The diffusion gap is a market, not a moat

Pat Grady recorded a 15-minute briefing for Boston College’s investment committee on 24 September. It escaped the committee room and travelled, as good venture narratives do, through LinkedIn posts and newsletters. The most useful phrase in it is Sequoia’s “diffusion gap”: model capability is moving faster than companies can absorb it.

This gap exists, but its implications are often misinterpreted. While a market gap generates immediate demand, it does not automatically lead to long-term business viability. Investors must therefore distinguish between startups that provide lasting solutions to bridge this gap and those that are merely temporary placeholders until the underlying model vendors expand their own capabilities.

The market opened before the moats did

Grady frames AI as a revolution in computation rather than another change in how information is distributed. He argues that the addressable market extends beyond software into services, while existing internet and mobile distribution lets new products spread unusually fast. In his telling, the capability curve bent three times: ChatGPT made broad model knowledge useful, reasoning models improved deliberation, and long-horizon agents began to plan, use tools and recover from mistakes.

Sequoia had already made the public version of this case in January. Pat Grady and Sonya Huang called long-horizon agents “functionally AGI” and defined AGI, for investment purposes, as the ability to figure things out. Their definition is deliberately commercial rather than scientific. The useful part is not the label. It is the shift from selling a tool that answers questions to selling work that gets completed (Sequoia, 14 January 2026).

That shift expands the application market. A product can now attack a labour budget, not only a software budget. It can also fail in more expensive ways. An answer in a chat window is easy to inspect. An agent that changes a production system, submits a filing or speaks to a customer needs permissioning, supervision, audit trails and reliable hand-offs. The model may supply more intelligence every quarter, but somebody still has to turn it into accountable work.

This is where the diffusion gap becomes investable. Sequoia’s own AI Ascent notes describe it as the distance between what models can do and what the Fortune 500 has deployed. The firm tells founders to build from the customer backwards, exploit the gap and design around what the user can safely delegate (Sequoia, 8 May 2026).

The dangerous half-life of an AI feature

The same dynamic that creates the opportunity can destroy it. Model providers are not passive infrastructure vendors. They cut prices, ship new capabilities and send deployment teams into large enterprises. Every improvement can erase a layer of application code. A startup may spend six months turning an unreliable model behaviour into a feature, then watch the next model release make the feature native.

Grady’s advice is unusually blunt: an AI company should be ready to reinvent itself every few months. That is sensible operating advice and an awkward investment proposition. Reinvention is not a moat. It is the cost of staying in the race.

A better test is to ask what compounds while the models change. Customer-specific workflow knowledge can compound. So can proprietary feedback, permission structures, integrations, trusted distribution and an evaluation system that measures whether the work was done correctly. A thin interface to one model does not become more defensible with use. A system that learns how a hospital escalates uncertainty, how a bank documents an exception, or how an engineering team rolls back a failed deployment might.

The distinction matters because open models are improving too. In the briefing, Grady argues that companies should be able to “own” intelligence for stable workloads by post-training an open model and running it cheaply. That will not fit every task, but it sets a useful architectural standard: models should be replaceable, evaluations repeatable, and customer value separable from the logo on the API invoice. A startup that cannot swap its model without rebuilding the product has outsourced part of its product roadmap.

The labs are also distribution

The diffusion gap is often described as if a capable model sits on one side and a confused enterprise sits on the other, waiting for a startup to build a bridge. The market is less polite. The labs want the token consumption on the far side of the bridge. They have capital, brand, model telemetry and direct access to major buyers. Their deployment teams can discover valuable workflows before an independent application company sees them.

This does not mean the application layer loses. It changes what founders must own. The strongest companies will control a customer relationship or a domain-specific operating loop that the model vendor cannot serve generically. The closer a workflow is to regulated decisions, messy systems of record, proprietary data and organisational trust, the harder it is to collapse into a general-purpose model interface.

Sequoia’s July investment note on Sable is a clean example of the firm’s thesis and its bias. It says the bottleneck is helping customers understand what AI can do, then presents a multimodal “AI employee” that joins customer calls and operates a browser as one answer (Sequoia, 16 July 2026). The product may be promising. The note is still an investor announcement, not independent evidence of durability. Its claims about demand and deployment should be treated accordingly.

A better diligence question

The LinkedIn circulation around Grady’s briefing focused on the headline claims: “AGI arrived,” four startup opportunity areas, open models and a spectacular step-up between Sequoia’s entry prices and later financing rounds. One detailed newsletter summary reports Grady saying that seven early deals averaged a $110 million post-money valuation before later investors priced them at an average $3.4 billion. That number comes from the speaker and has not been independently verified (The AI Opportunities, October 2026). Even if exact, it says as much about the market for allocation as it does about company value.

For seed investors, the briefing suggests a more useful set of questions:

• If the underlying model becomes ten times better and ten times cheaper, which part of the product becomes more valuable?
• Does usage create proprietary evidence about outcomes, or merely more token spend?
• Can the company change models without asking customers to relearn the product?
• Who owns the permissions, audit trail and failure recovery when the agent acts?
• Is the startup selling completed work, and can it prove that work was completed correctly?
• Is the valuation based on present evidence or an expected markup from the next capital provider?

Grady’s original recording is worth watching because it is candid about both sides of the market: the capability is extraordinary, and the financing environment is frothy (Pat Grady, “AI for BC IC 2026-09,” Loom; LinkedIn circulation supplied for this review). The diffusion gap may support a generation of valuable application companies. It will also support plenty of temporary features sold at permanent-company prices.

The gap is where a startup can begin. The moat has to be built somewhere else.