← Weekly
Lab — early draft from Era Haus

Nvidia loses its CEO as Anthropic sets its IPO price range

14-Nov-2026Weekly

This week showed how much of a company's AI plan rests on decisions made by someone else. The suppliers and platforms underneath a business's AI tools can change a cost, a risk or a sales channel in a single announcement, and the businesses that depend on them rarely get a say. Knowing who those suppliers are is now part of the plan.

Jensen Huang resigns as Nvidia CEO; finance chief Colette Kress steps in

On November 12 Jensen Huang resigned as chief executive of Nvidia, the company that makes most of the chips AI runs on, after a board inquiry into his alleged ties to a suspected Chinese spy. Chief financial officer Colette Kress is interim CEO, and the board named its first independent chair while it searches inside and outside the company.

Nvidia's chips are the scarce input behind most AI services, and Huang had led the company since founding it. A leadership search at the peak of demand adds uncertainty at the one supplier nearly every AI provider depends on. Its quarterly results on November 17 are the first read on whether customers see any change.

Source: NPR

Anthropic sets its IPO price range at $70 to $78 a share

On November 10 Anthropic, the maker of Claude, launched its roadshow to list on Nasdaq under the ticker ANT, offering 825 million shares at $70 to $78 each, up to about $64 billion. It came after the company retracted earlier terms it had sent to banks as inaccurate. Seven of its suppliers, including Google and Nvidia, have agreed to buy shares at the IPO price.

For the first time, a public market will put a daily price on a company that many businesses buy their AI from. Part of the early demand comes from Anthropic's own chip and cloud suppliers, so the first trading weeks say more about its staying power than the opening price does.

Source: Anthropic

Google launches Gemini 4 Carbon at a discount; Anthropic cuts Claude Fable 5.2

On November 10 Google made Gemini 4 Carbon, its new top model, generally available at an introductory $2 per million input tokens and $10 per million output (tokens are the units AI usage is billed in), rising to $4 and $20 on December 1. A day earlier Anthropic released Claude Fable 5.2 at $8 and $40, down from $10 and $50, The Next Web reported.

Top-tier prices now change with each release, and launch discounts expire within weeks. A budget built on one supplier's list price, or on an introductory rate, needs a date next to every number.

Source: Google

OpenAI lets businesses claim profiles and run their own agents in ChatGPT

On November 11 OpenAI opened verified business profiles in ChatGPT. Companies can see how often ChatGPT mentions them and how many users click through, and can run their own customer-facing agent inside it. US online sellers can also pay to make their agent the destination of a ChatGPT ad.

This is the listing, analytics and paid-placement layer that search engines spent two decades building, arriving in one release. For any business that gets customers through search, ChatGPT is now a channel with its own dashboard and its own ad budget. Claim the profile before a competitor's agent answers questions about you.

Source: Quartz

Notion's AI chat transcripts exposed in a breach at its supplier Fireworks AI

On November 12 Notion, the workplace notes app, said attackers who breached Fireworks AI, a company that runs open models for other software firms, reached chat transcripts from Notion's AI Insights feature. Fewer than 1% of Notion workspaces were affected; Notion switched the feature off and stopped using Fireworks for those models.

Your AI vendor's security includes every supplier it sends your data to. Ask which model hosts sit behind the tools you use.

Source: TechCrunch

United Airlines bars outside AI agents from booking its flights

On November 12 United said third-party AI agents may no longer book flights or manage reservations for customers, citing the risk that shared passwords expose personal data or trigger unwanted purchases. Travelers can still use AI to compare fares.

Big consumer brands are writing the rules for AI agents before that traffic is significant. A company that plans to sell through agents should expect some large sellers to refuse agent bookings.

Source: CBS News

SoftBank posts a $7 billion loss as US data center plans cost more

On November 10 SoftBank, the Japanese investor that is OpenAI's largest backer, reported a loss of about ¥1.1 trillion ($7 billion) for July to September. A $3.7 billion charge came from changing how it buys power equipment for its US AI data centers, and financing costs more than doubled.

Building AI capacity is costing its biggest financiers more than they planned, and those costs eventually reach the price of the service.

Source: The Star

EU leaders hold their first AI summit in Dublin and promise new financing

On November 13 European Union leaders met in Dublin for their first summit on AI. Commission President Ursula von der Leyen said the first part of a wider financing plan arrives in about two weeks, on top of the up to seven AI gigafactories (very large data centers for training models) the EU put out to tender in July.

For European buyers, this is the start of a local alternative to US cloud capacity. Of the roughly €30 billion behind the gigafactories, only about a third is public money; the rest still has to be found.

Source: The Next Web

Next week brings two tests of the same supply chain: Nvidia reports results under an interim chief on November 17, and Anthropic is expected to price its shares on November 18, CNBC reported. The move for this week is a short list. Write down which chips, models, hosts and platforms your AI work depends on, and who you would call if any one of them changed its terms tomorrow.

This week’s analysis

United Bars Outside AI Agents as Brussels Drafts Rules for Them

United barred outside AI agents, OpenAI let US sellers run their own in ChatGPT, and Brussels drafted agent rules that leave sellers free to refuse.

On November 11 United Airlines changed its contract of carriage, the terms every one of its tickets is sold under, to bar outside AI agents from booking or changing its flights, per Skift, a travel-industry publication. The same day the European Commission proposed rules for agents that buy on a consumer's behalf, and they oblige no seller to accept one. A day earlier OpenAI had let US online sellers run their own agent inside ChatGPT. Buying through an AI agent is becoming something the seller permits.

The assumption: an agent can go wherever its customer can

The companies building shopping and booking agents started from a reasonable premise. The agent acts for a real customer, with that customer's login and payment card, so a seller who turns the agent away turns away a sale. Sellers have lived with intermediaries they did not choose for decades. Airlines have sold seats through online travel agencies and price-comparison sites since the early days of the web, often reluctantly, and those sites stayed.

Two things made the premise hold. Refusing the tool a paying customer picked is expensive, and telling an agent from a person at checkout is hard. On that reading, agent commerce would grow the way comparison shopping did: sellers would grumble, then adapt, and the agents would sit between them and the buyer.

A seller, a platform and a regulator in one week

United went furthest. Its new rule covers agents arriving through travel agencies and aggregators (sites that list fares from many sellers), applies even when a person supervises or approves the agent, and lets United cancel agent-made bookings without notice, per Skift on November 11. The airline's stated reasons are security and privacy: handing a password to an outside tool can expose personal data or trigger purchases the traveler never meant to make, CBS News reported on November 12. Travelers can still use AI to compare fares, but the purchase has to happen on United's own conditions.

A day earlier, OpenAI had built the opposite route for sellers. Its beta of verified business profiles in ChatGPT lets a company see how often ChatGPT mentions it and how many users click through. Eligible US online sellers can also run their own customer-facing agent inside the chat, free for now, per Digiday, a marketing trade publication, on November 10. In a limited test, some US advertisers can pay to make that agent the destination of a ChatGPT ad, Quartz reported on November 11. The sanctioned route to an AI shopper is the seller's own agent, on the platform's rules.

Brussels regulated the agent and left the choice of letting it in to sellers. The Digital Fairness Act, proposed by the European Commission on November 11, would replace the EU's 2005 directive on unfair commercial practices. Companies offering AI shopping agents would have to let consumers set purchase limits and decide which purchases need their approval, per MLex, a regulatory news service. The draft is aimed at the agent's maker and, according to law-firm briefings on it, says nothing that would oblige a seller to let an outside agent in.

What the three moves have in common

Each actor had a different motive: United cited security, OpenAI is building a revenue line, and the Commission is protecting consumers. Each treated the agent as a separate party that needs permission, distinct from the customer it serves. This week's Newsletter, Nvidia loses its CEO as Anthropic sets its IPO price range, read the week as a lesson in how much of a company's AI plan rests on someone else's decisions.

The pattern started before this week. On October 21 Walmart said US customers would be able to give each outside agent its own profile in their account, with a spending cap such as $100 or approval for every purchase, per Digital Commerce 360, an e-commerce trade publication. TechCrunch's headline that day set it against Amazon, which blocks AI agents. Walmart chose to admit agents on its conditions; Amazon and now United chose to keep them out. In October we also wrote, in Wikimedia and Google Start Sending AI Agent Costs Back, that the organizations on the receiving end had started billing agent mistakes to whoever sent them. The receiving end now also decides whether the agent gets in.

Where the evidence is thin

United is so far the only airline in this story, and its enforcement is untested: an agent booking that arrives through a travel site may be hard to tell from a human one. The Digital Fairness Act is a proposal that the European Parliament and member governments will negotiate, and its details can change before anything applies. No seller in this story has published how many purchases agents make. The evidence shows the direction of seller policy, and says little yet about volume.

Who should be uncomfortable, and the move

The founders who should worry most are those building booking and shopping agents whose plan assumes access to large sellers' checkouts without an agreement. Travel goes first, because United has already written the refusal into its contract. Sellers with no stated agent policy should worry too: in the US, a competitor with a verified profile and its own agent can now answer customers inside ChatGPT.

For builders: list the twenty sellers your agent most needs to transact with, read their conditions this month, and get written access (an agreement to connect to the seller's booking or checkout system directly, or a partnership) from the ones that matter before you scale. For sellers: decide your agent policy and put it in your terms (allow outside agents, route them to your own, or refuse them), and if you sell in the US, claim your ChatGPT profile. For anyone building an agent for EU consumers, consumer-set purchase limits and approvals are the design the draft points to.

Whoever writes the terms holds the advantage

Agent commerce was pitched as a tool the customer brings. The last month's evidence describes a channel the seller opens, on conditions it writes, and in Europe on conditions a regulator will review. A company with a large customer base and a direct channel can decide which agents reach its checkout, and it can charge, filter or refuse. A company whose product is an agent sitting between buyers and those sellers has to negotiate access with every seller it wants to reach, and its margin depends on how many say yes. The sellers who write their agent terms first will set the conditions everyone else signs. The agents that sign early will hold checkout access their rivals have to wait for.