In this interview Sonya Huang, a partner at Sequoia Capital, shares her perspective on why OpenAI’s approach is succeeding and how the broader AI ecosystem is evolving. She observes distinct generational usage patterns for ChatGPT—older users treat it as a Google replacement, millennials use it as a life‑coach, and younger users see it as an operating system. Huang argues that the next major breakthrough will come from scaling inference‑time compute rather than further pre‑training, citing examples from Go where allowing a model to “think” for a minute dramatically raises its Elo rating. She notes that image, video and robotics models are increasingly built on transformer‑diffusion hybrids, and that the real progress driver is the surrounding ecosystem of tools, memory, and APIs (MCP). The value, she says, will accrue in the application layer, with horizontal foundation‑model platforms winning large‑scale consumer adoption while vertical agents target specific industries. She discusses the nascent enterprise AI market, the imminent arrival of humanoid robots powered by synthetic data pipelines, and the massive capital intensity of training such systems, likening it to the autonomous‑vehicle era. Huang warns that venture capital is already abundant, but AI companies are showing unprecedented growth rates and a trillion‑dollar total addressable market. She sees private‑equity‑style cost‑reduction playbooks becoming AI‑enabled, and explains Sequoia’s “seed‑to‑IPO and beyond” strategy that uses multiple capital pools to back founders from the earliest conviction. Finally, she highlights energy and data‑center power as the current binding constraint on AI scale, pointing to massive renewable over‑builds in places like West Texas as the next frontier for AI infrastructure.
Chapters
How people are using ChatGPT was really interesting to me. If you're old, you're using it as a Google replacement. If you're in your 20s or your 30s, people tend to be using it as more of a life advisor life coach. If you're really young, like the Youths are using ChatGPT as an operating system, that use case seems really interesting to me.
What are the top kind of discussions that people are debating right now? This idea of like the pre training wall needing to move into more RL focused techniques to kind of get the next level of breakthroughs. Is that the right question to even be asking? Were people debating it? And do you have a take on the idea of hitting a pre training wall?
The top humans are like at 3500 ELO at go. The best bots are like 3000 before you give them access to inference time compute. But if you actually let the model actually sit and think for a minute before it places its piece, you can get that elo up to 5,500 points. So like way better than human superhuman to get roughly the same order performance. Once you've like, it's diminishing marginal returns on multiple vectors. Right. But once you've kind of like hit diminishing marginal returns on one vect pre training, if you start scaling post training from there and sorry specifically inference time compute, you can get like a hundred thousand x scale up in performance there. So we're just starting to climb that second curve. I think OpenAI deserves a ton of props for like seeing that and like investing decisively behind it. Because the hard thing is like scaling up reasoning infrastructure is different from scaling up pre training infrastructure.
Are you seeing that in image and diffusion as well? It feels like the images in ChatGPT is doing something different. It feels, feels like they're layering a few different techniques together. Are we evolving past the big diffusion model in image generation as well?
I'm not a researcher, but I get to talk to a lot of smart researcher friends. My understanding is that it's a, it's a combination of a transformer diffusion architecture and I think that most people don't believe that diffusion models will fully get us there. Whereas transformers have a lot more juice in them.
Sure.
And so even if you look at it's not just image, it's video, it's robotics. A lot of those have transformers as their backbone. There's so much that's happening in like the harness around it. Right. What is the, you know, what's the for loop that you run the Model in what, what tools do you give it access to? We did a little pull at AI essence of like what innovation is going to drive the most amount of progress in the AI ecosystem in the next 12 months. And like biggest answer was MCP and, and tool use and forming an ecosystem around that. So the models themselves get smarter but they're also surrounded by a big ecosystem.
The big story this week around OpenAI obviously after the event was the new CEO coming in to focus on applications. I'm curious what the general kind of vibe was around, around that, you know.
The mimetic crime in the venture ecosystem right now is just like the value is in the application layer. The values in the application layer there's going to be a place at least in the near term for vertical agents applied to a very specific sector. And like we've, we have a bunch of those companies in our portfolio. I think that what the foundation models have proven though, which was debatable a couple of years ago, is that they have every right to win the application layer. ChatGPT is like a, to me it's like a runaway freight train in terms of consumer adoption. Some of the Metrics they've published 300 to 500 million weekly actives year to date. It's just, it's phenomenal user growth. One of the things that Sam shared at our conference in terms of how people are using ChatGPT was really interesting to me. If you're old, you're using it as a Google replacement. So I'm old. If you're in your 20s or your 30s, people tend to be using it as more of a life advisor, life coach type thing. And then if you're really young, like the Youths are using ChatGPT as an operating system. And I found that framework really interesting and you know, especially in combination with like the curly building things around memory, around tool use, around connecting to your other applications. It really does feel like you know, if you're a young person and like really, really connecting ChatGPT and like mind melding with it. That use case seems really interesting to me. If you, you know, if you look at Google as front door to the Internet, it feels to me that ChatGPT is in many consumers minds at front door to AI. Your question, where will value accrue? We think it's in the app layer. Yeah, I think a lot of the horizontal app layer opportunity will be, will be won by foundation models like, like OpenAI, like xai. And then a lot of the companies we're backing are going after a very Specific vertical opportunity.
Are you particularly bullish on enterprise application as independent? I mean I'm thinking about like Google very much one consumer search. But you know, in legal there was LexisNexis and then Palantir helps, you know, the government search through data sets and there's, there's all these different enterprise use cases that Google wasn't able to go after.
Enterprise war feels like it is like, you know, we're in the first inning still. It's not clear who's going to win yet. I think the enterprise AI battlefield is very much a work in progress at this time. I think the shape of workflows and problems is so diverse in the enterprise that my guess is a lot of these companies will be successful.
How are you thinking about humanoid robotics? I want to see the satellite photo of the data center that's getting built out to do a massive training run for an end to end robotics model. And I haven't seen that yet. Is that the right signal to be for, for takeoff in humanoid robotics?
I think the humanoids are a lot closer than, than we may think they are. I would guess in like two or three years this stuff will be in tens of thousands of households at least. Maybe households. Yeah. A huge part of the breakthrough is, is going to be synthetic data pipelines. Robotics, unlike LLM, like you just don't have Internet scale data to train on. Yeah, but like one of the amazing things about what's happening, what's happening in AI right now is just like with LLMs but also more specifically with these generative world models, you can generate, you know, tens of thousands of variations of the same environment to simulate and these robots can get better and break through that data barrier extremely quickly. And so I, you know, I was personally like a robot bear for the longest time and you know, talked to a bunch of very smart people in robotics and over the last week and I've flipped. I feel like the humanoids are coming.
Do you think we're also going to see raises from humanoid robotics companies where you know, a ton of the raise goes into Nvidia GPUs to build a huge data center to train some massive model?
I think it's going to be extremely capital intensive and it reminds me me of autonomous vehicles five plus years ago where like, you know, at the end of the day like we have, we have Waymos and we have Teslas driving around those. Those companies had enormous economic engines to support the development. I think the same is happening with humanoid right now. And so my guess would be it's like it's going to, it's not just Nvidia GPUs, it's everything, right, because you're co developing the hardware, you still have to collect a ton of data on actions. And so it's, it's, it's a ton of spend everywhere.
What's your overall take on the venture market? Are we in a bubble right now? Can there ever be too much venture capital? All the key questions.
Never, John.
Never.
There is, there's too much venture capital already. I think there's like a cohort of companies that are like growing high quality revenue at the highest pace that we've ever seen. And that includes like I mentioned OpenAI before, but it's also companies like Glean and Harvey and Cira and all these companies. And so to the extent like valuation is a function of like how much of you de risked, how much product market fit do you have? Like what is your, what is the growth rate of your, of your business and what is the ultimate TAM potential? I think these AI companies are demonstrating just, you know, growth rates outside of what we've ever seen before and then TAM potential because it is very much you're selling into, you know, if you're able to get outcomes based pricing, you're selling into a, into a services replacement, not a tools replacement. And it's a, it's a TAM in the trillions. Right. And so.
Okay, but yeah, sorry, sorry to interrupt you on that point. Something, something I'm curious about is, is so yes, if you have an AI tool that can replace services spend, you can capture some, you can whatever, ideally capture a lot of that market. But the thing I keep coming back to, and maybe this isn't the right way to think about it, but you're, you're not simply competing with end humans that are delivering that services. A company will also be competing with other AI tools that have, you know, a similar cost structure. So does that not over time just drive the sort of dollar amount that you can capture just down to something that looks more like a software market?
We have this debate all the time, so I'm glad you bring it up. I think it really depends like if, if what you're doing is like, you know, really low switching costs, really low differentiation above what the models provide? Like yeah, I think that margin is going to get competed down. And so I think that's why we've historically debated a lot of these GPT wrapper companies. I think that if you're building something that's really hard to build or that's integrating into a customer base that wants to choose an AI champion and move on with life.
What is your take on private equity? Dipping their toe into more venture scale opportunities and venture investors starting to look more at private equity style roll up deals.
I think it makes a ton of sense and it's a continuation of the private equity play.
Right.
A lot of the investments I did in PE only worked because you took 20% of the costs out. And so like now you have a much better tool to go and do that. But like it very much is the playbook and it's what they're best in class at. And so do I expect that they'll be great at adding, you know, AI to the arsenal for how they get those margins up? Absolutely. When I think of the businesses that I'm excited to invest in, it's who's creating gross profit dollar creation. And you can choose to do that by investing in the billion dollar revenue company and taking their cost down 10% or you can choose to do that by backing that amazing Dev Tools founder that knows how to build the AI native dev tools company that's going to create $100 million of gross profit off the bat. And so I very much personally like to invest in the founders that are kind of creating net new revenue dollars and net new gross profit dollars. But multiple ways to play.
How do you think about the different businesses that Sequoia's running right now from early stage growth stage beyond? What do you think Sequoia does best and what are you excited about in the future?
If you think of our strategy, it's like it's seeds, IPO and beyond for like the most ambitious entrepreneurs in the world. And so sometimes we're able to catch them early at the seed and sometimes we catch them later on in their journey. But like the, the point of adding additional kind of pools of capital to our fund strategy has been when we find a winner in our portfolio, we want to be able to invest a lot of money behind that company. The reason for the Sequoia Capital Fund is like even after these companies go public, we think a lot of that return is still to be had. We've modified our structure over the years to be able to kind of support these companies as they, as they grow and become later stage and go public. But ultimately it's, you know, it's invest at the earliest point of conviction and ideally that's at the seed.
Sam had some interesting sort of quotes over the last week or so. I don't know exactly when they were happening something to the effect of the cost of AI or the cost of intelligence will just converge on the cost of energy or electricity. And I'm curious, you guys had Chase from Cruiseau. Yeah. Talk. And I'm, I'm curious if you think that that is a potential area that you expect to see more net new early stage startups exploring.
If you look at like typical data centers today, like 20, 40 megawatt data centers, like the biggest data centers of the world are in Northern Virginia or sorry, in the US or in Northern Virginia, the aggregate capacity is four and a half gigawatts there. Chase at Crusoe himself has 20 gigawatts in pipeline right now, more than 2 gigawatts built out. And so like the sheer scale of the build out right now is just like nothing that we've supported in the past. And the bottlenecks are moving around. So like people, it's actually impossible to get chips now. A lot of that is easing and power is a new bottleneck. And so this is why there's so much happening in West Texas right now in Abilene, where they just have this massive overbuild of renewables, especially with wind. And so I think very much you'll see a lot of the AI buildouts following power and energy because that ultimately is the binding constraint right now.
