Microsoft put $2.5 billion this week into a "Frontier Company" that embeds 6,000 AI engineers inside client organisations. Amazon launched a $1 billion forward-deployed-engineering org two days earlier. OpenAI and Anthropic already run one each. So four of the biggest names in AI now sell the same product: people who come to your office and make the models work.
The verdict: the model stopped being the product. Frontier access is turning into a commodity, and the money is moving to the two places a commodity business always defends. The meter, and the last mile.
Why sell engineers when you sell intelligence?
Because raw model access no longer differentiates. Claude, GPT, Gemini, GLM and Kimi are all good enough for most tasks, so API pricing races toward marginal cost. What still carries a margin is deployment: wiring models into a real business. Microsoft's 6,000 embedded engineers are not an AI product. They are Accenture with better model access.
The meter is the second defence. Gartner warned this week that AI coding token costs are heading toward the size of engineering payroll. And Amazon's own engineers are reportedly distilling Anthropic models to cut costs before new token pricing starts next year. Amazon is Anthropic's biggest backer. When the insiders hedge against the meter, the price signal is real.
Buyers route around it already. Coinbase says it moved much of its workload to Chinese models like GLM 5.2 and Kimi 2.7 with automated routing: always the cheapest model that clears the quality bar. That is commoditisation seen from the buyer's side. The model becomes an interchangeable input, picked per request, on price.
I run the same trade at a smaller scale. This whole news operation runs on local open-weight models on one Mac. Not because they beat the frontier. They don't. But for a defined, repeatable workload, good enough at near-zero marginal cost beats excellent at a metered price someone else controls.
Concentration makes it everyone's problem
J.P. Morgan counted 42 AI-linked companies producing 65 to 80% of the S&P 500's profits this week, a concentration the bank itself calls exuberant. If you have a pension fund, that concentration is your problem too. And everyone is hedging vertically: Anthropic is talking to Samsung about custom chips a week after OpenAI's Broadcom deal, and Meta is building a cloud business for its spare compute. Custom silicon protects the bottom of the stack, embedded engineers capture the top, and the metered API in the middle gets squeezed from both ends. Nobody builds a $2.5 billion services org if the API alone carries the valuation.
For a small business the reading is simple. None of those 6,000 engineers are coming to a twelve-person company; you are explicitly not the customer. So keep prompts and workflows portable and treat model choice as a routing decision, not a marriage. Cap token spend like a utility bill. And stay sober on autonomy: in a 500-day simulated software company this week, only three AI models finished above their starting capital. Buy narrow automation for defined tasks and keep a human on the P&L.
One claim you can check against me: this four-way race won't stay four-way. I expect Google, the only frontier player absent from the deployment-services news, to announce its own embedded-engineering or outcome-priced offer within weeks. If I'm wrong, I'll say it here.