GPU、TPU 与 AI 经济学详解｜Gavin Baker 访谈

I’d love to talk about how you, in the nitty-gritty process, approach new things that come out in this whole AI world, because it’s happening so constantly. I’m extremely interested in it, and I find it very hard to keep up. I have a couple of blogs that I go read and friends that I call, but maybe let’s take Gemini 3 as a recent example.

When that comes out, what—literally, take me into your office—what are you doing? How do you and your team process an update like that, given how often these things are happening?

I think the first thing is, you have to use it yourself. I would say I’m amazed at how many famous and august investors are reaching really definitive conclusions about AI based on the free tier.

The free tier is like you’re dealing with a 10-year-old, and you’re making conclusions about the 10-year-old’s capabilities as an adult. You could just pay, and I do think you need to pay for the highest tier, whether it’s Gemini Ultra, SuperGrok, whatever it is. You have to pay the $200-per-month tier, whereas those are like a fully fledged 30- or 35-year-old. It’s really hard to extrapolate from an 8- or 10-year-old to the 35-year-old, and yet a lot of people are doing that.

The second thing is, there was an Insider post about OpenAI, and they said, to a large degree, OpenAI runs on Twitter vibes. I just think AI happens on X. There have been some really memorable moments, like there was a giant fight between the PyTorch team at Meta and the JAX team at Google on X, and the leaders of each lab had to step in publicly and say, “No one from my lab is allowed to say bad things about the other lab. I respect them, and that is the end of that.”

The companies are all commenting on each other’s posts. The research papers come out. If on planet Earth there are 500 to 1,000 people who really understand this and are at the cutting edge of it, and a good number of them live in China, I think you have to follow those people closely. I think there is incredible signal. Everything in AI is just downstream of those people.

Everything Andrej Karpathy writes, you have to read 3 times, minimum.

Yeah. He’s incredible. I would say anytime somebody from one of the 4 labs that matter—OpenAI, Google, Anthropic, and xAI, which are clearly the 4 leading labs—goes on a podcast, it’s so important to listen.

For me, one of the best use cases of AI is to keep up with all of this. Just listen to a podcast, and then, if there are parts that I thought were interesting, talk about them with AI. I think it’s really important to have as little friction as possible. I’ll bring it up. I can either press this button and pull up Grok, or I have this.

Oh, wow. I don’t touch that. That just brings it right up.

Yeah, it brings it right up. What do you think of Patrick O'Shaughnessy?

“Patrick O'Shaughnessy is one of my favorite voices in investing. His Invest Like the Best podcast is straight fire. He does deep dives with folks like Bill Gurley—”

Gurley?

Yes.

It’s so—can you believe we have this?

I know. It’s like we have—I think somebody said on X—we’ve imbued these rocks with crazy spells, and now we can summon superintelligent genies on our phones over the air. It’s crazy.

1. The Blackwell Transition

Crazy. So when something like Gemini 3 comes out, the public interpretation was, “Oh, this is interesting. It seems to say something about scaling laws and the pretraining stuff.” What’s your frame on the state of progress in frontier models in general? What are you watching most closely?

I do think Gemini 3 was very important because it showed us that scaling laws for pretraining are intact. They stated that unequivocally, and that’s important because no one on planet Earth knows how or why scaling laws for pretraining work.

It’s actually not a law. It’s an empirical observation, and it’s an empirical observation that we’ve measured extremely precisely and that has held for a long time. Our understanding of scaling laws for pretraining—and maybe this is a little bit controversial with 20% of researchers, but probably not more than that—is kind of like the ancient British people’s understanding of the sun or the ancient Egyptians’ understanding of the sun.

They could measure it so precisely that the east-west axes of the Great Pyramids are perfectly aligned with the equinoxes, and so are the east-west axes of Stonehenge. Perfect measurement. But they didn’t understand orbital mechanics. They had no idea how or why it rose in the east, set in the west, and moved across the horizon.

The aliens.

Yeah, our god in a chariot. It’s really important every time we get a confirmation of that. Gemini 3 was very important in that way.

I think there’s been a big misunderstanding, maybe in the public-equity investing community or the broader, more generalist community, based on the scaling laws of pretraining. There really should have been no progress in 2024 and 2025.

The reason for that is, after xAI figured out how to get 200,000 Hoppers coherent, you had to wait for the next generation of chips. You really can’t get more than 200,000 Hoppers coherent. Coherent just means you could think of it as each GPU knowing what every other GPU is thinking. They’re sharing memory, they’re connected, they scale up networks and scale out, and they have to be coherent during the pretraining process.

Everything in AI is a struggle between Google and NVIDIA. Google has a TPU, and NVIDIA has its GPUs. Google only has a TPU, and it uses a bunch of other chips for networking. NVIDIA has the full stack, and Blackwell was delayed.

Blackwell was NVIDIA’s next-generation chip, and the first iteration of that was the Blackwell B200. A lot of different SKUs were canceled, and the reason for that is it was by far the most complex product transition we’ve ever gone through in technology.

Going from Hopper to Blackwell, first you go from air-cooled to liquid-cooled. The rack goes from weighing, in round numbers, 1,000 lb to 3,000 lb. It goes from, in round numbers, 30 kilowatts—which is 30 American homes—to 130 kilowatts, which is 130 American homes.

I analogize it to imagining if, to get a new iPhone, you had to change all the outlets in your house to 220 volts, put in a Tesla Powerwall, put in a generator, put in solar panels—that’s the power—put in a whole-home humidification system, and then reinforce the floor because the floor can’t handle this. It was a huge product transition, and the rack was so dense that it was really hard for them to get the heat out.

Blackwells have only really started to be deployed in scaled deployments over the last 3 or 4 months. Had reasoning not come along, there would have been no AI progress from mid-2024 through essentially Gemini 3. There would have been none. Everything would have stalled, and can you imagine what that would have meant to the markets? For sure, we would have lived in a very different environment.

Reasoning bridged this 18-month gap. Reasoning saved AI because it let AI make progress without Blackwell or the next generation of TPUs, which were necessary for the scaling laws for pretraining to continue.

The reason we’ve had all this progress is that we have these 2 new scaling laws of post-training, which is just reinforcement learning with verifiable rewards. Verifiable is such an important concept in AI. One of Karpathy’s great things was, “With software, anything you can specify, you can automate.”

With AI, anything you can verify, you can automate. It's such an important concept, and I think an important distinction. Then there's test-time compute. All the progress we've had—and we've had immense progress since October 24 through today—was based entirely on these 2 new scaling laws.

Gemini 3 was arguably the first test since Hopper came out of the scaling law for pretraining, and it held. That's great because all these scaling laws are multiplicative. So now we're going to apply these 2 new scaling laws—reinforcement learning with verifiable rewards and test-time compute—to much better base models. Google came out with the TPU v6 in 2024 and the TPU v7 in 2025.

And in semiconductor time, imagine Hopper is like a World War II-era airplane. It was by far the best World War II-era airplane. It's the P-51 Mustang with the Merlin engine.

And 2 years later, in semiconductor time, that's like—you know, you're an F-4 Phantom. Okay. [Laughter]

Because Blackwell was such a complicated product and so hard to ramp, Google was training Gemini 3 on 2024- and 2025-era TPUs, which are like F-4 Phantoms. Blackwell is like an F-35.

It just took a really long time to get it going.

So I think Google for sure has this temporary advantage right now. From a pretraining perspective, I think it's also important that they've been the lowest-cost producer of tokens. And this is really important because AI is the first time in my career as a tech investor that being the low-cost producer has ever mattered.

Apple is not worth trillions because they're the low-cost producer of phones. Microsoft is not worth trillions because they're the low-cost producer of software. NVIDIA is not worth trillions because they're the low-cost producer of AI accelerators. It's never mattered.

Okay.

And this is really important because what Google has been doing as the low-cost producer is sucking the economic oxygen out of the AI ecosystem, which is an extremely rational strategy for them and for anyone who's a low-cost producer. Let's make life really hard for our competitors.

So what happens now, I think, has pretty profound implications. One, we will see the first models trained on Blackwell in early 2026.

Yeah.

I think the first Blackwell model will come from xAI. The reason for that is that, according to Jensen, no one builds data centers faster than Elon. Jensen has said this on the record.

Even once you have the Blackwells, it takes 6 to 9 months to get them performing at the level of Hopper because Hopper is finally tuned. Everybody knows how to use it. The software is perfect for it. Engineers know all its quirks. Everybody knows how to architect a Hopper data center at this point.

By the way, when Hopper came out, it took 6 to 12 months for it to really outperform A100, which was the generation before. So if you're Jensen or NVIDIA, you need to get as many GPUs deployed in 1 data center as fast as possible, in a coherent cluster, so you can work out the bugs.

This is what xAI effectively does for NVIDIA because they build the data centers the fastest. They can deploy Blackwells that scale the fastest, and they can help work with NVIDIA to work out the bugs for everyone else. So because they're the fastest, they'll have the first Blackwell model.

We know that the scaling laws for pretraining are intact, and this means the Blackwell models are going to be amazing. Blackwell is not an F-35 versus an F-4 Phantom, but from my perspective it is a better chip. Maybe it's like an F-35 versus a Rafale. [Laughter] Now that we know pretraining scaling is holding, we know that these Blackwell models are going to be really good.

And based on the raw specs, they should probably be better.

Then something even more important happens. So the GB200 was really, really hard to get going. The GB300 is a great chip. It is drop-in compatible in every way with those GB200 racks. Now, you're not going to replace the GB200s. No new power walls.

Yeah. Just any data center that can handle those. You can slot in the GB300s, and now everybody's good at making those racks. You know how to get the heat out. You know how to cool them.

You're going to put those GB300s in, and then the companies that use the GB300s are going to be the low-cost producers of tokens, particularly if you're vertically integrated. If you're paying a margin to someone else to make those tokens, you're probably not going to be.

I think this has pretty profound implications because I think it has to change Google's strategic calculus. If you have a decisive cost advantage and you're Google, and you have search and all these other businesses, why not run AI at a negative 30% margin?

It is by far the rational decision. You take the economic oxygen out of the environment. You eventually make it hard for your competitors, who need funding unlike you, to raise the capital they need. Then, on the other side of that, maybe you have an extremely dominant share position.

Well, all that calculus changes once Google is no longer the low-cost producer, which I think will be the case. The Blackwells are now being used for training. Then, when that model is trained, you start shifting Blackwell clusters over to inference, and then all these cost calculations and these dynamics change.

I do think it's very interesting—the strategic and economic calculations between the players. I've never seen anything like it. Everyone understands their position on the board, what the prize is, and what play their opponents are running. It's really interesting to watch.

So I just think if Google changes its behavior, because it's going to be really painful for them as a higher-cost producer to run that negative 30% margin, it might start to impact their stock. That has pretty profound implications for the economics of AI.

And then when Reuben comes out, we'll know—the gap is going to expand significantly versus TPUs.

Versus TPUs and all other ASICs.

Now, I think Trainium 3 is probably going to be pretty good. Trainium 4 is going to be good.

Why is that the case? Why won't TPU v8 or v9 be every bit as good?

A couple of things. For whatever reason, Google made more conservative design decisions. I think part of that is that, in round numbers, there's a front end and back end of semiconductor design, and then there's dealing with TSMC.

You can make an ASIC in a lot of ways. What Google does is mostly the front end for the TPU, and then Broadcom does the back end and manages TSMC. It's a crude analogy, but the front end is like the architect of a house. They design the house. The back end is the person who builds the house, and managing TSMC is like stamping out that house, like Lennar or D.R. Horton.

For doing those 2 latter parts, Broadcom has a 50% to 55% gross margin. We don't know what the gross margin is on TPUs. Let's say in 2027, TPU consensus estimates are maybe somewhere around $30 billion. Again, who knows?

Yeah, yeah, yeah, but $30 billion, I think, is a reasonable estimate.

At 50% to 55% gross margins, Google is paying Broadcom $15 billion.

Okay, that's a lot of money.

At a certain point, it makes sense to bring a semiconductor program entirely in-house. In other words, Apple does not have an ASIC partner for their chips. They do the front end and back end themselves, and they manage TSMC. The reason is they don't want to pay that 50% margin.

At a certain point, it becomes rational to renegotiate this. Just as perspective, the entire operating expense of Broadcom's semiconductor division is, in round numbers, $5 billion. So it would be economically rational. If Google is paying them $15 billion on $30 billion, Google can go to every person who works in Broadcom's semiconductor division, double their compensation, and make an extra $5 billion.

In 2028, let's just say it does $50 billion. Now it's $25 billion. You could triple their compensation. And by the way, you don't need them all.

Yeah.

Of course, they're not going to do that because of competitive concerns. But with TPU v8 and v9, all of this is beginning to have an impact because Google is bringing in MediaTek.

This is maybe the first way you send a warning shot to Broadcom: “We're really not happy about all this money we're paying.” They did bring MediaTek in, and the Taiwanese ASIC companies have much lower gross margins. So this is kind of the first shot across the bow.

Then there's all this stuff people say: “Oh, but Broadcom has the best SerDes.” Broadcom has really good SerDes, and SerDes is extremely foundational technology because it's how the chips communicate with each other. You have to serialize and deserialize.

But there are other good SerDes providers in the world. A really good SerDes is not a moat. Maybe it's worth $10 billion or $15 billion a year, but it's probably worth about $25 billion a year.

Because of that friction, and because of what I think are conservative design choices on the part of Google, maybe the reason they made those conservative design choices is because they were going to a bifurcated supply. TPUs are slowing down. I would say GPUs are accelerating.

This is the first competitive response of Lisa and Jensen to everybody saying, “We're going to have our own ASIC.” It's, “Hey, we're just going to accelerate. We're going to do a GPU every year, and you cannot keep up with us.” And then I think what everybody is learning is like, “Oh, wow, that's so cool.”

You made your own accelerator as an ASIC. Wow, what’s the NIC going to be? What’s the CPU going to be? What’s the scale-up switch going to be? What’s the scale-up protocol? What’s the scale-out switch?

What kind of optics are you going to use? What’s the software that’s going to make all this work together? Then it’s like, “Oh, shit, I made this tiny little chip.” Whether it’s admitted or not, I’m sure the GPU makers don’t love it when their customers make ASICs and try to compete with them. And then it’s like, “Whoops, what did I do? I thought this was easy.”

How do you know?

It also takes at least 3 generations to make a good chip. TPU v1—I mean, it was an achievement that they made it.

Yeah.

It really wasn’t until TPU v3 or v4 that the TPU started to become even vaguely competitive.

Is that just a classic learning-by-doing thing?

100%.

Yeah.

Even if you’ve made the first one, from my perspective, the best ASIC team at any semiconductor company is actually the Amazon ASIC team. They were the first ones to make the Graviton CPU. They have Nitro. It was the first—it’s called a SuperNIC. They’ve been extremely innovative, really clever.

With Trainium and Inferentia 1, they may be a little better than TPU v1, but only a little. Trainium 2 is a little better. Trainium 3, I think, is the first time it’s like, “Okay.” Then I think Trainium 4 will probably be good. I’ll be surprised if there are a lot of ASICs other than Trainium and TPU.

And by the way, Trainium and TPU will both run on customer-owned tooling at some point. We can debate when that will happen, but the economics of success that I just described mean it’s inevitable. No matter what the companies say, the economics and reasoning from first principles make it absolutely inevitable.

If I were to zoom all the way out on this stuff—because sometimes I find these details unbelievably interesting, and it’s like the grandest game that’s ever been—what is the global human dividend of all this crazy development? We keep making the loss lower on these pretraining scaling laws, but who cares? It’s been a while since I’ve asked this thing something where I wasn’t blown away by the answer, for me personally.

What are the next couple of things that all this crazy infrastructure war allows us to unlock because it’s so successful? If I were to posit an event path, I think the Blackwell models are going to be amazing. The dramatic reduction in per-token cost enabled by the GB300, and probably more by the MI450 than the MI355, will lead to these models being allowed to think for much longer, which means they’re going to be able to do new things.

I was very impressed that Gemini 3 made me a restaurant reservation. It’s the first time it’s done something for me, other than research something and teach me stuff. If you can make a restaurant reservation, you’re not that far from being able to make a hotel reservation and an airplane reservation and order me an Uber. All of a sudden, you’ve got an assistant.

2. The Prisoner's Dilemma

Yeah. You can just imagine—it’s on your phone. I think that’s pretty near-term. Some big companies that are very tech-forward already have 50% or more of customer support being done by AI, and that’s a $400 billion industry. What AI is great at is persuasion. That’s sales and customer support.

If you think about the functions of a company, they’re to make stuff, sell stuff, and then support the customers. Right now, maybe in late 2026, you’re going to be pretty good at 2 of them. I do think it’s going to have a big impact on media. I think robotics, which we talked about last time, are finally going to start to be real.

There’s an explosion in exciting robotics startups. I do still think that the main battle is going to be between Tesla’s Optimus and the Chinese, because it’s easy to make prototypes and hard to mass-produce them. But then it goes back to what Andrej Karpathy said about AI: it can automate anything that can be verified. Any function where there’s a right or wrong answer, a right or wrong outcome, can use reinforcement learning to make the AI really good at that.

Yeah. What are your favorite examples of that so far, or theoretically?

Does the model balance? They’ll be really good at making models. Do all the books globally reconcile? They’ll be really good at accounting because it’s double-entry bookkeeping. It has to balance. There’s a verifiable “You got it right or wrong.”

Did you make the sale or not? That’s very clear. It’s just like AlphaGo: did you win or did you lose? Did the customer convert or not? Did the customer ask for an escalation during customer support or not? Its most important functions are important because they can be verified, right?

So I think if all of this starts to happen, there’ll be an ROI on Blackwell, and then all this will continue. Then we’ll have Rubin, and that’ll be another big quantum of spend: Rubin, the MI450, and TPU v7. I think the most interesting question is, what are the economic returns to artificial superintelligence?

All of these companies in this great game have been in a prisoner’s dilemma. They’re terrified that if they slow down, they’re just gone forever. Their competitors don’t slow down, and it’s an existential risk.

Yeah.

Microsoft blinked for about 6 weeks earlier this year.

Yeah.

I think they would say they regret that.

Yeah.

But with Blackwell, and for sure with Reuben, economics are going to dominate the prisoner’s dilemma from a decision-making and spending perspective, just because the numbers are so big. This goes to the ROI-on-AI question, and the ROI on AI has empirically, factually, and unambiguously been positive.

I always find it strange that there’s any debate about this, because the largest bidders on GPUs are public companies. They report something called audited quarterly financials, and you can use those things to calculate something called a return on invested capital. If you do that calculation, the ROIC of the big public spenders on GPUs is higher than it was before they ramped spending.

You could say, “Well, part of that is opex savings.” At some level, that is part of what you expect the ROI to be from AI. Then you say, “Well, a lot of it is actually just applying GPUs”—moving the big recommender systems that power advertising and recommendation systems from CPUs to GPUs—and you’ve had massive efficiency gains. That’s why all the revenue growth at these companies has accelerated.

But so what? The ROI has been there. Every big internet company has people responsible for the revenue who are intensely annoyed at the amount of GPUs being given to the researchers. It’s a very linear equation: “If you give me more GPUs, I will drive more revenue. Give me those GPUs, we’ll have more revenue and more gross profit, and then we can spend money.” It’s this constant fight at every company.

3. The Bear Case: Edge AI

One of the factors in the prisoner’s dilemma is that everybody has this religious belief that we’re going to get to ASI. At the end of the day, what do they all want? Almost all of them want to live forever. They think that ASI is going to help them do that.

Right. Good return.

That’s a good return. [Laughter]

But we don’t know. And if, as humans, we have pushed the boundaries of physics, biology, and chemistry—the natural laws that govern the universe—I’m very curious about your favorite sort of throw-cold-water-on-this-stuff takes that you think about sometimes.

One would be the things that would cause this demand for compute to change, or even cause its trajectory to change. I’m curious what you think those things would be.

There’s 1 really obvious bear case, and it’s just edge AI, and it’s connected to the economic returns to ASI. In 3 years, on a bigger and bulkier phone to fit the amount of DRAM necessary—and the battery probably won’t last as long—you’ll probably be able to run a pruned-down version of something like Gemini 5, Grok 4, Grok 4.1, or ChatGPT at, I don’t know, 30 to 60 tokens per second.

Then that’s free. This is clearly Apple’s strategy: “We’re going to be a distributor of AI, and we’re going to make it privacy-safe and run it on the phone.” Then you can call one of the big models—the god models—in the cloud whenever you have a question.

If 30 to 60 tokens a second at a 115 IQ is good enough, I think that’s a bear case.

Other than just the scaling laws breaking?

But if we assume scaling laws continue, we now know they’re going to continue for pre-training for at least 1 more generation. We’re very early in the 2 new scaling laws for post-training, mid-training, RLVR—whatever people want to call it—and then test-time compute inference.

We’re so early in those, and we’re getting so much better at helping the models hold more and more context in their minds as they do this test-time compute. That’s really powerful, because everybody’s like, “Well, how’s the model going to know this?” Eventually, if you can hold enough context, you can just hold every Slack message and Outlook message and company manual in a company in your context.

Yeah.

And then you can compute the new task and compare it with your knowledge of the world—what you think, what the model thinks, all this context. It may be that really, really long context windows are the solution to a lot of the current limitations. That's enabled by some of these cool tricks, like KV-cache offload and stuff.

But I do think that, other than scaling laws slowing down and other than there being low economic returns to ASI, edge AI is by far the most plausible and scariest bear case.

I like to visualize different S-curves. I invested through the iPhone, and I love to see the visual of the iPhone models as it sort of went from this clunky, bricky thing up to what we have now. Obviously, we've sort of petered out on its form factor.

If you picture something similar for the frontier models themselves, does it feel like it's at a certain part of that natural technology-paradigm progression to you? If you're paying for Gemini Ultra or SuperGrok and you're getting the good AI, it's hard to see differences.

I have to go really deep on something like, “Do you think PCI Express or Ethernet is a better protocol for scale-up networking, and why? Show me the scientific papers.” If you shift between models and you ask a question like that, where you know it really deeply, you know the answers. Then you see differences.

I do play fantasy football. Winnings are donated to charity, but these new models are quite a bit better at helping with, “Who should I play?” They think in much more sophisticated ways.

By the way, if you're a historically good fantasy football player and you're having a bad season, this is why: you're not using it. I think we'll see that in more and more domains.

But I do think they are already at a level where, unless you are a true expert or just have an intellect that is beyond mine, it's hard to see the progress. That's why I do think we need to shift from getting more intelligent to more useful, unless more intelligence starts leading to these massive scientific breakthroughs and we're curing cancer in '26 and '27.

I don't know that we're going to be curing cancer, but I do think from an ROI S-curve, we need to hand off from intelligence to usefulness. Then usefulness will have to hand off to scientific breakthrough; that creates whole new industries.

What are the building blocks of usefulness in your mind?

Just being able to do things consistently and reliably. A lot of that is keeping all the context. There's a lot of context if someone wants to plan a trip for me. I've acquired these strange preferences. I follow that guy Andrew Huberman, so I like to have an east-facing balcony so I can get morning sun.

The AI has to remember that being on a plane with Starlink is important to me. Here are the resorts I've historically liked. Here are the kinds of areas I've liked. Here are the rooms that I would really like at each. That's a lot of context, and keeping all of that and weighting it is a hard problem.

So I think context windows are a big part of it. There's this METR task-evaluation thing—how long it can work, how long it can work for. You could think of that being related in some way to context, although not precisely.

Task length needs to keep expanding, because booking a restaurant is economically useful, but it's not that economically useful. Booking me an entire vacation and knowing the preferences of my parents, my sister, my niece, and my nephew, and what that means—that's a much harder problem. That's something that a human might spend 3 or 4 hours optimizing. If you can do that, that's amazing.

But then again, I just think it has to be good at sales and customer support relatively soon. I do think we're going to see an acceleration in the awesomeness of various products, just because engineers are using AI to make products better and faster.

We both invested in Fora, the hearing aid company, which is just absolutely remarkable. I think it's something I never would have thought of.

And we're going to see, I think, something like that in every vertical. That's AI being used for the most core function—

Yeah.

—of any company, which is designing the product. There are already lots of examples of AI being used to help manufacture the product and distribute it more efficiently, whether it's optimizing a supply chain or having a vision system watch a production line. I think a lot of stuff is happening.

The other thing I think is really interesting in this whole ROI part is that Fortune 500 companies are always the last to adopt a new technology. They're conservative; they have lots of regulations and lots of lawyers. Startups are always the first.

So let's think about the cloud, which was the last truly transformative new technology for enterprises: being able to have all of your compute in the cloud and use SaaS. It's always upgraded, it's always great, et cetera, and you can get it on every device. Those were dark days before the cloud.

The first AWS re:Invent, I think it was in 2013. Every startup on planet Earth ran on the cloud. The idea that you would buy your own server and storage box and router was ridiculous. That probably happened even earlier; it had probably already happened before the first re:Invent. The first big Fortune 500 companies started to standardize on it maybe 5 years later.

You see that with AI. I'm sure you've seen this in your startups, and I think one reason VCs are more broadly bullish on AI than public-market investors is that VCs see very real productivity gains. There are all these charts that, for a given level of revenue, a company today has significantly fewer employees than a company of 2 years ago.

The reason is AI is doing a lot of the sales, the support, and helping to make the product. ICONIQ has some charts. a16z, by the way—David George is a good friend, great guy—has his Model Busters thing. There's very clear data that this is happening. So people who have a lens into the world of venture see this.

I do think it was very important in the 3rd quarter. This is the first quarter where we had Fortune 500 companies outside of the tech industry give specific quantitative examples of AI-driven uplift. C.H. Robinson went up something like 20% on earnings. Should I tell people what C.H. Robinson does?

Let's just say a truck goes from Chicago to Denver. The trucker lives in Chicago, so it's going to go back from Denver to Chicago. There's an empty load. C.H. Robinson has all these relationships with truckers and trucking companies, and they match shippers' demand with that empty-load supply to make trucking more efficient.

They're a freight forwarder. There are actually lots of companies like this, but they're the biggest and most dominant. One of the most important things they do is quote price and availability.

Somebody, a customer, calls them up and says, “Hey, I urgently need 3 18-wheelers from Chicago to Denver.” In the past, they said it would take them 15 to 45 minutes, and they only quoted 60% of inbound requests. With AI, they're quoting 100% and doing it in seconds.

4. Meta, Open Source, and Model Depreciation

And so they printed a great quarter, and the stock went up 20%. It was because of AI-driven productivity that's impacting the revenue line, the cost line—everything.

I was actually very worried about the idea that we might have this Blackwell ROI air gap because we're spending so much money on Blackwell. Those Blackwells are being used for training, and there's no ROI on training. Training is making the model; the ROI comes from inference.

I was really worried that maybe we'd have this 3-quarter period where capex is unimaginably high, those Blackwells are only being used for training, margins are staying flat, and AI is going up.

Those Blackwells are only being used for training, margins are staying flat, and AI is going up.

Yeah, exactly. And so ROIC goes down. You can see that with Meta. Because Meta has not been able to make a frontier model, Meta printed a quarter where ROIC declined, and that was not good for the stock. I was really worried about this.

I do think those data points are important in suggesting that maybe we'll be able to navigate this potential air gap in ROIC.

Yeah, it makes me wonder about this market. I'm like everybody else: it's the 10 companies at the top that have all the market cap and get more than all of the attention. There are 490 other companies in the 500. You studied those too.

What do you think about that group? What's interesting to you about the group that now nobody seems to talk about and no one really seems to care about because they haven't driven returns and they're a smaller percentage of the overall market?

Well, I think people are going to start to care if more and more companies print these C.H. Robinson-like quarters. Companies that have historically been really well run—the reason they have a long track record of success is that you cannot succeed without using technology well.

So if you have a kind of internal culture of experimentation and innovation, I think you will do well with AI. I'd bet on the best investment banks to be earlier and better adopters of AI than maybe some of the trailing banks. Sometimes the past is prologue.

One thing that I have a strong opinion on: all these VCs are setting up these holding companies, and we're going to use AI to make traditional businesses better. They're really smart VCs, and they have great track records.

But that's what private equity has been doing for 50 years. You're just not going to beat private equity at their game. What Vista did in the early days, right?

Yeah. Private equity's maybe had a little bit of a tough run. Multiples have gone up, so private assets are more expensive. The cost of financing has gone up, and it's tough to take a company public because the public valuation is 30% lower than the private valuation.

Private equity's had a tough run. I actually think these private equity firms are going to be pretty good at systematically applying AI. We haven't spent much time talking about Meta, Anthropic, or OpenAI, and I'd love just your impression of everything that's going on on this infrastructure side that we talked about.

These are 3 really important players in this grand battle, this grand game. How does all of this development that we've discussed so far impact those players specifically, do you think?

The first thing, let me just say about frontier models broadly.

Yeah.

In 2023 and 2024, I was fond of quoting Eric Vishria, our friend, a brilliant man. Eric would always say, “Foundation models are the fastest-depreciating assets in history.” I would say he was 90% right. I modified the statement: “Foundation models without unique data and internet-scale distribution are the fastest-depreciating assets in history.” But reasoning fundamentally changed that in a really profound way.

There was a loop, a flywheel, to quote Jeff Bezos, that was at the heart of every great internet company. You made a good product, you got users, and those users using the product generated data that could be fed back into the product to make it better. That flywheel has been spinning at Netflix, Amazon, Meta, and Google for over a decade, and that's an incredibly powerful flywheel.

That's why those internet businesses were so tough to compete with. It's why they had increasing returns to scale. Everybody talks about network effects much more, and network effects were important for social networks. I don't know to what extent Meta is a social network anymore. It's more like content distribution, but they just had increasing returns to scale because of that flywheel.

That dynamic was not present in the pre-reasoning world of AI. You pre-trained a model, let it out in the world, and it was what it was. It was actually pretty hard. They would do RLHF—reinforcement learning with human feedback—and try to make the base model better. Maybe you'd get a sense from Twitter vibes that people didn't like this, and so you'd tweak it. There were the little up and down arrows, but it was actually pretty hard to feed that back into the model.

With reasoning, it's early, but that flywheel has started to spin, and that is really profound for these frontier labs.

So, 1, reasoning fundamentally changed the industry dynamics of frontier labs. Just explain why specifically that is. What is going on?

Because if a lot of people are asking a similar question and they're consistently either liking or not liking the answer, then you can use that as a verifiable reward. That's a good outcome. Then you can feed those good answers back into the model. We're very early in this flywheel spinning—it's hard to do now, but you can see it beginning to spin.

5. Geopolitics and Rare Earths

This is important fact number 1 for all of those dynamics. Second, I think it's really important that Meta—Mark Zuckerberg, at the beginning of this year, in January, said, “I anticipate—I’m highly confident”—I’m going to get the quote wrong—“that at some point in 2025, we're going to have the best and most performant AI.” I don't know if he's in the top hundred.

Okay.

So he was as wrong as it was possible to be. I think that is a really important fact because it suggests that what these 4 companies have done is really hard to do. Meta threw a lot of money at it, and they failed. Yann LeCun had to leave. They had the famous billion-dollar offers for AI researchers.

By the way, Microsoft also failed. They did not make such an unequivocal prediction, but they bought Inflection AI. There were a lot of comments from them that they anticipated their internal models quickly getting better and that they were going to run more and more of Copilot on their internal models. Amazon bought a company called Adept, and they have models called Nova. No, I don't think they're in the top 20.

So clearly, it's much harder to do than people thought a year ago. There are many, many reasons for that. It's actually really hard to keep a big cluster of GPUs coherent. A lot of these companies were used to running their infrastructure to optimize for cost rather than performance.

Yeah.

And if you're running—because of the laws of physics, maybe the most anybody can get is 200,000 or 300,000 Blackwells coherent. We'll see. But if you have 30% uptime on that cluster and you're competing with somebody who has 90% uptime, you're not even competing.

So, 1, there's a huge spectrum in how well people run GPUs. 2, then I think there are these AI researchers. They like to talk about taste, and I find it very funny. “Why do you make so much money?” “I have very good taste.” What taste means is that you have a good intuitive sense for the experiments to perform.

This is why you pay people a lot of money, because it actually turns out that as these models get bigger, you can no longer run an experiment on a 1,000-GPU cluster and replicate it on 100,000 GPUs. You need to run that experiment on 50,000 GPUs, and maybe it takes days. So there's a very high opportunity cost, and you have to have a really good team that can make the right decisions about which experiments to run on this.

Then you need to do all the reinforcement learning during post-training well, and the test-time compute well. It's complicated. It's really hard to do, and everybody thinks it's easy, but all those things are really hard.

I used to have this saying—I was a retail analyst long ago: “Pick any vertical in America. If you can just run 1,000 stores and have them clean, well-lit, stocked with relevant goods at good prices, and staffed by friendly employees who are not stealing from you, you're going to be a $20 billion company, a $30 billion company.” Like 15 companies have been able to do that. It's really hard.

It's the same thing. Doing all of these things well, and then reasoning with this flywheel, is beginning to create more separation.

What's even more important is that every one of those labs—xAI, Gemini, OpenAI, and Anthropic—has a more advanced checkpoint internally of the model. A checkpoint is just—you’re continuously working on these models, and then you release a checkpoint. The reason these models get better is that the one they're using internally is better, and they're using that model to train the next model.

If you do not have that latest checkpoint, you're behind. It's getting really hard to catch up. Chinese open source is a gift from God to Meta because you can use Chinese open source to try things, and that can be your checkpoint. You can use that as a way to bootstrap this, and I'm sure that's what they're trying to do, and everybody else.

The big, giant swing factor, I think, is that China has made a terrible mistake with this rare-earths thing. China has the Huawei Ascend, and it's a decent chip. Versus something like the deprecated Hopper, it looks okay, so they're trying to force Chinese open source to use their Chinese chips, their domestically designed chips.

The problem is that Blackwell is going to come out now, and the gap between these American frontier labs and Chinese open source is going to blow out because of Blackwell. Actually, DeepSeek, in its most recent technical paper, DeepSeek-V3.2, said one of the reasons it struggles to compete with the American frontier labs is that it doesn't have enough compute.

That was their very politically correct, still a little bit risky way of saying, because China said, “We don't want the Blackwells,” right? And they're saying, “Guys, that might be a big mistake. That might be a big mistake.”

If you just play this out, these 4 American labs are going to start to widen their gap versus Chinese open source, which then makes it harder for anyone else to catch up because that gap is growing. So you can't use Chinese open source to bootstrap.

Geopolitically, China thought it had the leverage. It's going to realize, “Oh, whoopsie-daisy. We do need the Blackwells.” Unfortunately for them, they'll probably realize that in late 2026. At that point, there's an enormous effort underway. DARPA has—there are all sorts of really cool DARPA and DoD programs to incentivize really clever technological solutions for rare earths, like using enzymes to refine them.

There are all sorts of really cool things happening. There are a lot of rare-earth deposits in countries that are very friendly to America, that don't mind actually refining them in the traditional way.

So, I think rare earths are going to be solved way faster than anyone thinks. They’re obviously not that rare; they’re just misnamed. They’re rare because they’re really messy to refine.

Geopolitically, I actually think Blackwell is pretty significant, and it’s going to give America a lot of leverage as this gap widens. Going back to the dynamics between these companies, xAI will be out with the first Blackwell model, and then they’ll probably be the first ones using Blackwell for inference at scale. I think that’s an important moment for them.

By the way, it is funny: if you go on OpenRouter, you can just look. It’s whatever it is—1% of API tokens—but it’s an indication. They process 1.35 trillion tokens. Google did 800 or 900 billion over the last 7 days or the last month, and Anthropic was at 700 billion. xAI is doing really, really well, and the model is fantastic. I highly recommend it. You’ll see xAI come out with this, and OpenAI will come out faster.

OpenAI’s issue that they’re trying to solve with Stargate is that they pay a margin to people for compute. Maybe the people who run their compute aren’t the best at running GPUs. They are a high-cost producer of tokens, and I think this explains a lot of their code red recently.

Patrick O’Shaughnessy

Yeah. Well, just the $1.4 trillion in spending commitments.

And I think that was just, “Hey, they know they’re going to need to raise a lot of money,” particularly if Google keeps its current strategy of sucking the economic oxygen out of the room. You go from $1.4 trillion to rough vibes, to code red, pretty fast. The reason they have a code red is because of all these dynamics.

6. Data Centers in Space

So then they’ll come out with a model, but they will not have fixed their per-token cost disadvantage relative to both xAI and Google, and almost Anthropic at that point. Anthropic is a good company. They’re burning dramatically less cash than OpenAI and growing faster, so I think you have to give Anthropic a lot of credit. A lot of that is their relationship with Google and Amazon for the TPUs and the Trainiums.

You can look at Dario and Jensen. Maybe there have been a few public comments made between them.

Jousting.

A little bit of jousting. Well, Anthropic just signed a $5 billion deal with NVIDIA.

That is because Dario is a smart man, and he understands these dynamics about Blackwell and Reuben relative to TPU. NVIDIA now goes from having 2 of the fighters—xAI and OpenAI—to 3 fighters. That helps in this NVIDIA-versus-Google battle.

If Meta can catch up, that’s really important. I’m sure NVIDIA is doing whatever they can to help Meta: “Let us know you’re running those GPUs this way. Maybe we should twist the screw this way or turn the dial that way.” If Blackwell comes back to China, which it seems like it probably will, that will also be very good, because then Chinese open source will be back.

What are the other poles of things? One pole would be the other breakthroughs that you have your mind on in the data center that aren’t chips, that we’ve talked about before, as one example.

I think the most important thing that’s going to happen in the world in the next 3 to 4 years is data centers in space. This has really profound implications for everyone building a power plant or a data center on Planet Earth. There is a giant gold rush into this.

I haven’t heard anything about this, so please explain.

Everybody thinks AI is risky, but they say, “I’m going to build a data center. I’m going to build a power plant that’s going to power a data center. We will need that.” But if you think about it from first principles, data centers should be in space.

What are the fundamental inputs to running a data center? There’s power, there’s cooling, and then there are the chips.

That’s the total, if you think about it from a total-cost perspective.

Yeah.

And just the inputs to making the tokens come out of the magic machines.

Yeah. So in space, you can keep a satellite in the sun 24 hours a day, and the sun is 30% more intense. You can keep it in the sun constantly. If the sun’s here, you can have the satellite always catching the light.

Catching the light.

Catching the light. The sun is 30% more intense, and this results in 6 times more irradiance in outer space than on Planet Earth. So you’re getting a lot of solar energy. Point 1.

Point 2: because you’re in the sun 24 hours a day, you don’t need a battery, and this is a giant percentage of the cost. The lowest-cost energy available in our solar system is solar energy in space.

For cooling in one of these racks, a majority of the mass and the weight is cooling. The cooling in these data centers is incredibly complicated—the HVAC, the CDUs, the liquid cooling.

It’s fucking gold.

And it’s as close to absolute zero as you can get. So all that goes away, and that is a vast amount of cost.

Let’s think about how these might work. Maybe each satellite is a rack. That’s one way to think of it. Maybe some people make bigger satellites that are 3 racks. How are you going to connect those racks? In a data center, the racks are connected over a certain distance with fiber optics. That just means a laser going through a cable. The only thing faster than a laser going through a fiber-optic cable is a laser going through absolute vacuum.

If you can link these satellites in space together using lasers, you actually have a faster and more coherent network than in a data center on Earth. For training, that’s going to take a long time—

Yeah, just because it’s so big.

But for inference—and I think even training will eventually happen—let’s think about the user experience. When I asked Grok about you and it gave the nice answer, a radio wave traveled from my cell phone to a cell tower. Then it hit the base station, went into a fiber-optic cable, and went to some sort of metro aggregation facility in New York, probably within 10 blocks of here. There’s a small little metro router that routed those packets to a big xAI data center somewhere.

The computation was done, and it came back over the same path. If the satellites can communicate directly with the phone, and Starlink has demonstrated direct-to-cell capability, you just go boom, boom. It’s a much better, lower-cost user experience. So in every way, data centers in space, from a first-principles perspective, are superior to data centers on Earth.

7. Power Constraints as a Governor

So if we could teleport that into existence, I understand that portion. What are the frictions to that? Why will that not happen? Is it launch cost? Is it launch availability?

We need a lot of SpaceX Starships. The Starships are the only ones that can economically make that happen.

We need a lot of those Starships.

Maybe China or Russia will be able to land a rocket. Blue Origin just landed a booster. It’s an entirely new and different way to think about SpaceX.

It’s interesting that Elon posted yesterday, or said in an interview, that Tesla, SpaceX, and xAI were converging. They really are. xAI will be the intelligence module for Optimus, made by Tesla. Tesla Vision has its perception system, and SpaceX will have the data centers in space that will power a lot of the AI, presumably, for xAI, Tesla, Optimus, and a lot of other companies.

It’s interesting the way that they’re converging, and each one is creating competitive advantage for the other. If you’re xAI, it’s really nice that you have this built-in relationship with Optimus. Now Tesla is a public company, so I cannot imagine the level of vetting that will go into that intercompany agreement. Then you have a big advantage with these data centers in space.

It’s also nice if you’re xAI that you have 2 companies with a lot of customers who you can use to help build your customer-support agents and your customer-sales agents, with built-in customers. They really are all converging in a neat way. I do think it’s going to be a big moment when that first Blackwell model comes out from xAI next year.

Hmm. If I go to the other end of the spectrum and I think about something that seems to have been historically endemic to the human economic experience—that shortages are always followed by gluts in capital cycles—what if, in this case, the shortage is compute?

Mark Chen is now on the record as saying they would consume 10 times as much compute if you gave it to them in a couple of weeks. There seems to still be a massive shortage of compute, which is all the stuff we’ve talked about today. But there also just seems to be this iron law of history that gluts follow shortages.

What do you think about that concept as it relates to this technology? Could this technology be a glut?

Yeah. You know, AI is fundamentally different from software in that every time you use AI, it takes compute in a way that traditional software just did not. I think it is true that every one of these companies could consume 10x more compute. What would happen is that the $200 tier would get a lot better, and the free tier would get like the $200 tier. Google has started to monetize AI Mode with ads.

And I think that will give everyone else permission to introduce ads into the free mode, and then that is going to be an important source of ROI. It seems like OpenAI is tailor-made to—

Yeah, absolutely. All of them. And actions, like, “Hey—”

Here are your 3 vacations. Would you like me to book one? And then they're for sure going to collect a commission.

Yeah. You know, there are many ways you can make money. I think we went into great detail on maybe a prior podcast about how inventory dynamics made these inventory cycles inevitable in semis. The iron law of semis is just that customer buffer inventories have to equal lead times, and that's why you got these inventory cycles historically.

We haven't seen a true capacity cycle in semis, maybe arguably since the late 1990s. And that's because Taiwan Semi has been so good at aggregating and smoothing supply. A big problem in the world right now is that Taiwan Semi is not expanding capacity as fast as its customers want. I think this is actually a pretty big mistake they're in the process of making, just because you do have Intel with these fabs, and they're not as good, and it's really hard to work with their PDK.

But now you have this guy Lip-Bu Tan, who's a really good executive and really understands that business. By the way, I think Pat Gelsinger was also a good executive, and he put Intel on the only strategy that could result in success. I actually think it's shameful that the Intel board fired him when it did. But Lip-Bu Tan is a good executive, and now he's reaping the benefits of Pat's strategy. Intel has all these empty fabs, and eventually, given the shortages we have of compute, those fabs are going to be filled.

So I think Taiwan Semi is in the process of making a mistake, but they're just so paranoid about an overbuild.

Yeah.

And they're so skeptical. You know, they're the guys who met with Sam Altman and laughed and said, “He's a podcast bro. He has no idea what he's talking about.” They're terrified of an overbuild. So it may be that Taiwan Semi, single-handedly, because they're cautious, is the brakes on the bubble—the governor.

And I think governors are good. It's good that power is a governor; it's good that Taiwan Semi is a governor. If Taiwan Semi opens up at the same time that data centers in space relieve all power constraints—but that's, I don't know, 5 or 6 years away, when data centers in space are a majority of deployed megawatts—yeah, I think you get an overbuild really fast. But we have these 2 really powerful natural governors, and I think that's good. Smoother and longer is good.

We haven't talked about power, other than alluding to it through the space thing. We haven't talked about power very much. Power was the most uninteresting topic because there was demand, and nothing really changed for a really, really long time. All of a sudden, we're trying to figure out how to get gigawatts here, there, and everywhere. How do you think about it? Are you interested in power?

I'm very interested. I do feel lucky that, in a prior life, I was the sector leader for the telecom and utilities team.

Okay.

I do have some base level of knowledge. One, having watts as a constraint is really good for the most advanced compute players because if watts are the constraint, the price you pay for compute is irrelevant. The TCO of your compute is absolutely irrelevant because if you could get 3x, 4x, or 5x more tokens per watt, that is literally 3x, 4x, or 5x more revenue.

If you're going to build an advanced data center, an advanced data center costs $50 billion. A data center with your ASIC maybe costs $35 billion, but if that $50 billion data center pumps out $25 billion of revenue and your ASIC data center at $35 billion is only pumping out $8 billion, well, you're pretty bummed. It's good for all of the most advanced technologies in the data center, which is exciting to me as an investor.

As long as power is a governor, the best products are going to win irrespective of price and have crazy pricing power. I think that's the first implication that's really important to me. Second, the only solutions to this are— we just can't build nuclear fast enough in America. As much as we would love to build nuclear quickly, we just can't. We just can't.

Yeah.

It's just too hard. NEPA, all these rules—it's just too hard. A rare ant that we could move, and it could be in a better environment, can totally delay the construction of a nuclear power plant. One ant. It's crazy, actually. Humans need to come first. We need to have a human-centric view of the world.

The solutions are natural gas and solar. The great thing about these AI data centers is that, apart from the ones that you're going to do inference on, you can locate them anywhere. I think we're going to see all this activity in Abilene, because it's in the middle of a big natural gas basin, and we have a lot of natural gas in America because of fracking. I think we're going to have a lot of natural gas for a long time. We ramp production really fast.

I think this is going to be solved. You're going to have power plants fed by gas or solar. I think that's the solution. And already, all these turbine manufacturers were reluctant to expand capacity. Caterpillar just said, “We're going to increase capacity by 75% over the next few years.” The system on the power side is beginning to respond.

One of the reasons that I always love talking to you is that you do as much in the top 10 companies in the world as you do looking at brand-new companies with entrepreneurs who are 25 years old trying to do something amazing. And so you have this very broad sense of what's going on.

If I think about that second category of young, enterprising technologists who are now the first generation of AI-native entrepreneurs, what are you seeing in that group that's notable, surprising, or interesting?

These young CEOs are just so impressive in all ways, and they get more polished faster. I think the reason is that they're talking to the AI. “How should I deal with pitching this investor? I'm meeting with Patrick O'Shaughnessy. What do you think the best ways I should pitch him are?” And it works.

“Do deep research.” And it's good. “Hey, I have this difficult HR situation. How would you handle it?” That's correct. And it's good at that. “We're struggling to sell our product. What changes would you make?” It's really good at all of that today.

And that goes to these VCs seeing massive AI productivity in all their companies. It's because their companies are full of these 23-, 24-, or even younger AI natives. I've been so impressed with young investment talent, and it's just part of it. Your podcast is part of that. Knowledge, and very specific knowledge, has become so accessible through podcasts and the internet.

Impressive young people come in, and they're just—I feel like they're where I was as an investor in my early 30s, and they're 22. I'm like, “Oh my God, I have to run so fast to keep up.” These kids who are growing up AI-native are just proficient with it in a way that I am trying really hard to become.

Can we talk about semiconductor VC specifically and what is interesting in that universe?

The one thing I think is so cool about it and so underappreciated is that your average semiconductor venture founder is about 50 years old.

Okay.

Jensen Huang, and what's happened with NVIDIA and the market cap of NVIDIA, have single-handedly ignited semiconductor venture. But the way it's ignited, it's ignited in an awesome way that's really good for NVIDIA and Google and everyone.

Let's just say you were the best DSP architect in the world and had done that for the last 20 years, every 2 years, because that's what you have to do in semiconductors. It's like every 2 years, you have to win—run a race. If you won the last race, you start a foot ahead, and over time those compound and make each race easier to win.

Maybe that person and his team—maybe he's the head of networking at a big public company—are making a lot of money and have a good life. Then, because he sees these outcomes and the size of the markets in the data center, he's like, “Wow, why don't I just go start my own company?”

But the reason that's important is that there are thousands of parts in a Blackwell rack, and there are thousands of parts in a TPU rack. In the Blackwell rack, maybe NVIDIA makes, I don't know, 200 or 300 of those parts.

And the same thing in an AMD rack. They need all of those other parts to accelerate with them.

So they couldn't go to this one-year cadence if the rest of everything was not keeping up with them.

The fact that semiconductor venture has come back with a vengeance is really important. Silicon Valley stopped being Silicon Valley long ago. My little firm may have done more semiconductor deals in the last 7 years than the top 10 VCs combined, but that's really important because now you have an ecosystem of companies who can keep up.

That ecosystem of venture companies is putting pressure on the public companies that also need to be part of this if we're going to go to this annual cadence, which is so hard. It's one reason I'm really skeptical of these AI startups that don't already have some degree of success. I do think that's a super important dynamic and one that's absolutely foundational and necessary for all of this to happen, because not even NVIDIA can do it alone.

AMD can't do it alone. Google can't do it alone. You need the people who make the transceivers. You need the people who make the wires, who make the backplanes, who make every—who make the lasers. They all have to accelerate with you.

One thing that I think is very cool about AI as an investor is that it's the first time when, at every level of the stack that I look at, at least, the most important competitors are public and private. NVIDIA has very important private competitors. Broadcom has important private competitors. Marvell has important private competitors. Lumentum and Coherent—all these companies have important private competitors.

There's even a wave of innovation in memory, which is really exciting to see, because memory is such a gating factor. By the way, something that could slow all this down and be a natural governor is if we get our first true DRAM cycle since the late '90s.

Say more about what that means.

If a DRAM wafer is valued like a 5-carat diamond in the '90s, when you had these true capacity cycles before Taiwan Semi kind of smoothed everything out and DRAM became more of an oligopoly, you would have these crazy shortages where the price would just go 10x—things that are unimaginable relative to the last 25 years.

A giant DRAM cycle—a good DRAM cycle—is when the price stops going down. An epic cycle is maybe it goes up 30%, 40%, or 50%, whatever it is. But if it starts to go up by Xs instead of percentages, that's a whole different game. By the way, we should talk about SaaS.

Yeah, let's talk about it. What do you think's going to happen?

Application SaaS companies are making the exact same mistake that brick-and-mortar retailers did with e-commerce. Brick-and-mortar retailers, particularly after the telecom bubble crashed, looked at Amazon and said, “Oh, it's losing money. E-commerce is going to be a low-margin business. How can it ever be more efficient as a business?”

Right now, our customers pay to transport themselves to the store, and then they pay to transport the goods home. How could it ever be more efficient if we're sending shipments out to individual customers? Amazon's vision, of course, was that eventually they were just going to go down a street and drop off a package at every house.

They did not invest in e-commerce. They clearly saw customer demand for it, but they did not like the margin structure of e-commerce. That is the fundamental reason that essentially every brick-and-mortar retailer was really slow to invest in e-commerce.

Now here we are, and Amazon has higher margins in its North American retail business than a lot of mass-market retailers. Margins can change, and if there's a fundamentally transformative new technology that customers are demanding, it's always a mistake not to embrace it.

8. The SaaS Mistake

That's exactly what the SaaS companies are doing. They have their 70%, 80%, and 90% gross margins, and they are reluctant to accept AI gross margins. The very nature of AI is that software you write once is written very efficiently, and then you can distribute it broadly at very low cost. That's why it was a great business.

AI is the exact opposite, where you have to recompute the answer every time. A good AI company might have gross margins of 40%. The crazy thing is that, because of those efficiency gains, they're generating cash way earlier than SaaS companies did historically.

They're generating cash earlier not because they have high gross margins, but because they have very few human employees. It's tragic to watch all of these companies. If you want to have an agent, it's never going to succeed if you're not willing to run it at a sub-35% gross margin, because that's what the AI natives are running it at.

Yeah.

Maybe they're running it at 40%. So if you are trying to preserve an 80% gross-margin structure, you are guaranteeing that you will not succeed at AI.

Absolute guarantee.

And this is so crazy to me because, one, we have an existence proof for software investors being willing to tolerate gross-margin pressure as long as gross-profit dollars are okay. It's called the cloud.

People don't remember, but when Adobe converted from on-premise to a SaaS model, not only did its margins implode, its revenues actually imploded too, because it went from charging up front to charging over a period of years.

Microsoft was less dramatic, but it was a tough stock in the early days of the cloud transition because investors were like, “Oh my God, you're an 80% gross-margin business, and the cloud is in the 50s.” And Microsoft was like, “Well, it's going to be gross-profit-dollar accretive. It probably will improve those margins over time.”

Microsoft bought GitHub and used GitHub as a distribution channel for Copilot for coding, which has become a giant business. It runs at much lower gross margins, for sure, but there are so many SaaS companies. I can't think of a single application SaaS company that could not be running a successful agent strategy.

They have a giant advantage over these AI natives in that they have a cash-generative business. I think there is room for someone to be a new kind of activist, or constructive activist, and just go to SaaS companies and say, “Stop being so dumb.”

All you have to say is, “Here are my AI revenues, and here are my AI gross margins. It's real AI because it has low gross margins. I'm going to show you that, and here's a venture competitor over here that's losing a lot of money. So maybe I'll actually take my gross margins to 0% for a while, but I have this business that the venture-funded company doesn't have.”

This is such an obvious playbook that you can run at Salesforce, ServiceNow, HubSpot, GitLab, and Atlassian. All of them could run this. The way those companies could or should think about using agents is just to ask the question: “What are the core functions we do for the customer now? How can we further automate that with agents effectively?”

100%. Just like if you're in CRM, what do our customers do? They talk to their customers.

Yeah, we're customer relationship management software, and we do some customer support, too. So make an agent that can do that, right? Sell that at 10% to 20%, and let that agent access all the data you have, right?

Because what's happening right now is another agent made by someone else is accessing your systems to do this job, pulling the data into their system, and then you will eventually be turned off. It's just crazy, and it's just because they want to preserve their 80% gross margins.

This is a life-or-death decision, and essentially everyone except Microsoft is failing it. To quote that memo from that Nokia guy long ago, their platforms are burning.

Burning platform. Yeah.

Yeah. There's a really nice platform right over there, and you can just hop to it. Then you can put out the fire in your platform that's on fire. Now you've got 2 platforms, and it's great.

Your data centers and space thing makes me wonder if there are other less-discussed, off-the-wall things you're thinking about in the markets in general that we haven't talked about.

9. Nuclear and Quantum

It does feel like, since 2020 kicked off and 2022 punctured it, we've had a series of rolling bubbles. In 2020, there was a bubble in EV startups that were not Tesla, and that's for sure a bubble. They all went down 99%.

There was also a bubble in more speculative stocks. Then we had the meme stocks, GameStop. Now it feels like the rolling bubble is in nuclear and quantum.

These are fusion and SMR. It would be a transformative technology. It's amazing. But sadly, from my perspective, none of the public ways you can invest in this are really good expressions of this theme, are likely to succeed, or have any real fundamental support.

The same thing is true with quantum. I've been looking at quantum for 10 years. We have a really good understanding of quantum, and the public quantum companies, again, are not the leaders.

From my perspective, the leaders in quantum would be Google, IBM, and Quantinuum. The public ways you can invest in this theme, which is probably exciting, are not the best. So you have 2 really clear bubbles.

I also think quantum supremacy is very misunderstood. People hear it, and I think they mean that quantum computers are going to be better than classical computers at everything. With quantum, you can do some calculations that classical computers cannot do.

The thought that I have had, which is maybe less related to markets than just AI, is that I have been fascinated that for the last 2 years, whatever AI needs to keep growing and advancing, it gets.

Have you ever seen public opinion change so fast in the United States on any issue as it has on nuclear power?

It just happened like that.

Like that. Why did that happen right when AI needed it to happen? Now we're running up against the boundaries of power on Earth. All of a sudden, we're talking about data centers in space.

It's just a little strange to me that whenever there is something—a bottleneck—

A bottleneck—

A bottleneck that might slow it down, everything accelerates. Reuben is going to be such an easy, seamless transition relative to Blackwell, and Reuben's a great chip. Then you have AMD getting into the game with the MI450. It's just whatever AI needs, it gets.

You're a deep reader of sci-fi, so—

Yeah, exactly.

You're making me think of Kevin Kelly's great book, What Technology Wants. He calls it the technium, the overall mass of technology that is supplied by humans to grow more powerful.

Absolutely. It just wants to grow more and more powerful. Now we're going into an insane state.

I have a selfish closing question. Speaking of young people, my kids are 12 and 10, but especially my older son is developing an interest in what I do, which I think is quite natural. I'm going to try to start asking my friends who are the most passionate about entrepreneurship and investing why they are so passionate about it, and what about it is so interesting and life-giving to them. How would you pitch what you've done, the career you've built, and this part of the world to a young person who's interested in it?

I do believe, at some level, investing is the search for truth. If you find truth first, and you're right about it being a truth, that's how you generate alpha. It has to be a truth that other people have not yet seen. You're searching for hidden truths.

The earliest thing I can remember is being interested in history. I loved looking at books with pictures of the Phoenicians, the Egyptians, the Greeks, the Romans, and pyramids. I loved history.

I vividly remember, in I think the 2nd grade, my dad drove me to school every day, and we went through the whole history of World War II in 1 year. I loved that. Then that translated into a real interest in current events very early.

As a pretty young person—I don't know if it was 7th, 8th, or 9th grade—I was reading The New York Times and The Washington Post. I would get so excited when the mail came because it meant that maybe there was an Economist, Newsweek, Time, or U.S. News. I was really into current events because current events is applied history: watching history happen and thinking about what might happen next.

I didn't know anything about investing. My parents were both attorneys. Anytime I won an argument, I was highly rewarded. If I could make a reasonable argument for why I should stay up late, my parents would be so proud, and they would let me stay up late. But I had to beat them.

I was just going through life. I really loved to ski, and I loved rock climbing. I went to college, and rock climbing was by far the most important thing in my life. I dedicated myself to it completely. I did all my homework at the gym. I got to the rock-climbing gym at 7 a.m. and would skip a lot of classes to stay there. I would do my homework on a big bouldering mat.

Every weekend, I went and climbed somewhere with the Dartmouth Mountaineering Club. As part of that, on climbing trips, maybe we'd play poker. The movie came out while I was in college, and we started playing poker. I also liked to play chess.

10. Gavin’s Investing Origins

I was never that good at chess or poker. I never really dedicated myself to either. My plan, after 2 or 3 years of college, was that I was going to leave. I was a ski bum at Alta in college. I was a housekeeper. I've cleaned a lot of toilets.

It was shocking to me how people treated me, and it has permanently affected how I treat other people. You'd be cleaning somebody's room, and they'd be in there reading the same book as you. You'd say, “Oh, that's a great book. I'm about where you are,” and they'd look at you like you were a space alien. Then they would get even more shocked that you read. It had a big impact on how I've treated everyone since then.

I was going to be a ski bum in the winters. I was going to work on a river in the summers, and that was how I was going to support myself. Then I was going to climb in the shoulder seasons. I was going to try to be a wildlife photographer and write the next great American novel.

I can't believe I never knew this.

That was my plan. This was my plan of record. I was really lucky. My parents were very supportive of everything I wanted to do. My parents had very strict parents, so of course they were extremely permissive with me. I'll probably end up being a strict parent. The cycle continues.

My parents were lawyers, and they had done reasonably well. They both grew up in very economically disadvantaged circumstances. My dad talks about how he remembers every person who bought him a beer because he couldn't afford a beer. He worked the whole way through college and was there on a scholarship. He had 1 pair of shoes all through high school.

They were very supportive of this plan, and I had been very lucky. They sent me to college, and I didn't have to pay for it. They paid for my college education. They said, “You know, Gavin, we think this plan of being a ski bum and river-rafting guide, being a wildlife photographer, climbing in the shoulder seasons, and trying to write a novel sounds like a great plan. But we've never asked you for anything. We haven't encouraged you to study anything. We've supported you in everything you've wanted to do. Will you please get 1 professional internship—just 1? We don't care what it is.”

The only internship I could get at the end of my sophomore summer at Dartmouth was with Donaldson, Lufkin & Jenrette, or DLJ. It was in the private wealth management division, and I worked for the guy who ran the office. Whenever DLJ published a research report, I would go through it and look at which of his clients owned that stock. Then I would mail the research to the clients.

That day, we wrote on General Electric, so I needed to mail the GE report to 30 people. Then I needed to email the Cisco report to 20 people. I started reading the reports, and I thought, “Oh my God, this is the most interesting thing imaginable.”

I conceptualized investing as a game of skill and chance, kind of like poker. There is obviously chance in investing. If you're an investor in a company and a meteor hits its headquarters, that's bad luck, but you own that outcome. There is irreducible chance, but there's skill, too. That really appealed to me.

The way you got an edge in this greatest game of skill and chance imaginable was by having the most thorough knowledge possible of history. You intersected that with the most accurate understanding of current events in the world to form a differential opinion on what was going to happen next in this game of skill and chance. Which stock is mispriced in the Perry Mutual system? That is the stock market.

That was day 3. I went to the bookstore and bought the books they had, which were Peter Lynch's books. I read those books in 2 days. I'm a very fast reader. Then I read all these books about Warren Buffett. Then I read Market Wizards. Then I read Warren Buffett's letters to his shareholders. This was during my internship. Then I read Warren Buffett's letters to his shareholders again.

Then I taught myself accounting. There's this great book, Why Stocks Go Up and Down. Then I went back to school. I changed my majors from English and history to history and economics. I never looked back.

Investing consumed me, although I continued to really focus on climbing. I would be in the gym and print out everything that the people at The Motley Fool wrote. They were early to talking about return on invested capital, and incremental ROIC is a really important indicator. I would read it and underline it. I'd read books, and then I'd read The Wall Street Journal.

Eventually, there was a computer terminal set up near the gym. I'd go there and read news about stocks. It was the most important thing in my life. I barely kept my grades up.

That's how I got into it: history, current events, skill and chance. I'm a competitive person, and I've actually never been good at anything else.

Okay, I got picked last for every sports team. I love to ski. I've literally spent a small fortune on private skiing lessons. I'm not that good of a skier. I like to play ping-pong, and all my friends could beat me. [laughter] I tried to get really good at chess, and this was before you actually had to play the games, before it was easy to do it on the phone. My goal was to beat one of the people. I'm sure there's a park somewhere.

It's literally right there. The famous one is right there.

Okay. Well, there's one in Cambridge, and I wanted to beat one of them. I never beat one of them. I've never been good at anything. I thought I would be good at this. The idea of being good at something other than taking a competitive test was very appealing to me. I think that's been a really important thing, too. To this day, this is the only thing I've been vaguely competitive at. I'd love to be good at something else. I'm just not. [laughter]

I think I'm going to start asking this question of everybody. The ongoing education of a person may be an amazing place to close. I love talking about everything so much.

This is great, man. Thank you. Thank you. Thank you.