Anthropic and OpenAI fund separate firms to station engineers inside customers
Anthropic and OpenAI have each capitalized a separate services firm, $1.5 billion and over $4 billion, to embed engineers in enterprises. Wing partner Tanay Jaipuria treats that as evidence of an integration bottleneck.
The Engineer · Build desk
OpenAI's deployment company raised over $4 billion. Anthropic announced a $1.5 billion AI services firm. AWS committed $1 billion to a forward-deployed engineering unit.
| Figure | Value | Note | Claim |
|---|---|---|---|
| OpenAI deployment company, capital raised, over | $4 billion | 4 | |
| Anthropic AI services firm | $1.5 billion | 1 | |
| AWS unit, committed | $1 billion | 6 |
What happened
- Anthropic's firm launched in July as Ode with Anthropic and was built on an acquisition of Fractional AI.
- OpenAI bought Tomoro to start its unit, which opened with at least 100 forward-deployed engineers on day one.
- AWS committed $1 billion to a forward-deployed engineering unit at the end of June.
- Accenture has booked about $11.5 billion of generative AI work over the last three years.
Why it matters
- What it limits constraint Ode targets midsize companies and starts with its private equity sponsors' own portfolio companies, so midsize firms outside those portfolios are not first in line.
- What doesn't add up contradiction The labs have a motive beyond customer adoption: deployment work produces traces usable as training data. The spending is therefore not a clean read on customer need.
- What it costs cost The three stated figures add up to at least $6.5 billion, though they describe a firm's size, capital raised and a commitment.
- Who has to decide decision A team that wants per-viewer data limits in Claude Dashboards has to build and run its own identity and data layer, according to one author's test.
Next story · Invest · yesterday
Anthropic spends about $10,000 a head on forward-deployed engineers while AWS and OpenAI spend billionsConfirmed2 publishers
The case for an integration bottleneck leans on one benchmark, and it is a claim about someone else's workload. Jaipuria, a partner at Wing, calls GDPval the best evidence [14][10]. It takes real work products across many occupations, has experts grade model outputs, and finds that models beat the experts on over half the tasks [10]. GDPval is OpenAI's, and OpenAI is one of the two labs selling deployment help [11]. For the figure to transfer to a given enterprise, that company's work would have to resemble those work products, and its own reviewers would have to grade as the experts did.
The second reason Jaipuria gives is organizational. Putting agents to work means getting access to the right systems, possibly redesigning workflows, winning buy-in across the company and changing how people work [12]. He writes that these cut across teams and that "it's not necessarily any individual's specific job" [13]. In our view this is the stronger of his two arguments.
The demand evidence comes from incumbents. Palantir, which Jaipuria says arguably pioneered the forward-deployed engineer over a decade ago, has accelerated revenue growth in each of the last 12 quarters [8]. CTO Shyam Sankar said on a recent earnings call: "Remember, only Palantir has FDEs. Everyone else has sparkling sales engineers." [9] That is also a CTO describing his competitors.
A smaller case shows what the work looks like. An author writing at thedatavist.substack.com says Claude has little ability to control who sees what on the dashboards it creates [20]. The author allowed "I could be wrong here" [21]. The workaround is an MCP server written in Python with FastMCP. Claude sends each data request with the logged-in viewer's token. The server checks the token against policies and returns only the rows that viewer is entitled to [15]. Auth0 and Google handle sign-in [16]. The author reports four logins and four different answers on one dashboard, with one user refused all data [17].
We think the sources support a narrower claim than the thesis. Enterprises are buying engineer time in volume, and three large vendors have put billions behind selling it [18]. The newsletter does not split the gap between integration work and model capability on a customer's own tasks.
What to watch
- Whether Ode with Anthropic or the OpenAI deployment company takes customers outside the private equity sponsors' portfolios.
- Whether either lab reports revenue or customer counts for its unit.
- Part 2 of the Datavist series, and whether Claude Dashboards gains native row-level controls.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption30
- Hype gap+25
- Incentives70
- Confidence45
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic announced in May a $1.5 billion AI services firm set up with Blackstone, Hellman and Friedman, Goldman Sachs and other private equity firms.
- [2]
Anthropic's deployment firm launched in July as Ode with Anthropic and was built on an acquisition of Fractional AI.
- [3]
Ode with Anthropic is aimed at midsize companies, starting with the private equity sponsors' own portfolio companies.
- [4]
The OpenAI deployment company raised over $4 billion, led by TPG, and is majority-owned by OpenAI.
- [5]
OpenAI acquired a firm, Tomoro, to get its deployment company started, with 100+ forward-deployed engineers on day one.
- [6]
AWS committed $1 billion to a forward-deployed engineering unit at the end of June.
- [7]
Accenture has booked about $11.5 billion of generative AI work over the last three years.
- [8]
Palantir, which arguably pioneered the forward-deployed engineer over a decade ago, has accelerated revenue growth every quarter for the last 12 quarters.
- [9]
"Remember, only Palantir has FDEs. Everyone else has sparkling sales engineers."
- [10]
OpenAI's GDPval takes real work products across many knowledge work occupations and has experts grade model outputs; models can do over half the tasks better than the experts, but most organizations are nowhere close to leveraging AI at that level.
- [11]
GDPval is OpenAI's benchmark, and OpenAI started its own deployment company.
- [12]
Putting AI into practice requires access to the right systems, potentially redesigned workflows, buy-in across the organization and changes in how people work.
- [13]
"it's not necessarily any individual's specific job"
- [14]
Tanay Jaipuria is a partner at Wing and writes a weekly newsletter on the business of the technology industry; he argues DeployCos address a diffusion gap, or capability overhang, between what models can do and what they are used for.
- [15]
In the author's design, Claude sends each request for dashboard data to a custom MCP server with the logged-in viewer's authentication token; the server verifies the token against a set of policies and returns the data that viewer's token entitles them to.
- [16]
The author's MCP server is written in Python using FastMCP, and Auth0 and Google handle authentication, with Auth0 adding the user's role.
- [17]
The same Claude dashboard artefact produced four different answers for four different logins, with one user outright refused access to any of the data behind it.
- [18]
The stated figures for Anthropic's firm ($1.5 billion), OpenAI's company (over $4 billion raised) and AWS's unit ($1 billion committed) sum to at least $6.5 billion, though they describe different things.
- [19]
Deployment work generates long-horizon traces of how real work gets done across industries, valuable as training data directly or as insight into where labs should focus model improvement; Jaipuria lists this as a secondary reason for the labs' DeployCos.
- [20]
Claude has little in the way of being able to control who sees what on the dashboards it creates.
- [21]
The author found no native way to do row-level security in the Claude application, and added "I could be wrong here".
Sources
2 independent publishers whose own reporting we read for this story.
- The Rise of DeployCos (6 minute read)
tanayj.com
1 article · October 15, 2026
- Need to Know: Implementing row-level security for Claude Dashboards via Model Context Protocol (11 minute read)
thedatavist.substack.com
1 article · October 15, 2026
Topics and entities
Topics
- Enterprise AI Adoption
- AI in business intelligence
- AI services firms
- Model Context Protocol
- Forward-deployed engineering
- Row-Level Security