Sequoia's generative AI thesis from 2022 to 2026
Sequoia has published a generative AI essay almost every year since 2022. Read in order, they trace a shift from model capability to customer value to selling finished work.
Sequoia Capital has written about generative AI in public since 2022. Its essays are useful because the firm revisits its own predictions and says where it was wrong. Read in sequence, they show a clear shift. The focus moved from what models can do, to whether customers stay, to what work an AI product can complete and charge for.
2022, A Creative New World
In September 2022, Sonya Huang and Pat Grady argued that generative AI could make billions of knowledge and creative workers at least 10% more efficient. On that basis, Sequoia wrote that generative AI "has the potential to generate trillions of dollars of economic value." The essay was a bet on a platform shift, written before ChatGPT launched.
2023, Act Two and the retention problem
A year later, "Generative AI's Act Two" reported that startups alone had passed $1B in generative AI revenue. It also named the problem. Sequoia found a median daily-to-monthly active user ratio of 14% for generative AI apps, against 60% to 65% for the best consumer companies. The firm concluded that the biggest challenge was proving value, not finding demand.
Sequoia described the first year as "technology-out" and predicted a second act built "customer-back." It also revised an earlier view on defensibility. The firm wrote that the moats are in customers, not data, and that workflows and user networks were proving more durable than data flywheels.
2024, the $600B question
In June 2024, David Cahn updated his math on the gap between AI infrastructure spending and AI revenue. His method was to take Nvidia's run-rate revenue forecast, double it for total data centre costs and double it again for the end user's margin. The result was a $600B revenue requirement, up from $200B the previous September. Cahn still concluded that company builders focused on delivering value to end users would be rewarded, and that falling compute prices are good for startups.
2024, Act o1 and service-as-a-software
In October 2024, Sequoia's "Act o1" focused on reasoning models that spend more compute at inference time. It introduced the phrase "service-as-a-software." The essay contrasted the cloud era, which sold software by the seat, with an AI era that sells work by the outcome. Its example was Sierra, which is paid per resolved customer issue. Sequoia called the application layer the most interesting place for venture capital.
2026, This is AGI
In January 2026, Grady and Huang argued that long-horizon agents meet a functional definition of AGI: the ability to figure things out. They wrote that the applications of 2026 and 2027 would be doers rather than talkers, and that selling work is now possible.
Three modules from the sequence
- Capability is not the moat. Sequoia moved away from data flywheels toward customer workflows as the durable advantage.
- Retention is the test. The 2023 essay judged products by daily use, not sign-ups.
- Pricing follows the work. By 2024 Sequoia was pointing founders at outcome-based pricing against the services budget.
Founder takeaways
Treat these essays as a record of what changed in four years, not as a forecast to follow blindly. The constant across all five is the customer. Measure daily active use, own the workflow end to end and price against the work your product replaces. Sequoia's own revisions show that the founders who did this stayed ahead of each shift in the models.
Sources
- Generative AI: A Creative New World
Sequoia Capital (Sonya Huang, Pat Grady) · Sep 19, 2022 - Generative AI's Act Two
Sequoia Capital (Sonya Huang, Pat Grady) · Sep 20, 2023 - AI's $600B Question
Sequoia Capital (David Cahn) · Jun 20, 2024 - Generative AI's Act o1
Sequoia Capital (Sonya Huang, Pat Grady) · Oct 9, 2024 - 2026, This is AGI
Sequoia Capital (Pat Grady, Sonya Huang) · Jan 14, 2026
AI
Sequoia's cognitive revolution essay compares AI to industry
Sequoia partner Konstantine Buhler argues the AI transition will follow the Industrial Revolution's pattern, at larger scale and higher speed. He sees coding agents as the first application, not the defining one.
· 3 min
AI
AI adoption by Canadian businesses in StatCan's 2026 survey
Statistics Canada's June 2026 survey found that 19.2% of Canadian businesses used AI to produce goods or deliver services, up from 6.1% in 2024. Two in five still say AI is not relevant to them.
· 3 min
AI
Sequoia's MAD framework from AI Ascent 2026
Sequoia's fourth AI Ascent brought more than 150 founders and researchers to San Francisco. Pat Grady's advice for builders on top of the labs was a three-part framework called MAD.
· 3 min