Sequoia Capital partner Pat Grady has released an internal presentation he delivered on Sept. 24 to the Boston College Investment Committee, outlining how the firm is reading the current AI cycle. Titled “What on earth is happening in AI right now?”, the presentation spans AGI, frontier model labs, AI applications, startup valuations, data-center spending and social risk.
Grady said Sequoia now believes AGI has already appeared, and that the real turning point was not ChatGPT or reasoning models, but the arrival of long-horizon agents in 2025. He said Sequoia had publicly predicted in January that “2026 will be the year of AGI,” a view he described as strongly contrarian at the time. By September, he said, that position had moved much closer to market consensus.
Grady calls this AI wave a computing revolution, not just another internet shift
Grady placed AI in the context of several decades of technology cycles, from chips and computer systems to networks, the internet, mobile devices and now AI. In his view, each wave has built on the layer before it. Still, he argued that the current AI cycle differs from previous ones in three basic ways.
First, he said the market is much larger. AI is not just taking aim at the traditional software market, in his account. It is moving into the far bigger services sector, which he said may be roughly two orders of magnitude larger than software.
Second, he said this is the fastest-spreading technology revolution on record. When cloud adoption started, fewer than 100 million people around the world were actually online. The smartphone wave also required consumers to buy new hardware. Today, nearly everyone is already connected and already has a phone, so AI does not need to build a new distribution channel from scratch. Once a product crosses a reliability threshold, it can spread globally very quickly.
Third, and most important in his telling, AI is not another information-distribution revolution. It is a computing revolution. Personal computers, the internet, cloud and mobile changed how information moved. AI changes how information gets processed. Grady said the last time humanity saw a computing shift this fundamental was during the silicon chip revolution in the 1960s and 1970s.
Three inflection points: ChatGPT, o1 and Claude Code
Grady’s framework highlights three major capability jumps over the past five years. The first was ChatGPT in November 2022, which he said proved the power of large-scale pre-training. The second was OpenAI o1 in late 2024, which showed the market what reasoning could look like.
He compared those two stages to System 1 and System 2 thinking in psychology: fast intuition in the first case, sustained reasoning in the second.
The third stage is the one he treated as decisive. Grady pointed to long-horizon agents represented by products such as Claude Code and Opus 4.5 in 2025, saying they marked the first time AI showed the ability to complete work autonomously over longer timeframes. On the surface, that may look like just a third point on a model-capability curve. In his presentation, though, he described it as a discontinuity. “We think that final one is when we finally arrived at AGI,” he said.
He used the analogy of “a faster horse” versus “a car” to describe the shift. In earlier years, AI was still a faster horse: still software at its core, only stronger. With agents, he argued, AI starts to look like a car. It is no longer just doing existing work faster. It is doing the work in a different way.
From app to assistant: agents are starting to finish jobs, not just tasks
Grady said that shift is already moving into the mainstream. He cited newer AI products such as Instinct and Muse as examples of systems evolving from apps into assistants. In his words, they are no longer just handling tasks; they are actually getting jobs done.
He compared the moment to Zoom. Video conferencing existed before Zoom, but users did not trust it. Zoom’s breakthrough, he said, came when it crossed a reliability threshold and people were finally willing to default to video meetings.
He argued that AI agents are now crossing a similar threshold. The theory that AI could execute work on behalf of users has been around for some time. Only now, he said, are people beginning to hand work over to AI in practice.
Healthcare is one of the clearest beneficiaries in his view, because doctors in remote areas may eventually carry the full body of medical knowledge in their pockets. Education has not broken out yet, he said, because public education systems move more slowly, though he pointed to Alpha School as an early signal of what could come.
A warning sign: hyperscalers are borrowing to build data centers
Grady’s presentation did not focus only on upside. He said 2026 has already produced a notable capital-markets signal: AI capital expenditures by hyperscalers have started to exceed what their own free cash flow can support.
That means major technology companies are no longer relying solely on internal cash flow to build AI infrastructure. They are beginning to use debt financing for capex.
As models become more capable, he said, security risks are also likely to rise quickly. He also listed mental health as another negative externality tied to AI. In the presentation, he referred to what he called “AI psychosis,” saying some users may fall into it after extended conversations with AI. White-collar work, he added, is already being affected across the board.
For Grady, the real line is not whether AI replaces white-collar workers. It is whether white-collar workers learn to use AI. Those who do may continue to thrive; those who do not could run into serious trouble.
Frontier labs are already looking past AGI toward ASI
From the perspective of frontier AI labs such as OpenAI and Anthropic, Grady said another major shift is underway: AGI is no longer the destination. Internal thinking at those labs has moved quickly from AGI, or artificial general intelligence, toward ASI, artificial superintelligence.
One core concept here is RSI, or recursive self-improvement. Grady said frontier models have already been heavily involved over the past year in the work of building the next generation of models. Once AI begins helping improve AI, labs become concerned about a capability takeoff that humans may no longer be able to match.
That feeds directly into p(doom), the probability that AI eventually goes out of control or threatens human survival.
Grady said labs have begun to recognize over the past few months that alignment is a real issue and one they had previously underestimated. He linked that reassessment to the recent cooling in AI progress and the rapid rise in safety discussions around frontier models.
What comes next: science models, in-house chips and continual learning
On the technical side, Grady said models are moving quickly beyond pure language. He identified several key directions for the next phase: scientific and mathematical capability, new model architectures, in-house AI chips and continual learning.
In this context, continual learning means future AI systems may no longer remain fixed after training as one-size-fits-all models. Instead, they may keep learning from their environment and from user interaction.
He also said model companies such as OpenAI and Anthropic are locked in intense token wars. They are cutting API prices to compete for the relatively small set of enterprise customers that consume tokens at massive scale. At the same time, he said they are creating what he called “DeployCo,” sending large numbers of consultants into Fortune 500 companies to build custom AI applications directly and drive more enterprise workloads onto their token infrastructure.
Even with huge capital spending, Grady said labs and AI startups remain broadly “severely compute constrained,” with demand still running ahead of supply.
Sequoia sees the biggest startup opening in the “Diffusion Gap”
Rapid progress in models does not mean enterprises are keeping up. Grady called the distance between the two the “Diffusion Gap”: the set of things models can already do is much larger than the set of things enterprises are actually doing with AI in production.
In Sequoia’s view, that gap is now the largest startup opportunity in the application layer.
Grady said the firm is especially interested in high-value knowledge work, including coding, cybersecurity, healthcare, financial services and accounting. He argued that each large knowledge-work category could eventually produce one or several giant AI companies.
Another important opening is the system of record. In the cloud era, some of the biggest application-layer companies were systems such as Salesforce, ServiceNow and Workday, which controlled core enterprise data and workflows. Grady said a new system of record is now taking shape in the AI era as well.
AI is starting to rewrite frontier tech, including drug design
Grady also highlighted the impact of AI on scientific research. He said companies are beginning to generate candidate drugs in something close to a one-shot process. Where scientists might once have spent years on molecule design, models may now cut that work down to days. The presentation cited Chi Labs as one example.
He also flagged domain-specific AI labs as an area worth watching, rather than building yet another general-purpose frontier lab. One example in the deck was a classifier called Jev. Grady said many jobs enterprises now hand to large language models are, at their core, simply multiple-choice decisions. If those are moved to a specialized classifier, cost can fall by about 100x and speed can improve as well.
He said the company behind Jev went from zero to $100 million in revenue in the past seven days after the product suddenly took off. He also stressed that this was “an outlier among outliers,” not a normal case.
Growth rates are breaking out of the old SaaS scale
Grady said some AI companies are growing at more than 10% per day. He also described another high-value knowledge-work AI company that generated about $200 million in revenue at the end of 2025 and is expected to reach $700 million by the end of 2026.
The presentation intentionally did not reveal the full names of these companies. The point, he said, was that AI-native businesses are already growing on curves that do not look like the old SaaS era.
“Own Your Intelligence”: workloads are moving off foundation models and onto in-house stacks
Another major trend in the presentation was what Grady called “Own Your Intelligence.” He said this shift is not mainly about distrust of foundation models, and not mainly about GDPR, data privacy or security either. The main driver, in his account, is economics.
OpenAI and Anthropic, he said, occupy only a small number of points on the price-performance Pareto frontier. Inside enterprises, there are many different types of workloads. For a large share of those tasks, the most efficient solution is not to call the highest-end foundation model every time. It is to use an open-source model and then post-train it on the company’s own data and task mix.
As a result, he said, there is a major trend of moving workloads away from foundation models and onto companies’ own open-source models.
That could mean application-layer companies gradually reduce dependence on any single model vendor even if OpenAI and Anthropic continue to lead on raw capability.
Jev as a new model-architecture example
Grady returned to Jev later in the presentation and framed it as one sign that new model architectures are beginning to emerge. He described Jev as a classifier and argued that many enterprise tasks currently sent to large language models are, in essence, multiple-choice decisions within a bounded set of options.
Those jobs do not always need an expensive general-purpose LLM. With a specialized model such as Jev, some workloads can be processed at roughly 100x lower cost and at higher speed.
His broader point was that the future may not belong exclusively to ever-larger general-purpose models. There may also be a large market for smaller, faster and cheaper systems optimized for specific tasks.
He repeated that Jev’s commercial growth was extraordinary even by current AI standards: after the product suddenly broke out, the company behind it went from zero to $100 million in revenue in seven days.
Why application startups now have to reinvent themselves every four months
For all the opportunity, Grady said the application layer has become exceptionally hard to build in. The companies that are working, he said, almost all have to “reinvent themselves every four months.”
The reason is simple: foundation model capability is improving too quickly. What stands as the core product of a company today may become a free feature in a new foundation model four months later.
That pressure is pushing successful AI companies toward structures that look more like research labs. Instead of moving at the pace of the weakest person in the organization, they protect a small group of top performers and give them very high autonomy to move product work forward quickly.
He also said corporate structure itself is shifting away from traditional hierarchies and toward a “network of agents.” AI is used to spread information through the company and help employees understand what is happening inside the organization, while people operate with more autonomy.
Seven Sequoia-backed deals saw valuation jump nearly 31x in about a month
Grady ended with financing data from seven companies Sequoia participated in over the past 12 months. Based on that sample, he said the market is developing a “two-step” funding structure.
The first step is the company-building partner: early investors such as Sequoia that actively help founders build companies. Across those seven deals, Sequoia’s average entry point came at a post-money valuation of about $110 million.
The second step is the capital partner, meaning later-stage investors whose role is primarily to provide large amounts of capital. In less than a month on average, Grady said, the next round valuation for those seven companies reached $3.4 billion.
That amounts to a jump from $110 million to $3.4 billion in roughly one month, or close to 31x on average.
Grady said he had never seen anything like it before and called it one of the clearest signs of current AI-market froth. Even so, Sequoia has not turned negative on the long-term trajectory. He closed the presentation by saying that no one knows what lies on the other side of this door, but two things look certain: costs will keep falling, and capabilities will keep rising. If both happen at once, he said, the world will enter a period of accelerating change.
He ended with a single line: “Hold on to your hats.”