← Weekly
Lab — early draft from Era Haus

Apple and Goldman ration AI while Anthropic commits $895 billion

07-Nov-2026Weekly

This week the cost of AI started to decide how much of it gets used. Companies that had urged staff to use as much as they liked began setting allowances, while one AI developer disclosed close to $900 billion of spending commitments to keep up with demand. For anyone planning next year, the useful question is how much AI the budget, the compliance team and the neighbors will allow.

Anthropic's IPO filing shows a $17 billion quarter and $895 billion of commitments

On November 2 Anthropic, the maker of the Claude models, updated its filing to list on Nasdaq with third-quarter results: revenue of about $17 billion for the quarter. It also disclosed $895 billion of purchase and lease commitments, almost all for computing capacity, a total that includes agreements signed after September 30.

The commitments are the number a customer should read. Anthropic has agreed to pay for capacity years before the revenue that will cover it, so its prices and reliability depend on those contracts holding. The revenue shows the demand is real; the obligations show how much of the coming years is already promised to its computing suppliers.

Source: The Next Web

Apple cuts its engineers' daily AI allowance from $300 to $200

Apple lowered the daily cap on AI tool spending for its engineering teams in October, from $300 to $200 a person, Bloomberg reported on November 4, and is weighing $25 a day for other staff. On November 2 Quartz reported that Goldman Sachs had limited its AI coding tools to software engineers after a review of productivity and risk.

Both companies had encouraged wide use. Once the bills arrived, AI became a budget line with an owner, and most companies will reach the same point.

Source: TechCrunch

OpenAI puts ChatGPT and Codex on one shared usage allowance

On October 31 OpenAI said that from December, chat and coding on its Plus, Pro and Business plans will draw on one shared pool, measured in tokens, the units AI tools are billed by. The pool drains faster with stronger models and harder tasks, and OpenAI has not yet published its size.

A flat subscription is turning into a metered budget, and the heaviest users will hit its limit first.

Source: OpenAI

US voters reject data centers in local ballots across ten states

On November 3 voters in ten US states decided 41 local measures on data center construction, and the results went almost entirely against development. In Ohio, at least 14 towns passed bans or limits on large facilities, according to unofficial counts.

Permission to build is becoming as hard to get as chips. A data center that cannot win local approval delays capacity that AI developers have already sold, so local planning fights will set part of the price of AI.

Source: Venable

OpenAI's president says the company could be ready to list in 2027

On November 5 Greg Brockman, OpenAI's president, said the company could be ready for a public listing in 2027. A week earlier the Financial Times reported that OpenAI had cut the target for its current funding round from about $30 billion to about $20 billion, after a fall in AI-linked stocks.

Investors are now choosier about the largest AI rounds, even for the best-known lab. A company choosing its main AI supplier should weigh how securely each one is financed.

Source: TechCrunch

Salesforce agrees to buy Cresta, whose AI agents answer customer calls

On November 2 Salesforce signed an agreement to buy Cresta, a startup whose AI agents handle live customer phone calls and coach human representatives during them. Salesforce did not disclose a price and expects the deal to close in the first months of 2027.

Software that answers customer calls is being folded into the large customer-management suites. Two days later NICE, a maker of contact-center software, reported that revenue from its AI products rose 46% while it cut about 5% of its staff. Buyers of call-center AI will have fewer independent vendors to choose from.

Source: The Next Web

Zalando plans an AI agent that buys from other shops inside its app

On November 3 Zalando, the Berlin-based online fashion retailer, announced Zalando Extended, a service that lets customers buy from other online shops without leaving its app, starting with the home-goods retailer Westwing. From 2027 its AI agent will complete the checkout for them.

Zalando is making itself the shopping agent before a chatbot can stand between it and its customers. Any retailer that sells through marketplaces should ask who will own the customer once agents do the buying.

Source: FashionNetwork

Anthropic's next step is a price range for its shares, expected as soon as next week, and it will show what investors will pay for a company carrying $895 billion of obligations. This week's move is in your own contracts: before January, find out how each AI service you pay for meters its use, because flat-fee plans are turning into allowances.

This week’s analysis

Texas Elects a Data Center Freeze Backer as Apple Caps Staff AI Use

Apple and Goldman capped staff AI use, but Texas's planned data center freeze and local bans may limit capacity AI suppliers already sold.

On November 3 Texas, which MIT Technology Review calls the center of the US data center buildout, elected a governor who has promised to keep new data center approvals frozen until state lawmakers write new rules. Voters in ten US states turned down data center projects in local ballots. On November 5 the Philadelphia Semiconductor Index of US chipmakers fell 12.3%, per CNN. Bloomberg reported this week that Apple and Goldman Sachs had cut staff AI use. The loud story was that AI spending hit its limits. The limit that binds is permission to build.

The case that the boom hit its limits

The strongest version of the week's narrative runs on three levels at once. Budgets first: Apple cut some engineers' daily allowance for AI tools from $300 to $200 a person in October, Bloomberg reported on November 4, and some teams may fall to $100. Two days earlier, per Bloomberg, Goldman Sachs had limited its AI coding tools to software engineers after a review of productivity and risk.

Capital next. OpenAI cut the target for its current funding round from about $30 billion to about $20 billion after a fall in AI stocks, the Financial Times reported at the end of October. Then the markets: on November 5 the S&P 500 fell more than 4%, per CNBC, and some companies that build and rent out data centers lost more than 14% of their value in a day.

Read together, as this week's Newsletter laid out in Apple and Goldman ration AI while Anthropic commits $895 billion, the pattern looks like spending running into its costs. Companies that told staff to use as much AI as they liked are setting allowances. Investors are choosing which AI rounds to fund. On that reading, 2027 is the year AI budgets shrink.

AI revenue is still rising

The suppliers' figures do not show customers buying less AI in total. Anthropic, the maker of the Claude models, reported revenue of about $17 billion for the third quarter in its amended stock-market filing on November 2, up from about $11.6 billion in the second. Growth is slowing: by the filing's own figures it added about $5.4 billion of quarterly revenue in the third quarter, against about $6.8 billion in the second. Revenue still rose by almost half in three months.

Money is also moving toward cheaper ways of running AI. Fal, which runs open AI models (ones any company can download and host) for other businesses, raised $1.3 billion at a $16 billion valuation, double its price in April, TechCrunch reported on November 5.

The rationing is real, and Bloomberg says Apple has considered going as low as $25 a day for some teams already capped at $100. It is also narrower than the headlines. Apple staff told Bloomberg that few engineers used the full $300, and senior leaders can grant exemptions. Goldman's stated reason was productivity and risk, with no mention of the bill. Both companies are deciding who uses the tools and how much one person may spend. That is ordinary budget control, and a manager can loosen it with an email.

Permission is the limit that binds

Texas already had a pause. On August 3 Governor Greg Abbott told the state's utility regulator and its grid operator to stop approving pending data centers while they audit projects waiting for a grid connection, the Associated Press reported; critics said it fell short of a true moratorium. Gina Hinojosa, the Democrat who beat him, campaigned on a formal moratorium that lasts until the state Legislature passes what she calls real protections, including for local water supplies. She takes office in January, when the Legislature also meets. Both parties in Texas are now rationing permission. They disagree on how long.

Locally, voters decided measures in 41 jurisdictions across ten states on November 3, almost all against development, according to the law firm Venable. In Ohio, at least 14 towns passed bans or limits. Markets read this as a supply problem: Joe Maher of the research firm Capital Economics told CNN that the sell-off appeared to be triggered by concern that Democrats in the US Congress and the states would restrict data center development. The Associated Press also cited borrowing costs that day, so the election explains part of the fall, and we cannot say how much.

What makes permission the binding limit is how much has been promised on capacity still to come. Anthropic's filing lists about $895 billion of purchase and lease commitments, substantially all for computing capacity. Only about $69.8 billion falls due in the twelve months after September 30, and about $447.5 billion by the end of 2030, per CNBC. About half falls due after 2030, and some of it pays for sites still in planning. Nscale, a British company that builds AI data centers, plans one for Anthropic in West Virginia. Nvidia is weighing a guarantee of billions of dollars of its debt, per TechCrunch on October 31. When a chipmaker considers backing its customers' loans, the scarce input is a finished building with power connected, and a moratorium delays exactly that.

Where this read could be wrong

Hinojosa's moratorium has not been issued, parts of her plan need lawmakers to act, and Texas officials may resist it. Capacity can also move. Brazil's Senate approved REDATA, a federal tax regime for data center construction, on September 1, per the law firm Cescon Barrieu. A study by FGV Projetos, the consulting arm of the Getulio Vargas Foundation, estimates it narrows the cost gap with building in the US from 26.7% to 17.7%. If projects shift to places that want them, the result is longer timelines and higher prices, with no hard stop. The market link to the election is traders' interpretation, so it is the thinnest leg here. The Texas pause and the ballot results are on the record.

Who should be uncomfortable

The companies that should worry are those whose 2027 plans assume AI gets cheaper per task every year and that capacity is always available on demand. That includes heavy users of coding agents. From December, OpenAI's Plus, Pro and Business plans will draw chat and coding from one shared allowance measured in tokens, the units AI tools are billed by, OpenAI said on October 31; it has not published the allowance's size. It also includes anyone who chose an AI supplier on price alone, without asking how that supplier's capacity is financed and where it sits.

The move, before renewals close in January: ask each AI supplier what share of the capacity it is selling you runs in data centers that exist today, and what happens to your price and service levels if a planned site slips. Ask for a cap on metered rates for 2027. The thing to watch is the Texas Legislature's first weeks in January, and whether Anthropic's share price range, which the Newsletter expects as soon as next week, discounts its commitments.

For two years the scarce input in AI was chips, and enough money could buy more of them. The scarce input now includes a planning vote, and money buys that slowly if at all. The budget caps at Apple and Goldman are the visible part of this week because they happened inside companies readers know. The part that will move prices in 2027 happened at the ballot box, in places where most AI customers have no vote. Operators should start treating AI capacity the way they treat electricity or office space: a supply with a location, a permit and local politics attached, priced in contracts that say what happens when the supply is late.