深度解析：Anthropic S-1 的迷思与 Mythos——Jamin Ball，Altimeter

The reality is that OpenAI and Anthropic have developed far more powerful models that they just haven’t released yet. If you compare their compute spend on serving models a year ago versus today, it’s gone up by 2.3x. Their revenue over the same period has gone up 14x. Embedded in that is basically a 7x efficiency gain in serving inference. This is a business that I think is almost accidentally profitable, right? Maybe my favorite stat in the whole S-1 was $141 billion of capital raised to date—that’s inclusive of equity, converts, and SAFEs—and a $134 billion cash balance today.

Cash in bank.

Right now, you want to go raise multibillion-dollar debt for GPUs? No bid. No bid. Debt markets are closed. There are big debt deals that aren’t happening.

Okay, we’re here in the studio with Jamin Ball from Altimeter. Welcome.

Thanks for having me. Excited to be here.

Anthropic is IPO-ing. I was thinking about who I know would be the perfect guest to cover the S-1. You’re the guy.

The alpha.

You’re—

Appreciate it.

You’re the guy. We met before, I think you started your Substack. Maybe I was just starting mine. How old is it?

Six and a half years now.

Okay. Actually, no, you’re older. What are the brag statistics about Clouded Judgment?

Ooh, brag statistics. Probably getting close to 100,000 subscribers. I think in six and a half years, I’ve missed one Friday.

Yeah.

I try to be pretty consistent.

And the general takeaway is: here are public market comps for the startup investing that you do, but Altimeter does everything, right?

Yeah. The idea was, “Hey, how can I provide some commentary on metrics around software companies?” The universe of companies that typically has public metrics is public companies. It’s a way to provide a blueprint to early-stage companies and founders: “Hey, here are some of the metrics required to be public one day.”

There are lots of tables, charts, and graphs on public metrics and valuation multiples. Maybe three or four years ago, I started writing a little ramble up front, which I treat more like a personal journal. What’s on my mind? What have I been thinking about? What’s top of mind? That’ll range from macro commentary to market commentary to trends in AI, or whatever’s trending at the moment.

Yeah. I think every generation has a VC blogger who’s the guy covering what’s going on. It used to be Fred Wilson, if you go back in the day. He used to write AVC, I think.

Yeah.

More recently, Tomasz Tunguz.

Yeah.

I mean, he’s still active, but I feel like you’re the king of charts and public stocks right now.

No, I appreciate it. Tomasz was a huge inspiration. I was at Redpoint for quite some time with Tomasz, and I always—

And now you started Theory.

Exactly.

And, yeah, so.

Exactly. But if you give us—

Is Redpoint okay? You guys are all leaving, like—

No, Redpoint’s great. They’re great. I owe everything I learned to those guys, and I have a ton of respect for them.

Yeah, amazing. Okay, we’re going to get right into Anthropic. You have some spiel, and you’ve had some public dust-ups with people that we won’t name. People are generally confused because a lot of our audience are engineers and researchers.

Yeah, yeah.

They’ve never seen an S-1.

Yeah.

They’ve never been through an IPO process. They don’t know what a roadshow is or how this is done.

Mm-hmm.

But more importantly than the mechanics, what does it mean about our general state of AI?

Yeah.

This is the most visibility we’ve ever gotten into Anthropic.

Yes. Yes.

And now we know a lot more things than we used to.

Yeah, totally. Just to say it up front, because I’ve been asked to say this: I work for Altimeter, and we have a big position in Anthropic. Everything I say is my personal view, not my firm’s view.

But yeah, it’s a big deal. It seems like every quarter the numbers for the big labs, OpenAI and Anthropic, leak. But it’s a subset of the numbers; it’s not everything. Now we get a full, open-kimono view of the financial profile of one of the most important and iconic companies, maybe of our lifetime.

Just a quick process refresher: they’ll file an S-1, probably wait about 15 days—there’s usually a minimum—before they start a roadshow. The roadshow is when they start pitching investors, the IPO investors who travel across the country and the world to hear the pitch. That’ll usually last a week and a half, and then the stock will start trading publicly. So we’re probably looking at—

End of November?

Yeah, around the 10th to the 20th of November for when it actually starts trading publicly.

Do people IPO past Thanksgiving? That’s a weird dead zone to—

Yeah, between Thanksgiving and Christmas is kind of a dead zone.

Yeah.

I would say post-Labor Day to Thanksgiving is generally a hotbed for companies to go public. People are around, people are engaged, not on—

Yeah.

—not on vacation yet.

And it’s Anthropic.

The reality is—

This is—

—I think people would make time for—

I’ll go to work for Anthropic. Yeah.

Yeah, exactly. Exactly.

Okay, so those are the normal mechanics. Off the top of your head, people should understand that an S-1 is just a filing that you have to do. OpenAI has done theirs confidentially as well, and then—

Yeah.

—and then they go public.

Yep.

People have filed S-1s and not IPO’d.

That’s true.

Cerebras famously did that.

That’s right.

And then they did it again. It’s just part of the routine process. You have investor day, whisper rumors, and all these things, right?

That’s right.

So, unless you’re literally a banker or fund manager, that’s all you need to know, basically.

Yeah.

Right?

Yeah.

There’s no—this is part of the process, and there are a lot of indications meant to make it a successful IPO.

Yeah.

“Successful” is a mixed or loaded word these days, because typically you’ll hear things like, “The IPO was so successful, the stock popped 50% on the first day.”

Mm-hmm.

That’s potentially a failure if you’re Bill Gurley.

Yeah, exactly. Exactly.

Who is like, “Well, you could have got 50% more.”

Yeah, exactly. And just for the audience, the mechanism is that an IPO is like a fundraise on steroids. You go out and raise a bunch of money at a certain price.

How much are they raising?

I don’t think that’s been decided yet.

Like $20 billion?

I don’t think they’ve decided yet.

Okay. I heard a rumor that—

But I think the rumor is $75 billion to $100 billion.

Fuck.

It’ll be a big raise. Typically, what you used to see in normal IPOs was companies floating about 10% to 15% of their company, which means what they would sell into the IPO would be 10% to 15%.

Yes, and there’s an employee lockup and blah, blah, blah.

Exactly. Now, as companies get bigger and bigger, they might float less of the company because they can take less dilution. It’s no different from a company raising a Series A, B, or C. You’re going to raise money at a certain dilution. And yeah, the Bill Gurley thing is: if the stock pops 50% on day one, why wouldn’t you have just raised money—

Priced it higher.

At a higher price?

Higher. Yeah.

Or taken on less dilution.

Yeah.

And that's maybe—

And we don't have to get into that.

A whole separate topic.

Yeah, yeah.

You've got lots of—

There's—

—nuance to public markets and IPO allocations—

There's tradition.

—and all that.

There's conflicting incentives—

Yes.

—and then there's the chance that you actually don't pop.

Yes.

So, at a high level, why now?

Yeah.

Why IPO now, right?

Mm-hmm.

They could do a Stripe. They could do a Databricks.

Yeah.

Right?

Yeah.

They're not. They're IPO-ing.

Yeah.

Why?

1. Why Anthropic Is Going Public

I mean, I think there are a couple of reasons. It's all speculation, so it's hard to know, but companies have historically chosen to go public for a number of reasons. There's the obvious one: providing liquidity to employees and investors in an unrestricted way. Of course, tenders are now a pretty regular thing for big companies like Databricks, Stripe, OpenAI, and Anthropic—

Like once a year.

—that they all do.

You'll get, like, “Well, if you want to cash out, here's the—”

Yeah.

—“currency.”

But you'll be limited in how much you can sell. Once a company has gone public and through the lockup—which used to be 180 days, or 6 months, and now is bespoke for every company, with shares dribbled out over shorter or longer periods—you have, in a sense, unrestricted access. You can generate liquidity for your employees and help morale. Or you need a liquid currency for acquisitions.

Yes.

A company might be more willing to accept an acquisition from a public company because the currency it gets back is liquid. In a private company, it might not be.

It's made up.

You don't really know the value.

Yeah.

Yeah, it's Monopoly money.

Yeah.

These quasi-public companies—OpenAI, Anthropic, Databricks—I think their currency has almost started to be viewed as cash, not Monopoly money, because they're very high quality and very late stage. That effect probably matters less. A bigger reason to go public used to be access to much larger amounts of capital. If you wanted to raise $100 million, the only way to do it used to be to go public. Now you can—

Huge.

—you can raise—

Very huge.

—a huge amount in a Series A, or a seed round. A seed round can be more than that. So it's not necessarily the amount of capital. There's a huge capital base available to private companies, whether it's U.S. institutions getting bigger, large pension funds and endowments that used to act as LPs investing directly in companies, or the emergence of big sovereign international funds investing directly. I don't know if the argument necessarily holds that if you want to raise a lot of money, you should go public. I think OpenAI raised a round of something like $120 billion or $130 billion—some ridiculous amount.

Biggest of all time.

Yeah, exactly.

Yeah.

I think a more important angle might be the debt piece. Once you go public—actually, let me take a step back. If you want to raise debt, or if you want someone else to raise it, not on your behalf, but as a neocloud, for example, that neocloud might want to raise debt to build a data center for Anthropic to use. Debt providers are going to look at Anthropic and underwrite its creditworthiness. They'll say, “We're going to syndicate however many billions of dollars of debt for you, Anthropic, or through a neocloud for you. But we're going to need full open kimono on your financials. We're going to need to know exactly how the business is doing so we can underwrite you, and everyone we syndicate the debt to on the back end also needs to see it.” Historically, companies have been hesitant to go full open kimono with their financials to the debt and credit markets because they don't want those financials to be fully public.

If you want to be rated by the rating agencies, they'll give you a credit rating. Investment grade is the highest, and your credit rating impacts the rate at which you can raise debt: the higher your rating, the lower the rate you'll receive. But you have to be open kimono with your financials. Going public is a great way for companies to get rated. I think SpaceX very famously got rated very quickly after going public. Then you can—

Too quickly?

—tap into... Arguably, yes. A lot of people said it was too fast and that the company needed to season more as a public company.

And it was added to the S&P 500, too.

Yeah.

Something like that?

Yeah.

Yeah.

New rules. Exactly.

Yeah.

A lot of these companies will be capital-intensive, whether it's for training or building out compute clusters, and they'll want to tap into the credit markets. I think they can do that much more efficiently as a public company.

Sure.

So it's probably a confluence of all those things adding up to why a company would go public. I think if you ask these executive teams and management teams, they want everyday Americans to be able to participate in the equity upside of these companies.

I was going to bring that up.

Yeah.

Yeah, up to and including giving shares to everyone.

Totally.

Every American.

Totally. The reality is that private markets are more of a cabal. They're a small subset of the population: qualified purchasers, accredited investors, big funds, pensions, and endowments get to participate in the equity appreciation of these businesses. Your doctors, dentists, and everyday Americans don't, and there's something that doesn't feel right about that. Everyone's—

Oh, I mean, straight up, you'll get social unrest, right?

Yeah, exactly.

And straight up, you'll get regulation voted in because, well, those guys have their AI, and I'm going to vote against it. I have my vote.

That's right. That's right.

That's a really interesting governing mechanism of democracy versus capitalism, where capitalism encourages centralization—

Yeah.

—and democracy encourages decentralization somewhat. Anyway, there's all that. I think it's very valid. I'll point out for people who don't know about capital structure that equity is junior to debt.

Yeah.

And so it's nice to have $2 trillion in junior cash in your enterprise value, right?

Exactly. Exactly.

And that's public. I'll add another reason for “why now,” which is a little more advanced than people know: interest rates are spiking for data center debt globally. Is there a source to look that kind of stuff up? Is it just Bloomberg, or is—

Hmm. I mean, I think most of these—

I mean, there are charts that people tweet around, but—

Most of these big debt deals will price, and there'll be an interest rate, an OID, and a bunch of things associated with the price. Oftentimes, those will be public once they've priced. But yeah, rates are going up. Risk-free rates are going up. The Fed paused and decided not to hike yesterday or earlier this week. I don't know if that's necessarily a reason to go public, but—

It's part of the “why now” if—

Yeah.

—if it does help with that. If my debt went from, I don't know, 3% to 5%, that's billions of dollars. If I can keep a lid on that, that's very important to me.

Yeah. A lot of the data center debt deals are now getting done over 10%, so it's getting expensive.

Okay.

It’s getting expensive.

Yeah. And then the other part of this discussion is the ROI on a data center, right? There are a lot of AI skeptics—

Mm-hmm. Mm-hmm.

—who are saying, “This is the biggest build-out in human history—”

Yeah, yeah.

“—and we are throwing money after this illusory dream of AI.”

Yes.

And on the other side, AI proponents are saying, “Your payback period is 1 year.”

Yeah, totally.

Why is there such a huge difference?

This was one of the best things about the Anthropic S-1: a lot of the metrics that were leaking were ARR or this or that, but you didn’t really get to see many below-the-line metrics.

That’s not ROI, yeah.

You didn’t get to see the cash.

Yeah.

How much are they burning?

It’s not profit.

Are they profitable? Are they not? That, I think, was one of the biggest surprises of the S-1. Here’s a business that showed 85% gross margins, over 20% free cash flow margins—

Mm-hmm.

—and 8% GAAP operating margins. So even on a GAAP—

That’s new.

—basis—

Yes.

—highly profitable. My favorite stat of the whole S-1 was $141 billion of capital raised to date. That’s inclusive of equity, converts, and SAFEs, and a $134 billion—

Cash in bank.

—cash balance today. So—

Cash and cash equivalents, but yes.

Exactly. You think, okay, you’ve raised this much and you have this much in the bank. That means you’ve burned $7 billion to get to the Q3 numbers that aren’t in the S-1 but were broadly reported a few weeks ago: $16.8 billion of quarterly revenue off of $7 billion of lifetime burn.

Hmm.

The burn is something we should get into, because that was a huge debate: It looks like they did raise debt, but are they really profitable? What is this net loss on the P&L? There are lots of—

Yeah. So, yeah.

—lots of things out in the world that aren’t fully true.

That is important. I do want to cover it. There’s a little bit more in the weeds than—

Yes.

—than at the start of this conversation.

Yes.

But that overall efficiency is by far—if you’ve been following the relative fates of, let’s call it, SpaceX, Anthropic, and OpenAI—the ones that are going public, they are the most efficient lab in the world—

Yeah.

—by an order of magnitude.

Yeah, totally.

When I wrote up the S-1, I thought, okay, this is the mystery. Why? What’s going on? How are they so much more efficient?

Yeah.

In their S-1, they attributed it to the loops.

Yep. Yeah.

The improvement loop, the investment loop, something like that.

Yeah, exactly.

People can go read that. I already wrote it up. Maybe they’re just smarter.

Yeah. I think one of the best numbers to look at is their growth relative to the change in serving compute spend.

Hmm.

They break out their compute spend into training and internal use—

And by the way—

—versus—

—if you have charts—

Yeah, let’s pull it. We do have a chart on that. Here’s a chart that shows the—

Let’s do full screen on this one.

Oh, yeah, yeah, yeah.

Just so it’s a little bigger.

Perfect.

Yeah.

Here’s a chart, broken down in their S-1, that shows the percentage of their compute that goes toward serving models—that is, revenue-generating inference—or internal R&D, training, maybe safety and alignment, or other non-serving tokens. What’s interesting is that only 27% of their compute today is going toward serving models.

Here’s why this is so interesting. If you compare their compute spend a year ago with today, their compute going toward serving models has gone up by 2.3x. Their revenue over the same period has gone up 14x. Embedded in that is basically a 7x efficiency gain in serving inference.

This is a business that I think is almost accidentally profitable. A 7x increase in the efficiency of serving tokens means that if you were monetizing a megawatt at $10 million to $15 million, now you’re monetizing it at $70 million to $100 million.

Here’s a bit of a random rant. Software companies have historically told the world and investors, “We’re an annuity of cash flows. Once we acquire a customer, our net retention is high, and they’ll stick with us.”

Customer lifetime value.

Exactly. “We’re going to be burning a lot of money early on, but we’re in growth mode. We’re in customer acquisition mode. Once we exit growth mode, we’ll be more in harvest mode, and that’s when we’ll become profitable, because we’re not spending a lot of money on customer acquisition costs.” The general model of a SaaS company was unprofitable over time, then profitable.

What’s happened with Anthropic—and let me pull up another chart here—I believe I have their free cash flow margins.

Oh my God.

I had to cut the chart, right? It was negative 4,000%, negative 1,500%—super, super negative. If I actually drew the axes to scale, it wouldn’t look real. The axis is broken. This is the free cash flow margin of Anthropic: 21% free cash flow margin in their most recent quarter.

What you’re seeing is a business that, because it was able to drive such efficiency in its compute—that’s why I went on that compute ramble—has been able to monetize at much higher rates and deliver positive free cash flow margins pretty early on.

This is a business that says, “We’ve proven that we can operate highly profitably. What we’ve probably done is underinvested in compute and compute build-out, and we need to be more aggressive.” My guess is that you’re actually going to see this level off and maybe go negative again as they say, “We’re going to reinvest, and we’re going to invest in a lot of compute. We’re going to build out a lot of compute.”

Yeah.

Because now they’ve shown themselves and the world—

The capability.

—“Hey, if we need to, we can flip the switch and be profitable.”

Yeah.

At the end of the day, cash is king. Your measure of profitability should be: Is your cash balance going up or is your cash balance going down? Now, if you wanted to poke holes in this, the reality of GAAP accounting is that there are two forms of—

For listeners, a couple of things on the accounting side, because we’ve got to educate people. FCF is typically the most honest version of what profit looks like.

Yeah.

In this case, you can mostly regard it as what a profit margin would look like, because FCF focuses on cash—

Mm-hmm.

—and so there are fewer tricks to play with accrued receivables and offsetting your liabilities. That is roughly what it means.

The reason people have adjusted revenue and adjusted profit is that there are nuances in running a business like this. GAAP, which stands for generally accepted accounting principles, is what you officially have to report, but it doesn’t recognize how you want to run the business—

Right.

—it doesn’t distinguish investment decisions from the profitability of core business decisions. It all gets smushed together. What I’m trying to guide people toward is that if you see a change in this margin profile, it is a conscious choice.

Yeah.

Actually—

Yeah.

—you should probably split it out further between what is actually the core business and whether that’s being affected, or whether this is more investment, and therefore FCF went down.

Yeah. At the end of the day, cash is king. Your measure of profitability should be whether your cash balance is going up or going down. If you wanted to poke holes in this, the reality of GAAP accounting is that there are two forms of expense. Very broadly, there are two forms of expense.

There’s the cash salary that you pay people, and then there’s also the equity—

Stock-based—

—and then the non-cash compensation that you pay people. Non-cash stock-based compensation, or equity compensation, doesn’t show up in free cash flow. It’s added back because it’s a non-cash expense. A lot of people look at free cash flow and say, “Well, it actually overstates your profitability,” because the reality is, if you didn’t grant equity to your hires, they wouldn’t join you. It is a real expense. Treat it as a real expense.

I don’t have a chart for this, but if you were to adjust free cash flow—

It’s still profitable.

—for stock-based comp, you’re still at an 8%—

Yeah.

—free cash flow, excluding stock-based comp. So you’re even profitable if you include the big equity grants.

By normal rules, yes.

Yeah, exactly.

And so, if you’ve never seen a software IPO, this never happens.

Exactly. Yeah, it’s very rare. I think there are a lot of software companies that are free cash flow positive today.

Yeah.

But if you factor in—

Not at IPO. Never.

Not at IPO.

Never at IPO.

But even today, the mature ones, if you factor in stock-based comp, most of them are free cash flow neutral or negative if you treat stock-based comp as a cash expense.

So all those companies are dying, and everyone’s just buying Anthropic, right? Is that the conclusion?

I think there’s a—well, they’re not quite that, but certainly—

Why would you do anything else?

Exactly.

Yeah.

Look, here’s a crazy stat. In the S-1, as a breadth-versus-depth measure, they called out that 50% of the Fortune 500 is using Anthropic.

Yeah.

However, if you look at what that cohort, the Fortune 500, is spending on Anthropic as a percent of their expenses—

Hmm.

—it’s 0.1%. So you ask yourself, where’s that headed? I think John Collison had Sam Altman on his podcast the other day. I think he mentioned that 1% of Stripe’s expenses are on tokens today, but he expects that to go to 15% to 20%—

Jesus.

—over the next few years.

Jesus.

Stripe is a very forward-leaning tech company. Your average business in the Fortune 500 isn’t going to show that level, but it could show the same order of growth, or the same magnitude of growth. If Stripe is saying, “We’re going to 15 to 20X the spend on tokens relative to other expenses in the organization,” maybe you see something similar in the Fortune 500, going from 0.1% to 1%. Ultimately, where do you think that settles out? How deeply can they penetrate their TAM? It’s an important question people should think about. To me, what it demonstrates is that despite the insane ramp in revenue, growing over 1,000% year over year in Q2, there’s probably still a lot of room for them to continue to penetrate their TAM.

Yeah. I think the word is “virtual collaborators.” I say “agents,” but you’ve got to make it a friendly term for people. I really start to tell people they should think about it in terms of their employee spend versus their—

Mm-hmm.

—token spend.

Mm-hmm.

Most people are at, let’s call it, 0.1%—

Yeah.

—or 1% to 5%.

Yep.

The median AI company is at 50%.

Mm-hmm.

And the top AI companies, let’s say the 90th percentile, an Anthropic person is at 10X tokens—

Yeah.

—tokens to salaries.

Yep.

And so—

Yeah.

—that’s probably the right comparison: what are you spending on humans?

Yeah, right.

Probably 30%, right?

Yeah.

Like—

Yeah.

—and, well, okay, so—

Could be more, even. Humans could be 40%. Total head count could even be—

Yeah.

—like, 40%, 50%.

Depends what kind of business you’re running, right?

Yeah, exactly.

And so if you’re a physical business, fine. Until we have good robots, that’s not going to happen. But what are you spending on knowledge work by humans?

Mm-hmm.

Knowledge work by humans includes coders, but also admin, ops, legal, and whatever you want.

Yeah.

All that. That percentage, the ratio of tokens spent to human spend, was zero to whatever—

Yep.

—and now it’s going to rise.

Yeah.

And—

Totally.

—how far can it go?

Totally, totally. It can go a lot. There are a lot of hyperbolic statements being made, both positive and negative, about OpenAI and Anthropic, the big labs. I think it really boils down to 3 main questions: Can Anthropic stay on the frontier? Is there a business—a durable business—to be built at the frontier? And if so, how much is it going to cost? The second and third questions relate to each other: Can it make enough money to cover all the costs of being at the frontier? That’s the one where I think I have a chart.

Yeah, please.

Let me see which one. What did I want to bring up? This one. No, let me pull up this one. This is a good one. This is kind of the second question: how much will it cost to be at the frontier? Or maybe let’s go to what kind of business you can build at the frontier.

The other interesting stat in Anthropic’s S-1 was that, for the 3 months ending in September, 80% of their revenue was driven by models that were 6 months old or younger. What does that tell you? The vast majority of their revenue is coming from their most recent, most powerful models, and I expect that to continue. When it comes to frontier tokens versus open-weight tokens, I think there’ll be a massive market for both. I think open-weight models will continue to eat share as a percent of total tokens. But as a percent of wallet share in dollars, frontier models will continue to take the lion’s share. I expect frontier tokens to be the most economically valuable and capture the most market share. I expect open-weight and cheap tokens to capture the most market share in terms of total tokens.

Yeah.

I think this stat—that 80% of revenue comes from models 6 months old or less, which is much higher than I would have expected—demonstrates the value of being on the frontier.

So let’s go to the next question: What is it going to cost? Here’s a chart. I’ll have to caveat it. A lot of people look at the Anthropic S-1 and say, “Oh my gosh, there’s $856 billion of—”

Off the books.

—of, like, not even off the books, but committed deals that they’ve signed.

Yes.

How are they going to pay for it?

Yeah.

They’re at whatever their ARR is—70, 75, 80, I don’t know. I think the last number was 65. The Q3 number was 16 point something. But whatever, 70, plus or minus. How are they going to pay for $855 billion?

You know, like—

Like—

Have you read the rest of the S-1?

What I tried to do in this chart is break down their commitments. They broke down their commitments, but they didn’t do it by year.

Is it normal, by the way, to have it year by year? I feel like this is unnecessary.

No. So this year-by-year breakdown is a little bit of Jamin’s best interpretation. They’ll give a dollar—

Oh, no, I’ve seen variations of this. Yeah.

Okay. Yeah.

Yeah.

I think what you have to ask yourself is, for the commitments they’ve made—and I’ve tried to annualize them—if they said this is a 5-year deal—

Yes.

—I’ve just made it over a 5-year period.

By the way, every banker in the world has done this already. They did this on the night of the S-1.

Mm-hmm.

Right?

So you have to ask yourselves: can you pay for it? At a fixed basis, the gray bars are maybe less hard commitments, but commitments they want to make.

The SpaceX one is in there, right?

Yeah, exactly.

That’s the one they can walk away from in 90 days.

Exactly.

Yeah.

Will they be able to fund $100 billion a year? It isn’t $850 billion a year. The question is whether they can fund $100 billion a year. There are a couple of ways they could do that. First, do you expect their growth to continue? That’s a question of whether there will continue to be a business at the frontier.

Or you go back to the fact that only 25% of their compute today is allocated to revenue-generating or inference-serving use cases. If they were to double that and add no compute, they could, in theory, double their revenue. When you look at the headline number of $855 billion, you think, “Oh my gosh, that’s a huge number. How are you going to do it?” Some of that is very back-weighted. But when I think about the trajectory of the business and the TAM that’s been penetrated today, I don’t think it’s far-fetched at all to say the business trajectory—and the revenue—will more than cover the expense. I think I have another chart in here. This is another great one: compute margin. This looks at taking their revenue and subtracting all compute, including internal model training.

Same-quarter matched, right?

Exactly.

Okay.

Exactly. What you’re seeing is that even as they’ve shifted more of their compute to non-revenue-generating use cases, they’ve still scaled their compute margin to well over 50%.

Mm-hmm.

They’re getting hyper-efficient at the compute they bring online. The compute they decide to use for revenue-generating use cases has become 7X more efficient over the last year. You have to ask yourself whether they’ll continue to ramp compute and capacity. I think the answer is yes. If you ask anyone at Anthropic—and this is all speculation—they’d probably say, “We wish we’d started doing it sooner. We wish we’d been more aggressive.”

They’re trying to commit to more compute.

That’s right.

If you don’t talk to any of these guys, you come away thinking, “Oh my God, these guys are absolutely crazy and fiscally irresponsible.”

Yeah.

But if you have any confidence, or if you see up close the demand for intelligence that people are snapping up and immediately consuming whatever you give them, then, yes, you’re actually trying to get more compute.

Yeah.

The game isn’t to reduce the commitments. It’s actually to get more commitments.

Yeah. Yeah.

They’d have a trillion if they could.

Yeah. I don’t think I’ve spoken to anyone in the AI industry, certainly not anyone at the big labs, who’s said, “Yeah, I feel good about the compute we have.” It’s almost always the opposite.

An underrated factor in future compute demand will be compute used for safety, alignment, and verification. I think more and more compute is going to go toward those functions, whether that’s for research or just for verification and alignment at inference time.

Yeah.

That will increase compute demand for all these companies.

Yeah, but we probably won’t break that out. It might be 5% now and 20% in the future.

Of course. Of course.

I’d just tell people: if you think this is bad now, wait till you see the OpenAI S-1.

Yeah.

It’s not bad. That’s the whole point: they’re really trying to get compute from every source they can get.

Mm-hmm.

OpenAI’s buying TPUs. There was a cold day in hell when that would happen.

They have their own TPUs, right? They have their own chip initiative, Jalapeño, which has pretty rave reviews.

Uh-huh.

Anthropic staffed up a compute team to build its own chips. Obviously, the hyperscalers are building their own chips. It’s a mad dash.

Mad dash. And the only winners are TSMC, NVIDIA, and ASML, and whatever else is in the supply chain.

Totally. And I think it’s getting harder and harder, especially in the United States. You have a lot of pushback. Data centers are too noisy; people don’t want them in their neighborhoods. They add a lot of air pollution.

The water.

People don’t want that. They say it uses too much water, which it doesn’t.

Oh my God, the water.

It’s the electricity. It uses a lot of electricity. People think it’s going to increase their electricity bill if they live nearby. Or it’s the simple idea that AI is going to kill us and they don’t want to enable that. There’s all kinds of pushback. It’s become hyper-political.

The unfortunate thing is that it has become a lot harder. You do see a lot more pushback. Midterms are coming up, right? Who takes control of the House and Senate could have a big impact on regulation, or the lack of it, around data center and intelligence center build-outs.

The good thing is that it’s getting harder and harder to build these data centers. Building behind the meter is harder and harder, too. So you have a natural governor on the build-out, and I think we’re a lot less likely to overbuild too quickly and create a supply glut. Think of the internet bubble in the early 2000s, when we built way too much fiber and there wasn’t as much demand for the internet. There was a lot of dark fiber—fiber with no bits and bytes traveling over it. We have no dark GPUs today.

It’s probably a good thing that there’s some rate limiter. It will stop us from entering a huge bubble, but it’s also bad because the industry needs more compute. Especially as these models get more powerful, security concerns go through the roof. You also need a lot more compute for the defenders. The defenders need more capabilities.

Oh, yeah.

I don’t know; it’s my personal belief that we’re going to be compute-constrained for far longer than the world thinks.

Yeah. I’d be curious what the market has priced in as the world’s expectations, and then, after doing the work on what we actually think we have, what the delta is. It could be 3 years; it could be 30 years.

Yeah. Yeah, yeah, yeah.

That’s a very material thing to figure out.

Yeah, of course.

One thing to guide people toward: there’s all this political stuff, and I honestly would say Meta does a lot of work, and I think increasingly OpenAI does a lot of work, on how to make sure they’re a positive presence in the communities they’re in, right?

Yeah.

Yes, you need to take care of that. Coming back to research relevance, ROI, and what it means about the industry, this profit motive here—I do wonder sometimes, and this has come up on our other podcasts, specifically tying in the research side: how much more is there to improve?

Yeah.

Was this—

Yeah.

—kind of a one-time upgrade of our inference stack—

Yeah.

—where, as you noted, we got something like a 7X improvement in inference efficiency?

Yeah.

Because there can’t be improvements forever in our inference stack.

Yeah.

Unless we fundamentally change the chip, we probably started out pretty unoptimized, and we’re a lot more optimized than we were 2 years ago.

Yeah.

There’s a limit to how optimized you can get, just because of the physical movement of joules and bits and bytes. That’s my question: how much efficiency is left to squeeze?

Yeah.

Maybe it might cap out in terms of that year-on-year growth.

Yeah.

Well, look, I think I have a chart I’ll get to in a second, but I think this is a debate that’s existed for quite some time.

Yes.

Right? You had the Ilya moment a couple of years ago: “Hey, the data and the internet are fossil fuels. We’ve exhausted them. There’s no more data left to train on, and we’re hitting a plateau.” Then we hit inference-time reasoning, inference-time compute, and reasoning models, and a new paradigm emerged.

Maybe embedded in your question is this: Is the gap between the frontier and the leading open-weight model shrinking?

I don’t mind the gap. It’s not something I talked about.

But I think it’s interesting—

Yeah.

—because the real way to think about or answer that question isn’t the gap between the best leading open-weight models today and the frontier models today. The reality is that OpenAI and Anthropic have far more powerful models they’ve developed that they just haven’t released yet. This is a chart Mark Chen published from OpenAI. Switching gears from Anthropic to OpenAI—

Yeah, he’s been on the pod, the cooking pod.

Oh, nice.

We made some soup for researchers.

Amazing. I love that pod. When they solved the Navier–Stokes problem, that big math problem, Mark came out and said, “Hey, here’s a model.” It may be dark blue; light blue is harder to see, but the top bar is an internal model they used to solve that problem. The blue bar is Astra. Look at how much more powerful the internal model is than what’s publicly out there. In some ways, that could scare you a little bit, right? These labs have even more powerful models.

So if you think about the gap between open-weight and frontier models today, I think it’s actually far wider than people expect, because there are far better models at the frontier that haven’t been publicly released yet. Whether that’s because of safety, testing, alignment, or some combination of everything is unclear, but I do think you’re still seeing the frontier accelerate—even if it may appear that what’s released is plateauing. I’m not even sure it looks like what’s released has been plateauing.

No, no one’s saying it is plateauing today. We’re all discussing whether it could plateau in the future.

Yeah.

Because that’s a pretty important thing to figure out for your investing, which is part of the deal here.

Totally.

Yeah, that’s a very well-phrased point. I thought you were going in a different direction, where you were going to argue about the bitter lesson, so kudos on that. I think it’s a very fair point that every model generation still shows quantum leaps.

Mm-hmm.

This is only in math, but OpenAI has a pretty good track record of generalizing beyond just math or coding.

Yeah.

Right?

Totally.

My point was a little bit more with my financial analyst hat on: you had this one-time 7× improvement in revenue efficiency, or whatever.

Mm-hmm.

And it’s not like they raised prices.

Yeah.

So where is it coming from?

Oh, here’s another good one.

Do you know what I mean? They didn’t raise prices.

I think I have a chart.

You know what I’m saying, right?

Yeah.

They have—

Yeah, it’s actually the opposite. This is an awesome chart.

Yes.

This shows the cost to complete a task with each model family.

They’re charging less?

Yes, exactly. The cost to complete the same task with each successive model family is reduced by about 25% to 40%. With Fable or Opus, and then with Fable 5.1 and Opus 5.5, you see that reduction.

Yeah.

The costs are going way down, their compute efficiency is going way up, and I think a lot of that demonstrates Jevons’ paradox.

Oh, my God.

I know, I know.

I knew it was going to come. It was going to come.

You can’t have a podcast on AI without that coming up.

Yeah, I know.

It is the appropriate time to bring it up.

Yes.

As things get cheaper, go back to the early days of the internet. How many bytes were traveling across the internet in 2000? How many bytes are traveling across the internet today? Many, many orders of magnitude more. And the price to transfer bytes over the network has dropped by many orders of magnitude.

It’s 2 to the N. It’s literally 2 to the N.

Exactly. It’s P and Q, right?

Yeah.

Price has gone way down—

Yeah.

—and quantity has gone way up.

Yeah.

I think you’re going to see the same thing with people buying and serving models and tokens.

I’m generally a supporter of that. I do want to remind people that this is an objective fact. This isn’t VC bullshit; this is actually how the businesses are running. It’s surprising to us that while we read about this in theory, seeing it in financials is really ridiculous.

Yeah.

Yeah.

And you’ll continue to see it. Think about Jevons’ paradox in models. If you want a binary classifier, you don’t need a language model to do that. Just use a decision model.

Yeah.

I think you’re going to continue to see efficiencies like that. Go back to S3 and EC2, AWS’s big initial products: storage and compute. You’ve seen 99% price reductions in both of those services since they launched. But what’s happened to the EC2 and S3 businesses, or to the broader storage and compute businesses, with more CPU compute? They’ve gone up exponentially.

All the hyperscalers reported today. They’re all over $100 billion run-rate businesses if they were individual companies. Google Cloud’s grown over 100%. It was just crazy. Azure’s growing almost 50%. It was unreported today; we’re recording on Thursday. Thirty-nine, almost 40%. These businesses are accelerating and growing at—

Beast mode.

—it’s crazy.

Beast mode.

It’s crazy.

You should drop everything you’re doing and start one of these things.

Yeah, yeah.

It’s October 29 for those keeping track. And a fun personal anecdote: I was at my hedge fund the day they broke out AWS from Amazon.

No kidding.

At the time, we weren’t long Amazon because we were worried about AWS margins, and we wondered whether it was just startups, and what if we went through a startup winter? It was so nonsensical. The funny thing about finance is that you can do all this work and make worse decisions.

Yeah, yeah.

We definitely did that.

Totally. That’s the thing: market timing is impossible.

And fixating on the wrong thing.

Yeah, that’s right.

A lot of people are having to encounter that for the first time. You have 2,000 data points that were released in this thing, and you’ll fixate on one thing as if that’s everything you need to know. But you’ve lost sight of the bigger picture.

Yeah. Again, there are a lot of things I believe that others would disagree with me on.

Yeah, yeah.

But I believe there’s one question that matters, and that’s: Can Anthropic stay at the frontier?

Yeah.

To me, if I believe the answer to that question is yes—and I think everything they’ve done historically has demonstrated that—I think they’ll build a great business and will continue to.

But then there’s a right price for that.

Of course. And this isn’t about whether it’ll be priced right or anything like that. If you’re able to take a longer-term view on whether the business will be successful over time, I think that’s the most important question. Whether you think the IPO price is cheap, expensive, or just right—we don’t even know what that’s going to be yet.

Prices can change in public companies for idiosyncratic reasons, based on individual company performance, or they can change because of macro-correlated risk. The reality is, if Anthropic had gone public a month ago, with the big rout you’ve seen in public markets recently, especially in semis and the AI trade, Anthropic would also be down, even if the business was doing well. Stocks can go up and down based on individual company performance as well as just—

Market data.

Market data. Exactly.

Yeah.

Exactly.

You should be prepared for those kinds of discussions if you’re getting into this. As a side question, there are a lot of non-public-market trackers of the stock. Here’s my proxy for OpenAI stock; here’s my proxy for Anthropic stock. Do you put any weight into that, or zero? You don’t even—

I don’t.

Yeah.

Yeah, I don’t.

I’m wondering if I should pay attention to those things more closely, because they do give some indication of the stock price reaction when a model just—

No, it might show some market sentiment. Ramp had some data—I think I copied a Ramp chart the other day. This was Anthropic spend on some of their most recent models. The dark blue is Opus 5.5. What the Ramp data shows is—

People switch.

Yeah, exactly. With Opus 5.5, you saw some reacceleration. Would it surprise me if you saw a reacceleration of revenue and run rate in October? October won’t be in Q3; it’s the first month of Q4. No, it wouldn’t surprise me, because I do think a lot of the third-party data is showing and suggesting that.

Well, this chart is super interesting, because you just said that when you say “stay at the frontier,” there are different levels to that.

Yeah.

“Stay at the frontier” could mean we’re one of the top 2, one of the top 6, or in the top 20—whatever. Usually people mean the top 2, but sometimes you can expand that to the top 6.

Google now has a claim to be number one in some metrics. Meta is doing super well. SpaceX is doing super well—not number one, but number three or number four. You could say that. So, look at that summer slump in Anthropic revenue. Was that because OpenAI was briefly ahead?

Yeah, I think Astra was a great model, and I think a lot of people started using Astra. It wasn’t just Astra, right? What were Google’s models? What was the one they just released, Carbon or Argon? Was it yesterday?

Carbon, yeah. Carbon, Argon. They haven’t released it yet. This is my gripe with Google: they never release anything. They just announce a waitlist.

Yes.

And then it’s like, “Goddammit, just—”

I think you also saw a massive explosion of Kimi K3, the GLM series, and DeepSeek models. Open-weight models got really powerful and got good quickly. It’s crazy: the far-right bar, the non-Opus 5.5 or Sonnet 5.5, has halved or more.

Maybe more.

In what, 2 or 3 months? Imagine if they didn’t have Opus 5.5 or Sonnet 5.5. Who knows what the revenue growth and trajectory could have been? That’s why I do think it’s important to have a frontier-class model.

I’ll say, from watching these labs for the past 3 or 4 years, they used to have more of a reserve. They’d have internal models they could release whenever they felt like it. Then, around the GPT-5 and GPT-6 era, everyone just released whatever they had when they had it.

Now I feel like the reserves are building up again. I don’t know why, but it definitely seems like, with the internal model, they have more wood behind the shed for stuff like this. They can say, “I’m sitting on Fable 5.5. I can release it the day before our IPO, and that’ll make my IPO go well.”

I think there are good reasons the models aren’t released right away. On one hand, there’s a lot of safety—

There’s safety checking, yes.

—and alignment and all that.

But as a PM, as an executive playing this game, I can play it pretty well if I have a backlog.

Totally. But also think about OpenAI’s Patch the Planet initiative, which is basically: let’s use these super-powerful models to go find vulnerabilities.

Not even. It was GPT-5.6. It wasn’t Frontier.

Yeah. Oh, I didn’t realize it was GPT-5.6.

It was just, “Let’s spam a bunch of last-gen models.”

Exactly. Let’s find all these vulnerabilities, like the PHP one, and create the patches so we can release the patch instantaneously while announcing the vulnerability. The challenge is that attackers are using these models and really powerful open-weight models, and they’re finding these vulnerabilities themselves.

Yeah, it’s a race.

It’s a race. It used to be that you had weeks or months from when a vulnerability was discovered to when there was an exploit. Now that’s collapsed, and it’s basically zero. You know, it’s zero days. The vulnerability is announced or discovered, and an exploit is available—it’s almost negative.

I think hidden in OpenAI’s announcement of the PHP thing was that they basically said, “There are more serious vulnerabilities that we know about that we’re working on patching right now.” Something more critical than something that touches 70% of the internet is pretty serious.

The 70% of the internet is just a proxy for old. Of course it’s older and more vulnerable.

That’s right.

That’s a good way to phrase it. I’m not that interested in dwelling on that stuff, only because there are going to be smarter people than me on it.

There’s one angle of this that’s kind of funny: we’re spending a billion tokens finding all the vulnerabilities, and we’re going to spend a billion more to fix them.

Mm-hmm.

What if we just didn’t do either? That would net out to the same position we’re at.

Yeah, yeah, yeah.

Anyway.

Well, the cat’s out of the bag on that one.

Yeah, the cat’s out of the bag.

You can’t put the toothpaste back in the tube.

But that is one argument people make: blame the AI instead of blame the people. They kill the messenger.

Right.

The other interesting part is that this might actually be an issue for nation-state actors who already had all these exploits.

Yeah.

Now civilians are finding them, and they have to work harder.

Yeah, exactly.

All my zero-days are gone. I had them for 10 years.

Yes.

But anyway, those are fun anecdotes or consequences of this. If you want to think about it in a really Straussian and Machiavellian way, if you’re a nation-state actor, what do you do to preserve your edge? You can’t have the channels you’re relying on lock up.

Yeah.

Then you send a Saudi decoy to put the thing behind your TV, so people get distracted by that and your real spy gets in.

That was crazy. That’s crazy.

This is—yeah, I watch too much TV to think that what we’re told is the full story. Right?

Yeah, yeah. I think that’s fair. Too much is at stake.

Right?

I think that’s fair. Too much is at stake.

Yeah, it’s too much at stake. Okay, coming back to the actual finance: I really want to be a primer for everyone who’s new to the economics of models and model labs. One thing people have said is that the R&D per model—let’s say, what does GPT-5, Sonnet 5, or Opus 5 cost? Let’s call it a billion dollars. That lasted me 6 months, right? Because, as you just said, it depreciates over 6 months. Isn’t that worrying? These are the fastest-depreciating assets in history.

Yeah.

This is something Sarah Guo and I have talked about before. I think the answer is no, but it’s a qualified no: as long as my next model run is better than the current state of the art, we’re good. But it’s a bit of a house of cards if the next one doesn’t work out, which briefly affected OpenAI. A lot of people know by now that GPT-4.5, the big run, wasn’t as successful. They’ve since fixed it, but a slump or a failed run can really screw with the ROI on the whole thing.

I think there’s a really good analogy to the chip industry. I remember speaking to someone much older and much smarter than me, who did a lot of semiconductor investing in the ’80s and ’90s, about all these chip companies.

They’ve seen it all.

They’ve seen it all, and they were highly skeptical. Sometimes industry veterans are highly skeptical, and they miss the next big opportunity because they never think the incumbent can be disrupted. But their point of view was that it’s very different to tape out a chip that’s competitive with the leading chips at the time versus building a machine that can continuously tape out chips that compete with the Vera Rubin chips.

Yes. It’s the model factory.

The chip factory.

Yeah.

This individual’s point of view was that maybe a chip startup can build one chip, but they won’t have the machine to constantly pump out a new chip every year and stay at the frontier. That’s why NVIDIA is king. It’s not because they have the best chip now; it’s because they have the best factory.

The process for chips.

Exactly.

Yeah, it’s the machine that builds the machine, basically.

That’s kind of the bear case on all these new labs. You talk to many smart people, and they might tell you, “Hey, if you want to really compete and have a SOTA model at the frontier, you can’t have an order of magnitude less compute than the labs. You just can’t compete without the compute of the big labs.”

As a new lab, the question isn’t whether you can build one model. It’s whether you can attract the compute and talent required to continuously pump out new generations of models every year and ultimately compete in your domain. I think it’s very hard for unknown players to catch up to the generalized frontier models. You’ll continue to see domain-specific models, or new types of models—decision models like Jev. But the big labs are on the treadmill of continuously pumping out new models. If Anthropic stops, OpenAI and Gemini will continue, and you’ll see revenue fall off. That’s why I think it’s so important to stay at the frontier.

What I’ve seen in the data from Anthropic is that, as their model families get better and better, they get more efficient, but they also capture more revenue. That could be because they’re doing more economically valuable work, or because they’re doing the work faster. But what you’re seeing now, in black and white, is that as Anthropic has doubled its compute spend to build and train these models, it has grown revenue 14-fold. I’m not really sure what would cause that trend to stop.

I think models are only going to take on longer-running, more economically valuable tasks. If you can’t tell, I’m a hyper-optimistic person, so maybe I’m drinking too much of the Kool-Aid. But I do think we’re nowhere close to hitting a limit on the economic value that leading models can capture.

What’s maybe more interesting is that OpenAI is also trying to compete at the low end of the market with much smaller, cheaper, Luna-class models, spanning everything from frontier models to distilled versions of their own models that cost cents per million input and output tokens, not dollars.

Yeah, I cosign all of that. I think that’s a good way to think about it. I’ll just comment on something because it’s in my head: Sonnet 3.7 has been pretty successful. Sonnet 4, that whole series, is very interesting.

Mm.

The approaches of the different labs are also super interesting. Anthropic to this day has no decision model. They’re like, “Well, we’re good. We use TypeSafe, whatever.”

Yeah, yeah, yeah.

They don’t have a standalone voice API.

Right, right.

They have a voice model; they just don’t market it. They’ve never done images. There’s a lot of enterprise focus and coding focus.

Yes.

OpenAI has been a little more like, “We’ll do everything. We’ll try Sora.” They shut down Sora. Remember when Sora was the number one app in—

Yeah, exactly. Exactly.

In the US?

Yeah, exactly. Yeah.

And then Gemini has been the most like, “We will explore everything.”

Yeah.

And we appreciate them because if they didn’t try it, no one else would.

Yeah.

Right?

Yeah.

They have the funds for it. But I think often about Anthropic in terms of, okay, enterprise and coding: they were first to it, and they nailed it. They continue to nail it, and everyone’s chasing after them.

Mm-hmm.

Is that the story going forward? Is this just more of the same, or what should we expect?

2. How Labs Choose Their Markets

I think a lot of this boils down to a new type of innovator’s dilemma, which I almost call the compute innovator’s dilemma. Let’s say you’re OpenAI, and you have Sora, many different model families, image models—you have everything. The reality is, how do you choose what to serve? Maybe Sora demand explodes. Are you going to allocate a lot of your Codex compute to Sora?

To swap videos? Yeah.

Because you probably will have to. If you actually want to make the Sora—

I mean, that’s exactly why they shut down Sora.

Yeah.

Yes.

—for the experience to be good, you need to. At the end of the day, a lot of people have decided that some of the highest-ROI, most economically valuable work has been in coding, so it’s hard to prioritize anything else. That’s probably why there are opportunities for other companies and startups to build models and products in those domains. What are OpenAI and Anthropic chasing? They’re not chasing the incremental $10 billion market cap. They’re chasing the next trillion. So for a startup or a new lab, you could build a company, product, or model and exit for $10 billion. That’s a home run, a grand slam. That’s awesome, and simultaneously, it’s too small to matter to OpenAI and Anthropic.

What has Anthropic done a bang-up job at? They said, “Coding is our thing. We’re going to be great at coding. We’re going to build the best coding harness in Claude Code, and we’re going to continue to nail it.” I think everyone is starting to ask themselves, “When do we tap out the coding TAM?” Maybe you asked a version of this earlier. I think one point of view is—

I have asked this question very, very many times.

—but I think one point of view is that coding is actually the best generalizing—

Knowledge work.

—model and framework.

Yes.

You think of it as coding, but when I use Claude Code, I’m not coding; I’m using Claude Code. When you use Codex, are you coding? No. These are generalizing, and coding is a good framework for generalizing knowledge work.

I do think you’re seeing a big push by both OpenAI and Anthropic into life sciences and drug discovery. They’ve both been quite public about their interests in capturing that domain as a new field.

Yeah.

And I do think those are just different—

Yeah, it would seem like Anthropic is investing more than OpenAI on—

I totally—

—specifically the life sciences stuff.

I agree.

Yeah.

Yeah, I’d agree. I’d agree. But now I can’t remember the question you asked. I do think you’re starting to see these companies evolve the scope of what they want to go after, and it has been very targeted on coding because it’s been the best generalizing. You want to generalize knowledge work, and these models have generalized best as they’ve gotten stronger in coding. But that can also give them a soft underbelly in other areas.

Yeah.

Maybe it’s visual thinking and reasoning. You’re seeing great models there, like Eleutian. Maybe it’s decision models like Jev and TypeSafe. There are a lot of different soft underbellies, which is great. They’re opportunities. You’d love to see better, cheaper models in other domains. You don’t need to drive the Ferrari engine to the grocery store. Maybe a Toyota will do just fine.

Yeah.

And I think that will continue.

And shout-out to Lorien Charot. We just interviewed Andrew on his benchmark and Glimpse. I think Glimpse is the new—are you guys invested?

Glimpse, yeah.

Oh, yeah. Oh, wow.

We are.

Wow, okay. All right.

Shameless promotion.

It’s weird—this is a whole different discussion—but with Altimeter, I don’t get how you lead the giant rounds in the Anthropics of the world, but also care about Lorien as something much smaller, probably worth only a minute of your time.

Yeah. Well, no. We spend a ton of time on them. I would say we barbell.

Yeah.

We’re lifecycle investors. We invest early, we invest late, and we invest in the public markets. And I think there’s a lot of value. The way we think about it is that some of the best founders today need late-stage capabilities early on.

Yeah.

Late-stage capabilities aren’t just capital. They’re relationships with executives and public companies. Companies are built in hyper mode now. It’s not, “Let me go raise a round, add $10 million of ARR in 18 months, and raise my next round a few years later.” You’re growing to $100 million super fast and raising nine figures super quickly, and it’s hard. In many ways, you have to break the laws of physics to figure out how cultures can be maintained—

Like, lawyers—

—at high—

Lawyers are not set up for this. They’re like, “What do you mean, why don’t we come in on a weekend? Because Monday might be different?”

Yeah, exactly. Exactly. You used to think, as a board member and investor, “Let me think ahead. Let me make sure we hire the VP of Sales and the CFO six months before we need to, so we’re not behind.” Now, six months before you need to may have been before the company was even incorporated. How do you hire the right executives so you’re not doing it too late? How do you hire ahead, but not too far ahead?

I feel like the pressure of building companies has never been greater. The size of the prize has never been greater, too. It might appear as easy as ever to build a company, but I actually think it’s the opposite. It’s almost as hard as ever to build a good company, and maybe more importantly, to build a durable company—which is what everyone aspires to do.

Yeah. I’ll also shout out, on the model factory thing, we interviewed Poolside, which got acquired by Nvidia.

Yes.

They were pitching this model factory really hard. I do think the business that builds the business is the general takeaway I want people to know.

One of the things I wanted to get into—and I didn’t quite start it yet, but I’ll start it now—is that typically, when you have an S-1, you have a clear sense of the business: market segments, product segments, the org chart, whatever.

Mm-hmm.

And we’ve got Claude Code—

Yeah.

—Claude Cowork—

Yeah.

—miscellaneous—

Yeah.

—or platform, whatever.

Yeah, yeah, yeah.

And I’m like, okay, what about legal? What about design? All this stuff that you’ve added. Claude.ai, the consumer business.

Yeah, yeah.

For a while, that was a thing.

Mm-hmm. Mm-hmm.

Do we care? Is this all one revenue bucket?

Mm.

What about the products, right?

Yeah.

There’s Claude—

Yeah.

—Code, and what else?

And this is where I do think—and there was a section of the S-1 where they broke out some of these vertical applications they want to go after. It’s kind of funny you bring this up. I had a hypothesis years ago that maybe the labs would stop releasing models and instead just build applications and products powered by their models.

Yes.

If they’re powered by the best models that are never released publicly, maybe they have the right to build all the best applications.

Yeah, you kill Harvey by only—

Mm-hmm.

—using Claude, Fable 6 Legal.

Right. Or it’s, “Hey, you know what? Cursor? Guess what? Years ago, we got the best model, and we’re not releasing it. So guess what? We’re going to have the best coding agent because our most powerful model isn’t released publicly.” I do think that’s why some of the leading application companies today are moving toward in-housing their models.

Yeah.

They want to control their own destiny. You need to. I think it’s going to be really hard for an application company to survive over the long haul without an internal research team. One of my investment theses is that if we’re investing in applications, I want to believe they have 2 things: a really strong product organization and a really strong research organization. I also want to believe they’re aligned, and that as the research gets better, the product gets better. That sounds counterintuitive, but a lot of research organizations are just doing research for the sake of research, rather than closing the loop and having a positive cycle of product improvements with research improvements.

Yeah.

I’m not in Anthropic, but I have the benefit of being friends with a fair amount of people there. I would say research is king at Anthropic.

Yeah, totally. Totally.

By far.

Totally.

Part of my whole thesis around model labs versus agent labs is that model labs will respect the model and AGI, and try to build AGI, and products will be the final last mile—

Yeah.

—the delivery mechanism. Whereas agent labs will respect the customer the most and build whatever they need, and then they’ll backward-integrate to models as and when.

Mm-hmm.

But they’ll probably never be at the frontier for those things.

Yeah.

And both can coexist.

Yeah, totally.

Both have a market.

Totally.

But, yes.

Maybe to your question, I do expect to see a much bigger push by both labs, Anthropic and OpenAI, into vertical applications and diversifying beyond just delivering models. I think—

I think Anthropic’s actions so far have shown no serious intent at it.

That’s true. Yeah.

Because Claude Design is just a tab inside of—

That’s right.

—Claude.

That’s right.

It’s not a whole Adobe replacement.

Mm-hmm.

Claude Legal is a Markdown README file, you know, Markdown repos.

Mm-hmm. Mm-hmm.

That’s the AGI pill: if your research is king and one model does everything, that’s what you do.

Yeah.

Right? So—

Sometimes there’s classic enterprise and organizational management.

Yeah.

You have a cash-cow product, and think: what do your sales reps want to sell? The thing they know how to sell, that they’ve been selling, that’s crushing it.

Yeah.

What’s harder to sell?

New things.

New things, new SKUs. So it’s inertia, momentum, and just classic big-company boring stuff.

But I think for someone like me, who’s been through maybe 15 years of IPOs, for an IPO-stage company—much less a $2 trillion company—you’ve never seen, “Well, we have 3 SKUs.”

Yeah, exactly. But I think to Anthropic’s credit, they’ve been really good at focus. They haven’t been enamored with the next thing or—

No robots.

Yeah.

No images.

Again, you look at OpenAI, and it’s tough. Who are they competing with? They’re competing with NVIDIA, Google, Meta, and Anthropic. They have different lines of business that are all competing with giants, including the biggest tech companies in the world. It’s a tall task to take on one at a time, let alone all of them at once. But that’s the game you take on when you’re a trillion-dollar company. The stakes are just higher.

Yeah. Okay, I want to round out the Anthropic discussion, and then we can go to general commentary on other stuff. We’ve cheated a bit with Ellurian and TypeSafe. Have we covered the main debates you’ve been engaged in?

There are 2 others. This is going to be very financially in the weeds. You’re going to talk about convertibles.

Yes.

Okay. On the balance sheet, earlier in the episode we talked about $140 million of capital raised and $134 billion of cash. That’s what the balance sheet shows you. What the balance sheet also shows is about $108 billion of convertible notes payable as a liability. So the question is: Is that debt? Did they raise debt? You also see a huge portion below operating income of other income and expenses that makes it look like they have a huge net loss. They’re both related, and let me explain. Any company—and any startup founder listening to this would understand and appreciate this—has different ways to raise capital. You can raise preferred equity, convertibles, SAFEs. There are a whole host of options. Anthropic, like every other startup, did everything: they raised preferred equity, convertibles, and SAFEs. There are different reasons certain investors or the company might prefer one or the other.

And—

Early on, Anthropic and Google, in 2023, 2024, and 2025, provided a lot of dollars to the company in the form of convertibles, SAFEs, and notes. The value those converted at was pegged more to the valuation the company was raising at at the time. Why might they prefer a convertible? Google and Amazon are public companies. If they have an ownership stake in a third party that exceeds a certain limit, they have to report more about that look-through ownership. If you own 20% of another company, you almost have to—

It’s consolidated financial statements.

Exactly.

Yeah.

And if that company’s losing a lot of money, you now have to report that loss on your books.

Oof.

Strategic investors don’t want to hit certain ownership thresholds. So, “Hey, give us a convertible, give us a note. We’ll give you cash. You haven’t given us shares yet, so we don’t have that equity ownership. We haven’t gone above those thresholds.” Maybe Anthropic didn’t want to give them voting shares. I don’t know. Maybe it was, “Let’s have your shares convert into non-voting shares, because we don’t want the governance of a supplier or a strategic.” I’m totally speculating and hypothesizing.

Those are all reasonable reasons to do it.

They’re reasonable reasons why they were convertibles. But the conversion price was at a much lower valuation relative to what it is today. When you look at the balance sheet, it has to balance: assets equal equity plus liabilities. When you raise a convertible, you get cash, so your assets go up. For the balance sheet to balance, your liabilities also have to go up, which makes sense. A convertible is a promise of shares in exchange for cash, so Anthropic took on that liability. At the time, the liability matched the cash raised. What’s happened to Anthropic’s valuation since 2023 and 2024?

200x.

It’s gone way up. Now the dollar value of your liability has skyrocketed because you promised shares at a certain price, and the valuation of the business has gone up. So now the liability looks big. It looks like $108 billion of debt that Anthropic raised. The reality is, they didn’t raise $108 billion of debt.

There was about $11 billion in those combined convertible notes. About $5 billion has already been converted to equity, and there’s $6 billion outstanding. The market value of that $6 billion is $108 billion because the company’s value has appreciated meaningfully. There’s this debate: “Well, they have $108 billion of debt on the balance sheet. They haven’t raised $141 billion; they’ve raised $141 billion plus $108 billion.” That debate is misguided. There is no debt on the balance sheet, and the $11 billion of converts is included in the $141 billion of total capital raised. That’s the first debate: have they really only burned—

Can I just say, good job summarizing this?

Yeah.

Uh—

But it actually goes deeper.

Yes.

That’s just the balance sheet.

Okay.

Now let’s look at the income statement, the P&L.

Yeah.

They have an operating profit. You could say, “Great, GAAP operating profit,” but what comes below operating profit? Net income. Someone could say, “Well, in the first 6 months of 2026, they had negative $80 billion of net income.” So you could say—and this is a debate, this is what people are saying—it’s very misguided: “They’re losing $80 billion. What a terrible business.”

Oh my God.

Technically, by GAAP accounting, it is.

It’s true.

But let’s go back to the balance sheet. The liability has gone way up. That liability is now $108 billion. Let’s think about the balance sheet: assets equal liabilities plus equity. As liabilities have gone up, 2 things have to happen: either your assets have to go down or your equity has to go down. Your assets haven’t gone down, so the value of your equity has to go down. This is where it might get a little complex, but how the 3 statements link is—

Yeah, people can take Accounting 101. It’s—

Net income—

Yeah.

—on the P&L flows to retained earnings on the equity side of your balance sheet. As that liability has gone up, retained earnings has to go down by the same amount, which basically means net income has to go down by the same amount, because net income on the P&L flows to retained earnings on the balance sheet.

It’s all mathematically linked.

And so—

You can’t break it, yeah.

It’s all mathematically linked. You can’t break it. But in order for those statements to link and tie, you need a huge decrease in value on your income statement. What happens is, guess what, Anthropic’s value has just skyrocketed recently. They’ve raised big rounds over the last 9 months, and as those converts have appreciated in value, that appreciation shows up as a loss on the P&L. Of the $81 billion of loss in net income in the first 6 months of this year, $72 billion is from a series of converts for Anthropic from AWS and Google, and $9 billion is from a separate set of SAFEs and notes. So 100% of the loss in the first 6 months is just the appreciation in value of their converts. If Anthropic had never issued converts to Amazon and Google and had just done straight equity, that loss wouldn’t be there.

Yeah.

So, technically, by GAAP accounting, they do show an $80 billion net loss. Do I view that as an accurate representation of the business? No, because 100% of that loss is just the appreciation of their equity—

Yeah.

—which everyone would say is a good thing. As those notes convert to equity, that goes away because you no longer have this dynamic of your liability going up as your stock price goes up, which means your net income has to go down over here, so all the financial statements tick and tie. Sorry if that got too technical and in the weeds.

No, that was a really good recap of the discussion. That should be a definitive answer for anyone who has come across that topic. There’s another discussion about the circular financing of Google and Amazon. We don’t have to go into that. I think that’s small in comparison to the numbers we’re talking about today.

Yeah.

In general, book value accounting is not that relevant compared to overall enterprise value.

Mm-hmm.

And—

Oh, and maybe one other thing related to this.

Yeah.

I think there are a lot of people who say, “Okay, well, the only reason the liability’s gone up is because the value’s gone up.” Does that mean Anthropic issued shares today and gave Amazon and Google $108 billion worth of value, but gave them those shares at the price—

Yeah, no—

—from 4 years ago?

They just—

Did they just give them shares at a huge discount? No, that didn’t happen. What happened is they raised that money years ago in normal financings, like a normal startup would, and over time it’s just appreciated in value.

I think a lot of people just aren’t financially educated to that level because they never had that exposure. So—

Yeah.

—it is what it is. You’re just educating people. What was the second debate you were referring to? You said you had 2 debates.

It was related. It was—

Oh, there’s the other one.

—the balance sheet.

Oh, wow, okay.

Have they raised more debt? And then there was the income statement: are they actually losing money?

As an investor, why do you get involved in these kinds of companies?

This is more of a… I mean—

You’re just having fun.

You know, yeah.

Okay, all right.

I think Anthropic… Look, Altimeter is a firm. We do venture investing, growth investing, and public investing.

Yeah.

I spend a lot of my time on venture investing, the early stage. However, I consider myself a student of finance, and Anthropic—

And—

—is such an iconic company.

You shouldn’t have to educate people on the 3 statements of accounting.

Yeah.

That’s like—

Yeah, but I think a lot of people look at the balance sheet or the income statement, the P&L.

Yeah.

But there’s a lot to learn from the balance sheet and the cash flow statement, too. This was more of a, “Hey, as someone who’s intellectually curious, Anthropic is a really important company to understand.” If the Anthropic S-1 came out and the financial profile looked doomy and didn’t look like a good business, I think we’d have a lot to be worried about.

Yeah.

If they’re not delivering profits and high ROI, what does that mean for everyone downstream of them? I think it’s really important to understand the Anthropic P&L. Here’s what you’re seeing about their gross margins: lots of leverage, and they’re improving rapidly.

One nuance on this is that this is the half that is not through cloud partners.

No, this is just the GAAP gross margin.

Right. Okay.

Is that… okay? Is that—

I don’t know. I read something about how, obviously, cloud partner margin would not go—

So—

—into gross margin at all.

Yeah, I believe a lot of the costs they pay out as revenue share are in sales and marketing—

Yeah, yeah.

—which is below the line, which I think is what you’re saying.

Which, by the way, is $3 billion—

Exactly.

—in sales and marketing.

I think 40-something percent of the revenue comes from cloud partners.

Yeah.

It’s not a trivial amount.

48%.

48%? Okay.

Yeah.

Yeah, I was close.

Insane.

You see the gross margin profile improving. Sure, maybe it’d be lower if you factored in the third parties, the cloud providers, but the trend is going up. You see operating margin going up, with positive GAAP operating margin. You see free cash flow margin going up. So, to me, what I see is a highly profitable business that’s probably going to shift to reinvesting in compute and building out. I think the way compute gets built is going to look different in the future. It’s going to be expensive, with more upfront payments. I do think you should expect to see the profitability profile of these labs go the other way in the short term as they shift to investing in growth and investing in compute.

But the biggest takeaway for me was that it’s a highly profitable business that’s still operating in growth mode, and that it’s seeing some reacceleration, probably in the third-party data from the new models. That gave me a lot more confidence in the rest of the industry.

Yeah.

So long as the revenues show up, the compute obligations will be met, the supply chains will continue to be saturated and constrained, and companies will continue to perform. OpenAI and Anthropic are a huge percentage of overall AI revenues, so everyone looks to them as a proxy for the rest of the industry. If they keep growing—which a lot of the bears would say they’re going to stop doing because open-weight models are going to eat their lunch, and this and that—I think those are the important questions.

Yeah. Very important, very good questions. Do you dabble in forward projections of numbers? What should we expect from FY ’27? What should the—

Oh, I think it’s too hard.

—Street expect? Okay. All right.

Yeah, I think that’s too hard.

Because the delta is very high, right? Given the year they just had.

It’s a business that’s so hard to project because the future is so unknown and unknowable. They’ll go public, and the sell-side research analysts at all the banks will build their models and come up with a forward number. I don’t know where that’s going to shake out. I do think you’ll see a lot of optimism about what the future of this business could look like.

But would it surprise me if it was much higher than expected? Anything could happen. Anyone trying to predict the future, especially in the short term—it’s as hard as ever. I think businesses at the frontier are the most predictable because they’re the least likely to be disrupted. But I don’t know. We live in a highly uncertain time.

Yeah. The A-plus financial analyst as a PM that I would have wanted in my hedge fund would do physical-limits calculations—

Mm.

—on, okay, you have all these inference optimizations. How much juice is left on the current thing?

Yeah.

Let’s say the next gen has this level of inference. How much can we charge for that, given the competitive profile and exhibited pricing history of these things?

Yeah.

And then finally, how much GPU supply has been locked in and secured, so that translates to this number of tokens, and that’s what it can sell for? Because those are the physical limits of where these things can go.

Mm.

I think a lot of financial analysts—and I used to do this—you’d have a spreadsheet of past performance, and you’d click the right corner and drag it out.

Yeah, yeah, yeah, exactly.

That’s not how it works.

Can’t do that. I mean—

That’s not good analysis, right?

Totally. I think there are limitations everywhere.

Yeah.

Everyone knows Nvidia can only produce so many chips.

Yes.

Challengers like Cerebras and TPUs can only produce so many chips. But I think there are other constraints. There’s only so much land, power, and shell that’s approved and permitted.

Correct. So you do that buildup, right?

And at the same time—

And then you’re like, “Okay, who has what share?”

3. Debt Markets Hit Capacity Limits

When we talked a little bit about debt and credit markets, if you want to go raise multibillion-dollar debt for GPUs right now, there’s no bid. No bid. Debt markets are closed. Big debt deals aren’t happening, almost the equivalent of an IPO getting pulled. Big debt deals aren’t happening.

So why is there $100 billion in equity but no $100 billion in debt? Because I go, I’m gonna turn around and put those—

Well, I think right now—

—money right in the data center.

These credit funds are all looking at the look-through exposure, and for any NeoCloud or hyperscaler that’s raising debt, the biggest question the credit providers are going to ask is, “Who’s the offtake?” The reality is that’s who I’m underwriting—

Yeah.

—because that’s the customer. That’s the one paying; that’s the end payer of the bill. If the end payer does poorly, the bills don’t get paid. Right now, the end payers are Anthropic and OpenAI. Many of the data centers being built by the NeoClouds are for OpenAI and Anthropic, and a lot of credit providers are saying, “I’m feeling kind of full. I just did a ton of debt for one of these.”

It doesn’t matter if it was CoreWeave, Nebius, an upstart like Volta, or someone else. It’s, “Hey, I’m feeling full.” I think debt markets are going to go through a little bit of indigestion. They need to figure out: Do I feel good? Do I feel great? Are some of these projects that were promised being built on time? Is the revenue flowing back to me?

I do think credit and debt markets will ebb and flow, because there are a thousand reasons why a project could get delayed. A NeoCloud could sign a contract with OpenAI, sign with Nvidia, have Dell package everything up, and raise the debt. Then there could be a local community uprising and picketing. I’m literally not even kidding. Then it’s delayed. What happens? The debt was raised, and someone has to pay the interest.

Yeah.

You have creative delayed-draw facilities where the debt is committed, but you don’t draw it down. But all of these have limits, and you get super in the weeds on SLAs and who was responsible for it. Was it a delay in the supply chain? Was it a delay in construction by the data center developer? Was it a delay by the operator? Was it something else, like community unrest? That’s everything credit markets are thinking through.

Yeah.

And then there are a lot of other technical reasons we don’t need to get into that make things difficult, with credit default swaps and hedges going through the roof. For debt providers, it’s tough to hedge the exposure they have. As the hedges get more expensive, your willingness to take on more risk goes down because you can’t hedge out of it. It gets super complicated.

These are all symptoms of the debt market running out of money.

Yeah. Well, I don’t know if debt markets are running out of money. Debt markets are super deep.

Yeah, yeah.

I think it’s more about appetite.

Yeah.

And of course, there’s a price for everything. Hey, great. You know what? Go take out $5 billion of debt at 20%.

But you see this at the equity layer. Again, how come there’s $100 billion of equity that’s going to go into Anthropic, and Anthropic cannot raise $100 billion or $50 billion?

I don’t know if I would say they can’t do it. They need to be rated. Do they want to?

Yeah.

My guess is Anthropic and OpenAI are happy that NeoClouds exist to take the balance-sheet risk. Maybe they don’t want to take that balance-sheet risk, because things could happen—

Exactly.

—out of their control that blow them up. You have a huge debt stack in your cap table, and things out of your control go wrong—

Someone else’s problem, yeah.

—you’re not the one holding the bag. And again, that’s the beauty and the curse of why NeoClouds can be great businesses or brutal businesses if the contracts aren’t airtight and things aren’t structured well. It’s a balance-sheet business. Just like fintech balance-sheet businesses, it can ebb and flow. There are a ton of—Oops, my computer may have died.

It’s fine. We’ll edit.

Okay. Balance-sheet businesses, right? There were a ton of fintech companies in 2020 that did super well when interest rates were low and got blown up when interest rates were high. Balance-sheet businesses can be really tricky and very cyclical, and you don’t want to own a balance-sheet business when a market cycle tips over.

At some level, I’m sure that, especially for internal chip efforts, the OpenAI team might want to develop its own data centers for its Jalapeño chips, and Anthropic might want to do the same for its chips. But I’m sure they’re also very happy for other people to take on that risk, because it’s someone else’s core competency, and a lot of debt could add a lot of blow-up risk.

There was no debt on Anthropic’s balance sheet. Now, they have facilities in place that they could draw down, so would it surprise me if you start to see debt on their balance sheet in the future? No. But again, you want to build a gigawatt, right? That’s $50 billion, $60 billion, $70 billion. That’s not a trivial amount for just 1 gigawatt. I think both OpenAI and Anthropic have said they want to get to 5 by the end of the year. Who knows if it’s that, or something more or less than that?

Are they about equal?

I don’t know.

Okay.

That’s interesting.

Yeah.

Because I think Anthropic—

OpenAI was ahead.

Correct.

Yeah.

Right? By a lot.

Yeah, exactly.

And I think Anthropic’s closed the gap a lot. OpenAI at least started with Stargate, which was a plan—the first stage in a plan to get to 30—

Mm-hmm. Mm-hmm.

—30 gigawatts.

Mm-hmm. Mm-hmm.

And so—

Yeah, by 2030.

Yeah. I haven’t heard that plan from Anthropic.

But again, let’s do some fun math really quickly, on the fly. Let’s see if I can do it. I’m getting tired. OpenAI, let’s say, when they came out with that plan, maybe a year ago, was at about 1.5 gigawatts, something like that. So whatever, let’s round that down to 0 for the sake of the argument.

Let’s say OpenAI and Anthropic each go from 0 to 30 over 5 years. So you’re going to add 60 gigawatts of AI capacity, just the 2 labs. What’s a good assumption for their total capacity? I don’t know, 40%, 50%. Let’s say it’s 40%. I’m getting too deep into the math in my head, but let’s say that gets you to 110 or 120 gigawatts of total capacity brought online over the next 5 years. Let’s say 100, just so the math is easier.

Yeah.

And let’s say it costs $50 billion to build out a gigawatt. That’s $5 trillion. $5 trillion over the next 5 years, so call it a trillion a year. You’re talking about a nontrivial percentage of GDP—real GDP. That’s about 5%, plus or minus, of GDP. That is a massive—

Bonkers.

—massive build-out. It rivals the railroad build-out. I’ve seen some charts that compare those expected numbers, but when you’re measuring things in whole-number percentages of GDP, that’s huge. The credit markets and debt markets need to be super, super, super deep and super, super open for that to happen. You need a global build-out.

It’s kind of funny. The SemiAnalysis number for next year is around 50, and I think Brad at the All-In Summit said, “Ah, maybe it’s lower.” I think there was some trolling going on back and forth. The end result was, maybe the 50-gigawatt number—I don’t want to speak for anyone—was more about what’s constructed. Brad was talking more about what’s constructed and live, generating tokens, because there can be a gap between what’s constructed and ready to go versus what’s actually live and generating tokens.

I think it’s a real debate in the world, and you could see a pretty wide range of outcomes. Do we really add 50 gigawatts of capacity next year as an industry, or is it 20 or 30? Who knows? I just know that everyone you talk to in the supply chain is saying, “We have orders or indications for orders that are well over 100% of our capacity.”

Yeah.

I don’t have—

That’s true.

—line of sight to that slowing down for the next few years.

Yeah.

And—

And the limits are land, shell, power, all these—

Yeah.

—but also silicon, straight up. Compute, memory. Every—

Totally.

—every part of the supply chain—

Totally.

—that’s the work. That’s why SemiAnalysis is so powerful right now, because—

Look at these recent Kimi models. Kimi K3. Do you know how many gigabytes of memory you need just to serve that? You’re not running that on your laptop. You need a rack to serve one model.

Yeah.

Again, those are just the big models, right? You’re also going to have tons of small models. But I know the big models are getting bigger, which only requires more memory and more compute to run.

I used to say they were capping out. I used to say the max size was 1 to 2, and then we’d just optimize from there.

Yeah.

What’s interesting is that it does feel like open-weight models are capping more around 100 billion active parameters.

Yes, that’s right. Yeah.

And—

That’s mostly an inference-speed thing. But I think—

Totally.

—it’s actually really interestingly tied to the generations of GPUs they have.

Yeah. Yeah, exactly.

Right? Because you—

Exactly.

—wait for the Vera Rubins to roll out—

Mm-hmm.

—and whatever China has.

Yeah, exactly.

And, well, then they’ll get it, you know?

Yeah. Totally.

They’re really memory-constrained at the inference layer, and that actually shapes the model architecture, which is very cool.

Totally.

Oh, so that is the Anthropic powerhouse tour.

Yes.

Congrats. I don’t know how you felt coming into this, but you did great.

Oh, this was fun.

And there’s obviously more to do and more to analyze, but I definitely wanted to do this on the podcast because I know you.

Yeah, of course.

When else were we—

You’ve been meaning to do this for a while.

—when else were we going to get this in a pod—

Glad we could finally make it happen.

Yeah, exactly.

Yeah.

Exactly. Any other commentary? When you don’t spend time on the labs, what are you spending time on?

4. Where Altimeter Sees Opportunity

In a world that’s highly unknown, unknowable, and uncertain, I fall back on the core beliefs I have conviction in: the world is going to consume more models, more power, more compute, more networking, and more storage, and we’re going to need to build a lot more data centers to serve the intelligence we’re building. Those are my 6 core truths. A lot of what I’m looking for are businesses that are a direct derivative of one of those core beliefs.

Yeah. Any other expressions of those core beliefs that you’ve done recently?

Recently? There’s a chip company still in stealth that’s super exciting. Maybe we can talk about it later.

Okay.

There’s a power company, American Terawatt. They want to build a new grid, an industrial grid to serve industrial offtake. Big, huge, bold vision. We talked about Alorian[?] and Glimpse, and visual thinking—

Visual reasoning.

—visual reasoning.

The compute market side, effectively slicing out compute to increase efficiency.

More of the—

Trading futures. Trading futures on compute, all that stuff.

Oh, I haven’t spent as much time there.

Yeah.

I haven’t spent as much time there.

We spent some time with Ansh, who’s doing National Compute, AMP Grid, and SF Compute and all those guys. In New York, we launched a compute trading track where—

That’s right.

—mostly it’s just to increase utilization—

Yeah.

—is the broader goal.

Yeah.

That’s how you can magically create a whole bunch more compute, by making better use of the compute we already have.

Yeah.

Totally.

If MFU—whatever choice of utilization you use—is at 30% to 40% right now, taking it up to 60%, you’ve doubled—

Yeah.

—the number of GPUs you have.

Yeah, totally.

And so I think it’s—

I think there’s going to be so much cool innovation. It’s why this is the most exciting time to be in tech, to be investing, to be around all of these startups. You’re seeing innovation at the silicon layer, the capital markets layer, and the product layer. It’s creative destruction everywhere.

Yeah. I notice you’re not doing much on the application layer.

You know, look, I would say—

The agent layer. Yeah.

No, certainly are. Again, I’m speaking more personally than necessarily on Altimeter’s behalf. I want to believe there’s a strong product-research-org combo. There are certainly a lot of great teams, and early on in market cycles, we use these triangle analogies where you have compute or infrastructure—hardware infrastructure, software infrastructure, and software applications.

In a mature market, the biggest TAM is in apps, and the smallest TAM is in hardware. Early on in cycles, it’s inverted. The biggest markets are at the top or bottom of the triangle: physical hardware, then infrastructure, then apps. Eventually, the triangles invert. At some point in the future, I think apps and agents will be the biggest part of the TAM. We’re just early on, and there’s lots of opportunity in the other parts of—

Yep. Objectively, that is correct. I do feel like—and by the way, you can see various versions of this in all of computer history.

Right down to when hard disks were getting built out or whatever. Anyway, I don’t think those are the best examples. Western Digital?

Western Digital.

Right. Anyway, I think that is true, but to me, domain specificity is very interesting, and the app layer, the agent layer, is supposed to capture that.

Mm-hmm. Yeah, totally.

The biggest outcome here is Cursor right now.

Yeah.

And that’s $60 billion. We’re here talking about $2 trillion. I don’t know.

Yeah, exactly.

Very, very different.

Yeah, totally. I think I also skew a little bit more toward infrastructure. Just personal bias.

Yeah, we’re infrastructure guys at heart.

Yeah, exactly.

Okay, great. My closing question was going to be: What’s your favorite Brad story? Brad’s a character, and we’ve worked with him for, I don’t know, three or four years now.

Six years.

Six years.

Yeah.

Okay. I feel like I’m getting my timeline wrong, or I met you way earlier than I thought I did.

No, I think we met earlier, because I think I was still at Redpoint when we first met.

Oh my God. Okay. All right.

Yeah. Where were you? Netlify? No.

Airbyte.

No, before that.

Before that was Temporal.

Before that.

Netlify.

That’s when we met.

Really?

Yeah.

Jesus. That’s ancient history.

Yeah.

Anyway, Brad’s a character. I just think he’s interesting. What’s a good story you have?

Oh man, you put me on the spot. Honestly, my favorite Brad story is just how much he’s worked on Invest America. I don’t know if people know this initiative. It’s the idea of how we can get people invested in capitalism early in their lives.

I think most people in America have a negative view of capitalism because they haven’t participated in it. A lot of people view capitalism as a great thing, and they’re the people who have participated in the equity upside of America and compounded their wealth. But most people don’t have that luxury. So how can we educate people with firsthand experience about the benefits of capitalism by helping them participate early on?

The idea is an investment account funded by the U.S. government at birth that anyone can contribute to along the way. That could be philanthropists like Michael Dell, who made a $6 billion donation. I think Elon Musk has pledged one share to every kid in America, which I think is $10 billion.

Anthropic has some version of the public benefit fund or something.

Exactly. I think just seeing him work on that behind the scenes for years, and seeing it come to fruition very quickly, is the classic overnight success. There isn’t one specific story. It’s more about seeing the grind that goes on.

It’s a cause he believes in. And is it done, by the way? I actually don’t know.

It’s passed into law.

It’s passed. So what? If I have a kid now, they get it?

You could open an account. Any new kid born gets $1,000 at birth. Parents can contribute to it, employers can contribute to it, and friends can contribute to it.

So it’s like a baby 401(k).

Exactly. You have people like Michael Dell saying, “Here’s my donation, and I want it to be gifted to”—I can’t remember the exact specifics, but I think it was for the bottom 75% by income and ZIP code. The money just—

That’s better.

Exactly. The money just—because, as Brad would say, 529 plans are education plans. I can fund money that can be used for my kid’s education one day. The reality is, the people using 529 plans are generally the wealthiest part of the country. 529 plans and the tax advantage weren’t made for the wealthiest part of the country, but that’s who uses them the most. What he wanted to do was create investment accounts for the people who actually need them.

I think it’s an awesome initiative that’s just getting started. I think you’re going to see a lot more direct philanthropy. A big issue people have with philanthropy is, “You’re going to donate a bunch of money, and then middlemen are going to take half of it.”

I have seen—

And there’s—

Literally, directly, this—

Tons of loss.

It’s used for corruption, straight up.

Exactly. But what’s better? Direct contributions, straight to the bank accounts of the people who need it most. No middleman. Direct accounts that people can see, appreciate, and value over time. You can’t touch it until you’re 18 years old. You get to see the appreciation happen over time, and maybe learn and participate in the upside of America along the way.

I think it’s really caught on, and seeing how passionate he is about it has been inspiring.

Okay. My one pushback is: I’m homeless on the streets. Food stamps, everything.

Yeah.

I have $1,000 sitting there.

Yeah.

You’re like, “Well, give it to me.”

It’s made more for kids now. Hopefully you’re not—but maybe there are plenty of homeless kids.

Yeah, exactly. So that’s the thing. It’s like, “Well, did you know you have a bank account and you’re participating—

So that—

—in the growth of capitalism?”

That is the biggest thing.

I don’t care, you know?

That’s the biggest thing now: people need to claim their accounts. The government opened up accounts for every kid in America, and parents need to claim them. The money is sitting there for newborns.

Yeah.

If you’re 15 years old, money doesn’t show up in your account. I think donations will, so you could be a 15-year-old with a $250 donation from Michael Dell sitting in your account. But those accounts need to be claimed. The focus has gone from getting this passed into law, getting it up and running, and setting up the infrastructure so people can donate, to converting those—

Yeah, this is execution.

—created accounts—

Yeah.

—to claimed accounts.

Yeah.

And that’s just a grassroots, word-of-mouth game.

Yeah.

Because there are people whose accounts are sitting there with money in them—

Yeah.

—and they just need to be claimed.

And you didn’t know. Yeah, yeah.

Yeah.

Yeah. And by the way, most people have money sitting with the government that has not been claimed. You can look it up.

Yeah, unclaimed property.

Unclaimed.

It’s a great use case for Muse and Instinct and—

Have an agent go look up your name.

—and have an agent. Exactly.

You probably have a few thousand dollars sitting somewhere that someone returned to you, and it’s just sitting there.

You just need to know your address.

Yeah.

Your prior addresses.

Yeah, yeah.

And then go have Instinct or Muse—

Yeah, the highest ROI on this podcast is that.

Exactly.

Not the Anthropic investment stuff.

Exactly.

Yeah. Cool. Thank you so much for spending your time with us.

Yeah, of course.

This was great. We’ll have you back. We’ll see how the IPO goes.

This was fun. Thanks for doing it.

Thank you.