In June, the AI industry reorganized around the engineers who deploy it.

The read

Strip the framing and this is an argument about forward deployment. Every defense on the list is a by-product of putting engineers inside the customer: you learn the exceptions because you handle them, you accumulate the feedback because you run in production, you own the execution layer because you built the integration. The moat was never the model. It is the residue forward deployment leaves behind, which is exactly why the labs are integrating downward into services, and why the squeezed middle, the undifferentiated app wrappers with no deployment surface, is the audience this memo is quietly written for.

The read

Watch the topology. In May, Anthropic bought a forward-deployed team outright to build a captive delivery core. In June, it rented one of the largest services benches on earth. Doing both at once tells you how acute the last-mile shortage is: even a frontier lab cannot hire deployment capacity fast enough, so it buys some and borrows the rest. It also sets up the year's central tension. The labs now compete with the very integrators they partner with, and both are tying delivery to a single model. The work that survives that squeeze is the kind no lab-owned services arm can credibly offer: model-agnostic, code left in the client's repos, paid on the outcome.

The read

Six months ago 'FDE' was a curiosity on job boards. In June it has a comp ladder that runs into the high six figures, a credentialing pipeline (a16z just opened an FDE fellowship), and now formal, P&L-bearing orgs inside public companies. Databricks naming the function and putting customer counts behind it is the tell that this is structural, not a hiring fad. The constraint it is built to relieve is the same one the labs are spending billions on: models that dazzle in a demo and stall in production. The work that closes that gap finally has a name, a number, and a place on the org chart.

The read

Two deals, two answers to where applied-AI value accrues. Cursor is the productizable end: a coding agent that scales like software and sells at frontier-adjacent prices. Fin is the other end, an autopilot that hands over a finished outcome, the thing investors now mean by service-as-software. What neither acquisition buys is the messy, non-productizable middle: the integration into one company's systems of record and its thousand exceptions. That is the work forward deployment exists to do, and the reason the labs are assembling services arms instead of simply shipping better models.

The read

The supply side is the whole game. Demand for forward-deployed engineers has run far ahead of the people who can do the work, which is why the labs acquihire entire teams instead of recruiting, why compensation has climbed into the high six figures, and why a venture firm is now manufacturing the talent itself. An eight-week fellowship does not close a structural shortage. But when a16z, Salesforce, and the labs all move to build supply in the same quarter, the scarcity stops being anecdotal and becomes the binding constraint on how fast the category can grow. Supply, not capability, is the ceiling now.