The Model Was Never the Hard Part. Nine Billion Dollars Just Proved It.

In eight weeks, Microsoft, AWS, OpenAI, and Anthropic each stood up a unit whose entire job is to embed their own engineers inside customer companies. Combined, they put more than nine billion dollars behind it.

The trigger is one uncomfortable number. An MIT study found that 95 percent of enterprises spending on generative AI got no measurable return. The models work. The deployment does not.

The scarce skill is no longer access to a good model. It is the person who can wire a model into a real workflow and make it survive production. Price yourself, and staff yourself, accordingly.

Four labs, one move, almost no gap between them

Start with the timeline, because the timeline is the story.

In early May, OpenAI and Anthropic announced billion-dollar deployment ventures within days of each other. OpenAI formed The Deployment Company, a joint venture it majority owns and controls, raising over four billion dollars from a TPG-led group of investors and folding in Tomoro and its roughly 150 forward-deployed engineers. Anthropic set up a 1.5 billion dollar joint venture with Blackstone, Hellman and Friedman, and Goldman Sachs to put engineers inside mid-sized companies.

On June 30, AWS committed a billion dollars to its own Forward Deployed Engineering unit, led by Francessca Vasquez. The model there is specific: pods of five or six engineers embed with a client for roughly 45-day cycles, and the work is priced on fixed outcomes, not billable hours.

Two days later, on July 2, Microsoft launched the Frontier Company, 2.5 billion dollars and around 6,000 engineers and industry specialists, led by Rodrigo Kede Lima. It is deliberately model-agnostic. A customer can run OpenAI, Anthropic, Microsoft, or open weights, with Accenture, EY, KPMG, and PwC brought in as delivery partners. Early engagements name LSEG, Land O'Lakes, Unilever, and Novo Nordisk.

Four of the biggest names in AI, four separate units, one strategy, all inside a two-month window. When companies that compete this hard converge this fast on the same answer, they are responding to the same problem. And the problem is not a modeling problem.

What they are actually admitting

For three years the pitch was that intelligence would arrive through an API. You would buy tokens, point them at your business, and value would follow. The model was the product, and everything downstream was your problem to figure out.

These nine billion dollars are a retraction of that pitch. The labs are now saying, with their capital allocation rather than their marketing, that the model is not the product. The working system is the product, and the working system does not assemble itself from a subscription.

That is a strange thing for a model company to admit. It means the thing they sell, raw capability, is necessary but nowhere near sufficient. The gap between a capable model and a deployed outcome turned out to be wide enough that they would rather staff it themselves than keep watching customers fall into it.

The 95 percent is the whole reason

The MIT Project NANDA number is the pressure behind every one of these announcements. Somewhere between 30 and 40 billion dollars of enterprise generative AI spend, and 95 percent of organizations report no measurable return. Only about one pilot in twenty produces real, trackable impact on the books.

Sit with what that implies. The models are not the failure point. The same models that clear hard benchmarks and write working code are sitting inside enterprises producing nothing the finance team can find. The failure is in the last mile: the integration, the data plumbing, the workflow redesign, the evaluation, the part where a demo becomes a system people actually use every day.

That last mile has a name now, and it is a job, not a feature. A forward-deployed engineer is someone who sits with the customer, learns the actual workflow, and builds the connective tissue between a general model and a specific business process. It is unglamorous work. It is also, apparently, the difference between the 5 percent and the 95 percent.

What this means if you build or buy

Three things follow from this, and none of them are abstract.

If you are an engineer, the market just told you where the scarce value sits. It is not in prompting a model, which everyone can do, and it is not in access to the model, which is a commodity you rent by the token. It is in the ability to take a capable model and make it hold up inside a messy, real system with real data and real users who will not tolerate a flaky output. That skill is what four labs are now paying a premium to hire and embed. If that is what you do, you are underpriced.

If you run a team, notice that the fix these companies chose was people, not another platform. They did not ship a new tool to close the 95 percent gap. They hired engineers and put them next to the problem. Your own version of that is a deliberate choice to staff deployment as a first-class function, not to treat it as something that happens for free once the license is signed.

If you are buying, read the pricing model as a signal. AWS charging on fixed outcomes instead of hours, and doing it in 45-day cycles, is a tell. It says the vendor now believes the risk lives in whether the thing works at all, and they are willing to hold that risk. When the seller starts underwriting the outcome, that is the clearest admission yet that the outcome was never guaranteed by the model alone.

The part not to overcorrect on

This is not a story about AI failing. The capability is real and it is still compounding. The story is narrower and more useful than that: capability and deployment are two different products, and the industry priced only the first one for three years.

It is also expensive in a way that will not scale to everyone. Embedded pods and 6,000-person units are how you serve LSEG and Novo Nordisk. They are not how a fifty-person company gets AI into its workflow. Most organizations will never get a forward-deployed engineer from Microsoft. What they get instead is the lesson those engineers embody, which they can apply themselves: the model is the easy 20 percent, and the other 80 percent is the workflow, the data, and the evaluation that proves it works.

The labs just spent nine billion dollars making that lesson impossible to ignore. The cheapest way to learn it is to watch them pay for it, and then go do the last mile yourself before someone bills you for it.