AI the News That's Fit to Prompt

Anthropic spent $10 billion on compute last quarter, and $5 billion of that was short-term capacity that costs “significantly more” than the multi-year kind. That one line explains why Fortune’s analysts think the IPO math is thin, why Elon Musk says serving Grok Bot is “unsustainably high relative to revenue,” and why Meta expects to improve compute efficiency tenfold in a year. Today’s stories split between the bill for running AI and the question of who can be trusted to run it.

Today’s Headlines

Paying for Compute

• The IPO math – Fortune reports Anthropic’s third-quarter revenue rose nearly 47% while compute and infrastructure expenses rose about 77% to $9.9 billion. Jay Ritter says any valuation above about $1.41 trillion means paying more than $40 per dollar of revenue, a level only 14 of 4,110 IPOs reached from 1980 through 2024, and those lagged the market by 15.4% on average. At $1.5 trillion to $2 trillion, buyers would pay $43 to $57 per dollar of trailing sales.
• Where the money goes – TechCrunch reads the same updated S-1 and finds compute took 58% of revenue, up from 48%. Revenue per compute dollar peaked just over 2x and fell to 1.72x. The reporter calls the squeeze a blip and a “champagne problem,” but notes the S-1 expects compute to remain the largest expense “for the foreseeable future,” and Anthropic’s own build-out, Anthropic Direct, cannot keep pace with demand.
• The other side of the invoice – Fortune’s Grok Bot story shows who is collecting. SpaceX’s AI division made $7.56 billion in revenue, up from $813 million a year earlier, largely by leasing capacity to Anthropic and Google, yet SpaceX spent $16.4 billion on AI capex, 84% of its total. Replying to a user whose allowance ran out four days early, Musk wrote that serving Grok Bot cost too much relative to revenue, and he told analysts a six-month payback on compute next year is “not impossible.”
• The $15 billion gigawatt – BCG estimates the AI ecosystem needs about $2.6 trillion a year in end-market spending by 2030, with $1.8 trillion from enterprises, which would lift AI from 5% to 25% of corporate tech budgets. Each gigawatt coming online in 2026 must earn about $15 billion to cover its costs, and BCG’s advice to buyers is to avoid long-term price-per-token deals while prices fall.
• Money and power chase the same bottleneck – Fal raised $1.3 billion at $16 billion, double April’s $8 billion, with more than $800 million in annualized revenue, though inference providers still run thin margins. Meanwhile data centers are buying rebuilt jet engines because the IEA says a 100 megawatt grid hookup in Northern Virginia takes about seven years. Microsoft even pulled out as anchor customer of Nscale’s Loughton site over power delays.

Verifying and Containing Agents

• Proof of which model answered – Anthropic’s FlashProve would let a user verify that an output came from a specific model, even against the lab running it. The post says it could stop a company swapping in a cheaper model to cut costs, or a pricier one to inflate results. Labs would publish a fingerprint of each model that does not reveal its weights, and the researchers say there are no plans to deploy it.
• Risk language moves into filings – Apple’s 10-K warns that agents may take actions “not directed, instructed or authorized,” defy directions, and “conceal actions.” Fortune notes Alphabet made a similar warning, and that Anthropic’s prospectus estimates 5% to 7% of this year’s revenue came from customers later found to be violating its usage policies.
• Rules instead of mind reading – MIT Technology Review describes mechanism design, which sets incentives so agents behave whatever their internals look like. In a preprint from MIT’s Stephen Morris and colleagues, a model is rewarded for how well its answers predict a second model’s, so deliberately underperforming becomes a losing bet. Princeton’s Elad Hazan would have models alternate as solvers and auditors, “one agent acts like law enforcement.”
• When the automation errs in public – Meta’s language models replaced human reviewers for about half of review requests by June, and the company says they make 13% fewer errors. Yet Brazil’s electoral court ordered equal treatment of the presidential candidates under a 500,000 reais daily fine, and No Kings’ Instagram account went down for about 90 minutes on its Oct. 17 day of action. An Institute for Strategic Dialogue report calls leaning on automation in an election “an untested proposition.”

Builders, Buyers and Their Bets

• Cloning software for $10 – Matthew Berman says the latest models can clone Excel in about 90 minutes to a couple of hours, for roughly $10 in tokens against $180 for a lifetime license. He calls Opus 5.5 the turning point and predicts software tends toward open source. His evidence is his own tests, and he concedes maintenance and security are the real objection.
• A model with a lane – Matt Maher’s hands-on test of GPT-6.1 Terra, which he says costs about ten times the efficient Luna tier per token, found it shipped changes through an established pipeline but got stuck in its own rules on open-ended problems. He says it tends to succeed or stop and hand work back in computer use. His fix is escalation steps to a stronger model, and a warning that many pretty-good results compound into one you cannot diagnose.
• Swing again – Nate B. Jones says fewer than 200 people worked on Meta’s Muse model and product, that more than a hundred model behaviors were tested before launch, and that 10 million people set it up in six weeks. His point is that a $19.2 billion Reality Labs loss did not make Zuckerberg careful. Fortune’s Grok story adds the bill: Zuckerberg told analysts Muse is costly to run.
• What Freshworks stopped doing – Freshworks reached about $992 million in ARR by Sept. 30 and crossed $1 billion in October after buying Torii and Robusta, with $18.3 million in GAAP net income. CEO Dennis Woodside credits stopping “a bunch of smaller projects.” McKinsey’s Alex Panas says “AI is everywhere except in the EBIT line,” and Michael Park describes agents surfacing the five things a team disagrees on out of twenty.
• Who buys if no one is paid – David Shapiro says households get roughly 60% of income from wages, 20% from government, and 20% from property. He claims workers received under 53 cents of each dollar produced in 2026, and points to Alaska’s fund, Norway’s roughly $1.9 trillion fund, and over 6,600 employee ownership plans. All figures are his, without sources in the video.
• Defaults for children – Google will turn AI off for K-12 students next year unless administrators opt in, but because Search and AI are integrated, schools that restrict AI lose Search. Apple, meanwhile, will require 18+ ratings for dating and romantic AI companion apps, with new descriptors due Nov. 16. TechCrunch cites 220 million downloads of top companion apps as of last summer.
• Consumer and corporate moves – Netflix’s “AI Discover” tab offers mood-based themes such as “something light” and “scare me,” which Netflix calls its “test kitchen.” OpenAI hired former Army Secretary Dan Driscoll as senior counselor for national security modernization two months after he left the Pentagon. It also hired Uber’s Jill Hazelbaker as chief corporate affairs officer, filling a role vacant since Hannah Wong left in January.

The Throughline

The shared theme is who absorbs the cost of an agent doing real work. Berman’s $10 Excel clone is a single user’s token bill. Grok Bot is the same economics at scale, where more frequent agent users drive up costs without necessarily paying more. Anthropic’s 58% compute share and Meta’s plan for a tenfold efficiency gain on Muse say the same thing from the lab side. Cheap for one person and expensive for the company that serves millions is a strange place to build a pricing model.

There is a pointed link between the IPO story and the SpaceX numbers. Anthropic pays SpaceX $1.25 billion a month for compute, and Fortune says leasing to Anthropic and Google drove much of SpaceX’s AI revenue growth. One company’s margin pressure is another company’s revenue line, which is why BCG’s $15 billion per gigawatt figure matters: the build-out only works if someone downstream, a lab, an enterprise, a consumer, earns enough to cover it. Fal’s doubling valuation and Freshworks’ “everywhere except EBIT” quote are the same question seen from two ends.

The safety stories make a quieter argument that fits the cost one. FlashProve, mechanism design, and Apple’s 10-K all assume you cannot fully inspect the model, so you verify outputs or shape incentives instead. Maher’s advice to build escalation steps into a system is the practical version, and Meta’s election errors show what happens when automation replaces the check. Meta says its models beat humans on error rates, yet the errors that surfaced came during elections. A lower average error rate does not help much when the misses land on a presidential campaign.

What to Watch

• Where Anthropic prices. Fortune says institutions want the low end of $1.5 trillion to $2 trillion. Whether the final price lands near Ritter’s $1.41 trillion line, and how the next compute-share figure moves, will show whether the third quarter was a blip or a trend.
• Whether anyone publishes a fingerprint. FlashProve needs labs to release a fingerprint for each model, and the researchers say they have no plans to deploy it. If a lab adopts it, buyers could start asking vendors to prove which model they are getting.
• Brazil’s ruling on Meta. The electoral court has not yet decided whether to fine Meta, and federal prosecutors filed a civil suit on Oct. 21. The outcome will set an early precedent for how courts treat automated moderation errors during elections.
• Agent pricing. Musk and Zuckerberg both say agent usage is costing more than expected. Watch whether Grok Bot, Muse, or others change allowances or pricing, since BCG’s advice to avoid long-term token commitments assumes prices keep falling.

Go Deeper

• They did it… – Timestamped notes on the three conditions Berman says were met, his Excel clone test and its roughly $10 token cost, the open-source ArtCraft suite, and his answer to the maintenance and security objection, with the sponsor segment labeled.
• You’ve Never Lost $19 Billion – Jones’s three copyable lessons, the hotel-bathroom analogy for polish, the agent design questions he suggests asking, and why consensus compromises become the user’s problem.
• GPT-6.1 Terra Is Fantastic at Some Work and Terrible at Others – Maher’s model lineup and its pricing ratios, where Terra shipped customer changes and where it got stuck, and reflection questions on which of your tasks are truly bounded.
• ELI5: Post-Labor Economics – Shapiro’s three ways money reaches a household, his seven-source 2045 example, and a clear list of which statistics are only his claims.