at least $44.9 billion with Microsoft between November 2026 and May 2033, at least $114.5 billion with Google between April 2026 and July 2033, and at least $110.0 billion with Amazon between May 2026 and April 2036. All three are owed "regardless of actual usage." If spending falls short, Anthropic pays Google and Amazon the difference as a true-up. The Microsoft and Amazon commitments are non-cancelable except for the supplier's uncured breach, and Microsoft's also falls away for any that is never delivered. Together that is about $269.4 billion of minimums on our arithmetic.

The second kind is owed to newer, smaller suppliers, and the terms run the other way. Anthropic agreed to spend up to $120.5 billion with SpaceXAI, the former xAI compute business, into 2029, about $3.1 billion a month once all capacity arrives, expected in December. After non-cancelable periods end in January and March 2027, either side can cancel on 90 days' written notice. With Nscale it is up to about $44.6 billion into 2033, "with fees accruing only as capacity is delivered to and accepted by us," and Nscale "is required to use best efforts to obtain qualifying financing for the NVIDIA and data center infrastructure." With Lambda it is about $41.3 billion into 2034, again paid on delivery, and either party may end the $34.9 billion August agreement "if Lambda does not timely secure the facility lease." That is about $206.4 billion on our arithmetic, owed only when the hardware shows up, and partly cancelable.

Add the two groups and the named contracts come to about $475.8 billion. On top sit the Broadcom chip lease, a $63 billion vehicle financed by banks and Blackstone with Broadcom supporting the equipment's , the right to issue up to $42.0 billion of convertible notes to Broadcom, and an AMD agreement with no fixed dollar amount. The filing's own total reads "aggregate committed spend of more than $ ," and the blank will be filled in the amendment that carries the price range. Outside estimates of the total run from about $542 billion to more than $850 billion depending on what they count, which is why 2026-11-01-F1 bets on the filed number rather than on any of them. SourcesA C C

Now put the week's financing news against that split. Most of it landed on the second group or on the lenders behind Anthropic itself.

On October 28th Bloomberg reported that banks delayed selling part of a $15 billion loan for a Texas data center leased to Anthropic, after the state's grid audit froze new projects. The lenders keep the risk on their own books while they wait. The same day Bloomberg counted at least 18 data-center financings in the and leveraged-loan markets at double-digit yields. On October 29th The Information reported that Nvidia is discussing credit support worth potentially billions for Nscale's West Virginia campus, the one Anthropic rents. That is the clause about "best efforts to obtain qualifying financing" being worked out in public: if Nscale cannot borrow on its own name, its chip supplier is offering to stand behind it. On the same Thursday Bloomberg identified Anthropic as the "Customer B" behind 34.5% of SpaceX's third-quarter revenue, on the contract Anthropic can leave on 90 days' notice from early next year. And on October 30th the Financial Times reported that Blackstone, lender in the Broadcom vehicle, now expects Anthropic to be rated double-B after listing, a notch of junk, where it had expected before 2027. SourcesB B B B B

None of this is a mistake on Anthropic's part. A buyer with options signs firm minimums with sellers who will exist in 2036 and keeps an exit from sellers who might not. The filing says plainly why the exits matter: "our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years." Customers can cut usage at any point. Anthropic cannot cut the hyperscaler minimums. So the flexibility it has goes into the second group.

Follow that through a bad year. Suppose usage growth slows in 2027. The Microsoft, Google and Amazon minimums are paid first because they cannot be avoided. The capacity that gets trimmed is the capacity that can be: SpaceX on 90 days' notice, Nscale and Lambda tranches not yet delivered or accepted. The companies that would absorb that are the ones that borrowed at double-digit yields to build it, and their lenders and guarantors behind them. That would include Nvidia, if the Nscale guarantee is signed, and SpaceX, whose AI division brought in about $7.56 billion in the quarter with Anthropic's share inside it.

The hyperscalers are the cheapest borrowers in the system. Microsoft's trade near 54 basis points. The capacity Anthropic can walk away from is financed by the most expensive money in the system. The way the contracts are written, a slowdown would hit the weakest balance sheets first. SourcesB

There are three good reasons to think this ends well. Anthropic's first-half operating cash flow was positive at $3.9 billion, and it holds $112.4 billion of cash and $19.3 billion of short-term investments. Customers paying $1 million or more a year grew past 1,900 by September 30th. And Anthropic's revenue is the reason these suppliers exist at all: SpaceXAI's quarterly revenue rose nearly 830% on the back of it. If revenue keeps compounding at anything like the first half's pace, the cancel clauses never get used, the Nscale guarantee is a cheap way for Nvidia to sell chips, and this section will have described a risk that stayed on paper.

What would show the risk moving: an amended S-1 that converts any of the second-group contracts to minimum commitments, which would mean Anthropic has taken the risk back. Or the reverse, Nscale or Lambda borrowing at investment-grade spreads with no supplier standing behind it, which would mean the market has decided Anthropic's word is good enough on its own. Watch the amendment first. The filing gives the date it was drafted for: the field reads "as of October 31, 2026," and it is blank. SourcesA

Finance

Four theses, each with what would break it.

Thesis 1: The market now pays for cash, not cloud growth. The week's reactions sorted by cash more than by growth, with Amazon the partial exception. Microsoft grew Azure 48% with positive free cash flow and its shares rose about 2.5% after hours. Alphabet grew Google Cloud 103% with a $9.1 billion quarterly outflow and closed down about 13% (some reports give 12.6%) ?. Amazon grew AWS 39% with a $42.4 billion trailing outflow, dropped 8% after hours and closed only 1.4% lower, helped by Jassy's call for AWS growth above 40% in 2027 and 2028. Meta reported its first negative free-cash-flow quarter, $2.12 billion, and fell 15% for October. Falsifier: Alphabet guides 2027 capital spending above $240 billion in February (2026-10-28-F1) and the stock rises on the day. SourcesA A B

One-day share moves after the week's AI earnings

Closing moves as reported by CNBC, Yahoo Finance and The National. Alphabet is also reported at -12.6%. Microsoft rose about 2.5% after hours on October 28th; no close is shown.

Thesis 2: Credit has split the buildout into two classes of borrower. Microsoft's $25 billion sale drew about $115 billion of orders and priced its 40-year bond 1.05 points over . Five-year credit-default swaps put Microsoft at about 54 basis points, Meta at 116 and Oracle at 275. Below that sit the developers: Volta's $5 billion loan cleared above 11%, Morgan Stanley says the best-rated data-center bonds went from under 5% to near 7% this year, and Barclays is asking insurers to take default risk on more than $10 billion of its data-center loans. The five biggest spenders have borrowed $326 billion this year by S&P Global Market Intelligence's count; Bloomberg counts $313 billion of investment-grade bonds from hyperscalers and developers together. The two measures cover different issuers and we do not reconcile them. Falsifier: Nscale or a comparable developer borrows at investment-grade spreads with no supplier guarantee before spring. SourcesB B B B

Five-year credit-default swap spreads, October 29th

ICE data as reported by Bloomberg. A higher spread means the market charges more to insure against default.

Thesis 3: The profit pool moved to memory, and memory is now taking revenue from its own customers. SK hynix earned 83.33 trillion won of on 104.62 trillion won of revenue, about an 80% margin, and says and are effectively sold out through 2029. Samsung's chip division made 106.5 trillion won while its phone and appliance business lost 800 billion won because the memory it buys from itself got expensive. Nikkei counts five memory makers out-earning five large US tech companies for the first time. Intel's chief executive said PC customers "began signaling a PC slowdown based on the memory prices and availability." Kioxia was the one memory maker to guide below estimates. Falsifier: any of SK hynix, Samsung or Micron describes 2027 HBM supply as balanced or cuts its contract-price outlook before year-end, which 2026-08-06-F5 already tests for 2026. SourcesB B B

Thesis 4: Private money left the most expensive name and went to , robots and states. OpenAI cut its round to about $20 billion from $30 billion, with roughly $11 billion committed at a $1.25 trillion valuation. In the same days Baseten sought about $2 billion at up to $30 billion, 2.3 times its June price; Physical Intelligence was reported in talks at about $20 billion with no revenue; Cognition drew offers at $70 billion to $80 billion. Canada took up to C$1.6 billion of Cohere's round at C$35.7 billion, Chinese state investors hold 27% of DeepSeek's, and Aramco paid about $2.6 billion in cash plus assets for 49% of HUMAIN. Falsifier: OpenAI closes $25 billion or more by year-end (2026-10-25-F1). SourcesB C B A B

The pipeline. Anthropic's S-1 is public with no price range; Bloomberg reported in early October that marketing could start the week of November 9th, and an amendment with a range has to come first. Agility's vote is November 2nd. Nscale has filed and investors were reported wanting about 40% below its ask. No has priced since Firmus withdrew. SourcesB A

Media

Ranked by how much a listener learns that the week's briefs could not give faster. All were filed in the Media library during the week, and the note on each says how much of it we heard or read.

Sam Altman with John Collison, "Why Sam Altman is against a single superintelligence", Cheeky Pint About 50 minutes. Altman puts numbers on the record that the week's finance stories needed: about $10 million of compute on the proof, a $70 billion five-year cost per that he wants down to $20 billion, and DRAM makers turning $20 billion of capital spending into $100 billion of free cash flow. Two days later he was in Yongin meeting 50 Samsung executives. He also says sandboxing a far-superhuman model is unsolved. Full transcript read, not a full listening. SourcesA

Nilay Patel with Joey Brookhart, "Inside Anthropic's trillion-dollar AI gamble", Decoder About 45 minutes. A SemiAnalysis analyst who modeled Anthropic's finances before the filing walks through the S-1's economics, governance and . Use it as a second reading of the document our Deep Read takes apart, and check whether it treats the hyperscaler and neocloud contracts as one number. Analyst interpretation of a public filing. Show notes only. SourcesA

George Kurtz, "The New Apex Predator", No Priors About 45 minutes. Recorded a week before his company published VAPOR PANDA, CrowdStrike's chief argues agents flatten the attacker pyramid by handing advanced tradecraft to low-skill operators, and covers open versus closed models for attackers and . A vendor framing the threat to fit the product, which the week's evidence happened to support. Show notes and chapters only. SourcesA

Ankur Moitra, Steven Brunton and Alex Kontorovich, "AI Crisis in Math", Lex Fridman Podcast #503 About 3h43, so use the chapters: the AI stretches run roughly minutes 18 to 74 and the last 40 minutes. Three working mathematicians on whether models create mathematics, recorded the week Axiom formalized the classification and two labs claimed the . Chapter list and guest list only. SourcesA

Cliff Sosin, "Owning the Complements to Intelligence", Invest Like the Best About an hour. A concentrated public-markets investor argues intelligence has limits in chaotic systems and explains why he avoids the buildout and owns what agents will route customers toward. The week's best counterweight to talk, from someone with a position. Transcript opening and chapter list only. SourcesA

Katelyn Lesse and Angela Jiang, "How Anthropic ships Claude with Claude", Latent Space About 1h45. Anthropic's heads of engineering and product on Managed Agents keeping out of the agent through an egress proxy. In a month when nine platforms lost their customers' stored keys, that design choice is the practical half of the episode. The host's S-1 figures are inconsistent on the page. Show notes and half the transcript read. SourcesA

Editorial

NHNour Haddad
Deployment & the Floor

Your router is the last to hear about the patch

Nour Haddad

On October 29th Epoch AI published what an exploit costs now. It took a model anyone can download, Z.ai's GLM 5.3, and gave it what an attacker gets for free every time Firefox ships: the list of what was fixed. Working alone in a , the model built a webpage that reads files off a visitor's computer, for each of six Firefox releases. About $1,000 a version. Twelve hours at its fastest. It never refused. Firefox has more than 200 million active users. SourcesA

A security fix used to be the start of a race the defender usually won. The fix went public, a skilled attacker needed weeks to turn it into a working exploit, and in those weeks most people's browsers quietly updated. That margin was never a policy. It was a fact about how scarce exploit writers were. This week it went away for at least one major browser, at the price of a used laptop.

So who is still on the old version when the exploit arrives? Not the security team at a bank. They are in Google's Fairwind Program, the group getting Gemini 4 Carbon before anyone else, and on Anthropic's Cyber Verification Program list. The people on the old version are the ones who never restart Firefox because forty tabs are open. The dental office running an unsupported Windows machine because the X-ray software only works there. The family whose router has not updated since they plugged it in. The parent on a three-year-old Android phone the manufacturer stopped patching. Epoch's own advice to vendors names them: prioritize "infrequently updated software such as router firmware and phone operating systems." SourcesA

The labs have spent the month gating their strongest models by who is asking. Google opened Carbon to defenders first. Anthropic kept the model that wrote its Chrome, Safari and Linux chains to itself. OpenAI rates GPT-Rosalind-6 Critical for cybersecurity and sells it only to vetted labs. All three are reasonable. None of them protects the dental office, because the model that did the Firefox work is open, it runs on hardware nobody can switch off, and CrowdStrike watched a state group run a DeepSeek model against 200 organizations this month. A commercial host is already selling a refusal-stripped DeepSeek build behind a sign-up form. The update is the only gate the people at the bottom of this chain can rely on. SourcesA A A

My position: when a model can turn a published fix into an exploit in half a day, any device that cannot install a security fix within days of its release is defective, and the cost of that defect belongs to the company that sold it, not the person who owns it. Browsers are close: Firefox and Chrome already ship weekly. Phones and routers are not, and their makers set the pace for hundreds of millions of people who never chose a patch policy. Regulators already know how to do this. Europe's Cyber Resilience Act requires security updates for a product's support period; the deadline that matters now is how many days each update takes to arrive. SourcesA

Zeabur's 321,810 accounts had their secrets read, and the company is closing, so every one of those small teams is migrating by New Year's Eve. One stolen enterprise key made more than a million successful calls in 13 hours, billed to whoever owned it. The banks and the labs had security staff and early model access; the small teams had neither, and they got the breach. I wrote this morning that an with no spending ceiling is a blank check. SourcesA

Here is what would prove me wrong. Attackers have never been short of exploits for old bugs; plenty of breaches still run on flaws years past their fix. If cheap model-written exploits only add to a pile attackers already had, then the bottleneck was always victims who do not update, and nothing about the changed. The test is whether anyone catches a model-written exploit for a recently patched browser bug in the wild. I am logging that call at 55% by the middle of next year, 2026-11-01-T3. If it does not happen, I will have mistaken a lab result for a change in the street.

Until then, two things for Monday. Restart your browser, the one with forty tabs. Then find out when your router last updated. If you cannot find the answer, that is your answer.

Winners & Losers

Technology

Winner: . Axiom's 24-day classification found two real errors in published mathematics; &Truth raised a $70 million seed to prove code correct; a scaffolded model cleared Putnam. SourcesA
Winner: Google. Gemini 4 Carbon topped Artificial Analysis's index at 59 and launched on a defenders-first plan that set the industry's terms for cyber capability. Its hallucination rate rose to 11% from Argon's 4%. SourcesB
Winner: Mistral. Large 4's arrived on October 29th, inside the month Mistral named, and settled our call correct. SourcesA
Loser: anyone a week behind on updates. Epoch priced the patch gap at $1,000 per Firefox release. SourcesA
Loser: the human reviewer as a safety control. 86% deference to a coin flip, and a 40% agent share of pull requests at GitHub. SourcesA
Loser: extension marketplaces. A fake Claude Code update ran under Anthropic's own extension ID on Cursor's store for eleven hours. SourcesB

Business

Winner: the memory makers. Sold out to 2029 at SK hynix, the chief executive of OpenAI visiting Samsung in person. SourcesB
Winner: Cohere. A government took up to C$1.6 billion of its round at a record Canadian valuation, which is a customer and a backer in one. SourcesA
Winner: Microsoft. It was the only hyperscaler whose cloud growth came with positive free cash flow, and the bond market priced it accordingly. SourcesA
Loser: Zeabur and its customers. A breach that reached root on 101 servers ended in a shutdown notice. SourcesA
Loser: Moonshot. A distillation report naming its newest model, the retirement of the open K2 line from its , and weights still owed by November 4th. SourcesA
Loser: customer-service staff. Comcast's 634 Philadelphia headquarters notices, tied to agents taking customer calls, begin December 31st. SourcesB

Finance

Winner: SK hynix shareholders. Up almost 9% on the print, while the companies buying its chips fell. SourcesB
Winner: inference companies' private marks. Baseten at up to 2.3 times its June price. SourcesC
Loser: Intel. Down 29% in a day; the US government's roughly 10% stake lost about $12 billion. SourcesB
Loser: Alphabet equity. The second-largest one-day loss of market value in US history, by Bloomberg's count, on a quarter in which its cloud doubled. SourcesB
Loser: Meta credit. First negative free-cash-flow quarter, S&P negative outlook, $95 billion of planned equity. SourcesB
Loser: OpenAI's round. A third smaller in the week of its own deadline. SourcesB

Threads

The threads table in data/threads.md gets a November 1st section. What moved, and what did not:

Thread	This week
Cost of AI borrowing	Moved hard. Two credit classes: Microsoft at 54 basis points, Oracle at 275, 18 data-center financings at double-digit yields, Blackstone expecting junk for Anthropic.
Circular and financing	Moved. Nvidia in talks to guarantee Nscale's debt; Broadcom's named; SpaceX's largest customer is Anthropic.
The IPO queue	Moved. Anthropic's public S-1, with commitments and run-rate blank; no price range.
Private-market repricing	Moved. OpenAI to about $20 billion; inference and robotics marks up.
Memory squeeze	Moved. Memory makers out-earned big tech; Intel blamed memory for a PC slowdown.
Cheap offense	Moved hard. $1,000 Firefox exploits on open weights; Anthropic's unreleased chains; VAPOR PANDA on DeepSeek; Binance's $6.3 billion theft attributed to North Korea.
Supply-chain compromise of AI tooling	Moved. Nine platforms breached, Zeabur closing, a fake Claude Code extension.
Self-reports and independent checks	Moved. OpenAI's monitoring account, with METR's switch-off headers; Anthropic's monitor beaten by its own agents.
Verification and credit disputes	Moved. The classification formalized with two real errors; two labs claim the same refutation.
Open-weight commoditization	Moved. Large 4 shipped; distillation reporting; Kimi K3.1 Pro weights due November 4th.
Entry-level work and labor substitution	Moved. The Hague's 25,000, Comcast's 634, CS enrollment down 12%. Payrolls print Friday.
Consent and access	Moved. Four labs back one training opt-out; Reddit forces logins.
Power as the constraint	Moved. Victoria's 79 councils call for a halt; Hermantown approved Google; Anthropic lists power delay as a risk factor.
Pacing and exposure	Moved slightly. The Justice Department is asking about Nvidia's purchase.
Google's talent drain	Dormant.

Ledger

Settled this week: three right, three wrong. Brier for the six is 0.235, a hair better than a coin flip. SourcesA

Call	Confidence	Result
`2026-09-29-F1` Anthropic's public S-1 confirms the leaked revenue figure	75%	Right: fiscal 2025 revenue of $4.585 billion
`2026-10-25-F2` Anthropic's public S-1 appears on by January 31st	55%	Right: October 26th
`2026-10-06-T1` Mistral publishes the Large 4 weights by October 31st	75%	Right: October 29th
`2026-10-11-F2` Anthropic's public S-1 misses October 31st	55%	Wrong
`2026-09-27-F2` DeepSeek's round misses its own October target	60%	Wrong
`2026-10-04-T3` Huawei publishes the Atlas 850E developer kit by October 31st	65%	Wrong; rolled forward as `2026-11-01-T1`

The table, read honestly. Across 32 resolved calls the book scores 0.232. Calls made at 70% have come in at 40% (two of five), so that bucket is overconfident, and the October wrap found why: we kept betting that someone would wait. Calls at 60% came in at 68%, slightly too cautious. The response is to keep taking risk and to price it lower, which is what the six calls below do. Four are at 60% or below, and one is a bet on an event nobody has yet reported seeing.

Watching. 2026-08-15-T1 (Apple's own China model reaches Chinese users this year) is now contested in its own terms: Apple Intelligence reaches China in December, but with Alibaba's Qwen answering part of Siri, and the call needs Apple's own China-registered model. We will score it against what Apple says ships in December. 2026-09-20-F1 (Anthropic's IPO misses its November target) looks stronger with no price range on November 1st. 2026-10-24-T1 (Kimi K3.1 Pro weights) settles Wednesday. SourcesB

New calls.

2026-11-01-F1 Anthropic's amended S-1 puts its total commitments above $600 billion (60%). The named contracts alone come to about $475.8 billion before the Broadcom lease and AMD. Wrong if the figure the amendment fills in is $600 billion or less, or if no amendment fills it in by January 31st.
2026-11-01-F2 Anthropic does not price its IPO at $1.5 trillion or more by March 31st (60%). Investors called $1.8 trillion to $2 trillion fair in October, before the selloff and the junk-rating report. Wrong if it prices at an implied valuation of $1.5 trillion or above.
2026-11-01-F3 Meta keeps its AA- rating at S&P through June 30th 2027 (75%). S&P's own downgrade test runs to 2028. Wrong if S&P lowers the rating before July.
2026-11-01-T2 Gemini 4 Carbon opens to all paying developers by December 31st (60%). Google promised "as soon as possible." Wrong if it is still gated to programs or trusted partners at year-end.
2026-11-01-T3 A model-written browser exploit is caught in the wild by June 30th 2027 (55%). The editorial's test. Wrong unless a vendor, CISA or a named security firm reports in-the-wild exploitation of a Chrome, Firefox, Safari or Edge flaw with an exploit it says a model built.
2026-11-01-B1 Two more AI developer platforms disclose stolen customer keys by year-end (65%). Wiz counted nine in October. Wrong if fewer than two platforms outside that list disclose such a breach in November and December.

Next Week

Monday, November 2nd. Apple reports, its first quarter under John Ternus, with Apple Intelligence's China beta two days away. Palantir reports. Agility's SPAC holders vote. SourcesB
Tuesday, November 3rd. US midterm elections, with OpenAI's election operations center staffed through the count and generated disinformation already circulating. AMD reports, the next read on demand after the capex selloff. Senator Warren's deadline for Anthropic's answer on existential risk.
Wednesday, November 4th. Kimi K3.1 Pro weights are due (2026-10-24-T1, 70%). Apple Intelligence testing starts in mainland China in the iOS developer beta. Qualcomm, Arm and AppLovin report. SourcesA
Thursday, November 5th. Cloudflare reports, two weeks after its Workers AI exposure. Agility is expected to begin trading as AGLT.
Friday, November 6th. October , the first of the two prints 2026-10-04-B7 needs for information-sector jobs.
Week of November 9th. The earliest marketing start Bloomberg reported for Anthropic. Look for an amended S-1 with a price range, the aggregate commitment figure and run-rate revenue filled in. SourcesB
November 14th. 2026-08-13-B1 (DeepSeek holds its August price increase) and 2026-08-16-F2 (CXMT stays China's most valuable listed company) settle.

Sources

Primary A

Established outlets B

Secondary C

Contested figures ?

Alphabet's October 29th close: about -13% in most reports, -12.6% in others.
Anthropic's total commitments: blank in the filing; about $475.8 billion of named contracts on our arithmetic; outside estimates from about $542 billion to more than $850 billion.
AI borrowing in 2026: $326 billion of debt raised by the five biggest spenders (S&P Global Market Intelligence) against $313 billion of investment-grade bonds from hyperscalers and data-center developers (Bloomberg). Different issuers and instruments; not reconciled.

October 25th 2026