【红杉资本】为何AI工业革命是万亿级投资机会

We believe in the AI revolution at Sequoia, and we think that this is going to be a transformation that is as big, if not bigger, than the Industrial Revolution.

• In this short presentation, we're going to share what this cognitive revolution is all about and why it offers a $10 trillion opportunity.
• We have four sections: our thesis around artificial intelligence, the commercial opportunity, investment trends, and investment themes for the next 12 to 18 months.
• AI is akin to the Industrial Revolution, with key points being the invention of the steam engine and the emergence of factory assembly lines.
• The cognitive revolution presents the chance to expand into a $10 trillion market with many untapped opportunities.
• We're observing five key investment trends and themes going forward.

We believe in the AI revolution at Sequoia, and we think that this is going to be a transformation that is as big, if not bigger, than the Industrial Revolution.

In this short presentation, we're going to share what this cognitive revolution is all about and why it offers a $10 trillion opportunity.

We have four sections. The first is our thesis around artificial intelligence. The next is the commercial opportunity. Then we'll get into some investment trends. These are what we see going on right now in the state of artificial intelligence. And finally, we'll talk about some investment themes—what we see in the next 12 to 18 months.

We believe that artificial intelligence is akin to the Industrial Revolution. In particular, we want to call out three points in time of the Industrial Revolution.

• The invention of the steam engine, which kicked it all off.
• The first factory system, which put all the requisite parts under the same roof.
• The first factory assembly line as we know it today.

What's curious about this slide is the time in between these points: 67 years from the first steam engine to the first factory. And that first factory didn't even use the steam engine. It was powered by water then. 144 years between that first factory and the factory assembly line as we know it today. The question is, why did it take so long?

In particular, what was going on in these 144 years? We believe the reason why it took so long is the specialization imperative beyond a certain scale.

For a complex system to mature, it must combine general purpose components and labor with highly specialized components and labor. In other words, what was happening is taking these general technologies and specializing them to produce specific outputs.

Now we live in the cognitive revolution. You might say that the steam engine was that first GPU, the GeForce 256 in 1999. The first AI factory, putting all the components necessary to produce these AI tokens in 2016.

So the question then becomes, who is the John Rockefeller, the Andrew Carnegie, the Westinghouse, and the Wedgwood of this cognitive revolution?

We believe that it is the startups of today that are carrying out this specialization imperative and the startups that have yet to be formed that will build those applications.

Now, we're not called Sequoia History Rec, we're called Sequoia Capital. So let's talk about the dollars and cents.

You may have seen this slide before. We used it in our AI Ascent conference. On the left is a $350 billion circle representing the spend on software at the beginning of the cloud transformation. That 6% sliver is $6 billion of software as a service spend.

Now, what happened is not only did software as a service increase its share of the on-prem software market, it actually expanded the entire market—growing to over $650 billion. Today, we believe something similar is going to happen in artificial intelligence, and in an even bigger opportunity. This is the $10 trillion US services market, of which maybe $20 billion or so today is automated by AI.

That is the $10 trillion opportunity to not only grow the share of this pie but grow the pie itself.

You may have seen the last slide, but you've never seen this slide. This is a piece of an internal memo at Sequoia. That's services jobs sorted by the rightmost column. That column is a multiplication of the number of employees in the job and the median annual wage according to the U.S. Census Bureau.

What you'll notice is there are really big TAMs. You'll also notice that Sequoia has made investments in these spaces. Consider Open Evidence and Freed in the registered nurses space, or Factory and Reflection in the software developer space, or Harvey and Crosby and Finch in the law space.

So at Sequoia, we like to think about TAM, and in particular, market. Our founder Don Valentine always talked about market, market, market, and its importance.

This chart shows the market cap of the S&P 500. What you'll notice is a few really big companies. That's Nvidia on the far left with over $4 trillion of market cap, and the percentage is the one-year performance.

There's no Kirkland & Ellis law firm on this slide. There's no Baker Tilly accounting firm, even though these are businesses that do billions in revenue. We believe that the cognitive revolution offers an opportunity to expand the market and expand this slide to include many large standalone public companies built around AI in the services space.

Next, let's talk about five investment trends that we're noticing right in the here and now of the artificial intelligence cognitive revolution.

The first is leverage over uncertainty. We've noticed that work is moving from us having minimal leverage on a task and 100% certainty of the outcome to 100+ percent leverage on the task and way less certainty on the exact manifestation of the outcome.

If you're a seller, if your job is to be a salesperson and you're managing a series of accounts, potential customers to work with, today you might manage those accounts all on your own—monitoring each one for an opportunity.

But in the AI agent future, you could be using Rocks, which allows you to have hundreds of AI agents, an AI agent for every single customer—tracking their progress, seeing what's changing there and showing you opportunities to re-engage and expand the partnership with that customer.

Now, this AI agent is not going to do things exactly like you are. It might actually miss something or make a mistake, and that's where the person comes in and corrects it. We see 100+ percent leverage in this case, maybe even 1000% leverage with some more uncertainty. It's not exactly the work that you're doing.

Next is, we've noticed that measurement has moved into the real world. For most of the history of artificial intelligence, we used academic benchmarks.

When I was an AI engineer over a decade ago, we used ImageNet as the benchmark for our computer vision research. But today, if you want to prove excellence, you prove it in the real world.

Consider Expo, which wanted to show that their artificial intelligence was the best AI hacker in the world. Rather than just proving this on academic benchmarks, they went out into the real world and they competed on HackerOne against every other hacker in the world that had registered on that website to find vulnerabilities.

They were able to show that on real world data they were able to compete and win, being the number one hacker in the world. We've noticed that this is the new gold standard—not just academic benchmarks, but real world measurement.

The third is reinforcement learning. The AI industry has been talking about reinforcement learning for a long time now. We've seen it come into center stage in the past year. Not only are the large reasoning labs benefiting from reinforcement learning, but we also see many of our portfolio companies benefiting from it.

Consider Reflection, which uses reinforcement learning to train some of the best open-source models in the coding space.

A fourth trend we've noticed is AI in the physical world. It's really coming to life. And this isn't just humanoid robotics; it also is creating processes and creating hardware with artificial intelligence.

Consider Nominal, which uses artificial intelligence to accelerate the hardware manufacturing process and actually also uses artificial intelligence to do the quality assurance after it's deployed in the field.

Finally, we've noticed that the new production function is compute—that's flops per knowledge worker. If you survey our portfolio, they'll say that they forecast at minimum a 10x increase in the flops per knowledge worker. That's at minimum 10 times more consumption of compute per knowledge worker.

Because as described, the knowledge worker might be using one or dozens or hundreds or thousands of these agents. And on the more optimistic side, we see a future where there's 1000x or 10,000x consumption of flops per knowledge worker.

This is really powerful for not only inference companies but companies that are protecting that inference and companies that are using this new production function to get to more workers.

Next, let's talk about five investment themes. These are important themes that we're investing in in the year ahead.

The first theme is persistent memory. Persistent memory can mean at least two things. The first is long-term memory, remembering the context that has been shared with an AI for a long period. And the second is the persistence of the identity of an artificial intelligence—the agent remaining its own personality and style for some period of time.

This is going to be critical to crack in order to get into more and more work functions for AI. An artificial intelligence that goes into the productivity space needs to have long-term memory, needs to understand the entire context of an organization and of a function.

We've noticed that there's no equivalent of scaling laws in the persistent memory space. There have been a lot of attempts, be it vector, DBS and rag or longer and longer contact windows, but this has not been cracked and it's a major opportunity ahead.

The next is seamless communication protocols. Now there's been a lot of excitement around MCP, rightfully so. But let's think back to the Internet revolution.

TCP/IP was not the finish line; it was the starting gun. And in the communication revolution, we have an opportunity to actually take AIs and help them seamlessly communicate with each other. We believe that this is going to yield many major applications.

One to consider is shopping. Right now, if you wanted to purchase a product with AI, you might go and do some research with AI and then execute that order with your favorite provider of a one-click checkout.

But in the future, we see that AI is going to be able to do that because of the seamless communication protocol. AI is going to be able to go through the process and find the best prices, execute that purchase for you, and complete it all while decreasing the costs of all those businesses that have made it so much easier to use them than other providers.

The third trend is AI voice. You might be surprised that I didn't put AI video, but that's intentional. I think that in a year, AI video might be here, but AI voice is right now. And that's because not only has its fidelity increased, the quality of voice increased to the point where it's useful day to day, but the latency has decreased where you could have real-time conversations with AI voice.

Now, there's many applications of AI voice that are exciting. These include AI Friends, AI Companions, and AI Therapists, all these business-to-consumer applications. I'm also personally very excited about the other applications of AI voice in the enterprise. If you're planning to ship something as an enterprise and you're working on the logistics, many of those logistics are still done by voice.

Now you can actually work for a future that automates many of those logistical coordinations. If you want to buy or sell a large block of fixed income, you're probably doing that over voice with an over-the-counter trading desk.

All this in the enterprise space can be accelerated by using AI Voice.

Next, there's AI security. We see an absolutely huge opportunity in AI security, and that's all the way from the development layer all the way through to the consumer.

At the development layer, we see an opportunity to help the large foundation model labs and AI labs develop their technology in a secure manner. Then there's the distribution, ensuring that it's distributed in a secure manner and that no bad actors insert themselves in that process.

Then for the user themselves, ensuring that when they're consuming this product or vibe coding a new app, they don't accidentally introduce vulnerabilities.

A tangible example here would be a consumer who is instructed by their AI to download a piece of software with Terminal. They might not be very familiar with Terminal, and the AI might not know that this piece of software could introduce a vulnerability with the rest of that consumer environment.

We're going to enter a world where the AI is actually protecting both the individual and the AI agents.

In fact, we see a future where unlike the physical world, in this digital world you could have hundreds of AI security agents for every human and even for every agent. Unlike in the physical world, you're not limited by physical space or even the same costs. You could have a massive number of AI security agents per person, per agent.

Finally, we find that open source is at a pivotal moment in the AI journey. If you asked us two years ago, we would say that it seems like open source has a shot at competing with and maybe even winning against the state-of-the-art foundation models.

Today that position seems much more precarious. We believe that it's very important that open source competes and offers some of the best state-of-the-art foundation models.

We think it's critical for a more free, open future where anyone can build, and we hope to help build that future so that open source models are available to everyone to build excellent products with and that the AI future is not just limited to the extremely well-funded giants.

These are five investment themes that we're thinking about now. The question becomes what happens if we're able to take these investment themes and translate them into progress?

We believe that that progress will be able to take the time to that cognitive assembly line and compress it materially for many years into just a few.

We thank you so much for listening and hope to get to build in the cognitive revolution together.