Have you found anybody?

No, and I ask everyone.

My standard question is, "Can you tell me one quantitative data point in your business that's getting worse? Just one." That's my standard question, and it's at least in July and August, I haven't been able to find a single person.

Now, if we're being honest, you know, Anthropic is, you know, in a quiet period, so maybe they've slowed down a little bit.

But I do think the rest of the world has accelerated.

OpenAI has clearly accelerated.

Open source, I think, has accelerated more.

[chuckles]

Gavin, uh, what do you think of this announcement vis-a-vis round tripping and the hand-wringing concerns we talked about on this very podcast a year or two ago?

Well, I do think what's important here is what is, what, what's essentially happened is, like, Blackstone, KKR, these firms, Goldman Sachs, they would not, would not have done this, they would have not have done that CNBC, um, episode with Jensen if they did not believe that NVIDIA GPU compute was a financiable asset.

And so what's really smart here is you're getting, you know, some of the smartest asset managers in the world, you know, who charge two and 20 for many products, to essentially validate this market.

And NVIDIA's being a matchmaker, and what they're basically saying is, "Hey, our compute, because it's so flexible, is going to have a long enough life that you can finance it at lower rates than other, um, kinds of compute." And I think, and I think that's, that, that's smart.

That's good for everyone.

And it is interesting.

If you look at the underlying architectures of the three, what I call big Chinese open source models, and y- maybe even throw a few, or four, you know, if we have, um, Qwen, if we have Kimi, if we have DeepSeek, um, and then we have GLM, they're actually evolving in very different ways.

Fantastic

... linear scaling.

All of the biggest spenders on GPUs, with the exception of, um, xAI, are public companies, and public companies report something called quarterly financials, and you can use those documents to calculate something called a return on invested capital.

Um, it's really easy to do.

And to date, the ROI on all of this spending has been strongly positive to be good.

So while the answer to the $200 billion question was yes, the ROI was awesome.

The answer to the $600 billion question was yes, the ROI has been awesome.

Now we're at the trillion dollar question, and the ROI has actually still been good.

Now we are in the age of compute.

Yeah.

I mean, however you cut it, whether you cut GPU availability, whether you cut GPU rental pricing, I mean, whether you cut, like, the spot price of DRAM this month, token growth, everything is actually accelerated.

And I do think a big part of the problem is, one, the market does not have visibility into Anthropic OpenAI, and then I would say these open source inference clouds that monetize inference here in America, Fireworks, Base ten, Modal together.

And the picture looks very different when you see that because open source has accelerated massively because of GLM five point two, Kibi K three, and then Nemotron continues to kind of chug along.

We had a great, very small American open-source model release.

OpenAI has accelerated.

Anthropic continues to grow really strongly and is almost certainly pumping out significant amounts of free cash flow.

Yes.

So then why do they, why do corporates have to pay so much more for anthropic?

Well, because it turns out that there is, and this has been very surprising to me over the last year, that there is an enormous return to being at the frontier of intelligence.

And I think the future is likely to be one in which maybe a majority of economic returns continue to accrue to these frontier tokens, but the majority of tokens processed are these very cheap open source tokens.

And you can think of it, you use, you know, if...

Every enterprise, they're going to have a symphony of AI models that they use.

And you're going to have the conductor.

Maybe that's Anthropic.

But a lot of people look at the valuation in this one.

So are we in an AI bubble?

I do not believe we're in an AI bubble today.

I had, depending on how you look at it, the privilege and the misfortune of being a tech investor during the year two thousand bubble, which was really a telecom bubble.

And I think it's really helpful to compare and contrast today to the year two thousand.

First, I think Cisco peaked at a hundred and fifty or a hundred and eighty times trailing earnings.

Nvidia's at more like forty times.

So valuations are very different.

So what do you think happens with SaaS and software?

then

a really formative experience for me, the two stocks that I liked were both pretty small stocks.

One was integrated circuit systems and the other was Nvidia.

And because, and I was the analyst, and integrated circuits was run by Hawk Tan, you know, who's now the CEO of Broadcom.

I am...

Hawk was a little intimidating to me.

And so I did not talk to him as much.

What does history tell us about living through this kind of transformation or transformation of this magnitude?

Your thoughts, Gavin, on Micron and the impact on the industry, and is this a temporary bottleneck, or does this mean everybody has to get into this business quickly?

No.

Well, one, DRAM is the most important bottleneck.

There's a whole segment of people on X who are very focused on bottleneck.

I ca-bottlenecks, they call them the bottleneck bros.

You know, they'll, they'll do some work with Claude, find some esoteric Japanese company.

The bottleneck that matters is DRAM, and DRAM and HBM DRAM, this is the most important bottleneck simply because memory capacity and bandwidth are foundational to the performance of every AI model.

Right.

But luckily, we do have technology that helps researchers keep track of larger debris pieces and helps us get a sense of what the space debris landscape looks like.

A lot of our detection or tracking is done on ground and in space, where we have, you know, on ground we have radars and optical telescopes, whereas in space we might have some surveillance networks or sensors, as well as analyzing spacecraft that return to Earth and analyzing the damage done to their bodies and using that to kind of back propagate and see what debris they went through and where it was more dense.

If you're looking at the smallest debris that's smaller than a centimeter, like paint flecks, a collision might create little craters or break through the surface.

Then the next up from that would be that one to 10 centimeter range.

So from here, if we have a 10 centimeter object traveling at LEO speeds, it's not uncommon for its collision energy to be similar to that of seven kilograms of TNT.

And so we start to look at some pretty serious fragmentation and some damage and possibly even lethal damage to our, uh, satellites or assets in space.

Right.

Right now, Facebook is, you know, Meta is telling investors there's, like, no chance that we're gonna, you know, you know, monetize GPUs externally.

Well, like, 10 days before they announced all those layoffs, um, which I think was at the end of '22.

If not, I don't remember when they started to announce the layoffs.

It was short, shortly after Brad, maybe two months after Brad wrote that letter to Mark where he said, "You're the king.

You can do whatever you want, but this is what I might respectfully encourage you to do." Um, you know, they went from saying, we're, "We're not gonna do that, we're gonna keep investing," to, you know, super focused on OPEX in a couple of days, man.

So it's like people, they meet with Meta, and they hear, "Oh, you know, we're not, um, we're not gonna monetize our GPUs externally." Well, I don't know.

Mm-hmm.

[laughs]

Yeah, 100%.

But I think if, um...

I'm pretty confident Mark Zuckerberg has proven to be a good entrepreneur over time and has been willing to pivot, you know, and it's just, it's so funny.

You know, right now Facebook is, you know, Meta is telling investors there's, like, no chance that we're gonna, you know, r- you know, monetize GPUs externally.

Well, like 10 days before they announced all those layoffs, um, which I think was at the end of '22, if not...

I don't remember when they started to announce the layoffs.

Mm-hmm.

What does Elon need to do? What does SpaceX need to do to get to this kind of valuation and profitable place from your perspective? Yeah.

Well, I think we need to take a step back.

And a month ago, SpaceX did not have a cloud computing business.

Now, by some measures, it is the fourth largest cloud ahead of Oracle.

And they did that in one month.

Let's see what they do in a year.

The owners of landed power are heavily incented to give it to people who can energize it first.

Gavin, it sounds like you're suggesting to the degree that there is scarcity in compute and the purveyors of compute sort of have to decide who gets what and when, that space over terrestrial, that space might command maybe incrementally more attention or preference?

quarter.Gavin, you had a question?

Yeah, sure.

Oleg, so I, in, in, in the distant past, I ran a biopharmaceutical fund.

Uh, and, you know, it's a very hard job.

Congratulations on those numbers.

But I ran that fund right after the human gene- genome had been sequenced, and there was an expectation that the sequencing of the genome was gonna lead to this explosion in therapies, personalized medicines, et cetera, et cetera.

And I don't think, broadly speaking, we've made as much progress over the last twenty-five years as maybe people thought in the early 2000s.

Really great.

[laughs]

Um, but my understanding is that Cursor and Anthropic have more tokens of proprietary coding data than anyone else, and they each have more tokens of proprietary coding data than exist on the public internet.

And so they fed, Cursor fed, um, used Kimi 8.25, used their own private data, did some RL, some supervised fine-tuning, and they got a really good model.

And then they spent three weeks in the Colossus-2 cluster, and they got a model that 12 days ago was Pareto dominant with C- Composer 2.5. Now, it's on their own benchmark, um, Cursor Bench, so maybe take it with a grain of salt, but I think this just suggests that the Cursor data is very valuable for coding, and when it is-Train, you know, Chinchilla optimal or beyond Chinchilla optimal with reinforcement learning, you know, I think it suggests that xAI and SpaceX AI has a shot of being a real player in coding.

No

... around this table shares that belief.

And in the same way that Elon was able to re-engineer a rocket from first principles and re- make it reusable, he engineered an electric car from first principles.

You know, everyone else was trying to, you know, make an electric car like an internal combustion engine car, and he thought about it differently.

[laughs]

You're like, you are as important to your board members, particularly if you're really successful.

You know, maybe as the board members' families or parents, you know, the board members think about you a lot.

Um, once you're public, you're one of thousands of companies, um, and that's its own dynamic.

But the consequence of this is, is that private investors are often selling to management teams.

And at some level, that can mean telling management teams what they need to hear, because you want to be able to keep participating in the rounds.

Once a company's public, you can buy or sell as you wish, and this means that investors feel freer to give companies management teams.

... that you've bought recently?

Gavin, your take on the S1, and I think specifically Elon Web Services.

Well, I think what's important about, um, Elon Web Services does, does make me laugh, but 15 billion, that means the AI business right there is going to quadruple.

It, it has already effectively quadrupled.

I think what's important to- about that is there's a stat in it that for, I think, the f- their first data center, it was 122 days.

For the second one, it took them 91 days.

The third one was, I think, 66 days.

They build data centers dramatically faster than anyone else.At a lower cost.

Yeah

Say just, like, a little bit more about the kinds of people that are capitulating.

I don't know anyone like me who's not really bullish on DRAM.

There's all these interesting things happening with AI right now.

One is cross-sectionally, the valuations do not make sense.

They just flat out do not make sense.

They cannot all be true.

In other words, you have semi cap equipment companies trading at 40 times next quarter's annualized earnings and DRAM companies trading at mid-single digit.

Maybe run through them and riff.

What was it about Fidelity that led you to believe that?

At Fidelity, you join out of college.

You don't work for anyone.

They give you a group of thirty stocks.

They give you the phone number of the IR, the email address of the IR, how much Fidelity owns in percentage and dollar terms of each name.

They give you three monitors, and they say, "Go." That's very different than anywhere else where you work for someone.

I don't think I would have done that great working directly for someone.

As you were exposed to all this, what became the investment philosophy that matches your own disposition?

Mm.

And similarly, RKGI-2, which is another benchmark, actually called the jagged frontier in AI, where things that are very easy for AIs, like maybe math...... are really hard for humans and vice versa.

And RKGI measures tasks that are relatively easy for humans, maybe, you could almost think of it as a, you know, there's a lot of, like, geometric puzzle-solving, but very hard for AIs.

And here again, Grok-4 is, you know, significantly better than GPT-5.

Now in fairness and, you know, to be balanced and accurate, it is narrowly, very narrowly ahead in artificial analysis on, on the right.

But this is, this is the first time they have not released a model that wasn't decisively better.

And the progress for Grok-4 was exceptional.

How do you calculate it, Gavin? Do you have a back of the envelope way to answer this question? And if we catch up, or do we n- even need to catch up? Is there enough capital in the world that they can be patient?

Here we go.

But, yes.

But it, the deficit never really mattered because interest rates kept going down, such that even as our debt grew, interest expense as kind of a percentage of the government's budget stayed relatively l- low.

Now that rates have gone up and don't seem like they're, you know, going down anytime soon, the deficit does matter, and it really matters.

And you, you can kind of run a, a couple of scenarios, but, like, pick your, pick your metric.

If the deficit kind of continues at current levels, and we were to refinance the debt at the prices that, uh, you know, David was talking about, you know, it's, it's only a few years before spending on interest is significantly larger than spending on Medicare and Medicaid or Social Security or the military.

So pick something you care about, but, you know, our current course and speed, it's, it's in the not too distant future where interest expense is the biggest line item for the government, and that is not healthy.

But this was pretty impressive, right?

What do you think about this company specifically, if anything, and are we getting to a point where we're gonna have the same ChatGPT 2.5 moment, but with agents? And then if we do have that, Gavin, w- what will that look like in terms of employment and how companies are run?

Look, I do think, uh, if agents materialize as a reality, and, you know, Manus is maybe a little bit of a ChatGPT moment for that, I would say OpenAI and Anthropic...

Anthropic developed something called the Model Contact Protocol that OpenAI just adopted.

I think it will become a standard, and it makes it really easy for an LLM, like Stripe can just integrate with MCP, and then any LLM that uses MCP can, you know, interact with Stripe.

And this is solving a big, big problem for agents, um, in terms of just making them, you know, much easier to use, much more standardized.

But if agents become a reality, one, the ROI on AI and Blackwell is gonna be very high.

And two, what will...

Gavin, you had some thoughts?

Oh, nice

... uh, with O1 and O3, this is Arcagi.

It's designed...

You know, we keep, we keep changing the goalposts for the Turing test and AGI, and I'm sure we're gonna change them again as we're gonna blow through this.

And I just think what has happened is we were scaling around, around one, uh, on one axis, which was pre-training, and then we started scaling around inference time compute, and it's very clear that we have now added a third axis of scaling performance, and that is reasoning.

And what this is is these models...

The Internet is composed of answers, people giving answers, and what the models really benefit from is kind of the internal monologue of somebody getting to that answer.

Do you believe those drones were s- aliens? What's going on?