Last Week Ignite - 5.17.26
The week the frontier labs crossed the counter
Tomoro is a 150-person engineering shop in London. Tesco hired them to retool a checkout workflow around a frontier model. Virgin Atlantic, Mattel, Supercell, and Red Bull did the same. On Monday, OpenAI bought them and folded them into a new majority-owned services arm with more than four billion dollars of outside money behind it. The Tomoro engineers are still doing the same job this week. They now do it as employees of the model company they were deploying for clients.
In the same seven days, PwC enrolled its three hundred sixty-four thousand consultants in a training program built around Claude and stood up an internal finance practice that runs on it. Anthropic and PwC made the joint pitch in a press release on Thursday: insurance underwriting cycles compressed from ten weeks to ten days, an HR transformation prototype delivered in a week. On Friday, OpenAI shipped a personal finance product that connects ChatGPT to twelve thousand banks and brokerages through Plaid. Mint shut down two years ago and the category never quite reformed. The default just got written by the model company.
Three different moves. One pattern. The frontier labs stopped selling tokens this quarter. They are running operations.
Three moves in one week could be coincidence. The capital lined up behind them is what makes it durable. OpenAI’s deployment unit launched with named partners that include McKinsey, Bain, and Capgemini, plus Goldman Sachs and SoftBank on the money side. Anthology Fund, the joint vehicle Menlo Ventures runs with Anthropic, led a thirty million dollar round into a social commerce company called Nectar the same week. The lab is no longer the supplier in the supply chain. The lab is the table where consulting firms, banks, and growth funds now sit.
A founder I was on the phone with on Wednesday put it this way before I could ask. “We used to pitch CIOs that we’d be the integration layer between the model and their workflow. That sentence does not work this week.” It does not. OpenAI just hired the integration layer.
Here is what gets compressed.
For the past eighteen months, a familiar category of early-stage company sold itself with a version of the same pitch: we will take a frontier model and put it inside your enterprise. Some of those companies were good. Some had a real wedge in a vertical. Most were betting on a margin pool between the lab and the customer that the lab had not noticed yet. The lab noticed.
The same is true for personal finance. Copilot, Monarch, Rocket Money, and Cleo have spent two years competing to be the chat surface where a millennial sees her credit card spend. They were also betting on a margin pool. ChatGPT now sits inside that pool with a Plaid connector and a hundred million weekly users. Some of those companies still have a moat in coaching, advice, or community. The default does not.
The SMB market got the same treatment. Anthropic opened a small business product on Wednesday with native connectors to QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. Until last week, “AI for SMB bookkeeping” was a sensible pitch deck slide. It now competes against the connector pack that comes free with the model.
The cap chart and the macro print
While the labs were doing all of this, the public data underneath turned against them.
April CPI printed on Tuesday at 3.8% annual, with the monthly read at six tenths of a point. Producer prices the next day were louder: one point four percent month over month, the largest single-month gain since March of 2022. Senate Republicans confirmed Kevin Warsh as the next Federal Reserve chair on Wednesday afternoon, fifty-four to forty-five. He was sworn in on Friday. The “we get rate cuts in the second half” thesis that growth-stage models have quietly been leaning on did not die forever, though it is closed for the conversation the new chair will have at his first FOMC on June 17.
Capital priced none of that this week. Anduril closed a five billion dollar round on Wednesday at a sixty-one billion dollar post-money. Thrive Capital and Andreessen Horowitz led. The valuation roughly doubled in ten months. The same day, the Department of Defense announced an agreement for ten thousand low-cost hypersonic weapons with Anduril, CoAspire, Leidos, and Zone 5 named as contractors. Two days later, Cerebras priced its initial public offering at one hundred eighty-five dollars per share, above its range, and opened the next morning near three hundred fifty. The book was reportedly twenty times oversubscribed. By Friday afternoon, Bloomberg reported that Anthropic had begun talks for another thirty billion dollar round at a nine hundred billion dollar valuation.
Read those three lines without the macro context and the cycle never tightened. Read them with the macro context, and the gap between what the public data said and what private capital priced is the story.
The disconnect is not irrational. Defense budgets are appropriated, not voted on each month, and the autonomy buyer is uncorrelated with a Fed cut. Frontier AI capex runs on multi-year compute contracts. The names that moved this week were the ones whose customers do not check ten-year yields before signing.
For everyone else, the hardest piece of news arrived more quietly.
The real ceiling
Anthropic re-metered third-party agent harnesses on its paid plans on Thursday. Each plan tier now has a separate monthly credit limit for Claude Code and the Agent SDK, starting around twenty dollars at the bottom. Sam Altman responded a few hours later with a two-month Codex promotion for new business customers. None of that sounds dramatic until you read what ServiceNow and Uber told The Information earlier this month, which is that both companies blew through their entire annual AI token budgets before half the year was finished.
The agent economy is hitting a real ceiling. The ceiling is the cost of a session.
An agent that writes a ten-minute report at five cents per thousand tokens is cheap. An agent that runs for nine hours to refactor a codebase, retries when it fails, and burns a hundred dollars in inference per attempt is not cheap. Companies signed contracts last year on the assumption that volume would scale gracefully. Volume is scaling faster than the contracts assumed.
A few categories get more interesting in that light. Orchestration software that routes tasks to cheaper models. Software stacks for non-Nvidia silicon, now that Cerebras has a public mark. Finance and operations software for AI workloads, the budget controls and per-task monitoring that have existed for cloud spend since the early 2010s and barely exist for inference. None of those is a glamorous category. All of them sit on top of a real and growing customer pain.
Microsoft made a related point this week with less fanfare. The company shipped a vulnerability research system called MDASH, an orchestration of more than a hundred specialized agents over an ensemble of frontier and distilled models. MDASH reported sixteen critical Windows vulnerabilities in Tuesday’s Patch Tuesday cycle, four of them remote code execution flaws. On CyberGym, a public security benchmark, Microsoft claims MDASH outscored Anthropic’s latest research-preview model, called Mythos, by about five points. The number is vendor-claimed for now. The structural read is interesting anyway. Defensive AI started outperforming offensive AI, and the gap closed through careful orchestration of cheaper models. No bigger training run involved. Underpriced category.
What changes from here
A few categories move up the list for any investor doing diligence next week. Forward-deployed engineering teams in markets OpenAI’s services arm will not reach first, meaning regulated healthcare, the European Union, and parts of Latin America. Software that controls the inference bill. Defense autonomy at Series A and Series B, where dual-use buyers and appropriated budgets are insulated from the rate path. Power and grid infrastructure for the data centers that are now displacing residential ratepayers in places like Lake Tahoe, where the utility told forty-nine thousand residents last week that it will stop serving them after May of 2027 in favor of hyperscaler load. Defensive cyber, after MDASH.
A few move down. Generic AI services consulting. AI-native finance and CFO startups that compete with what PwC stood up. Personal finance chat assistants that compete with the new ChatGPT default. SMB bookkeeping AI that competes with Anthropic’s connector pack. Mobile-first coding agents after Codex moved into the ChatGPT mobile app. Anything sold to enterprises last year on the assumption that token prices would fall faster than usage rose.
The questions that experienced investors will ask founders this week are sharper than usual. If a Big Four firm stood up a vertical practice around a competitor’s model on Wednesday, what part of the wedge survives? What does monthly agent burn per active customer look like, and at what contract value does the unit economic flip? If the founder is quoting benchmark numbers, which independent lab has reproduced them?
For our late-stage book, the actionable read is narrow. The cleanest secondary mark for Anthropic this week is the Forge institutional price of two hundred sixty-four dollars and fifty-seven cents, set on Friday. Tokenized SPVs trading on retail platforms imply a much higher number; Anthropic publicly disavowed them on Tuesday, which makes them an unenforceable price signal. The Bloomberg report of a thirty billion dollar primary above nine hundred billion would reset the book if it closes. OpenAI’s last verified mark of eight hundred fifty-two billion from April held quiet. Cerebras gave the public market its first non-Nvidia AI silicon comp at roughly fifty-six billion fully diluted. Anduril at sixty-one billion is the new reference for autonomy. SpaceX is the same watch path it has been for two months, with the S-1 still the catalyst that matters.
The week ahead has its own. Google I/O on Tuesday, where the new Gemini is expected to land short of the frontier ceiling. Nvidia and Walmart earnings on Wednesday and Thursday. Independent reproduction of the MDASH and Mythos benchmark numbers when it comes. And the slow approach of the June 17 FOMC, the new chair’s first.
That meeting is the one that will tell us whether the disconnect between the cap chart and the macro print closes from the top, from the bottom, or for a while not at all.