Dinner at 270 Park

I had dinner last night at 270 Park, JP Morgan’s new NYC tower. The building is a destination now. The bar at the top is a destination. The lobby is a destination. The whole steel column is doing the work of telling you that finance is not going away even when every magazine cover this year says it might.

Across the table sat a principal at one of the major Silicon Valley investment shops. Unprompted, before the entrees, he was talking about services as the moat. In fintech. The category his firm has spent fifteen years arguing should be eaten by software. A year ago that same shop would have flagged my deck for the services line. Last night the services line was the thesis.

A few weeks before that dinner I had watched the Masters with a friend who runs coverage at one of the Wall Street shops not named Goldman. He was worried, in the half joking way smart people worry when they actually mean it, that Claude was going to take his job. I told him for free what his own desk bills clients a retainer to hear. The retainer is not for the work. It is for the trust people place in him to do the work. Claude can write the deck. Claude cannot walk into the lobby at 270 Park and have the CFO clear an hour on the calendar. The deliverable has never been the deck. The deliverable has always been the trust that gets the deck taken seriously in the room where the money moves. Knowledge comes with a brand. The brand is the trust. The trust is why the talent does not get automated.

What kind of week makes a fintech investor say that?

The kind of week that just happened.

Tuesday, May 5. Dario Amodei walked on stage with Jamie Dimon. First time the head of a frontier AI lab and the head of an American money center bank had ever shared a stage. By the end of the morning Anthropic had announced ten Claude finance agents in production at JPMorgan, Goldman, Citi, AIG, Carlyle, and Visa, a Microsoft 365 add-in for Excel, PowerPoint, and Word, a Moody’s data partnership, and Claude Opus 4.7. The day before, a $1.5 billion joint venture with FIS to push Claude into the systems running half of US retail banking.

Six days later, Monday, May 11. OpenAI launched the OpenAI Deployment Company with more than $4 billion of initial backing from TPG, Bain Capital, McKinsey, Capgemini, Brookfield, Advent, and a dozen other firms. It is staffed on day one by roughly 150 Forward Deployed Engineers acquired wholesale through the purchase of Tomoro.

Forward Deployed Engineer is Palantir’s coined term for an engineer who lives inside the customer’s operation, owns the integration, owns the outcome, and absorbs the political risk of the deployment. Every industry already had a version of that role. Oil and gas called them field engineers. Banks called them solutions architects. Manufacturing called them integrators. The job has existed for fifty years. OpenAI just bought 150 of them and gave the title trademark protection.

The two companies that built the technology that was supposed to replace services pivoted to services. In the same week. With branding. With a tower of lawyers and consultants standing behind them. The model is the easy part. The work is everything else.

Seven years of the wrong question

For seven years of fundraising at Andium every meeting closed the same way. What is your software ARR? What does gross margin look like net of services? Net of hardware? Net of the licensable detection IP? Net of the cellular and cloud costs? Net of the field hours? What is left, and what is the multiple on what is left?

Hardware was 1 to 3x revenue. Services were 1 to 2x revenue. Software ARR was 8 to 15x. As of February, 5 to 7. The discount was supposed to be permanent. Then the premium was. The spreadsheet wanted everything in the software column, and if it could not fit there it took a haircut.

The implicit message in every one of those meetings was that one day, when AI was real enough, services would collapse into pure software and the multiple would be uniform across the line items. I do not think anyone making that bet realized it would collapse the other way.

The market just rerated

In February, roughly $285 billion of public SaaS market value vanished in 48 hours. The sector has lost more than $2 trillion since late January. Forward P/E on public software is now 22.7 times earnings. The 2021 peak was 84.1. For the first time since the S&P 500 existed, the broad market is more expensive than software.

The mechanism is not subtle. AI agents collapse the seat. The seat was never the value. The work was the value. What survives is the work an agent cannot do alone: install at a site, sit inside a regulated workflow, own the outcome, hold the IP, take the phone call after hours in year three.

Palantir reported Q1 on May 4. Revenue grew 85 percent. US revenue grew 104 percent. Rule of 40 hit 145, a number matched in public markets only by NVIDIA, Micron, and SK Hynix. The company every SaaS purist spent a decade dunking on as “glorified services” is now compounding faster than the chip companies powering the AI build. Accenture booked $3.6 billion of GenAI consulting in a single quarter, more than double the year before. IBM Consulting carries a $6 billion AI services book. Sequoia’s Julien Bek published a thesis in April called “Services: The New Software.” It went viral inside a category that spent fifteen years treating services as the room you stood in while you waited to be invited into software.

Bek’s framing was the cleanest version of the argument. A copilot sells the tool. An autopilot sells the work. For every dollar spent on software, six are spent on services. AI is turning the six into software. He calls the winners autopilots. The labs are calling them deployment companies. The buyers will call them whatever gets the procurement office to sign. His landing line was the loudest: the next trillion dollar company will be a software company masquerading as a services firm.

The companies that got the PS discount are the premium tier now. The companies that priced services at zero are explaining the discount on their own multiples.

The buyers always knew

In 2019 it took us three years of patient work to get inside the large independents in Oklahoma. A board member introduced me to a friend named Clay. Clay knew everyone. The meetings happened everywhere; at a kitchen table in Edmond, on a back porch outside Pawhuska, in an office building downtown where the lobby still had the original wood paneling. Two of the largest US independents. The state’s environmental regulator. A handful of state senators.

None of those rooms wanted a pitch on software ARR. None of them wanted a pitch on the box. They wanted to know whether I would be on the other end of the phone in three years. They wanted to know whether I understood what they had been bolting to that wellhead for thirty years. They wanted to know whose problem the 2 AM surprise was going to be.

The pilots fell out as a formality. The decision had already been made by the time procurement got the email.

Every operator we have worked with since runs that play the same way. They already know what they want. They are evaluating the vendor, not the technology. The question that decides the contract is whether you will pick up in year three. The vendor who fails that test becomes a forever pilot. Trust here is not a credit score. It is a person another person vouches for.

The other word for it is CYA. The operations VP who signs the contract is making a career bet. If the tech fails, the question in the boardroom is who picked it and why. Software ARR does not absorb that question. A vendor on the other end of the phone does. The reason consultants get paid the way consultants get paid is that they sell cover. Now the labs are selling cover too.

The patchwork is the part nobody likes to admit. Industrial stacks run on disparate tools, half homemade and half off the shelf, and they all get implemented by consultants. Salesforce included. The pure software story was a fundraising story. The work was always going to be done by people. The only question was whose payroll the people sat on.

At Andium the answer is that the contract is the company. The box is the wedge. The patent is the lock. The person who picks up in year three is the moat. Industrial was never going to bend toward SaaS. SaaS bent toward industrial.