← Weekly
Lab — early draft from Era Haus

Anthropic's record IPO and the new rules on who carries AI's risk

21-Nov-2026Weekly

This week the market put a record price on the companies that supply AI, and banks and governments started deciding who pays when it goes wrong. A business that uses AI now inherits terms it never negotiated, from its bank, from its suppliers' governments and from its vendors. Knowing where those terms leave the loss is now part of buying AI.

Anthropic prices its IPO at $83 a share and raises $81.3 billion

On November 18 Anthropic, the maker of Claude, priced its shares at $83, above the range it set a week earlier, and raised $81.3 billion in the largest US listing on record. On its first day of trading on Nasdaq the stock closed 17% higher, CNN reported.

Businesses that run on Claude now buy from a company that must publish its results every quarter. Buyers get a clear view of its costs and margins before they renew, and Anthropic gets shareholders who will expect those margins to rise.

Source: CNBC

Nvidia's revenue doubles to $115.8 billion and its forecast excludes China

On November 17 Nvidia reported quarterly revenue of $115.8 billion, more than double a year earlier, in its first results since Jensen Huang resigned as chief executive. Its $130 billion forecast for the next quarter assumes no data center sales in China.

Demand for AI chips held through a leadership crisis, with a forecast that counts no data center sales in China. Computing power will stay scarce, and priced by the seller, for at least another quarter.

Source: Nvidia

US requires licenses for AI chips sent to Brazil and four Asian countries

On November 16 the US Commerce Department issued a rule requiring a license to ship advanced AI chips to Brazil and four Southeast Asian countries it sees as routes to China. It also reaches data centers anywhere that may rent computing power to Chinese companies. Shipments to companies headquartered in the US are exempt if the exporter reports them each month.

Brazil is the only Latin American country on the list. ABES, the Brazilian software association, warned that the rule raises costs and puts Brazilian-owned clouds and data centers behind US subsidiaries in the same market. For a business in Brazil, its cloud provider's nationality now affects how fast it gets new AI capacity.

Source: TechCrunch

Citi and Citizens make customers liable for what their AI agents do

By November 18 Citi and Citizens, two large US banks, had changed their account terms. A customer who gives an outside AI assistant access to an account now carries the loss if the assistant leaks a password or makes a payment nobody intended, even when it was hacked.

Until courts and regulators settle the question, the user carries the risk. On November 16 four UK regulators, including the Financial Conduct Authority, told ministers that companies launching personal AI agents must already comply with consumer, financial and data protection law. A business that lets an agent pay suppliers should cap what it can spend.

Source: TechCrunch

KPMG tests more than 100 AI agents on a full company audit

KPMG's UK firm ran more than 100 AI agents through every stage of an audit of an unnamed mid-sized company listed in London, alongside its normal audit, the Financial Times reported on November 18. KPMG did not rely on the result for sign-off and plans to extend the trial to as many as 20 larger audits.

The agents asked better questions than the accountants, by the pilot's own findings, but nobody has measured accuracy or cost yet. Two days earlier the UK audit regulator warned firms against using their AI tools to win extra work from audit clients, so the savings will have to show up in the work itself.

Source: The Irish Times

OpenAI cuts its top model's price as Google adds a cheaper Gemini

On November 17 OpenAI released GPT-6.2 Astra at $8 per million input tokens (tokens are the units AI use is billed in), about 20% less than the model it replaces. A day later Google launched Gemini 4 Neon, a middle tier below its top model, at $3 per million input tokens.

OpenAI's top model now lists at the same input price as Anthropic's Claude Fable 5.2. Google's top model, Gemini 4 Carbon, will cost half that even after its introductory price ends on November 30, so at the top Google is now the one competing on price.

Source: OpenAI

Attackers used an AI tool to take down the European Commission's payroll

On November 19 the European Commission confirmed that attackers used an AI tool in a cyberattack, first detected on October 22, that kept its payroll, leave and staff systems offline for more than 10 days. It has found no evidence that staff data as a whole was taken.

An AI-assisted attack kept the EU executive's back office offline for over a week. In a warning reported on November 16, the US FBI said ARTEX, a free AI toolkit built for testing network security, lets attackers reach administrator-level access in under three hours. Find out whether you could pay staff with payroll down for 10 days.

Source: The Record

Next week Anthropic's shares face their first full week of trading, the first read on how investors judge an AI supplier between results. The move this week is a list: for every AI tool or agent your business uses, write down who carries the loss when it makes a mistake (the vendor, the bank or you), and renegotiate or replace the ones where the answer is you.

This week’s analysis

AI Stocks Fell as OpenAI and Anthropic Cut Flagship Prices to $8

AI stocks fell while OpenAI and Anthropic cut flagship prices to $8; buyers should lock today's rates into 2027 contracts.

On November 19 the S&P 500 and the Dow, the main US stock indexes, closed in correction territory (10% below a recent peak), per CNBC. Anthropic had raised $81.3 billion the day before in the largest US listing on record, per CNBC, and Nvidia had reported revenue more than doubled. Outlets from NPR to Al Jazeera called it the week the AI boom broke. OpenAI cut its top model's price by about 20% the same week. Investors are repricing AI suppliers while customers pay less, and buyers should plan around that gap.

The case that the boom broke

The strongest version of the narrative starts with how broad the fall was. By November 19 the Nasdaq had erased its gain for 2026, per CNBC, and NPR called it the index's worst week since the COVID lockdowns of March 2020.

Investors had said as much before the week began. Bank of America's November survey of fund managers, taken from November 6 to 12, found a record net 38% saying companies are overinvesting in AI, per Fortune on November 17. The same managers made the largest monthly cut to their technology holdings since 2006.

Regulators read the same risk. The US Federal Reserve's financial stability report, published November 13, said 87% of the market contacts it surveyed named AI as a top risk, up from 50% in the spring. On November 18 the European Central Bank's Financial Stability Review listed stretched AI valuations among the main risks of an abrupt repricing of markets. On that reading, AI suppliers will find money harder to raise and will build less, and AI spending slows in 2027.

Demand held while prices fell

The suppliers' own numbers do not show customers buying less. Nvidia's quarterly revenue reached $115.8 billion, and it forecast $130 billion for the next quarter while assuming no data center sales in China, per its November 17 results. Its shares fell after the report anyway, noted Semafor, a news site, the next day. Anthropic sold its shares at $83, above the range it had marketed, and they closed their first day of trading 17% higher, CNN reported. This week's Newsletter, Anthropic's record IPO and the new rules on who carries AI's risk, noted that Claude's buyers now deal with a company that must publish its margins every quarter, to shareholders who will expect them to rise.

What buyers pay fell. Models are billed per million tokens, a token being roughly three-quarters of a word. OpenAI's GPT-6.2 Astra, released November 17, lists at $8 per million input tokens, about 20% below the model it replaces, per OpenAI. Anthropic's Claude Fable 5.2, released November 9, matches that rate, down from $10 for Fable 5.1, per Anthropic's published rates. In September, in OpenAI and Anthropic Cut Prices Below Their $10 Flagships, we found the discounts stopped one step below each lab's top model. Both top models have now come down.

The market marked down AI companies while the price of AI kept falling. The two can diverge for a while because they measure different things. A share price is a bet on future margins and on what the supplier's capital will cost. A price list is what a lab charges this month to win a contract.

Part of the low price has an end date

Some of the cheapest offers carry an expiry. Google's top model, Gemini 4 Carbon, sells at an introductory $2 per million input tokens until November 30 and $4 from December 1, per 9to5Google, which doubles the rate for anyone who budgeted on November. Microsoft's new MAI Code 2 coding model costs $0.50 per million input tokens inside GitHub Copilot, its coding assistant for developers, on a promotional rate that doubles after May 17, 2027, per InfoWorld.

Heavy users are already managing the bill. Business Insider reported on November 17 that Salesforce, which sells AI agents to businesses, was cutting jobs again and had put monthly limits on what staff spend on AI tools after some ran up thousands of dollars in a quarter. Salesforce says its teams manage their own spending and denies a company-wide limit per person. Either way, staff at a company that sells AI were spending enough on it, at falling per-token prices, for managers to step in.

The price list depends on where you buy

For buyers in Brazil the direction is less comfortable. On November 16 the US Commerce Department issued a rule requiring a license to ship advanced AI chips to Brazil and four Southeast Asian countries, per TechCrunch that day. Companies headquartered in the US are exempt from the license requirement as long as shipments are reported, according to the law firm Morgan Lewis. ABES, the Brazilian software association, said in a statement that the rule raises costs and puts Brazilian-owned data centers behind US subsidiaries in the same market. In Malaysia, shares in two data center builders fell nearly 20% on November 17, per Fortune. A company in São Paulo renting capacity from a local provider now carries a wait for licenses that a contract with a US cloud does not.

Where this read could be wrong

The causes of the fall are contested. NPR also pointed to the Federal Reserve's rate rise in September and to political pressure on data center construction. Nvidia is also searching for a new chief executive after Jensen Huang resigned on November 12. The prices here are list prices: no data in this piece shows what companies actually paid in November. And the claim that suppliers will push prices back up is a forecast, which nothing yet confirms. If investors' concern fades by January, the gap may close from the market side instead.

Who should be uncomfortable, and the move

The companies most exposed are those writing 2027 budgets on November's list prices, especially introductory ones, and those on monthly, uncapped usage terms with any one supplier, along with Brazilian and Southeast Asian firms that rely on locally owned AI computing capacity.

The move, before renewals close in December and January: turn the current price into a contract. Ask each AI supplier for a capped rate through 2027, written notice before any price change, and the rate that applies after any introductory period ends. The thing to watch is Anthropic's first quarterly results as a listed company, and whether it reports margins rising while its list prices fall.

The cheap model is a supplier decision

This week's prices tell you what three labs need to win business today. They say nothing about what a public company under pressure from its shareholders will charge in a year. The stock market and the price list disagreed in November, and for now buyers benefit. Suppliers whose shares fell, and the one whose new shareholders expect margins to rise, have reasons to close that gap through their customers. Operators who treat November's price as a permanent fact will be negotiating 2027 from the wrong number. The ones who write it into a contract now will be holding a price that does not move.