What this covers

This is a 29-minute keynote from Sequoia's AI Ascent 2025, delivered by three partners: Packer, Sonia, and Constantine. Packer carries most of the market and strategy argument. He frames it with the Physics of Technology Distribution: people must know about a product, want it, and be able to buy it. He then compares AI to cloud. Cloud began at roughly $400 billion of revenue, with about 200 million people online. Today 5.6 billion are connected, and ChatGPT reached global attention at launch. The keynote treats the AI opportunity as at least an order of magnitude larger than cloud, and because technology waves stack, it expects that opportunity to arrive faster.

Sonia covers the technical ground. Pre-training scaling is slowing after nine to ten orders of magnitude since AlexNet, so progress now leans on reasoning (AI)), synthetic data, tool use, and scaffolding. She argues the most interesting work sits where research and product blur, citing Deep Research and NotebookLM. She also points to 2024 milestones: coding reaching "screaming" product-market fit after Claude 3.5 Sonnet, and voice crossing the uncanny valley. She notes that Jensen Huang won the earlier infrastructure round of the stack.

The strategy section is where the claims get sharper. AI is said to attack both services and software profit pools, as companies move from selling tools into software budgets to selling outcomes into labor budgets. That shift is expected to expand pricing power and margins even as token costs fall about 99%. The keynote says defensibility comes from vertical focus, as with Harvey and OpenEvidence, and from a data flywheel that improves a measurable business metric. It warns against "vibe revenue." Foundation models are moving into the application layer, and macroeconomic noise is dismissed as minor next to AI's pull. The closing section looks toward an agent economy that needs persistent identity, protocols such as MCP, and security infrastructure. The advice for founders is to adopt a stochastic mindset and move at maximum velocity.

Sharpest takeaway

AI has transitioned from inevitable to imminent, with value accruing primarily to the application layer where startups can build defensible businesses through vertical integration, data flywheels, and customer-back approaches, while the next wave centers on agent economies requiring persistent identity, seamless protocols, and security infrastructure.

  • Precedent conditions for AI are fully in place (compute, networks, data, distribution, talent) and distribution physics have fundamentally changed with 5.6B internet users and 1.2-1.8B social media users creating adoption barriers similar to cloud and mobile.
  • Application layer companies like ChatGPT, Harvey, and Cursor are emerging as the value winners, with foundation models increasingly competing at the application layer, making vertical-specific, function-specific solutions with data flywheels the sustainable competitive strategy.
  • Agent economies represent the next major wave, requiring technical solutions for persistent identity (both agent consistency and user memory), seamless communication protocols (MCP), and security infrastructure to enable agents to transfer resources, make transactions, and understand trust.

The claims · ranked77 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.84

The physics of technology distribution require three conditions: people have to know about your thing, they have to want your thing, and they have to be able to buy your thing.

definitionhigh valueestablishednovelty 2/4durability 4/4· Packer

“If you think about the physics of distribution, you just need three things. People have to know about your thing. They have to want your thing. And they have to be able to buy your thing.”

0.79

AI adoption barriers have fundamentally changed compared to prior tech waves: ChatGPT launched November 30, 2022 with immediate global attention (versus cloud which required guerrilla marketing), with distribution rails of 5.6 billion internet-connected people and 1.2-1.8 billion monthly active social media users already in place, removing all barriers to rapid adoption.

causalhigh valueestablishednovelty 2/4durability 3/4· Packer

“When the cloud transition got going, nobody was paying attention. Ben off had to do these crazy gorilla marketing tactics to get anybody to pay attention. AI is very different. November 30, 2022, chat GPT comes out. The entire world is paying attention to AI.”

0.78

Industry-specific knowledge allows startups to speak customer language and be 'of the industry for the industry', as demonstrated by OpenEvidence in medical industry and Harvey with law firms.

causalhigh valueestablishednovelty 1/4durability 4/4· Packer

“You can be of the industry for the industry, kind of like what open evidence does for the medical industry. You can speak their language. Harvey sends lawyers in to talk with law firms.”

0.75

AI has gone after both services and software markets, meaning both profit pools are under attack as companies progress from selling tools into software budget to selling outcomes into labor budget.

causalhigh valueestablishednovelty 2/4durability 3/4· Packer

“It's not just the services market that AI has gone after. It's both services and software. Means that both of these profit pools are under attack. We've seen plenty of companies that start with software, gets a little bit smarter, becomes more of a co-pilot, gets a little bit smarter, becomes more of an autopilot, and they actually progress from selling a tool into software budget to selling an outcome to selling work into labor budget.”

0.75

Technology waves are additive and stacking up, meaning the opportunity from AI is much bigger than prior waves and is coming much faster.

causalhigh valueestablishednovelty 2/4durability 3/4· Packer

“These waves tend to be additive and so the opportunity is much bigger than it's been for prior waves. It's also coming much faster.”

0.75

Foundation models are getting better at penetrating the application layer through scaling laws, test-time compute, reasoning with tool use, and inter-agent communication, creating intense competition in the application layer.

causalhigh valueestablishednovelty 2/4durability 3/4· Packer

“But guess what? You got competition. We got a second scaling law. We got test time compute. We got reasoning with tool use and inter agent communication that allows the foundation models to get pretty far into the application layer.”

0.75

Data flywheels built on product usage data that competitors cannot access create sustainable moats, but only if that data improves a measurable business metric; data collection without business impact is worthless.

causalhigh valueestablishednovelty 2/4durability 3/4· Packer

“You can build data flywheels with the usage data of your own product. That's something that nobody else has. It needs to tie to a business metric or it just doesn't matter. And this is really important because this is one of the best modes that you can build.”

0.75

The blurry boundary between research and product is where the most exciting AI innovation is happening, with Deep Research and Notebook LM as breakthrough examples of research becoming product.

factualhigh valueestablishednovelty 2/4durability 3/4· Sonia

“A lot of the most exciting technology innovation in AI right now is happening at that blurry boundary between research and product. And I think the two breakthrough examples of this over the last year were Deep Research and Notebook LM.”

0.75

Voice generation technology crossed the uncanny valley in 2024, transitioning from 'almost there' to fully natural-sounding AI voices, representing a Her-moment equivalent for voice and dramatically closing the gap between science fiction and reality.

factualhigh valueestablishednovelty 2/4durability 3/4· Sonia

“2024 gave us what I would call the her moment for voice. And voice generation went from almost there to fully crossing the uncanny valley. And it's crazy how quickly the gap between science fiction and reality is closing on us. And it feels like the touring test really just snuck up on us.”

0.75

Pre-training scaling is slowing down because 9-10 orders of magnitude of progress since AlexNet means most low-hanging fruit has been picked, requiring research breakthroughs like reasoning, synthetic data, tool use, and AI scaffolding to continue intelligence scaling.

causalhigh valueestablishednovelty 2/4durability 3/4· Sonia

“From the technology back, the bad news is that pre-training does seem to be slowing down. We've scaled pre-training by nine or 10 orders of magnitude since the Alexet days, and that means a lot of the lowhanging fruit has been picked. The research ecosystem is finding new ways to break through. The most important breakthrough was reasoning from OpenAI.”

0.74

Building a successful AI company is 95% the same as building any company (solving important problems uniquely, attracting great people) and only 5% AI-specific, making foundational business skills the primary success factor.

normativehigh valueestablishednovelty 1/4durability 4/4· Packer

“So 95% of building a company in AI is just building a company. It's the same old stuff. Solve an important problem in a unique and compelling way. Get great people to follow you. All that good stuff. The other 5% is AI specific.”

0.73

The cloud transition began at $400 billion of revenue, which was already bigger than the global software market at the time, and by analogy the starting point for AI services market is at least an order of magnitude bigger, with end states 10-20 years out having potential to be absolutely massive.

causalhigh valueestablishednovelty 1/4durability 3/4· Packer

“The cloud at 400 billion of revenue is bigger than the global software market when that transition began. If we reason by analogy, the market we're going after in the world of AI services, the starting point is at least an order of magnitude bigger.”

0.73

Macroeconomic factors (tariffs, interest rate noise) are irrelevant compared to the structural pull of AI adoption, which creates a 'tremendous sucking sound' in the market that overwhelms any market volatility, making it imperative to move at maximum velocity to capture opportunity before competitors do.

causalhigh valueestablishednovelty 1/4durability 3/4· Packer

“Nature hates a vacuum. There is a tremendous sucking sound in the market right now for AI. All of the macroeconomic stuff, tariffs and interest rate noise doesn't matter. The rising tide of technology adoption absolutely swamps any of the volatility that you see in the markets. Ignore it. Okay? There is a tremendous sucking sound in the market and if you don't get in front of it, somebody else will because nature hates a vacuum.”

0.71

Pre-training progress is slowing down as scaling laws hit diminishing returns after nine to ten orders of magnitude of scaling since AlexNet, forcing new research approaches.

factualhigh valueestablishednovelty 2/4durability 3/4· Sonia

“From the technology back, the bad news is that pre-training does seem to be slowing down. We've scaled pre-training by nine or 10 orders of magnitude since the Alexet days, and that means a lot of the lowhanging fruit has been picked.”

0.71

AI has shifted from targeting just services TAM to attacking both services and software profit pools, with companies progressing from selling tools into software budgets to selling outcomes and work into labor budgets, putting both TAMs under attack.

causalhigh valuecontestednovelty 2/4durability 3/4· Packer

“We've actually updated our thinking. It's not just the services market that AI has gone after. It's both services and software. Means that both of these profit pools are under attack. We've seen plenty of companies that start with software, gets a little bit smarter, becomes more of a co-pilot, gets a little bit smarter, becomes more of an autopilot, and they actually progress from selling a tool into software budget to selling an outcome to selling work into labor budget.”

0.70

Vibe revenue (feeling good about revenue) can be dangerous because it leads founders to delude themselves about real product-market fit; actual metrics (adoption, engagement, retention) are what matter.

normativehigh valueestablishednovelty 1/4durability 4/4· Packer

“Vibe revenue can kill you. Everybody loves five revenue. Feels so good. Oh my god, we have so much revenue. Go look at it. Okay. Is it tire kicking or are you actually creating durable behavior change? Okay. Oh, I don't have the metrics to do that. Yes, you do. Inspect the adoption, the engagement, the retention.”

0.69

Anthropic's Model Context Protocol (MCP) created a strong ecosystem and network effects enabling agentic tool use, and Sequoia is excited about its potential to accelerate tool use adoption.

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“Enthropics MCP created strong ecosystem and networks and we're also excited to see what that does to accelerate agentic tool use.”

0.69

The preceding conditions for AI imminent impact are in place: compute, networks, data, distribution, and talent are all available, meaning we have all the ingredients needed to proceed.

factualhigh valueestablishednovelty 1/4durability 3/4· Packer

“AI is in fact imminent, not just inevitable. The precedent conditions are in place. compute, networks, data, distribution, talent. We have all the ingredients we need.”

0.69

Internet connectivity expanded from 200 million at cloud transition to 5.6 billion today, representing effectively every household and business in the world and eliminating adoption barriers.

factualhigh valueestablishednovelty 1/4durability 3/4· Packer

“Right side of the page, let's imagine you listen to Benoff. There are only 200 million people connected to the internet. Today, it's 5.6 billion. It's effectively every household and every business in the world.”

0.69

Voice generation in 2024 crossed the uncanny valley, creating what Sonia calls 'the her moment for voice' and closing the gap between science fiction and reality.

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“2024 gave us what I would call the her moment for voice. And voice generation went from almost there to fully crossing the uncanny valley.”

0.69

AI is enabling profound use cases across advertising (accurate ad copy), education (concept visualization), and healthcare (patient diagnosis) that are just beginning to be explored.

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“For example, advertising, the ability to create stunningly accurate and beautiful ad copy. Education, the ability to visualize new concepts with the snap of a finger. Healthcare, the ability to diagnose patients better with an app like open open evidence. We're just scratching the surface of what's possible.”

0.69

Coding reached 'screaming product-market fit' in 2024, with Anthropic's Claude 3.5 Sonnet triggering a vibe shift and enabling diverse users to create software efficiently.

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“The breakout application category of the year was coding, which reached screaming product market fit. Anthropics Claude 35 sonnet dropped last fall, triggering a rapid vibe shift in the coding landscape. And people are now using AI coding for really impressive things.”

0.69

Jensen Huang (Jensen Hang) was the real undisputed king of the AI stack who made 'all the dollars' through infrastructure sales, winning the narrative despite application layer being attractive.

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“the joke's actually on all of us because the real undisputed king of the stack who was making all the dollars was the goat himself, Jensen Hang, uh who we're excited to hear from shortly.”

0.69

Cost per token has declined 99% over the last 12-18 months and cost curves will continue downward, though test-time compute may add some costs, the overall COGS trajectory is downward.

factualhigh valueestablishednovelty 1/4durability 3/4· Packer

“cost per token is down 99% in the last 12 to 18 months. That cost curve is going to continue. those cogs are going to go down. I know it's stepping up with test time compute and all that, but that's going to go down too.”

0.69

Startup pricing power is likely to increase as they progress from selling tools (competing on cost) to selling outcomes (competing on value), enabling gross margin expansion even if COGS decreases.

forecasthigh valueestablishednovelty 1/4durability 3/4· Packer

“The price component of that, if you are successfully marching from selling a tool to selling an outcome and you're going up that value chain and you can capture more of that value, your price point is probably going to go up too. And so your gross margins may not be great today, but you should have a good path to really healthy gross margins over time.”

0.69

The precedent conditions for AI are all in place—compute, networks, data, distribution, and talent—meaning AI is imminent, not just inevitable.

factualhigh valueestablishednovelty 1/4durability 3/4· Packer (Sequoia partner)

“AI is in fact imminent, not just inevitable. The precedent conditions are in place. compute, networks, data, distribution, talent. We have all the ingredients we need.”

0.69

Seamless communication protocols (like MCP) are essential for agent economy because without them, you have no infrastructure for agents to transfer information, value, and trust—just as TCP/IP was essential for personal computing.

causalhigh valueestablishednovelty 1/4durability 3/4· Constantine (Sequoia partner)

“The second major technology shift is we need to have seamless communication protocols. Now the great news is everyone seems to be focused on this now. But imagine personal computing without seamless communication protocols. No TCP IP, no internet. We're just now building that protocol layer. And there's certainly a lot of excitement around MCP. It's so great to see the big players collaborating and working together to come up with this is just one but a series of protocols that are going to allow for transfer of information, transfer of value, and transfer of trust.”

0.69

Memory and self-learning in agents face major unsolved challenges despite attempts using RAG, vector DBs, and long context windows—agents must maintain consistency where it matters while only differing in areas where variation is appropriate.

factualhigh valueestablishednovelty 1/4durability 3/4· Constantine

“Now, we've been trying everything from rag and vector DBs to really long context windows, but everyone in this room knows that there are still major challenges in true memory and self-learning on that true memory and actually getting the agents to be able to be consistent where it matters and and only differ in the areas where they should be differing.”

0.69

Coding reached screaming product-market fit as the breakout application category of 2024, with Anthropic's Claude 3.5 Sonnet triggering a rapid vibe shift that enabled both expert engineers and non-coders to use AI for impressive software creation feats.

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“the breakout application category of the year was coding, which reached screaming product market fit. Anthropics Claude 35 sonnet dropped last fall, triggering a rapid vibe shift in the coding landscape. And people are now using AI coding for really impressive things. So, for example, this person vibe coded their own docend alternative. And so whether you're a veteran 10x engineer or somebody who has no idea how to code, we do think that AI is fundamentally changing the accessibility, the speed, and the economics of software creation.”

0.69

Vertical agents trained end-to-end on specific workflows using reinforcement learning and user data can outperform the best humans at those tasks, with evidence emerging in security (Expo agents outperforming penetration testers), DevOps (Traversal vs troubleshooters), and networking (Meter vs engineers).

factualhigh valueestablishednovelty 1/4durability 3/4· Sonia

“we see companies creating agents that are trained end to end to excel in a very specific workflow using techniques including reinforcement learning on synthetic data and user user data to make AI systems very performant at very specific tasks. And the evidence we've seen so far makes us really optimistic in security. Expo is now showing that they have an agent that can outperform human penetration testers. In DevOps, Traversal is showing that they can create an AI troubleshooter that is better than the best human troubleshooters. Um, same goes in networking with meter and networking engineers.”

0.66

ChatGPT's November 30, 2022 launch created unprecedented awareness, with the entire world paying attention to AI within days, contrasting sharply with the cloud transition where Ben Hoff had to use guerrilla marketing tactics to gain attention.

factualhigh valueestablishednovelty 1/4durability 4/4· Packer

“November 30, 2022, chat GPT comes out. The entire world is paying attention to AI. Middle column, these are the combined monthly active users for Reddit and the artist formerly known as Twitter. didn't exist at the start of the cloud transition, barely existed at the start of the mobile transition. Somewhere between 1.2 and 1.8 billion people on these platforms today.”

0.64

AI native applications showed poor engagement ratios (low daily/monthly active user ratios) in 2023, but by 2024 ChatGPT's engagement ratio has climbed significantly and is now approaching Reddit levels, indicating that AI is delivering real value.

factualhigh valueestablishednovelty 1/4durability 2/4· Sonia

“Back in 2023 we showed this chart of the ratio of daily to monthly active users for AI native applications versus for traditional mobile apps. And the punchline at the time was that AI apps had terrible engagement ratios. Hype exceeded reality in the data. We're very pleased to report that that punchline has now changed dramatically. It's been stunning to see, for example, the daily month to monthly active ratio for chat GPT climb the curve and now getting close to Reddit levels.”

0.64

Sequoia predicted that application layer companies would capture most value in the AI stack, and this prediction has been validated by companies like Harvey and OpenEvidence creating customer-back value.

factualhigh valueestablishednovelty 1/4durability 2/4· Sonia

“I remember my partners, Pat in particular, were adamant that the value was going to acrue to the applica uh to the application layer. And I remember thinking like, okay, Pat, sure, good luck with that. Uh but having seen how the last few years have played out, I think you were right, Pat. Uh if you see where value has been created, if you see companies like Harvey and Open Evidence really creating that customer back value, we very much believe that the application layer is where the value will ultimately acrue.”

0.64

AI native application engagement has dramatically improved—the daily-to-monthly active user ratio for ChatGPT is now approaching Reddit levels, indicating hype has been replaced with real utility and that AI is successfully integrating into daily user behavior.

factualhigh valueestablishednovelty 1/4durability 2/4· Sonia

“Back in 2023 we showed this chart of the ratio of daily to monthly active users for AI native applications versus for traditional mobile apps. And the punchline at the time was that AI apps had terrible engagement ratios. Hype exceeded reality in the data. We're very pleased to report that that punchline has now changed dramatically. It's been stunning to see, for example, the daily month to monthly active ratio for chat GPT climb the curve and now getting close to Reddit levels.”

0.64

Gross margins in AI companies will improve substantially over time despite potentially poor margins today, because cost-per-token has decreased 99% in 12-18 months and will continue declining, while price points should increase as companies move from selling tools to selling outcomes and capturing more value up the chain.

forecasthigh valueestablishednovelty 1/4durability 2/4· Packer

“We don't necessarily care where your gross margin is today. The COGS component of that is probably going to keep going down. You know, cost per token is down 99% in the last 12 to 18 months. That cost curve is going to continue. those cogs are going to go down...The price component of that, if you are successfully marching from selling a tool to selling an outcome and you're going up that value chain and you can capture more of that value, your price point is probably going to go up too. And so your gross margins may not be great today, but you should have a good path to really healthy gross margins over time.”

0.63

An abundance era is emerging where code is the first category to tip, raising the question of what happens when labor becomes cheap and plentiful—whether we get AI slop or whether taste becomes the scarce asset that determines value.

forecasthigh valuespeaker onlynovelty 3/4durability 4/4· Sonia

“We're entering an abundance era. code as the first market category to tip will offer us a preview into what that abundance era actually means. What happens when labor is cheap and plentiful? Are we going to get a bunch of AI slop? And what happens when taste becomes the scarce asset?”

0.62

Macroeconomic conditions like tariffs and interest rates are irrelevant compared to the structural force of technology adoption; a rising tide of AI adoption swamps market volatility.

causalhigh valuecontestednovelty 1/4durability 3/4· Packer

“All of the macroeconomic stuff, tariffs and interest rate noise doesn't matter. The rising tide of technology adoption absolutely swamps any of the volatility that you see in the markets. Ignore it.”

0.61

The cloud transition started with $400 billion in revenue, while the AI services market starting point is at least an order of magnitude bigger, with the endpoint 10-20 years from now having a chance to be absolutely massive.

causalhigh valueestablishednovelty 1/4durability 3/4· Packer

“The cloud at 400 billion of revenue is bigger than the global software market when that transition began. If we reason by analogy, the market we're going after in the world of AI services, the starting point is at least an order of magnitude bigger. The end point 10 or 20 years from now has a chance to be absolutely massive.”

0.61

Approaches like RAG, vector databases, and long context windows partially address memory problems, but there are still major challenges in true agent memory and self-learning.

factualhigh valueestablishednovelty 1/4durability 3/4· Constantine

“Now, we've been trying everything from rag and vector DBs to really long context windows, but everyone in this room knows that there are still major challenges in true memory and self-learning on that true memory and actually getting the agents to be able to be consistent where it matters.”

0.59

An agent economy is the next wave where agents will not just communicate information but transfer resources, make transactions, track each other, understand trust and reliability, and have their own economy while working with humans.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Constantine

“In the years to come, we think that this matures even further into an agent economy. An agent economy is one in which agents don't just communicate information. They transfer resources. They can make transactions. They keep track of each other. They understand trust and reliability. And they actually have their own economy. Well, this economy doesn't cut out humans. It's all about humans. The agents work with the people and the people work with the agents in this agent economy.”

0.59

Agent economy is the next major wave following agent swarms, characterized by agents that transfer resources, make transactions, track each other, understand trust and reliability, and maintain their own economy while working with humans in a human-centered system.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Constantine

“An agent economy is one in which agents don't just communicate information. They transfer resources. They can make transactions. They keep track of each other. They understand trust and reliability. And they actually have their own economy. Well, this economy doesn't cut out humans. It's all about humans. The agents work with the people and the people work with the agents in this agent economy.”

0.59

Companies will scale faster with fewer people than ever before, approaching leverage levels never seen in the economy, where individual functions cluster and merge into AI-driven processes.

forecasthigh valueestablishednovelty 1/4durability 1/4· Constantine (Sequoia partner)

“And we have seen companies scale faster than ever before with fewer people than ever before. And we do think we're going to get to the highest level of leverage that we've ever seen in this economy. Eventually, these processes and agents are going to merge. You're going to have neural networks within very large complex neural networks. A network of these neural networks.”

0.57

The first cohort of AI killer apps has emerged, including ChatGPT, Harvey, Glean, Sierra, Cursor, and A Bridge, with new companies rising across diverse end markets including Listen Labs and Open Evidence.

factualhigh valueestablishednovelty 0/4durability 2/4· Sonia

“we now think that the first cohort of AI's killer apps has emerged. Whether it's chat GPT, Harvey, Glean, Sierra, Cursor, A Bridge, but there's a whole set of new companies on the rise across a very rich and diverse set of end markets uh including Listen Labs, including Open Evidence.”

0.56

The stochastic mindset represents a departure from determinism in computing; AI systems are probabilistic, so asking an agent to remember the number 73 may result in remembering 37, 72, 74, the next prime, 79, or nothing at all.

definitionhigh valuespeaker onlynovelty 2/4durability 4/4· Constantine

“The stochcastic mindset is a departure from determinism. You know, a lot of us fell in love with computer science because it was so deterministic, right? You program the computer to do something, it does it. Even if that results in a sag fault. Now, we're entering an era of computing that's going to be stochastic. If you ask a computer to remember the number 73, it'll remember that tomorrow, next week, next month. If you ask a person or an AI, well, it might remember 73. It might remember 37, 72, 74, the next prime, 79, or nothing at all.”

0.56

Vibe revenue (non-sticky, low-engagement revenue that feels good in the moment) can kill AI companies, distinguishable from real revenue through inspection of adoption, engagement, and retention metrics showing whether the product creates durable behavior change.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Packer

“Revenue. Vibe revenue can kill you. Everybody loves five revenue. Feels so good. Oh my god, we have so much revenue. Go look at it. Okay. Is it tire kicking or are you actually creating durable behavior change? Okay. Oh, I don't have the metrics to do that. Yes, you do. Inspect the adoption, the engagement, the retention.”

0.56

Data flywheels only matter if they tie to a business metric—having usage data without a clear business outcome is either not a real flywheel or it doesn't matter, and this distinction is critical because data flywheels are among the best defensible modes in AI.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Packer

“Data flywheel. Raise your hand if you got a data flywheel. Okay, what business metric does that data flywheel move? Okay, I see less certainty. I got good news and I got bad news. Good news is if you can't answer that question, I still love you. The bad news is your data flywheel is either you don't have a data flywheel or it just doesn't matter. It needs to tie to a business metric or it just doesn't matter. And this is really important because this is one of the best modes that you can build.”

0.56

The stochastic mindset represents a fundamental shift from deterministic computing (where instructions produce guaranteed results) to probabilistic computing (where agents may output different results on the same input, like remembering '73' as '37', '72', '74', or nothing).

definitionhigh valuespeaker onlynovelty 2/4durability 4/4· Constantine

“The stochcastic mindset is a departure from determinism. You know, a lot of us fell in love with computer science because it was so deterministic, right? You program the computer to do something, it does it. Even if that results in a sag fault. Now, we're entering an era of computing that's going to be stochastic. If you ask a computer to remember the number 73, it'll remember that tomorrow, next week, next month. If you ask a person or an AI, well, it might remember 73. It might remember 37, 72, 74, the next prime, 79, or nothing at all.”

0.56

The agent economy will create a combination of greater leverage with significantly less certainty—humans will be able to accomplish more but must manage increased uncertainty and risk, requiring new risk management capabilities.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Constantine

“The third major change for each of us is a combination of the previous two. way more leverage with significantly less certainty. We're entering a world where you can do more, but you have to be able to manage that uncertainty and manage risks. And in this world, everyone in this room is extremely well suited to thrive.”

0.55

We are entering an 'abundance era' where labor becomes cheap and plentiful, raising the question of whether we'll get 'AI slop' and making taste the scarce asset.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Sonia

“We're entering an abundance era. code as the first market category to tip will offer us a preview into what that abundance era actually means. What happens when labor is cheap and plentiful? Are we going to get a bunch of AI slop? And what happens when taste becomes the scarce asset?”

0.55

The combination of greater leverage and significantly less certainty means the economy will feature more potential upside but more risk, requiring new skills in uncertainty management.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Constantine

“The third major change for each of us is a combination of the previous two. way more leverage with significantly less certainty. We're entering a world where you can do more, but you have to be able to manage that uncertainty and manage risks.”

0.55

Most companies that achieved $1 billion+ revenue in prior tech transitions (cloud, mobile, etc.) are concentrated at the application layer, and Sequoia believes the same pattern will hold for AI—the value accrues to applications, not infrastructure.

causalhigh valuecontestednovelty 1/4durability 3/4· Packer

“These logos represent the companies that got to a billion plus of revenue. We don't care about unicorns. We care about revenue and free cash flow. Got to a billion plus of revenue during prior transitions. Most of them are at the top of the page. Most of them are at the application layer. We have believed and we continue to believe that the same will be true for AI. The value is at the application layer.”

0.55

Startup AI companies face intensifying competition from foundation models that are penetrating the application layer through scaling laws (second scaling law, test time compute, reasoning with tool use, inter-agent communication), requiring startups to differentiate through vertical integration, function specificity, and handling complex problems requiring human-in-the-loop.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Packer

“But guess what? You got competition. We got a second scaling law. We got test time compute. We got reasoning with tool use and inter agent communication that allows the foundation models to get pretty far into the application layer. What do you do if you're a startup and you're not building a vertically integrated business? Go from the customer back. Think vertical specific. Think function specific. Deal with complex problems that might require a human in the loop.”

0.52

Startups should build vertically integrated businesses, going from customer backwards, thinking function-specific and dealing with complex problems requiring human-in-the-loop solutions to compete against foundation models in the application layer.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Packer

“What do you do if you're a startup and you're not building a vertically integrated business? Go from the customer back. Think vertical specific. Think function specific. Deal with complex problems that might require a human in the loop.”

0.52

Many new AI companies will be agent-first, and agents will evolve from today's simple prototypes pieced together to robust systems through two paths: orchestration with rigorous testing, or end-to-end task tuning.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Sonia

“Another prediction is that many of these new companies will be agent first and that the agents these companies are selling will evolve from prototypes that are often just pieced together today to ones that are really robust. And we see two paths that companies are taking to build that. Uh path one orchestration with rigorous testing and eval path two agents tuned on endto-end tasks.”

0.52

Vertical agents trained end-to-end on domain-specific workflows using reinforcement learning on synthetic and user data can outperform the best humans at highly specialized tasks.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Sonia

“Vertical agents are a wonderful opportunity for startup founders who deeply understand a domain. And we see companies creating agents that are trained end to end to excel in a very specific workflow using techniques including reinforcement learning on synthetic data and user user data to make AI systems very performant at very specific tasks.”

0.52

Persistent identity of the agent itself is critical for business relationships, requiring agents to maintain consistent personality and understanding across interactions.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Constantine

“Well, the first is the agent itself needs to be persistent. If you're doing business with someone and they change dayto-day, you you probably won't be doing business with them for very long. That dramatically different experience will take its toll. The agent is going to have to be able to persist on its own personality and its own understanding.”

0.52

Management mindset must shift from managing individual engineers to understanding what agents can and cannot do, paralleling the shift from IC engineer to engineering manager.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Constantine

“Second change is a management mindset. And this management mindset is going to be all about understanding what your agents can and can't do for you. Everybody knows that being a really good IC engineer is pretty different from being a great engineering manager. And this is going to be the transition that most of the economy is going to make into more complex managerial decisions like blocking processes and giving feedback.”

0.52

Security will form an entire cottage industry in the agent economy because transacting face-to-face (the traditional basis of trust) is impossible with agents; trust and security become even more important and will drive significant business opportunities.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Constantine

“If you can't meet the person you're doing business with palm to palm, face to face, that importance of security and trust is even further elevated. You can't with agents. So, we're going to have an entire cottage industry formed around that trust, that security, and it will be even more important in the agent economy than it already is in our current economy.”

0.52

The management mindset shift required for agent economies involves understanding what agents can and cannot do, similar to the difference between being an excellent individual contributor engineer versus a great engineering manager—managing agents requires this type of role transition.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Constantine

“Second change is a management mindset. And this management mindset is going to be all about understanding what your agents can and can't do for you. Everybody knows that being a really good IC engineer is pretty different from being a great engineering manager. And this is going to be the transition that most of the economy is going to make into more complex managerial decisions like blocking processes and giving feedback.”

0.52

Good vibes with customers (trust in the founder and product roadmap) are more important than product quality at this point in the AI adoption cycle, because the product will improve over time but trust determines whether customers stay to experience that improvement or leave.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Packer

“The good vibes piece of this is important, too. You need good vibes with your customers. What does it mean? Your customers have to trust you and you have to earn that trust. Trust is more important than your product at this point in time where we are in the cycle. The product will get better. If they trust you to make it better, you're in good shape. If they don't trust you to make it better, you're in bad shape.”

0.52

Foundation model improvements are occurring at the blurry boundary between research and product, with examples like Deep Research and Notebook LM demonstrating that the most exciting innovation combines research breakthroughs with immediate productization.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Sonia

“a lot of the most exciting technology innovation in AI right now is happening at that blurry boundary between research and product. And I think the two breakthrough examples of this over the last year were Deep Research and Notebook LM.”

0.52

Persistent identity requires two technical capabilities: (1) the agent itself must maintain consistent personality and understanding across interactions, and (2) the agent must remember users and understand their history, both essential for trust and reliability in the agent economy.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Constantine

“When when we talk about a persistent identity, we really mean two things. The first is the agent itself needs to be persistent. If you're doing business with someone and they change dayto-day, you you probably won't be doing business with them for very long...The second type of persistence is understanding you. Similarly, if you're doing business with a person and they they don't remember anything about you, they can hardly remember your name, that's also a big challenge to trust and reliability.”

0.52

Building successful AI companies is 95% traditional business building (solving important problems uniquely, attracting great people) and 5% AI-specific, with the AI-specific part involving capturing value across the entire value chain through customer-back design, data flywheels, vertical expertise, and end-to-end solutions.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Packer

“95% of building a company in AI is just building a company. It's the same old stuff. Solve an important problem in a unique and compelling way. Get great people to follow you. All that good stuff. The other 5% is AI specific.”

0.52

The Leone merchandising cycle encompasses the complete value chain from idea to product to engineering to market sales and support, and competitive moats can be built across every stage by taking a customer-back approach.

definitionhigh valuespeaker onlynovelty 1/4durability 4/4· Packer

“This is the Leone merchandising cycle. Our partner Doug Leone, the goat, spent 40 years lovingly crafting this little piece of content. This represents everything you need to take an idea in your brain and turn it into a product in the hand of your customers. Okay, that idea has to turn into a product that has to be built by an engineering team that has to be taken to market and sold and supported. This is the value chain.”

0.49

Agent-first companies represent a major 2025 prediction, with agents evolving from prototypes pieced together today to robust systems through two implementation paths: orchestration with rigorous testing/eval, or agents tuned on end-to-end tasks.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Sonia

“Another prediction is that many of these new companies will be agent first and that the agents these companies are selling will evolve from prototypes that are often just pieced together today to ones that are really robust. And we see two paths that companies are taking to build that. Uh path one orchestration with rigorous testing and eval path two agents tuned on endto-end tasks.”

0.48

Customer trust is more important than product quality at this stage of the AI cycle; products will improve, but trust determines if customers adopt those improvements.

causalhigh valuespeaker onlynovelty 1/4durability 3/4· Packer

“You need good vibes with your customers. What does it mean? Your customers have to trust you and you have to earn that trust. Trust is more important than your product at this point in time where we are in the cycle. The product will get better. If they trust you to make it better, you're in good shape. If they don't trust you to make it better, you're in bad shape.”

0.48

Companies can build moats by putting a 'big bear hug around customers' in ways foundation models probably won't, via end-to-end solutions, data flywheels, domain expertise, and customer intimacy.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Packer

“You can put a big bear hug around your customers in a way that foundation models probably won't. By the way, we love foundation models, too, but we're assuming that most of you are not building foundation models. Most of you are building applications.”

0.48

Companies can build moats by operating 'customer back' through end-to-end solutions (vs throwing tools over the fence), data flywheels using proprietary usage data, domain expertise ('of the industry, for the industry'), customer embedding ('big bear hug'), and deployed engineers.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Packer (Sequoia partner)

“Your customer is not sure what they want out of AI. You can have an opinion. You can give them an end toend solution that just solves the problem as opposed to throwing a tool over the fence. You can build data flywheels with the usage data of your own product. That's something that nobody else has. You can be of the industry for the industry, kind of like what open evidence does for the medical industry. You can speak their language. Harvey sends lawyers in to talk with law firms.”

0.45

There is tremendous market demand for AI (a 'tremendous sucking sound') and if startups don't move at maximum velocity to capture it, competitors will, because nature abhors a vacuum.

causalhigh valuespeaker onlynovelty 1/4durability 2/4· Packer

“Nature hates a vacuum. There is a tremendous sucking sound in the market right now for AI. And so all the stuff we just said about modes and metrics and whatever notwithstanding, you are in a run like heck business right now. Now is the time to go at maximum velocity all of the time.”

0.45

Companies will scale faster and with fewer people than ever before due to AI agents, potentially approaching the 'one-person unicorn' that Sequoia predicted a year ago.

forecasthigh valuespeaker onlynovelty 1/4durability 2/4· Constantine

“we predicted that we'd have the first oneperson unicorn. Now, that hasn't happened yet, but we have seen companies scale faster than ever before with fewer people than ever before. And we do think we're going to get to the highest level of leverage that we've ever seen in this economy.”

0.41

The Leone merchandising cycle (Sequoia partner Doug Leone's framework) represents the complete value chain from idea to product to market to support, and companies should build moats across the entire chain via customer-back thinking.

definitionhigh valuespeaker onlynovelty 0/4durability 3/4· Packer (Sequoia partner)

“This is the Leone merchandising cycle. Our partner Doug Leone, the goat, spent 40 years lovingly crafting this little piece of content. This represents everything you need to take an idea in your brain and turn it into a product in the hand of your customers.”

0.39

A cohort of AI's first killer apps has emerged, including ChatGPT, Harvey, Glean, Sierra, Cursor, and A Bridge, with many new companies rising across diverse end markets including Listen Labs and Open Evidence.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Sonia

“we now think that the first cohort of AI's killer apps has emerged. Whether it's chat GPT, Harvey, Glean, Sierra, Cursor, A Bridge, but there's a whole set of new companies on the rise across a very rich and diverse set of end markets uh including Listen Labs, including Open Evidence.”

0.35

Anthropic's MCP (Model Context Protocol) created a strong ecosystem and network effects, and Sequoia is excited about how this accelerates agentic tool use.

factualhigh valuespeaker onlynovelty 0/4durability 2/4· Sonia

“Enthropics MCP created strong ecosystem and networks and we're also excited to see what that does to accelerate agentic tool use.”

0.30

OpenAI's O3 reasoning model represents the most important breakthrough in overcoming pre-training scaling limits, going beyond the 'Strawberry' (O1) team's initial reasoning work.

factualestablishednovelty 1/4durability 2/4· Sonia (Sequoia partner)

“The most important breakthrough was reasoning from OpenAI. We were lucky enough to have Noam Brown from the Strawberry team give us a preview of what was going to happen with reasoning at AI Ascent last year and this year we're happy to have Dan Roberts in the audience who's going to be giving another talk on what's happening with O3 and reasoning ahead.”

0.28

AI applications are creating value across advertising (accurate ad copy), education (concept visualization), and healthcare (patient diagnosis), with each domain still only scratching the surface of what's possible as models become more capable.

factualestablishednovelty 0/4durability 2/4· Sonia (Sequoia partner)

“So for example, advertising, the ability to create stunningly accurate and beautiful ad copy. Education, the ability to visualize new concepts with the snap of a finger. Healthcare, the ability to diagnose patients better with an app like open open evidence. We're just scratching the surface of what's possible. And as the AI models become more and more capable, the things we can do with them through this front door become more and more profound.”

0.24

Jensen Huang (NVIDIA CEO) is 'the undisputed king of the stack' making all the dollars, despite Sequoia's focus on application companies.

factualestablishednovelty 0/4durability 2/4· Sonia (Sequoia partner)

“the joke's actually on all of us because the real undisputed king of the stack who was making all the dollars was the goat himself, Jensen Hang, uh who we're excited to hear from shortly.”

0.17

Packer was initially skeptical of the application-layer thesis but has been proven right by subsequent events showing that application companies (Harvey, OpenEvidence) are the ones creating the most customer value.

factualspeaker onlynovelty 0/4durability 2/4· Sonia

“I remember my partners, Pat in particular, were adamant that the value was going to acrue to the applica uh to the application layer. And I remember thinking like, okay, Pat, sure, good luck with that. Uh but having seen how the last few years have played out, I think you were right, Pat. Uh you belong over here.”