Last reviewed:

Why DeployCos Exist: The AI Deployment Gap, the Capability Overhang and the Labs' Motives

If frontier models are as capable as their makers claim, why does an industry of embedded engineers need to exist at all? The answer is a measurable gap between what models can do and what organizations have actually put into production - what this site calls the AI deployment gap. This page lays out the evidence for that gap, the organizational reasons it persists, and the additional reasons the AI labs in particular chose to build their own deployment companies rather than leave the work to partners.

The AI deployment gap: adoption is high, scale is rare

McKinsey's 2025 global survey of 1,993 respondents found that 88% reported their organization used AI in at least one business function, up ten percentage points from 2024. Only 7% said AI had been fully scaled across the organization 12.

Adoption is high, scale is rare

Share of McKinsey 2025 survey respondents (1,993 respondents).

Using AI in at least one function: 88%; Fully scaled across the enterprise: 7% Using AI in at least one function88% Fully scaled across the enterprise7%
Source: McKinsey & Company, "The state of AI in 2025" and "AI at work but not at scale," November 2025 12.

The pattern - near-universal experimentation, rare enterprise-wide deployment - is the demand signal behind every DeployCo. An organization that has tried AI in one team and stalled is the typical customer described in the launch materials of the OpenAI Deployment Company, Ode with Anthropic and AWS's FDE unit.

The capability overhang: models can do more than users ask

OpenAI uses the term capability overhang for the gap between what its models can do and how typical users use them. Its January 2026 report, "Ending the Capability Overhang," found from ChatGPT usage data that the typical power user (95th percentile) uses about 7x more "thinking capabilities" - measured as reasoning tokens - than the median user, and that some countries use about 3x more per person than others 3. The finding is OpenAI's own and based on its own usage data; it has not been independently verified, and it measures compute consumption, not productivity.

The capability overhang

Relative volume of reasoning used per monthly active user.

Power user (95th percentile): 7x; Typical user (median): 1x Power user (95th percentile)7x Typical user (median)1x
Source: OpenAI, "Ending the Capability Overhang," January 2026 3. OpenAI usage data; measures compute, not productivity.

GDPval: expert-level output, unused

The strongest evidence that the overhang is real at the task level is OpenAI's GDPval benchmark (September 2025). GDPval covers 44 occupations drawn from the nine industries that contribute most to US GDP, with 1,320 tasks written by professionals averaging 14 years of experience; a 220-task "gold" subset is public 45. In the initial results the top-scoring model, Anthropic's Claude Opus 4.1, produced work that expert graders rated as better than or on par with the human expert's on close to half of tasks, per The Decoder's summary of the paper 6. OpenAI also reported that models completed tasks roughly 100x faster and 100x cheaper than experts when counting only inference time and API cost, while cautioning that this excludes real-world oversight and integration 6.

Jaipuria's conclusion from GDPval is the core of the DeployCo thesis: "most organizations are nowhere close to leveraging AI at that level" 7.

The diffusion gap

Sequoia Capital's Pat Grady used the term diffusion gap at the firm's AI Ascent 2026 event to describe the same phenomenon from the investor's side: capabilities diffuse into the market far more slowly than they are created, and the gap between what models can do and what large enterprises have deployed is the opportunity for companies that build on the labs 78. DeployCos are one answer to the diffusion gap; AI-native rollups, which buy the business and deploy AI from inside, are another. See Market and Economics.

Why the gap persists: organizational alignment

The second reason Jaipuria gives for DeployCos is not technical. Putting AI into a real workflow requires access to the right systems, often a redesign of the workflow, buy-in across teams, and changes in how people work - and none of that is any individual's job 7. AWS's Brian Mitchell made the same point at the launch of its FDE unit: "the complexity isn't the model, it's the integration, the governance, and the operational effort" 9.

This is why the deployment gap looks like a consulting problem as much as an engineering problem, and why demand for implementation services from incumbents has grown alongside the new entrants. Accenture reported about $11.5 billion in cumulative advanced AI bookings across roughly 11,000 projects through its first fiscal quarter of 2026 10. Palantir's US commercial revenue grew 148% year over year in Q3 2026 11. The bookings data is charted in Market and Economics.

Why the labs built their own

Several reasons apply specifically to OpenAI and Anthropic, beyond the general demand 7:

  1. Training signal. Deployment work produces long-horizon traces of how real work gets done across industries. That is valuable either as training data or as insight into where to focus model improvement.
  2. Lock-in in a world of similar models. Model performance across the labs is close for most enterprise workflows. Being the firm that embedded the agents in a customer's workflows creates switching costs that the model alone does not. (The buyer's view of this is in Risks and Debates.)
  3. A bigger addressable market. Building and deploying agents allows charging for outcomes delivered rather than tokens consumed.
  4. Distribution through private equity. Both ventures are capitalized by private equity firms whose portfolio companies are the initial customers. OpenAI's partners sponsor more than 2,000 businesses 12; Blackstone conceived Ode after noticing the gap in its own portfolio rollouts 13.
  5. Demand exceeding delivery capacity. Anthropic's CFO Krishna Rao said at the announcement that "enterprise demand for Claude is significantly outpacing any single delivery model" 14.

Counterarguments

A neutral treatment should note the objections:

  • Services dilute lab economics. Services revenue carries lower margins than API consumption, and forward deployed engineering does not scale like software. Lab-backed DeployCos are separately capitalized partly for this reason.
  • The gap may close on its own. Better agent tooling, cheaper inference (Fournier says the price of intelligence has fallen "100x" in 18 months 15) and more capable models could make embedded engineers unnecessary for routine workflows. The counter-view is that the 7% figure has barely moved in three years of cheaper models.
  • Consultancies already do this. Accenture's $11.5 billion of bookings predates the DeployCo label. The question is whether engineer-led, model-native teams do it better, which is examined in DeployCo vs Consulting vs SaaS.

Frequently asked questions

What is the AI deployment gap?

The difference between how many organizations use AI somewhere (88% in McKinsey's 2025 survey) and how many have scaled it across the enterprise (7%) 1.

What is a capability overhang?

OpenAI's term for the gap between what a model can do and what typical users ask of it; power users draw on about 7x more reasoning capacity than the median user 3.

What is the diffusion gap?

Pat Grady's term for the lag between the creation of AI capability and its spread into enterprise use 8.

Will the gap close without DeployCos?

Possibly in part, as tooling improves. As of 2026 the evidence is that the gap is wide enough to support billions in dedicated capital.

References

  1. McKinsey & Company, "AI at work but not at scale" (chart), November 2025. mckinsey.com
  2. McKinsey & Company, "The state of AI in 2025: Agents, innovation, and transformation," November 2025. mckinsey.com
  3. OpenAI, "Ending the Capability Overhang," January 2026. cdn.openai.com
  4. OpenAI, "GDPval," September 2025. openai.com
  5. OpenAI et al., "GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks," arXiv 2510.04374. arxiv.org
  6. The Decoder, "OpenAI says top AI models are reaching expert territory on real-world knowledge work," September 2025. the-decoder.com
  7. Tanay Jaipuria, "The Rise of DeployCos," 10-13-2026. tanayj.com
  8. Sequoia Capital, "This is AGI," AI Ascent 2026 keynote (Pat Grady), as transcribed. rosetta.to
  9. Constellation Research, "AWS launches forward deployed engineering unit with $1 billion investment," 06-30-2026. constellationr.com
  10. Benzinga, "Accenture says AI demand is rising but scale remains limited," 12-2025 (Q1 FY26 results). benzinga.com
  11. Palantir Technologies, Form 8-K exhibit 99.1, Q3 2026 results, 11-02-2026. sec.gov
  12. OpenAI, "OpenAI launches the OpenAI Deployment Company," 05-11-2026. openai.com
  13. TechCrunch via The Next Web, "Anthropic and Blackstone bet the next fortune in AI is implementation, not models," 07-16-2026. thenextweb.com
  14. Anthropic, "Building a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs," 05-04-2026. anthropic.com
  15. The Decoder, "OpenAI's deployment chief on Codex growth, falling AI prices, and the ROI question," 2026. the-decoder.com