Sequoia's framing of the next trillion-dollar company is a sentence that has been circulating among the operators I talk to: a software company masquerading as a services firm. Not a tool that helps a professional do the work better. The firm that delivers the work itself. Konstantine Buhler has been developing the thesis publicly for the last year; Pat Grady and Sonya Huang's earlier essay Generative AI's Act Two laid the groundwork for it. The most recent summary by Charles Mei at The AI Opportunities sharpens the architectural implication.
The line lands harder the longer you sit with it. It is not a marketing pivot. It is an architectural one. For the last twenty years, software companies sold software, and services companies sold services, and the boundary between them was the boundary of what code could automate. AI dissolves that boundary. What you sell stops being a tool sitting next to the work. It becomes the work itself. Aaron Levie has been making the same case from the operator side for months, arguing that services-as-software is the only category large enough to absorb the next decade of AI deployment.
The interesting part is what this implies for product shape. If the unit of sale is no longer the seat, the API call, or the workflow license — if it is the closed book, the resolved ticket, the filed demand letter, the shipped pull request — then the surface of the product changes from the inside out. The UI shrinks. The brain expands. The company information layer becomes the moat. Headless is what this architecture looks like when you draw it.
The $6-to-$1 prize
The numerical premise of the thesis is simple, and Buhler has repeated it across enough talks that it has become the canonical statement of the opportunity. For every dollar the world spends on software, it spends roughly six dollars on services. Accounting software is a market. Accounting is a market six times larger. Legal software is a market. Legal work is a market many times larger. The same multiple holds across customer service, recruiting, sales, claims handling, software development itself. Jared Friedman at Y Combinator has been pushing the same shape at the seed stage, telling founders to skip the SaaS layer entirely and start at the deliverable.
Software companies have been competing for the smaller of the two budgets for two decades. They were doing it because services budgets were structurally out of reach: you needed humans to deliver them, and software companies were not in the business of hiring delivery armies. Big Four firms were in that business, and they captured most of the upside.
What AI changes is not the size of the prize. The $6 was always there. What AI changes is the reachability. A product that delivers the work itself, with the AI doing most of the doing, can now credibly bill for outputs without needing the labor stack underneath. The economics that used to require a partnership of 800 senior associates can now be approached by a company of 80 engineers and 20 domain experts. The labor budget becomes addressable for the first time.
The companies that read this correctly are not the ones racing to add a copilot to existing software. They are the ones starting over from the output side. The product is not the chat box. The product is the deliverable.
Headless is the architecture of selling work
Headless, in this context, does not mean no UI. It means the UI is no longer the product. The UI exists for inputs (briefing, context, oversight) and for audit (review, approval, dispute), not for the work itself. The work happens behind the UI, in a layer the customer does not see, and does not need to see.
Three layers do all the work. The AI brain is the foundation model, almost always rented from a frontier lab. The company information layer is the proprietary substrate the firm has built up over time: customer data, integrations into the systems where the work lives, workflow knowledge encoded as prompts and evaluation harnesses, the regulatory clearances and trust contracts that let the firm be in the loop at all. The input and audit layer is whatever surface the buyer uses to brief the work and approve the output.
Compare this to the SaaS stack of the last decade. In a traditional SaaS product, the UI is the product — the screen is where the value is delivered, the rest of the stack exists to make the screen possible. In a copilot SaaS, the UI is the product plus an AI helper bolted onto the side; the work still happens on the screen, with the AI in support. In a headless AI firm, the screen is the thinnest part of the stack. The work happens in the brain and the context. The screen is just where the deliverable lands.
This shape has a knock-on effect: the team you need to build it is different. Traditional SaaS rewards design and frontend depth. Headless AI firms reward model evaluation discipline, data engineering, integration breadth, and domain expertise that can be encoded into the context layer. The hiring chart is not the same chart.
Tools decay, work compounds
The architectural choice has an economic consequence that compounds in one direction and decays in the other.
A company that sells the tool is structurally short the model lab. Every model generation makes the user more powerful with less product. The tool becomes less differentiated; the same task gets done in fewer clicks, with less reliance on the specific surface the tool offers. Eventually the model lab itself ships the workflow as a feature, or a thinner wrapper does, and the original tool's margin compresses.
A company that sells the work is structurally long the model lab. Every model generation makes the same output cheaper to produce. The price the firm charges the customer is anchored to the value of the deliverable, not to the cost of the compute behind it. Margin expands with each model improvement, until competition catches up — at which point the firm reinvests the margin into deeper context, broader integrations, or new categories of work.
This is not a theoretical asymmetry. Anthropic moved from a $9B to a $30B run-rate in twelve months. None of that came from existing customers saving 5% on email drafting. It came from customers building businesses on top of the API, billing for outputs that the API made possible, and reinvesting the margin into the next layer of context. The lab captured the model layer. The work-firms captured the layer on top. Dario Amodei's framing of "the country of geniuses in a datacenter" is the same observation read from the model lab's side of the boundary: the brain is now a rented commodity at unprecedented capability, and the value capture happens wherever that brain gets pointed at a domain.
The companies that read the next twelve months correctly are the ones that figure out which side of this asymmetry their architecture sits on, and rebuild if they are on the wrong side.
The shapes that are shipping
The thesis is no longer aspirational. The shapes are visible in the market.
Harvey, under Winston Weinberg, sells legal work to BigLaw firms. Not legal software. Not a copilot bolted onto Microsoft Word. The unit of sale is the memo, the diligence summary, the contract review. The product is the deliverable.
Sierra, the company Bret Taylor and Clay Bavor started after Taylor's run at Salesforce and OpenAI's board, sells resolved customer service tickets to enterprises. 40% of the Fortune 50 already run their agents in production. The last round priced the company at $15B. Taylor has been explicit in public about the pricing model: the buyer does not pay for seats, they pay per resolution. The unit of value is the closed ticket, not the seat.
Cognition, founded by Scott Wu, sells software engineering work. Devin is not an IDE feature. It is an engineer that takes tickets and produces pull requests. The competitive comparison is not Cursor or Copilot. It is a junior developer hire.
11x.ai sells SDR output — meetings booked, leads qualified — instead of sales tooling. EvenUp sells legal demand letters to plaintiff firms instead of legal templates. Decagon sells front-line CX resolution to product companies.
Read the pricing pages of each of these companies. None of them charge per seat. Most of them charge per outcome, with a floor. The contract that they negotiate with the buyer is structurally a services contract, not a software contract. The legal and procurement category they get filed under, increasingly, is "vendor" or "delivery partner", not "SaaS subscription."
This is the part that has caught most incumbents off guard. The buyer-side procurement motion for $50K of SaaS is different from the procurement motion for $50K of delivered work. Different stakeholders, different SLAs, different escalation paths. Headless products force the entire commercial layer to change, not just the architecture. Even Marc Benioff's positioning of Agentforce as "the third wave of AI" — past copilots, into autonomous agents priced per conversation — is, read carefully, an incumbent's attempt to reach the same revenue line from the SaaS side. The destination is the same. The starting position is harder.
The moat is the company information layer
The strategic question, the one that determines whether a headless AI firm is a real company or a thin wrapper over a model lab, is this: what does the firm own that the lab cannot replicate?
The answer is not the model. The model is rented from a lab that is structurally better at building models than any vertical firm could be. That is fine. That is the right factoring. The firm should not be in the business of training frontier models. It should be in the business of putting frontier models to work.
What the firm owns, and what the lab does not, is the company information layer.
That layer is composed of four things, in roughly increasing order of how hard they are to copy: integrations into the systems where the work actually lives (CRMs, ERPs, ticketing platforms, court e-filing systems), proprietary workflow knowledge encoded as prompts, evaluation harnesses, and review patterns, the data signal accumulated from past runs that nobody else has access to, and the trust and regulatory standing earned with the buyer over time, including the contractual right to be in the loop.
None of these compound on rented infrastructure. They compound on owned infrastructure. The firms that figure this out spend their second and third years building this layer, not building features. The firms that do not, find their margin eroding as the lab below them ships the next model and the lab next to them ships a thinner wrapper that does the same job for less. Andrej Karpathy's line about software 3.0 — that the prompts and the context are the new code — is the same idea pointed at a different audience. The defensible IP is migrating into the part of the stack that the model lab does not see.
This is the architectural decision I have spent the most time on. Building anything that puts proprietary differentiation into the model layer — fine-tuning the lab's frontier model with your data, hoping it stays differentiated — is building on rented land. Building proprietary differentiation into the context layer is building on owned land. The two strategies look similar in a slide deck. They produce very different companies three years out.
The middle position is the worst
The category that loses hardest in this shift is the one that does not commit. Copilot SaaS — bolt a model onto existing software, ship the result as a productivity feature, charge a premium tier — is the worst seat in the house.
It does not capture the services budget, because the product is still being bought as a tool, with seat-based pricing and a software procurement motion. It does not defend the tool margin, because the next model generation makes the copilot a smaller and smaller share of the value the user receives, and the lab will eventually offer something similar natively. Copilot SaaS is the position that loses to both directions at once.
The companies in this seat have, broadly, eighteen to twenty-four months to pick a side. Either they go up the stack, build the context layer, and start selling work — accepting that they are a different kind of company now, with a different team, a different sales motion, and a different category. Or they accept that the tool is going to keep compressing into a smaller and smaller share of value, and they price accordingly, and they hold their position by being deeply useful in a narrow corridor.
Most will pick neither cleanly. That is the cohort that gets quietly acquired or quietly winds down across 2027 and 2028. The clean picks are the ones that survive.
The bet I am making
The architectural decision in front of any Head of AI right now is the one this thesis forces. You can keep building products where the screen is the product and the model is a feature. Or you can start building products where the deliverable is the product, the model is a rented brain, and everything you own lives in the context layer.
The bet I am making, with the time and capital I have, is on the second one. The model is interchangeable. The context, the integrations, the institutional knowledge of how the work actually gets done, the trust earned with the buyer who is willing to outsource the work to a system — those are what compound. Building anywhere else is building on land you do not own.
The next trillion-dollar company will not have a homepage you visit to log in. It will have an SLA you sign. The product will be the company. The company will be the product. The screen will be the smallest part of either.