Enterprise AI's Decider Shifts from Model Strength to Whether the Last Mile Lands — July 2, Microsoft launched Frontier Company: $2.5B and 6,000 engineers embedded in customers, co-designing/co-deploying and billing on measurable business outcomes; the trigger is the MIT NANDA stat that 95% of enterprise GenAI pilots deliver zero P&L impact. The same forward-deployed engineering (FDE, Palantir's old playbook) has Amazon in for $1B and OpenAI/Anthropic launching their own in May — the default move for every giant within six months. Judgment: (1) Microsoft's $2.5B buys delivery, not tech — four giants betting on embedding means the model-layer fight is over; what blocks enterprise AI isn't intelligence but the demo-to-production gap, and the 95% dies in the last mile; (2) value migrating to the delivery layer is the next stop after model commoditization ("model from moat to utility" in March, telco-ification in ENTRY 83) — pure-wrapper SaaS gets marked down, "vertical + embedded delivery + outcome-based" becomes the new scarce asset; (3) the trap: FDE is a heavy-headcount, low-margin, hard-to-scale services business — pricing a consulting firm on software multiples is a mismatch; the watershed is who can distill embedded delivery into reusable workflows and agents so marginal delivery cost trends to zero — turn consulting into product and you earn software multiples, fail and you're Accenture with an AI label; (4) for founders: stop pitching "our model is stronger" (95% of failed pilots educated the market) — those who can show delivery count, reusable agents, and falling marginal cost get the next round.