Sequoia Partner: The market opportunity for AI is 10 times that of cloud computing, and the next form of AI agents is vertical domain intelligences

Recently, Sequoia Capital held the AI Ascent 2025 themed event, where three partners of Sequoia Capital, Packer Radio, Sonya Huang, and Konstantine Buhler, attended and delivered opening speeches.

In the speeches, Sequoia Capital outlined why the market opportunities represented by AI are at least ten times that of cloud computing, and discussed which areas startups should focus on to achieve victory, as well as how the rise of AI agents will create a whole new economic paradigm.

Here are the highlights from the speeches:

1. When the cloud transformation began, cloud service revenue reached $400 billion, larger than the global software market. If we draw an analogy, the starting point of the AI service market is at least an order of magnitude larger, and the endpoint in 10-20 years could be very substantial.

2. AI targets not only the service market but also the software market, both of which are facing disruption.

3. We do not care about unicorns; we care about revenue and free cash flow. Most successful companies are at the top of the page, at the application layer. We have always believed that AI will be the same, with value lying in the application layer.

4. For AI companies, 95% of the standards are the same as how we view other companies, the unique 5% for AI companies lies in revenue atmosphere, profit margins, and data flywheels.

5. The participation rate of AI applications has significantly increased, with the daily active/monthly active ratio of ChatGPT climbing close to Reddit levels, meaning more and more people are deriving value from AI, and we are all climbing the ladder together, integrating AI into our daily lives.

6. The most prominent application category this year is programming, which has achieved an astonishing product-market fit. We believe AI is fundamentally changing the accessibility, speed, and cost-effectiveness of software creation; we are entering a rich era, with code as the first turning point market category.

7. Pre-training seems to be slowing down. Since the Alexnet era, we have scaled pre-training by 9 to 10 orders of magnitude, meaning many easily obtainable results have already been achieved. The research ecosystem is looking for new breakthrough methods.

8. The first batch of AI killer applications has emerged, including ChatGPT, Harvey, Glean, Sierra, Cursor, A Bridge, and a whole set of emerging companies are rising in various rich and diverse terminal markets, including Listen Labs, Open Evidence, etc.

9. By 2025, the next form of AI agents is expected to be vertical domain intelligent agents, which are agents created by companies that excel in specific workflows through end-to-end training, performing reinforcement learning on synthetic data and user data, enabling AI systems to excel in specific tasks

10. In the agency economy, agents not only convey information but also transfer resources, conduct transactions, track each other, understand trust and reliability, and possess their own economic systems. Three technical challenges are particularly critical: persistent identity, seamless communication protocols, and security.

Below is the full summary of the speech:

Packer Radio:

My name is Packer Radio, and I am a partner at Sequoia. Sonia, Constantine, and all our partners at Sequoia will serve as today's hosts.

Before diving into the main content today, Sonia, Constantine, and I will share some insights we've learned over the past year or so. We are very aware that we are just the appetizers, not the main course. Yesterday, I received an email from a co-founder saying, "Hey, I might be late, probably around 9:35." I thought to myself, "That's an interesting time, just when Jensen is going on stage." So, we understand, but we still want to share some thoughts before getting into the main topic.

First, let's calibrate and understand what is happening in the world of AI. We use a simple framework to analyze the market: Don Valentine's questions—What is it? Why is it important? Why now? It may be inevitable, but is it imminent? Finally, what should we do? How can we leverage this? How can we win? We have discussed these questions over the past few years, but in the next few minutes, we will update some thoughts.

To be honest, I had a brilliant explanation for the question "What is it?" but Constantine tactfully pointed out that explaining AI to a room full of AI professionals is not a good idea. So, we will jump straight to the "Why is it important" part.

Who remembers this slide from last year? Thank you.

The top row represents cloud transformation, the bottom row represents AI transformation; the left side is the past, the middle is the present, and the right side is the future. It indicates that when cloud transformation began, cloud service revenue reached $400 billion, larger than the global software market. By analogy, the starting point of the AI service market is at least an order of magnitude larger, and the endpoint in 10-20 years could be extremely vast.

We have now updated our thoughts, AI is not only targeting the service market but also the software market, both of which are facing disruption. We see many companies starting with software, becoming smarter as co-pilots, then becoming smarter as fully autonomous systems, transitioning from selling tools to selling software budgets, and then to selling outcomes, entering labor budgets. Both target markets are worth competing for.

Who remembers this slide from last year? Only three or four people? That's a bit unfortunate. Don't be shy, you can raise your hand

This layered cake represents the technological waves that have accumulated over the past few decades. This slide has two points: first, AI is indeed imminent, not just inevitable. The prerequisites are all in place: computing power, networks, data, distribution channels, and talent; we have all the necessary conditions. Second, these waves are often superimposed, so the opportunities are much greater and come much faster than previous waves.

I hate these charts with time on the horizontal axis and vanity metrics on the vertical axis, which people use to justify various mistakes. But the observation is correct; things are happening at an increasingly rapid pace. Few people delve into the reasons, so we want to briefly discuss the physics of distribution: people need to understand your product, want your product, and be able to purchase your product, that’s all.

Do you remember this logo? When the cloud transformation began, no one was paying attention, and Benioff had to use guerrilla marketing strategies to attract attention. AI is completely different; when ChatGPT was released on November 30, 2022, the whole world began to pay attention to AI. The middle column represents the total monthly active users of Reddit and former Twitter, which were almost nonexistent at the start of the cloud transformation, just emerging at the start of the mobile transformation, and now about 1.2 to 1.8 billion people use these platforms, which are important channels for understanding new things. On the right, if we listen to Benioff, there were only 200 million people connected to the internet at that time, now it’s 5.6 billion, effectively covering every household and business globally. This means the infrastructure is in place, and when the starting gun fires, there are no adoption barriers. This is not just a phenomenon unique to AI, but a new reality of technology distribution; the physical rules have changed, and the tracks are in place.

Another slide from last year:

What should we do? Where can we win? Two points: first, there are still many blanks; this is last year's slide, and now there are fewer blanks, but the opportunities remain vast. Second, these logos represent companies that have achieved over 1 billion in revenue during previous transformations; we do not care about unicorns, we care about revenue and free cash flow. Most successful companies are at the top of the page, at the application layer. We have always believed that AI will also be like this; the value lies in the application layer.

However, you face competition. We have the second expansion law, with calculations during testing, reasoning with tools, and communication between agents, enabling the foundational model to delve into the application layer. If you are a startup and do not plan to build a vertically integrated business, you need to think from the customer's perspective, considering specific vertical fields, specific functions, and addressing complex problems that may require humans in the loop This is a competition, and value will be generated here; this should be everyone's top consideration.

How can one win? 95% of AI company building is the same as building a regular company: solving important problems in a unique and compelling way to attract top talent to follow you. The remaining 5% is unique to AI, and in the competition for the application layer, several points need to be considered.

This is the Leone marketing cycle, content carefully crafted by our partner Doug Leone over 40 years, representing everything needed to transform the ideas in your mind into products in the hands of customers. Ideas must be converted into products, built by engineering teams, then pushed to the market and supported by sales. This is the value chain, with the bottom being the technical perspective and the top being the customer perspective, allowing you to build a moat across the entire value chain.

Your customers are uncertain about what they want from AI; you can have your own perspective and provide end-to-end solutions rather than just throwing a tool at them. You can build a data flywheel using the usage data of your own product, which others do not have. You can belong to the industry and serve the industry, just like Open Evidence does for the healthcare sector. You can speak their language, like Harvey's lawyers talking to law firms. Honestly, we wouldn't advise Ford to deploy engineers, but you can do that; it's difficult but feasible. You can fully embrace your customers, which basic models may not be able to do. By the way, we also like basic models, but we assume that most people are not building basic models but are building applications.

I have two more slides, and then we will hand over.

We are often asked: what do you look for in AI companies? In fact, 95% of the criteria are the same as how we view other companies. Here are the 5% that are unique to AI.

The first point is revenue atmosphere. The revenue atmosphere can be detrimental to you. Everyone loves the revenue atmosphere; it feels great: "Wow, we have so much revenue!" But analyze it carefully: is this a temporary attempt or has it truly created lasting behavior change? You might say, "I don't have metrics to measure this," but you actually do. Check the product adoption rate, engagement, and retention to see what users are actually doing with your product. Don't deceive yourself by treating the revenue atmosphere as real revenue; this will ultimately hurt you.

The good atmosphere part is also very important. Is Andrew Reed in the room? Atmosphere check, how does everyone feel? How does everyone feel? I heard someone say "perfect atmosphere," which is great. You need to maintain a good atmosphere with your customers; what does that mean? Your customers must trust you, and you must earn that trust. In the cycle we are in, trust is more important than the product If they believe you can make the product better, you are in a favorable position. If they don't believe, you are in an unfavorable position.

The second point, profit margin. We don't necessarily care what your gross margin is today. The cost part may continue to decline. Over the past 12 to 18 months, the cost per token has decreased by 99%, and this cost curve will continue to decline. I know the cost is rising during testing, but that will also decrease. As for the price part, if you successfully shift from selling tools to selling outcomes, moving up the value chain and able to capture more value, your price point may increase. So your gross margin may not be great today, but you should have a good path to a healthy gross margin.

The third point, data flywheel. Please raise your hand if you have a data flywheel. What business metric can this data flywheel drive? I see the certainty decreasing. The good news is, if you can't answer this question, I still like you. The bad news is, your data flywheel either doesn't exist or isn't important. It needs to be associated with business metrics; otherwise, it doesn't matter. This is critical because it's one of the best moats you can build.

On the last slide, who can tell me how these two things are connected? If you can answer that, it's truly impressive because it makes no logical sense at all. Nature abhors a vacuum. There is a huge pull in the market right now attracting AI. All macroeconomic factors, such as tariffs and interest rate fluctuations, don't matter; the upward trend in technology adoption absolutely overwhelms any fluctuations you see in the market, so ignore them. There is a huge pull in the market, and if you don't act first, others will, because nature abhors a vacuum. So, even though we just talked about moats and metrics, you are now in a business that needs to move at full speed. Now is the time to move at maximum speed.

Sonya Huang:

Thank you, Pat. I will focus on the current development of AI, providing an annual review from both customer feedback and technology perspectives.

In 2023, we presented a chart comparing the daily active users to monthly active users ratio of AI-native applications versus traditional mobile applications. The conclusion at that time was that the engagement rate of AI applications was poor, with hype exceeding actual data.

We are pleased to report that this conclusion has now changed significantly. Seeing the daily/monthly active user ratio of ChatGPT rise and approach Reddit levels is astonishing. I think this is great news, indicating that more and more people are deriving value from AI, and we are all climbing the ladder together, integrating AI into our daily lives.

Sometimes this way of using it is good and fun. Personally, I burned an embarrassing number of GPUs trying to "jiblify" everything. While the "jib" moments are fun and go viral, what’s more exciting is that we are just scratching the surface of deeper applications, such as:

• Advertising field: Able to create incredibly accurate and beautiful ad copy
• Education field: Able to visualize new concepts in an instant
• Healthcare: Able to better diagnose patients through applications like Open Evidence

We have only scratched the surface of possibilities; as AI models become more powerful, the things we can do through this entry point become increasingly profound.

How many people here have seen the movie "Her"? We have Brendan in the audience today. While we still don't have an AI version of Scarlett Johansson, 2024 has given us what we call the "Her" moment—voice generation technology has moved from "almost there" to fully crossing the uncanny valley. Some of my friends say this, but please be cautious; let me see if I can truly astonish you.

Finally, the most prominent application category this year is programming, which has achieved an astonishing product-market fit. Anthropic's Claude 3.5 Sonnet, released last fall, has sparked a rapid shift in the programming field. People are now using AI programming to create impressive things. For example, this person used AI to write their own Docent alternative. Whether you are an experienced 10x engineer or someone who has no idea how to code, we believe AI is fundamentally changing the accessibility, speed, and cost-effectiveness of software creation.

From a technical perspective, the bad news is that pre-training seems to be slowing down. Since the Alexnet era, we have scaled pre-training by 9 to 10 orders of magnitude, which means many easily obtainable results have already been achieved. The research ecosystem is looking for new breakthroughs.

The most important breakthrough is OpenAI's reasoning capabilities. We were fortunate to get a preview from Noam Brown of the Strawberry team at last year's AI Ascent, showcasing the direction of reasoning capability development. This year, we are excited to have Dan Roberts in the audience, who will give another talk on O3 and reasoning progress. Not just reasoning, but also synthetic data, tool usage, AI-assisted scaffolding, all of which are coming together to create new ways to expand intelligence.

Anthropic's MCP has created a powerful ecosystem and network, and we are also looking forward to seeing how it accelerates the use of agent tools. All these larger foundational models, reasoning time reasoning, and tool usage are coming together to create AI capable of completing increasingly complex tasks. The Meter benchmark is a good quantitative measure, but I think what's more powerful is communicating with each of you to understand the things made possible by O3, Operator, Deep Research, or Sonnet

Currently, the most exciting technological innovations in AI are happening at the blurred boundaries between research and products. Two groundbreaking examples from the past year are Deep Research and Notebook LM. We are thrilled to have the creators of these two products in the audience today—Risa and Jason from Notebook (who are creating a new company called Hu) and Issa Hulford from OpenAI.

Let's discuss where value will be generated in the AI technology stack. I remember discussing this with the excellent partners at Sequoia, at that time I personally was the "middle intelligence" in this chart, saying "ah, I'm not sure about the GPG wrapper." I remember my partner, especially Pat, insisting that value would accrue to the application layer. I remember thinking at the time, "Well, good luck, Pat." But seeing the developments over the past few years, I think you were right, Pat; you belong on this side, well done.

If you see the creation of value, if you see companies like Harvey and Open Evidence truly creating customer value, we firmly believe that the application layer is ultimately where value will concentrate, and competition at this layer is intensifying, with foundational models also competing here.

By the way, the one who is actually profiting the most from this is industry leader Jensen Huang, and we look forward to hearing his speech soon.

Back to the application layer, we believe the first batch of AI killer applications has emerged, including Chat GPT, Harvey, Glean, Sierra, Cursor, A Bridge, and a whole set of emerging companies are rising in various rich and diverse end markets, including Listen Labs, Open Evidence, etc. We are excited to showcase many of these companies today.

Another prediction is that many of these new companies will initially adopt an agent model, where the agents sold by these companies will evolve from often simple prototypes today into truly powerful systems. We see companies taking two paths to build: Path one is orchestration through rigorous testing and evaluation; Path two is tuning agents for end-to-end tasks. We look forward to hearing more from Harrison of Langchain and Issa of OpenAI on this today.

Regarding 2025, we predict that the next form of AI companies will be vertical agents. For entrepreneurs who have a deep understanding of a specific field, vertical agents present an excellent opportunity. We see agents created by companies performing exceptionally well in specific workflows through end-to-end training, utilizing technologies that include reinforcement learning on synthetic and user data, allowing AI systems to excel in specific tasks. The evidence so far makes us very optimistic: in the security field, Expo demonstrated that their agents can surpass human penetration testers; in the DevOps field, Traversal showed they can create AI troubleshooting tools better than the best human troubleshooters; in the networking field, Meter also outperformed human network engineers. All these data points, although still in the early stages, lead us to be very optimistic that vertically trained agents solving specific problems can surpass today's best humans.

The last prediction about agents in 2025: we are entering a rich era, with code as the first turning point market category previewing the actual meaning of this rich era. What happens when labor becomes cheap and abundant? Will we get a large amount of low-quality AI output? What happens when taste becomes a scarce asset? We look forward to seeing the continued progress of coding agents and their impact on the technological landscape, while also serving as a precursor to how AI will change other industries.

Now I will hand over the time to Constantine. Thank you, Sonia.

Konstantine Buhler:

Good morning everyone, thank you Sonia, thank you Pat. We just discussed very important topics: why these are so important, what is happening in the world right now, and the current state of AI and its near future. Now we will take a step back and consider mid-term and long-term predictions. In this section, we will divide it into three parts: first discussing what we see as the next major wave, then exploring the technologies needed to achieve that wave, and finally discussing what this means for each of our daily lives.

A year ago, AI Scent was discussing intelligent agents, at a time when intelligent agents were just beginning to form business models. The core of the discussion was that these machine assistants would eventually converge into machine networks. These machine networks are now widely referred to as "agent swarms," and they play roles in many companies, beginning to become a key part of the AI technology stack. Agents cooperate, compete, collaborate, and reason with each other. We believe this will further develop into an "agent economy" in the coming years.

In the agent economy, agents not only convey information but also transfer resources, conduct transactions, track each other, understand trust and reliability, and possess their own economic systems. This economy does not exclude humans; rather, it is human-centered—agents collaborate with people, and people cooperate with agents. However, to achieve this important next wave, we face many technical challenges, three of which are particularly critical:

First, persistent identity. This has two layers of meaning: first, the agent itself needs to maintain consistency. If you do business with someone who changes every day, you may not want to collaborate with them long-term. This drastically different experience can have negative impacts. Agents need to be able to maintain their personality and understanding. The second type of persistence is understanding the user. Similarly, if you do business with someone who knows nothing about you, or cannot even remember your name, this poses a significant challenge to trust and reliability.

We have been trying various technologies from RAG and vector databases to ultra-long context windows, but there are still significant challenges in achieving true memory and self-learning, especially in keeping agents consistent in important aspects while only differing in areas where differences should exist.

Second, seamless communication protocols. The good news is that it seems everyone is now focused on this. Imagine if personal computing had no seamless communication protocols, no TCP/IP, no internet—what would that be like? We are building this protocol layer, and there have already been many exciting developments around MCP. It is fantastic to see large companies collaborating to establish these protocols; this is just one of a series of protocols that allow for information transmission, value exchange, and trust transfer.

Third, security. This topic will become increasingly important and is certainly a focus for many. If you cannot meet with business partners face-to-face, the importance of security and trust becomes even more pronounced. This is especially true when dealing with agents. Therefore, a whole new emerging industry will form around trust and security, which is more important in the agent economy than in the current economy.

After discussing the technologies needed to realize the agent economy, let’s explore what this means for each of us:

First, it will change our mindset. The people in this room have already adopted what we call "random thinking." Random thinking is a departure from certainty. Many people love computer science because it is very deterministic—you program a computer to do something, and it will execute, even if the result is a segmentation fault. Now we are entering an era of random computation. If you let a computer remember the number 73, it will remember it tomorrow, next week, and next month. But if you let a person or AI remember it, they might remember 73, or they might remember 37, 72, 74, the next prime number 79, or nothing at all. This way of thinking is drastically different from the past few decades

The second change is the management mindset, which will focus on understanding what your agents can and cannot do for you. Everyone knows that being an excellent IC engineer is different from being an outstanding engineering manager, and most economics will shift towards more complex management decisions, such as blocking processes and providing feedback. I sincerely hope this does not lead to year-end reviews for agents—let's try to avoid that.

The third major change combines the first two points: more leverage accompanied by significantly reduced certainty. We are entering a world where more can be done, but you must be able to manage this uncertainty and risk. In this world, everyone in this room is well-suited to thrive.

A year ago at the AIScent conference, we discussed this chart, at which time we were talking about leverage. We believed that various functions within the organization would begin to have AI agents, and then we predicted that these functions would start to merge, coming together, and the entire process would be completed by AI agents. We even predicted the emergence of the first "one-person unicorn" company.

Although this has not yet happened, we have already seen some companies expand at an unprecedented speed while requiring fewer personnel than ever before. We believe we will reach the highest leverage level ever seen in this economy. Ultimately, these processes and agents will merge, and you will see neural networks nested within larger, more complex neural networks, forming a network of neural networks.

This will change everything: it will reshape individual work, reorganize companies, and rebuild the economy. Thank you all for attending; we will have a wonderful AIScent conference today, and we greatly appreciate your participation