Sequoia Capital global partner Pat Grady told the Boston College Investment Committee in a recent closed-door session that the AI cycle should be read as a rebuild of computing itself, not as another chapter in the long history of information distribution.
The presentation, compiled from the video “AI for BC IC” and translated by AGI 2050, focused less on broad public narratives and more on the mechanics of commercialization, research priorities, and capital formation. Grady’s argument was that AI is expanding beyond the software market and into professional services, that frontier labs have already shifted their center of gravity from AGI to ASI and recursive self-improvement, that a clear diffusion gap separates model capability from enterprise adoption, and that private-market pricing has developed a new structure under overheated conditions.
Not an information-distribution cycle, but a computing reset
Grady said venture investors often use a stacked chart to explain how technology waves build on one another: chips at the base, then computing systems, then network connectivity and the internet, then mobile devices and the application layer that followed. AI now sits on top of that stack and gives software a new level of processing power.
Even so, he said this cycle differs from earlier ones in three ways.
First, the total addressable market is much larger. Previous cycles were mostly contests over software TAM. AI, in his telling, is moving directly into services TAM, a market one to two orders of magnitude larger than software alone.
Second, diffusion is happening faster. Grady said that speed is not unique to AI so much as a function of infrastructure readiness. In the early cloud era, fewer than 100 million people worldwide were connected to the internet. In the early mobile era, users still had to buy smart devices one by one. Today, network infrastructure and smart hardware are already widespread, so once a technology crosses the usability threshold, the barriers to scale are far lower.
Third, and most important in his view, the physical logic of this transition is different. Personal computers, local area networks, the internet, and cloud computing were all, at their core, revolutions in information distribution. They improved how efficiently information moved through space. AI changes how information is processed and recombined.
Grady said industry has not seen a foundational shift of this kind since the rise of semiconductor integrated circuits in the 1960s and 1970s. For application-layer founders, that means the technical ground beneath them is still moving at high frequency.
Three turning points: ChatGPT, OpenAI o1, and long-horizon agents
Grady broke the last few years into three key inflection points.
- ChatGPT at the end of 2022 established the technical viability and scaling potential of unsupervised pre-training.
- OpenAI o1, released at the end of 2024, validated the reasoning path and marked an early move from intuitive “System 1” style responses toward slower, more logical “System 2” style thinking.
- Claude Code and Opus 4.5 showed that long-horizon autonomous agents could work in complex task environments.
He said those stages belong to the same technical continuum, but by the time the third one arrived, the industry had crossed a critical threshold. AGI, in his framing, had moved from technical hypothesis to industrial reality.
Grady noted that Sequoia had written earlier this year that 2026 would be “year one” for AGI. At the time, that claim was still debated. He said it has since become a broad consensus across technology and investment circles.
From a better carriage to a car
Borrowing a question associated with Sequoia founder Don Valentine — “So what? What real value did this create?” — Grady said the industry has moved from the stage of an improved carriage to the stage of a car.
For several years, many products labeled as AI were still incremental upgrades to traditional software, just better digital tools. Now, he said, the technology is changing how productivity is delivered.
He pointed to several areas where that shift is already visible.
- Software is moving from a toolbox to a collaborator. Products such as Instinct and Muse, he said, are no longer just task-processing tools. They are becoming business collaborators that deliver outcomes end to end. He compared this to video conferencing before and after Zoom crossed the threshold of usability and trust.
- Advanced expertise is becoming more widely accessible. In healthcare, workers in primary-care settings and remote regions can use mobile devices to access higher-level diagnostic knowledge. In education, systems such as Alpha School are showing how AI-native models could reshape personalized learning.
- Infrastructure spending is spilling into the physical economy. The buildout of hyperscale data centers is creating sustained industrial demand in engineering, electromechanical equipment, and energy infrastructure.
He also described several second-order effects.
- Capital expenditure pressure is rising for hyperscalers. Grady said free cash flow alone is no longer enough to cover compute-related capex, and debt issuance has become a common way to sustain the race.
- Cybersecurity threats are spreading. As foundation models improve, the barrier to building advanced automated attack tools falls, putting pressure on the balance between offense and defense across digital infrastructure.
- Knowledge work is being repriced. White-collar roles centered on text and basic information processing are already under pressure. Workers who use models to amplify output gain a clear productivity edge, while those who fail to build a collaboration layer with these tools face ongoing substitution pressure at the margin.
- Distribution and governance frictions are increasing. Power consumption tied to compute is becoming a geopolitical and regional governance issue, and the gap between returns to capital and returns to labor in the AI era is adding to existing inequality.
Frontier labs have moved on from AGI to ASI and RSI
Inside leading labs, Grady said, the conversation is already highly focused. The main research priority is no longer the engineering realization of AGI. It has shifted to artificial superintelligence, or ASI, and recursive self-improvement, or RSI.
He defined RSI as a loop in which models can participate in, or even lead, the coding, testing, and architectural optimization of the next generation of models. That possibility has pushed research teams to think more carefully about “takeoff” scenarios and alignment failure risk, often referred to as p(doom).
Grady said this is the technical backdrop to recent public calls from figures such as Anthropic co-founder Dario Amodei for the industry to recalibrate the pace of frontier development.
On the engineering and commercialization side, he listed four clear trends.
- Models are moving beyond pure natural language. Training and inference are shifting toward rigorous domains such as scientific discovery and advanced mathematical conjecture testing, which creates a fresh need to optimize the underlying compute architecture.
- Compute supply and demand will remain mismatched. Shortages are still a central bottleneck for both top labs and application companies, and the push toward in-house ASICs and other specialized silicon is getting stronger.
- Competition in inference APIs is intensifying. The fight over token consumption is fierce, and API pricing keeps falling. To lock in usage from large enterprises including Fortune 500 companies, leading labs are building their own enterprise deployment teams and using heavy customization and consulting to drive adoption.
- Acquisitions are becoming routine. As cross-domain competition settles, frontier labs are putting regular mechanisms in place to acquire talent, teams, and assets, and their commercial footprint is starting to resemble that of large enterprise companies.
The diffusion gap is where application-layer companies can win
Grady argued that there is a visible lag between what foundation models can do and what industry has actually absorbed. He called that lag the “technology diffusion gap,” and said it is the main opening for application-layer startups.
The most commercially reliable entry points in this cycle, he said, are in high-value knowledge work: software engineering, cybersecurity, healthcare services, finance and investment banking, audit, and compliance. Each vertical with a strong professional knowledge barrier has room, in his view, to produce a next-generation platform company worth tens of billions of dollars.
He compared the moment to the cloud era, when ServiceNow, Workday, and Salesforce established themselves as systems of record. A new generation of systems of record is now forming across professional verticals, with AI as the interaction layer and logic engine.
Some vertical architectures, he said, are already showing efficiency levels well beyond those of general-purpose large models.
- In biopharma, models tuned for specific scientific computing tasks can generate candidate molecules in very short cycles, compressing the early-stage design work in drug development.
- In minimalist architectures, Grady pointed to Jev as a recent example. He described it as a high-throughput “decision classifier.” In production, many requests that appear to require complex LLM calls are really discrete classification tasks. By aggressively trimming compute, Jev cut per-inference cost by about 100x and reached an unusual ARR jump in a short period.
That pattern, he said, shows how AI-enabled software can produce abnormal growth curves in high-margin, knowledge-intensive niches.
What durable application companies look like
Grady said companies that can build lasting moats at the application layer tend to share three traits.
First, they control their own intelligence stack. More companies are moving core workloads away from closed-model APIs and toward deep post-training on open-weight foundations. The reason, he said, is cost and efficiency. General-purpose models sit on a limited Pareto frontier for general tasks. For a specific commercial workflow, the best economics often come from owning an open model that has been tuned for that job.
Second, they organize around exceptional strengths. Because the underlying stack and model capabilities can change every few months, surviving application companies need very short product rebuild cycles. That pushes them away from traditional collaboration models constrained by the weakest link and toward structures where a small number of top operators drive technical breakthroughs.
Third, governance is becoming more distributed. Hierarchical command-and-control systems are giving way to decentralized structures in which AI agents act as information hubs. Dense automation and autonomous coordination let these organizations run efficiently with very small teams.
Private-market pricing is splitting into a two-step structure
On capital markets, Grady said AI asset pricing over the past year has become sharply polarized and highly structured.
To balance risk and return in an overheated market, some early-stage companies are now using a two-step financing model. Strategic investors that help build the business are separated from purely financial investors that provide expansion capital later.
Based on several transactions Sequoia was deeply involved in over the past year, he described a pattern.
- At the early stage, institutions acting as strategic co-builders lead rounds at what he called relatively reasonable valuation baselines, such as around a $110 million post-money valuation.
- After a short period of business validation, outside financial capital moves in quickly and pushes valuations several times higher in follow-on rounds. In some samples, the average moved above $3 billion.
Grady said that rapid repricing reflects intense demand from private-market capital for companies with real technical and commercial moats. It also shows how much liquidity premium has built up in the market.
Sequoia’s view: the rebuild of global business systems is still in its first half
Grady said no institution has a fully certain map of how the industry will evolve. Still, one rule from past technology cycles remains intact: when the marginal cost of a key productive input falls by orders of magnitude while overall processing capability keeps improving exponentially, structural change across industry becomes irreversible.
By his reading, the technical reconstruction of the global commercial system is still only in the first half.