AWS and Microsoft Just Stood Up Consulting Arms Three Days Apart. The Hyperscalers Are Copying the FDE Playbook, Not the Cloud One.

AWS announced its Forward Deployed Engineering unit on Tuesday, June 30. A $1 billion commitment, thousands of engineers on the roster, 45-day embed cycles, pods of five or six per client, first named accounts already running (Allen Institute, Cox Automotive, NBA, Ricoh, Southwest, NFL). Two days later, on July 2, Microsoft answered with Microsoft Frontier Co.: $2.5 billion, 6,000 employees, run by Rodrigo Kede Lima, announced by Commercial Business CEO Judson Althoff. Same premise, larger check.

Three days, two hyperscalers, roughly $3.5 billion of freshly ring-fenced payroll pointed at customer sites. Neither company invented the mechanic they just committed to. Palantir has been running Forward Deployed Software Engineers since 2005 for CIA, NSA, and DoD contracts. Anthropic and OpenAI have been quietly building their own Applied AI groups on the same template since 2024. What happened this week is the hyperscaler admission that the model is not the product, the workflow is, and the workflow does not install itself.

The Two Announcements, In One Table

$3.5 billion between the two hyperscaler moves is small on the balance sheets it came out of. Microsoft's 2026 capex is running past $80 billion, AWS past $105 billion. But the payroll headline is deceptive. What actually got booked is a change in cost of revenue category, from "infrastructure we own and rent out" to "engineers we own and rent out." That is a category the hyperscaler income statement has, historically, refused to grow.

Why the FDE Model Beat the AWS Model, on the AWS Balance Sheet

The Palantir mechanic is unfashionable and specific. One engineer scopes the workflow on day one, ships the first cut inside a week or two, sits with the customer through the production incident six months later, and gets rotated only when the customer can run the workflow without them. It scales poorly by design, because the point is accountability across the full loop, not headcount efficiency. Palantir spent fifteen years being told this was a services business dressed as a software one and would never get a software multiple. Then GenAI turned every enterprise buyer into a Palantir customer.

The MIT NANDA study from earlier this year put a number on the demand: 95 percent of enterprise generative AI pilots deliver zero measurable P&L impact. AWS and Microsoft did not need MIT to tell them that. Their sales teams have been sitting inside the failed pilots for eighteen months. What MIT gave them was public cover for rewriting the go-to-market from "buy an API and figure it out" to "buy the API and we will send six engineers to figure it out with you."

It is the same admission Anthropic just made on the product side. Ten days ago Anthropic launched Claude Science, and TF wrote it up as a harness product wearing a science skin. The model was not the news. The coordinating agent and the reviewer agent and the connectors into 60-plus databases were the news. AWS Frontier Deployed Engineering and Microsoft Frontier Co. are the human-scale version of the same move. If the customer cannot install the workflow, the vendor will install it for them.

What This Does to the Buyable Frontier Model Business

Two things at once. First, it pulls the deployment revenue that was going to Accenture, Deloitte, and Capgemini onto the hyperscaler income statement. Accenture's generative AI backlog crossed $6 billion inside FY 2025. The hyperscalers just told their largest customers that they can source that engagement inside the vendor relationship instead of through a third party. Microsoft's Copilot business alone has been leaking to Global SIs for the last two years, and Frontier Co. is the reclaim.

Second, it changes the shape of the sales motion for Anthropic and OpenAI. Both labs have been running the FDE playbook themselves and both have been sitting inside the same customer accounts as their hyperscaler distribution partners. That was tolerable when AWS and Microsoft were pitching pure infrastructure. It becomes competitive when AWS and Microsoft start sending their own engineers into the workflow. The federal buyer that just signed the first-of-its-kind California Anthropic deal is going to run the same procurement calculus, and Microsoft's new consulting arm is going to bid on the implementation.

The near-term revenue read is bullish for the labs. Bigger implementation teams mean bigger seat counts and bigger inference bills, and the token meter that just landed on the developer through GitHub Copilot (which TF covered in the first 30-day cycle post) is a much easier sell once a Microsoft-badged engineer is on site vouching for the workflow. The medium-term read is harder. The customer relationship that Anthropic and OpenAI thought they owned now has a hyperscaler-badged engineer sitting in it full time, and that engineer is not neutral on which model gets called.

The Margin Question

Forward Deployed Engineering is a low-gross-margin business. Palantir runs roughly 55 to 60 percent operating margins across the FDSE program, but that is after 20 years of reusable Foundry and Gotham tooling underneath. A fresh FDE arm carries the same headcount cost with none of the platform amortization, and hyperscaler income statements have been running 30-plus percent operating margins on rented compute.

Nobody at Microsoft or AWS believes Frontier Co. and AWS FDE will match those margins in year one. The bet is that the FDE seat is a customer acquisition cost for a much larger downstream compute contract, and that the compute margin absorbs the services drag. The same bet Anthropic is making with Applied AI, and the same bet the Blackstone and Hellman & Freeman joint venture with Anthropic (reported at $1.5 billion, with Anthropic, Blackstone, and H&F sharing a $300 million founding commitment) is structured around. The financing gets shifted off the primary balance sheet where it would drag margin, and stays close enough to influence the compute pull-through.

The IPO windows for both frontier labs are inside this same window. Anthropic filed confidentially on June 1, OpenAI is inside its own preparation cycle for a listing that keeps sliding right of 2026. Both S-1 drafts have to explain a services line that is now sitting next to a competing services line from the labs' largest distribution partners. The customer concentration table is going to have to disclose whether a hyperscaler FDE sale that also pulls Claude or GPT-5.5 tokens counts as an Anthropic or OpenAI account, or as a Microsoft or AWS account. The answer probably splits the revenue, and that split is a new footnote the buy side has not modeled.

The Federal Gate Underneath It

Both announcements arrived on top of active federal procurement plumbing. Governor Newsom's California partnership with Anthropic (announced June 29, giving state agencies plus cities and counties Claude at half price) sits in the same procurement window as the AWS FDE unit standing up. Microsoft Frontier Co. is going to bid on the same implementation work. The federal gate that pulled Anthropic's Fable 5 model for 19 days (which TF covered around the Sonnet 5 empty-room launch) is now the same gate every hyperscaler FDE deployment has to clear.

What that means in practice is the security review that used to be a bottleneck on model selection is now a bottleneck on implementation partner selection too. If a California state agency wants Claude and can only source implementation help through Microsoft Frontier Co., the procurement runs through Microsoft's clearance profile, not Anthropic's. Compliance-cleared distribution partners just became a moat the hyperscalers can rent to the frontier labs.

Three Signposts in the Next 90 Days

One. Utilization on the AWS FDE pods. Vasquez said engagements run in 45-day cycles. Two 45-day rotations get us to mid-September. If the first named accounts (Allen Institute, Cox Automotive, NBA, Ricoh, Southwest, NFL) publicly reference a shipped workflow before Q3 earnings, the model works. If the second rotation looks like the first (scope, deploy, hand off, roll to next customer), the utilization math starts justifying the payroll line. If the second rotation looks like extended residencies instead, the services drag is going to show up in Q3 free cash flow before it shows up in the analyst notes.

Two. Whether Google Cloud stands up its own FDE arm. Google has been sitting on the Applied AI expertise inside DeepMind for the same 18 months and has kept it internal. A Google Cloud FDE unit at Big Tech scale would make it a three-way race and would confirm the category has broken out from "pilot support" to "strategic services line." Silence past July would tell you Google Cloud is still running the pure-platform playbook the other two just left.

Three. Whether Anthropic and OpenAI harden or dissolve their own Applied AI groups. The reasonable bet is they harden them, because losing the customer workflow to the hyperscaler badge is worse than the margin drag. Watch for hiring signals on the Anthropic Applied AI page and the OpenAI Forward Deployed Engineer page inside the next 60 days. A hiring surge means the labs are treating this as a defense. A hiring pause means they are ceding the workflow layer to the hyperscaler distribution partner and betting that the model quality gap keeps them anchored in the account.

The Bottom Line

We have been writing for a quarter that the model is not the product and the workflow is. TF has argued it through the tokenmaxxing cliff, through Claude Science, through the Copilot bill cycle, and through the AWS-at-the-origin versus Cloudflare-at-the-edge monetization gateway piece. This week two hyperscalers put $3.5 billion of payroll behind that read. AWS committed on Tuesday, Microsoft answered on Thursday, and both explicitly borrowed a model the labs they distribute have been running for 18 months and Palantir has been running for 21 years.

The buyable frontier model business now has a service layer on top of it, and the service layer is being sold by the same three or four companies that own the distribution. If you are a customer, that is a shorter path to a workflow that actually ships. If you are Anthropic or OpenAI, it is a partner conversation you have to run every quarter instead of every year. If you are Google Cloud, the next 60 days decide whether you sit this category out or match the check. The workflow just became a hyperscaler product line, and the hyperscalers just told the market that is where the next dollar of AI revenue is coming from.