Tech-out
Foundation model labs advance capability and push into applications from the model layer down.
Three Sequoia partners, one arc: a roughly $10 trillion services opportunity, the year of agents, and a human anchor
Guillermo Flor breaks down Sequoia's analysis of the future of AI from its annual AI Ascent summit in three parts. Pat Grady calibrates the moment: this wave is the biggest and fastest yet, aimed at a roughly $10 trillion services market, and the durable advantages for builders are moats, affordance and diffusion. Sonya Huang names 2026 the year of agents and argues that services, not software, is the new addressable market.
Constantine Buhler closes with a long view. Cognitive work is following the arc of physical work, which machines now perform at more than 99%, and he tells four stories (aluminum, alien design, emerging sciences, the art of unreason) before settling on a human anchor: man is the measure of all things.
View this analysis as a KG entity“In a revolution in computation the things that you build might today might be irrelevant tomorrow.”
Every figure is stated by the source; nothing here is estimated by this page.
Grady calibrates the moment: the waves of technology, why this wave is the biggest and fastest yet, what makes it different, and the three pillars (moats, affordance, diffusion) on which to build on top of models.
Modern technology has come in layered waves, each building on the one before. All of these waves are additive.
ChatGPT revealed the power of pre-training to a mainstream audience.
o1 surfaced a second scaling law around inference-time compute: reasoning models.
Claude Code and the Opus 4.5 / 4.7 generation have demonstrated the power of long-horizon agents. Between the second and third inflection there is a hard break, a discontinuous shift in what software can do.
Knowledge workers more productive, without changing the shape of their work.
Productivity gain of the faster-horses eraApplications that fundamentally restructure the work, the workflow, and the organization.
Improvement delivered by the cars eraFoundation model labs advance capability and push into applications from the model layer down.
Startups, largely Sequoia's portfolio shape, build specialized applications on top of the labs and wrap themselves around real customer workflows.
In a computation revolution, capabilities change faster than customers do. Wrap tightly around specific customers and their problems: customer intimacy, workflow integration and earned trust are the moats that compound over time.
The quality of an object that signals how it should be used without explanation. Raw foundation model access has low affordance; the application-layer opportunity is to package immense capability into surfaces so intuitive that the customer reaches the outcome with minimal effort.
The widening gap between how fast capabilities are created in the labs and how fast they diffuse into real enterprises. Closing it is the core mission of the application layer.
“The degree you wrap around your customers is going to be way more valuable.”
Huang argues that if 2022–2024 was defined by chat interfaces and 2024–2025 by reasoning models, 2026 is defined by agents, moving from systems that respond to a prompt to systems that pursue a goal.
In 2022, projects like AutoGPT and BabyAGI became overnight GitHub sensations by wrapping GPT-3 in a planning loop and pointing it at a goal. The concept was right; the execution was not. The agents looped, hallucinated, and failed repeatedly: agents were a known destination years before they were viable.
An AI suggestion in the editor, useful but not transformative.
A human directs an agent or coordinates a small team of them.
Long-running workers that operate without continuous human supervision, often spawning sub-agents of their own.
Pipelines in which human review has been removed entirely. They sound implausible until they appear in production, which is already happening in narrow domains such as cybersecurity.
An agent can analyze a genome, recommend interventions, prescribe medications, and identify clinical trials.
An agent can negotiate contracts, perform litigation work, and settle disputes.
Agents are solving Erdős-class math problems and proposing novel superconductors.
Personal agents are taking over inboxes, calendars, financial accounts, and tax filings.
Buhler pairs physical and cognitive work, treats the industrial revolution as precedent for a cognitive revolution, and tells four stories about the future before closing on Protagoras.
Physical work moves matter through space: the package on the Pony Express, the satellite on a Falcon 9. Cognitive work moves ideas through minds: Pythagoras formulating his theorem, DeepMind solving protein folding. The two are profoundly different in substance, but their revolutions follow a strikingly similar arc.
In the mid-1800s the Washington National Monument, then the tallest building in the world, was capped with 100 ounces of aluminum, then the most precious metal known and displayed at Tiffany's. Within a few decades electrolysis made aluminum cheap, and it became candy wrappers and sandwich foil.
Aluminum is intelligence, electrolysis is artificial intelligence. Capabilities that once required decades to acquire, such as PhD-level expertise in narrow domains, are becoming instantly invocable and effectively disposable.
In 2006, NASA used an evolutionary algorithm to optimize an antenna for a satellite mission. The conventional antenna was clean, symmetric and geometric; the machine-designed antenna was contorted and asymmetric, and dramatically outperformed it.
As AI takes over the design of chips, vehicles, buildings and software, the outputs will increasingly look alien by human standards. The aesthetic and structural intuitions that have governed human design for centuries may not apply.
For nearly a century steam engines were perfected through pure tinkering, by engineers like Newcomen and Watt. Then, in the early 1800s, Sadi Carnot formalized thermodynamics.
AI is currently in its tinkering phase. Within the next two decades a new fundamental science of intelligence, at the level of thermodynamics in scope and rigor, will emerge. It will be taught in high schools and may help master AI and possibly understand consciousness itself.
For tens of thousands of years visual art trended toward realism, from cave paintings to Renaissance masterworks. The daguerreotype and early photography made that pursuit obsolete. Instead of ending, painting reinvented itself by asking how the heart and the soul see the world, and Impressionism, Expressionism, Cubism and Neo-Expressionism followed.
As machines take over the literal execution of cognitive work, human creative and interpretive activity will shift toward the dimensions machines cannot replicate.
“man is the measure of all things”
Three forces define this cycle. A technology floor that moves weekly, an addressable opportunity roughly 10x larger than prior platform shifts, and a playbook that breaks from cloud and mobile precedent.
Sequoia's annual AI Summit, which Flor calls probably one of the most important AI events in the year. It brought together more than 150 leading founders and researchers, including Demis Hassabis, Andrej Karpathy, Greg Brockman, Boris Cherny of Anthropic, Dmitri Dolgov of Waymo and Jim Fan of Nvidia.
Affordance, borrowed from design, is the quality of an object that signals how it should be used without explanation. A hammer has high affordance; raw foundation model access has low affordance, because the average enterprise employee has no idea how to make use of a terminal interface with Claude Code.
Reasoning and planning, action-taking, and iteration. Together they produce agency: the practical ability to get things done.
Using coding as the example: tab autocomplete in 2023, then agentic development, then background and async agents, and at the frontier dark factories where human review has been removed entirely.
Because capable agents make services, not software, the addressable opportunity. In medicine, law, the sciences and consumer life, work once captured by professional service firms is being absorbed into the application layer.
Cognition is following the arc of physical work. More than 99% of physical work is now done by machines, and within a generation 99.9% of cognition on Earth will be too, in a revolution larger in scope and faster in tempo.
Aluminum and the Washington Monument (intelligence becomes disposable), the Age of Alien Design (machine-made forms look strange), Emerging Sciences (a science of intelligence on the scale of thermodynamics), and the Art of Unreason (human creativity shifts to what machines cannot replicate).
Because, in Buhler's reading, nothing in a vacuum has value in itself. Value is conferred by human experience and meaning: AI can perform the work, but only human relationships and human attention supply the reason any of it matters.
Three Sequoia voices: Pat Grady on calibration, Sonya Huang on agents as the story of 2026, and Constantine Buhler on what comes next.
Sequoia estimates a roughly $10 trillion services market is now potentially addressable by machines, though the figure could be $5 trillion or $50 trillion. For scale, US legal services alone is a $400B market, equal in size to the entire global software industry.
It is the biggest yet, the fastest yet, and it is a computation revolution rather than a communication one, so the strategic playbook breaks from cloud and mobile precedent.
November 2022, when ChatGPT revealed pre-training; roughly two years later, o1 and the second scaling law of inference-time compute; and now Claude Code and the Opus 4.5 / 4.7 generation, which demonstrate long-horizon agents.
If a system can be dispatched to perform a job, recover from failure along the way, and persist until the job is complete, it functionally qualifies. Practical capability matters, not the label.
Faster horses made knowledge workers 10–40% more productive without changing the shape of their work. Cars deliver 10–40x improvements and fundamentally restructure the work, the workflow, and the organization.
Foundation model labs advance tech-out, from the model layer down. Startups advance customer-back, building specialized applications that wrap around real customer workflows.
Moats, affordance and diffusion. Capabilities change faster than customers do, so customer intimacy, workflow integration and earned trust compound; affordance packages capability into intuitive surfaces; and diffusion is the gap between lab capability and enterprise adoption that the application layer exists to close.
The quality of an object that signals how it should be used without explanation. Grady rates raw foundation model access as low affordance.
Defined commercially in the talk: a system that can be dispatched to perform a job, recover from failure, and persist until the job is complete.
A revolution in how information is processed, as opposed to a communication revolution in how information is distributed.
The foundation model labs' strategy of expanding capability and pushing into applications from the model layer down.
The startups' strategy of building specialized applications that wrap around real customer workflows.
The widening distance between how quickly capabilities are created in the labs and how quickly they reach real enterprises: foundation models advance daily, the average Fortune 500 absorbs change in months or years.
A durable advantage built from customer intimacy, workflow integration and earned trust, which compounds even as the underlying technology shifts.
An agent able to sustain a task across a long time horizon, the capability marking the third inflection.
A pipeline in which human review has been removed entirely; already appearing in production in narrow domains such as cybersecurity.
The framing that capable agents make services, not software, the largest addressable commercial opportunity.
Buhler's term for the transition in which machines take over thinking, paralleling the industrial revolution's displacement of muscle.
Machine-optimized forms that look strange by human standards, illustrated by NASA's 2006 evolved antenna.
Buhler's forecast that human creative activity, like painting after photography, will shift toward what machines cannot replicate.
Grady's contrast between AI that made work 10–40% more productive and AI that delivers 10–40x improvements and restructures the work.
Interactive graph visualization derived from the companion RDF. Click nodes to resolve, drag to explore. Graph data embedded from companion RDF at generation time.
Query this knowledge graph on URIBurner. The editor opens on the canonical SAMPLE entity-type summary (DAV named graph). Pick a recipe, edit freely, then run live or copy.
Reproduced verbatim from the companion RDF. Execute loads the query into the workbench below and runs it live.
PREFIX schema: <http://schema.org/>
PREFIX : <https://www.theaiopportunities.com/p/sequoia-ai-ascent-2026-the-future#>
SELECT ?metricIri ?metric ?statedValue
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/sequoia-ai-ascent-2026-the-future-claude_sonnet_5_5-1.ttl>
WHERE {
?metricIri a :Metric ;
schema:name ?metric ;
:statedValue ?statedValue .
}
ORDER BY ?metricRuns against the named graph this file is loaded into on URIBurner; every row projects the IRI of the entity it describes.
PREFIX schema: <http://schema.org/>
PREFIX : <https://www.theaiopportunities.com/p/sequoia-ai-ascent-2026-the-future#>
SELECT ?stageIri ?position ?stage ?description
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/sequoia-ai-ascent-2026-the-future-claude_sonnet_5_5-1.ttl>
WHERE {
:codingSpectrum schema:hasPart ?stageIri .
?stageIri schema:position ?position ;
schema:name ?stage ;
schema:description ?description .
}
ORDER BY ?positionRuns against the named graph this file is loaded into on URIBurner; every row projects the IRI of the entity it describes.
PREFIX schema: <http://schema.org/>
PREFIX : <https://www.theaiopportunities.com/p/sequoia-ai-ascent-2026-the-future#>
SELECT ?waveIri ?position ?wave ?baseIri ?base
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/sequoia-ai-ascent-2026-the-future-claude_sonnet_5_5-1.ttl>
WHERE {
?waveIri a :Wave ;
schema:position ?position ;
schema:name ?wave .
OPTIONAL { ?waveIri :builtOn ?baseIri . ?baseIri schema:name ?base }
}
ORDER BY ?positionRuns against the named graph this file is loaded into on URIBurner; every row projects the IRI of the entity it describes.
PREFIX schema: <http://schema.org/>
PREFIX : <https://www.theaiopportunities.com/p/sequoia-ai-ascent-2026-the-future#>
SELECT ?talkIri ?talk ?claimIri ?claim
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/sequoia-ai-ascent-2026-the-future-claude_sonnet_5_5-1.ttl>
WHERE {
?claimIri a schema:Claim ;
schema:name ?claim ;
schema:isPartOf ?talkIri .
?talkIri a :Talk ;
schema:name ?talk .
}
ORDER BY ?talk ?claimRuns against the named graph this file is loaded into on URIBurner; every row projects the IRI of the entity it describes.