| |
Since last week: Anthropic filed to go public, and Google's parent, Alphabet, lost roughly $500 billion in stock market value in one day. Microsoft offered buyouts to about 60,000 employees, and attackers are now confirmed to be using the PHP flaws from our last issue. |
|
In this issue
01 · The One Thing · 02 · Three Numbers · 03 · Product · 04 · Do This Week · 05 · The Signal · Skim ends here.
06 · The Margin-Proof Tracker · 07 · The Synthesis · 08 · Where the Minds Disagree
Then: What We're Watching · Worth Your Time · Corrections
|
|
01 · The One Thing
Anthropic, the maker of Claude, filed to go public on October 26. It is the first public look at a top AI lab's finances.
Anthropic reported $16.3 billion in revenue for the first half of 2026, up from $1.2 billion a year earlier. Customers who paid it a year ago now spend about nine times as much as a group, it says (Pro and Max plans excluded). Its operating loss was $693 million for the half. The second quarter alone made a $974 million operating profit, after five losing quarters. Most of its $80.7 billion net loss was an accounting charge, not cash. Its co-founders will keep 50.1% of the vote. (Anthropic S-1 · Anthropic)
|
| |
The executive shift: Your AI vendor is locked into years of compute bills. More buyers are locking in too. Its compute deals "generally extend over multiple years," and The Register puts those obligations at more than half a trillion dollars. Most of its revenue rises and falls with use. But "a growing number of enterprise customers" now sign non-cancelable contracts with minimum spend, the filing says. Blackstone, the anchor investor in a $63 billion deal to finance Anthropic's chips, now expects Anthropic to get a "junk" credit rating after the listing, the FT reported. A vendor in that spot needs your usage to grow. Know your minimums, how fast your use is growing, and what switching would cost. (Anthropic S-1 · The Register · TechCrunch on the FT) |
| $500 billion | Roughly the stock market value Alphabet lost on Thursday, October 29, in its worst day since it began trading.Why it matters: Google's search ad growth slowed to 12%, and Google said ads in AI Mode, its chatbot-style search, earn less. Even an AI leader now has to show that AI will not eat its core profit. ( AP · TechCrunch · CNBC) |
| 3% | The share of organizations that exceed their cost-savings goals through productivity, Gartner says, while 77% of board members expect AI to deliver those savings. Also, 38% of companies are over budget on AI, and 85% plan to spend more next year.Why it matters: this measures savings, not every kind of value from AI. But boards are counting on savings, and very few companies are beating their targets. ( Gartner) |
| 60,000 | The Microsoft employees offered buyouts, about 27% of its staff, AP reported.Why it matters: a top seller of workplace AI wants to cut people costs while it spends heavily on AI. Its own post never mentions AI, and it is an offer, not a layoff (see 08 · Where the Minds Disagree). ( Microsoft · AP · TechCrunch) |
The Gemini app reached one billion users, Google says (Oct 28). A day later, every user could create and edit documents, spreadsheets, slides and PDFs inside the chat. For leaders: the free assistant your staff already carry now makes finished work files. Decide which company files may go into it. ( Pichai's remarks · Google)
Google put ads in AI Mode in 10 more countries, including Canada, Australia and India, and added ads to follow-up questions in the US (Oct 28). Brazil, Mexico and the UK are planned for later this quarter. For marketers: the AI answer is becoming paid space. ( Google)
Claude Enterprise now shows what each team's AI spend produced, with one shared budget per team (Oct 28). For finance: a start on tying AI spend to output, though the reports come from the vendor. ( Anthropic)
Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, built for cybersecurity and long, multi-step tasks (Oct 28). For now only a small group of cyber defenders can use it. For security leaders: ask your vendors if they will get access. ( Google DeepMind)
1 · Ask finance for one page on AI spending by Friday: each vendor, the last 12 months, the growth rate, any minimum commitment and the contract terms. Owner: your chief financial officer. Our read: add one line per vendor on how you would switch, and get a competing quote before the next renewal. Stakes: more enterprise buyers now sign minimum-spend contracts, Anthropic's filing shows (see 01 · The One Thing). Rival offers are reportedly getting cheaper (see 05 · The Signal).
2 · Before any AI tool switches on as "included," name who owns its budget and set a cap. Owner: your chief information officer, with finance. Our read: ask each vendor what stops working when the budget runs out. Stakes: GitHub and Oracle are moving AI to metered billing (see 05 · The Signal).
3 · List every AI coding tool, add-on and agent that can reach your code, cloud accounts or passwords, and who approved each one. Owner: your chief information security officer. Our read: keep an approved list, limit what each add-on can reach, and get alerts on new versions so you can roll one back fast. Blocking all updates would also delay security fixes. Stakes: this week a tampered Claude Code add-on reached up to 5,000 users (see 05 · The Signal).
Horizon Wall Street sent the AI bill back. Investors want proof that the spending will return cash. Primary: Alphabet CEO remarks. Reported: AP, TechCrunch, CNBC. Alphabet's revenue grew 25%, Sundar Pichai said. But free cash flow was negative $9.1 billion as quarterly capital spending hit $51.4 billion, TechCrunch reported. Amazon's cloud sales rose 39%, yet it burned $42.4 billion of cash over 12 months, AP reported. Microsoft plans to raise $25 billion in bonds. S&P cut Meta's credit outlook, doubting its "ability to monetize its AI investments." Last week we asked whether the big three would raise spending and put a number on what AI earns. Amazon will "spend meaningfully more in 2027." Alphabet held its 2026 forecast but expects more in 2027, and Microsoft trimmed its forecast. Our read: expect vendors to push harder to make AI pay. Your bill is one place they will look. ( Pichai's remarks · TechCrunch · AP on Amazon · TechCrunch on Amazon · CNBC on spending · CNBC on Meta · AP on Friday)
Trust The newest way into your company is an AI tool your staff installed. Reported: Ars Technica, The Verge, WIRED, The Register. Primary: CISA. A tampered copy of Anthropic's Claude Code add-on sat in Cursor's store for about 11 hours on Thursday. It came through Anthropic's official account there, researchers said, and add-ons there update on their own by default. It tried to plant a hidden remote login. Up to 5,000 users may have installed it, Cursor said. Separately, a contractor at Accenture linked a Defense Department system to a personal Claude Code account, and the agent appears to have sent out login details, WIRED reported, citing sources. Accenture said the claims were "found to be without merit." A poisoned version of LiteLLM, a popular AI gateway, was downloaded about 49,000 times in under two hours, The Register reported. And CISA now lists both PHP flaws from our last issue as exploited. Our read: what an AI tool can reach matters more than which model it runs. ( Ars Technica · The Verge · WIRED · The Register · CISA)
Work AI now comes built in, and you pay by the meter. Primary: GitHub. Reported: The Register. On October 28, GitHub began including its Copilot coding assistant with every GitHub Enterprise seat, for customers who work with an account team. Use is "metered from the first request against your enterprise budget." When the budget runs out, chat and code review stop; basic code completion keeps working. Oracle will add charges for AI work on top of its subscriptions. "We don't have to add more seats in order for us to now generate more revenue," an Oracle executive said. Our read: when AI comes included, someone has to own the budget. ( GitHub · The Register on Oracle)
Deploy OpenAI is reportedly using discounts to win Anthropic's customers. Buyers can use that. Reported: The Information, via TNW. Vendor claim: OpenAI. Replit moved most of its AI spending to OpenAI after OpenAI offered more than 30% off, The Information reported. The contract is worth over $100 million a year. Replit's chief executive said cost, partnership and performance all drove the choice. At Rippling, OpenAI's share of coding work rose from 16% in May to 52%, while Anthropic's fell from 57% to 18%. TNW said it had not verified the figures. OpenAI also released GPT-6.1 Terra, a mid-priced model at half the price of the previous Terra model, it says. Our read: AI prices look negotiable right now. Anthropic's revenue ran nearly double its compute costs in the first half, its filing shows, so there may be room to bargain. ( TNW · The Register)
▼ Below the Cut
AI agents will do our shopping, the pitch goes. At DoorDash, they barely have. Reported: Ars Technica. Outside AI agents have produced "several hundred orders" on DoorDash since they started, its chief executive said on Thursday's earnings call. DoorDash handled just under 1 billion orders in the quarter. Its link for outside agents is less than a month old and open only to a few beta partners, mostly for office lunch orders. "When it comes to a task that's as simple as ordering food, we tend to see that come directly to DoorDash," he said. Our read: one small test cannot settle how people will shop. For now, treat agent shopping as an experiment, not a sales channel. ( Ars Technica)
| End of skim · deep read begins |
| 06The Margin-Proof Tracker |
| |
No row added, none moved. N = 21. Two tests came due, and neither cleared the bar. Duolingo is still the only Stage 3 row, and no row has reached Stage 4. |
This tracks public AI value claims and the evidence behind them. A low rung is not a verdict on whether a claim is true. The evidence ladder: 0 · Narrative (a story, no numbers) · 1 · Operational (activity counted, no money attached) · 2 · Financially linked (a money figure tied to AI, mixed with other causes or too narrow to be the whole cost) · 3 · P&L-attributed (a margin change credited to AI in a filed document) · 4 · Sustained (Stage 3 held four quarters).
| Company |
Evidence |
Grade |
Next test |
| Microsoft (test came due, not cleared) |
EvidenceThe test was whether any AI figure would reach a filed document. One did, but it measures sales to a partner, not savings. Microsoft's quarterly filing shows $10.0 billion of revenue from its commercial deals with OpenAI, including revenue sharing, in the quarter ended September 30. That is nearly 11% of its $92.3 billion total. The filing does not say how much was cloud computing and how much was revenue sharing, or what it cost to earn. It does discuss AI costs: it spreads spending on AI infrastructure and training across its business units, and plans more capital spending for them. It does not credit any saving or margin gain to Microsoft's own use of AI. The row's earlier evidence is unchanged (Issue 014). Microsoft 10-Q, Oct 28 · TechCrunch on the filing |
Grade2, unchanged |
Next testNext quarterly filing: does Microsoft credit any saving or margin change to its own AI use? |
| Alphabet (test came due, not cleared) |
EvidenceThe test was whether Alphabet would separate AI revenue. It did not. Sundar Pichai said Google Cloud revenue grew 103% and called AI infrastructure and solutions "the major drivers," with no separate AI figure. Usage is counted instead: the Gemini app has one billion users, Google says, and its models, sold to developers, handle nearly 32 billion tokens (small units of text) a minute. Pichai's remarks, Oct 28 |
Grade2, unchanged |
Next testQ4 results: does Alphabet report AI revenue, or what AI Mode earns per search, on its own? |
Next tests. Duolingo's Q3 is the next step toward Stage 4. Customers Bancorp's next filing: does it credit AI for the savings and give AI's cost? Bank of America's next filing: does it price and explain its AI figure?
Who pays when the AI bill comes due
Three groups looked at the same bill this week, from three sides.
The seller has fixed costs and mostly flexible revenue. Anthropic's filing says its revenue is "primarily consumption-based," and much of it is not under long-term contract. Its compute deals run for years, on terms that "obligate us to pay for resources regardless of actual utilization." Meeting those bills "depends on our ability to generate significant revenue growth," the filing warns. Some of that risk is passing to buyers: a growing number of its enterprise customers now sign non-cancelable contracts with minimum spend. (Anthropic S-1)
Investors are losing patience with the builders. Alphabet, Amazon, Meta and Microsoft plan to spend more than $740 billion on capital this year, TechCrunch reported. Investors sold. Alphabet, Amazon and Meta spent more cash than they brought in, and Microsoft turned to bonds.
Few buyers are beating their savings targets yet. Just 3% of organizations exceed their cost-savings goals through productivity, Gartner says. Yet 85% plan to spend more next year, and more of that spending is moving to a meter. GitHub now meters Copilot use for some large customers. Oracle will charge for AI work on top of seats.
Analysis: the meter is how the bill reaches you. A seller with fixed costs needs usage to grow, and a minimum-spend contract locks some of that growth in. As a group, Anthropic's customers from a year ago now spend about nine times as much. That is great for the seller. For a buyer, it is good only if the work is worth more than the bill. Many companies cannot yet tell.
The counter-case. Buyers have more power than this suggests. Prices per unit keep falling: OpenAI's new mid-priced model costs half as much as the one before it. OpenAI is reportedly offering big discounts to take Anthropic's customers. And a meter can protect you. You pay for work done, not for seats no one uses. The danger is not the meter itself. It is a meter no one on your side is watching.
A test for any AI contract. Ask three questions:
What are we billed on: seats, usage or results, and what minimum have we promised?
Who on our side watches the meter, and what caps it?
How fast could we switch, and what would it cost?
If no one can answer the second question, your vendor is setting your AI budget.
| 08Where the Minds Disagree |
| |
Should companies cut staff now to pay for AI?
The case for cutting now. Microsoft offered buyouts to 60,000 people as it spends heavily on AI. It has not tied the offer to AI. Satya Nadella described the plan on a podcast almost two years ago: "Our total people costs will go down, our cost per head will go up, and my GPU per researcher will go up," The Verge recalled. (GPUs are the chips that run AI.) On this view, budgets are fixed, and computing power is the better buy. (The Verge)
The case for waiting. The gains are hard to pin down. In Atlassian's survey, the share of developers saving 10 or more hours a week with AI fell from 68% to about a quarter, InfoQ reported. The two surveys were a year apart with different samples, and the answers are self-reported. Gartner predicts that 75% of organizations that cut middle managers because of AI will rehire at least half of those roles by 2028. "Benefits appear immediately; costs appear later," Gartner says. Amazon told its engineers that measuring AI by token use or code output is "deeply flawed," CNBC reported. (InfoQ · Gartner · CNBC)
Our read: redesign the work first, then decide the headcount. Microsoft can afford to experiment, and says so: "The Voluntary Exit Offer is new for us. We'll learn from it." Most companies cannot yet measure what AI saves them. Cut before you can measure, and you may pay twice: once in severance, and again to rehire.
|
Dated tests, each with the source for its date.
Data centers on the ballot · Tuesday, November 3. Voters in 10 states decide 41 data center measures, from new rules for builders to outright bans, Axios reported. All six decided earlier this year went against development, Axios said, citing Ballotpedia. Campaigns have spent $173 million on almost entirely negative data center ads, according to AdImpact. The next test: how many pass, and whether they slow the build-out behind your vendors' prices. ( Axios · Axios on ads)
Warren's deadline for Anthropic · Tuesday, November 3. Warren asked how Anthropic will manage the risks its own filing describes before it sells shares to the public. The share count and price range are still blank. The next test: Anthropic's reply, and an updated filing with a price range. ( CFO Dive)
Moonshot's model becomes downloadable · Wednesday, November 4. The weights of Kimi K3.1 Pro (the trained model itself) are due under an open license, VentureBeat reported. Moonshot says it cut the model's hacking skills. The next test: whether outside testers confirm the cut. ( VentureBeat)
Microsoft's buyout window closes · December. Eligible employees have until December to decide, The Verge reported. The next test: how many accept, and whether job cuts follow. ( The Verge)
Was the fear about open models and hacking overblown? An AI researcher tests his own view after a month of security scares. Interconnects, Oct 30
What AI agents are doing to the world's code. GitHub's yearly report: merged code changes tripled in under a year, to 220 million a month. Simon Willison, Oct 28
No corrections this issue.
How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis. Written and edited by Mario Suarez · Independent analysis. Anthropic and OpenAI models assisted in creating this newsletter.
|