Four AI labs spent $9 billion in eight weeks hiring the engineers you compete with

Between May 4 and July 2, 2026, four companies that sell you AI models each announced they are now also in the business of building the systems that run on top of them.

Anthropic went first, on May 4: a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs, capitalized around $1.5 billion, to deliver AI implementation services directly to enterprises. The same day, Bloomberg reported OpenAI was assembling something similar. OpenAI made it official on May 11 — the OpenAI Deployment Company, raising more than $4 billion from TPG (lead), Advent, Bain Capital, and Brookfield, alongside Goldman Sachs, SoftBank Corp., Warburg Pincus, and consulting firms Bain & Company, Capgemini, and McKinsey — launched alongside its acquisition of Tomoro, which brought roughly 150 forward-deployed engineers on day one. On June 30, AWS announced Forward Deployed Engineering, a $1 billion commitment with named early customers including the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh, and Southwest Airlines. On July 2, Microsoft announced Frontier Company: $2.5 billion, roughly 6,000 people, led by Rodrigo Kede Lima, formerly president of Microsoft Asia.

Add the four figures and you get $9 billion, almost exactly, across roughly eight weeks. That number is real. What it is made of is not the same in each case, and the difference matters more than the total.

Two very different kinds of $9 billion

Anthropic's and OpenAI's units are joint ventures with outside capital. Private equity and investment banks are putting real money at risk on the bet that enterprise AI implementation will be a large, recurring, billable business — Anthropic split its roughly $300 million commitment with Blackstone and Hellman & Friedman as near-equal partners, with Goldman Sachs, Apollo, General Atlantic, Leonard Green, GIC, and Sequoia filling out the rest. OpenAI's structure is similar, and Bloomberg separately reports the joint venture's total valuation at $10 billion.

AWS's and Microsoft's units carry no outside investors at all. The billion and the $2.5 billion are internal commitments, funded entirely by Amazon and Microsoft, with no reported breakdown of how much is new spending versus reallocated budget, and no clarity on whether the thousands of engineers involved are new hires or existing staff moved onto a new org chart.

That split tells you what kind of business each company thinks it is running. A joint venture with private equity money attached is a bet on a durable, separately monetizable services line — investors do not put $300 million into something they expect to be a loss leader. An internal-only commitment, by contrast, reads more like a distribution strategy: AWS and Microsoft do not need forward-deployed engineering to make money on its own terms, because its job is to get more workloads running on AWS and Azure. AWS's own materials describe pods of five to six engineers running roughly 45-day engagement cycles with a client — fast, cheap-relative-to-scale, and structured to end with the client's workload already dependent on AWS infrastructure.

What actually changed

None of this is a new idea. Palantir built an entire company on the forward-deployed engineer model — send skilled engineers to live inside a customer's operations, ship production software fast, iterate on-site. What changed between May and July is that the four largest AI vendors decided the model was worth billions of their own capital (or, in Anthropic's and OpenAI's case, someone else's) rather than leaving it to consultancies and system integrators.

The reason is straightforward: model quality has stopped being the main thing enterprise buyers argue about. Sol, Terra, Opus, and Gemini variants are all credible on a spec sheet. What is still genuinely hard, and what enterprise buyers keep failing at on their own, is turning a capable model into a working system against real data, real compliance constraints, and real legacy software — the part no API call fixes. All four companies concluded that whoever owns that last mile owns the customer relationship, the renewal, and a much larger share of the AI budget than API tokens alone would ever capture.

What this means if you build or sell AI implementation

If you are an independent consultant, a small agency, or a solo builder doing exactly this kind of work today, the honest read is that you are now sharing the room with vendor teams backed by nine- and ten-figure capital, a direct line to the model provider's own roadmap, and, in AWS's and Microsoft's case, no requirement to be profitable on the engagement itself. That is real, and it will cost some deals, particularly the large, single-vendor, greenfield builds that look good in a press release.

But the structure of these programs also tells you where they will not go. A vendor's forward-deployed team sells one company's stack, is optimized for the engagements that make a good case study, and is built around fixed-length cycles — AWS is explicit about roughly 45 days. That leaves three kinds of work these teams are structurally weak at, and where an independent builder's position gets stronger, not weaker:

• Anything genuinely multi-vendor. A client running Claude for one workflow, GPT-5.6 for another, and a fine-tuned open model for a third has no single vendor whose forward-deployed team can advise on the mix without a conflict of interest.
• Anything after the engagement ends. A 45-day pod does not stick around for the year two migration, the incident at 2 a.m., or the slow accumulation of technical debt nobody flagged during the sprint. That ongoing ownership is exactly the work an independent builder can commit to and a rotating vendor team cannot.
• Anything too small to be worth a vendor's time. These programs are built for logos like the NBA and Southwest Airlines. A 40-person company with a real but modest AI implementation need is not the target account for a $1–4 billion program built to move enterprise contracts.

The question to ask before any of these teams walks in the door

If a client — or you — is evaluating whether to bring in a vendor's forward-deployed team, ask which kind of $9 billion you are looking at. If it is joint-venture money with outside investors attached (Anthropic, OpenAI), the team has a genuine, separately-measured incentive to deliver a system that works on its own terms, because outside investors are tracking the P&L. If it is an internal-only commitment (AWS, Microsoft), the real incentive is consumption of that vendor's cloud, and the engagement's success metric may quietly be measured in committed spend rather than in whether the system still runs cleanly eighteen months later. Neither is dishonest, but they are different contracts, and the pitch deck will not spell out which one you are signing.

The AINews build-vs-buy framework is a reasonable starting checklist for the underlying decision — lock-in, hidden costs, and who owns the result — applied here to a specific new option: build it yourself, hire an independent builder, or let the model vendor's own AI agent team build it for you, on their infrastructure, on their clock.