OpenAI 想获得美国联邦政府的帮助｜Sharp Tech with Ben Thompson

Have a dissenting voice. But first and foremost, there's Aaron, who says:

“It seems like OpenAI CFO Sarah Frier finally said what we've all been thinking. The money for the trillion-plus dollars of deals OpenAI has promised is simply not there. It seems our largest, most successful companies are now content to openly ask the U.S. taxpayer to subsidize their risk. There certainly exists an argument for public assistance for certain parts of the AI stack, particularly power. But are we really already at the point in the bubble where we're asking for bailouts?

“Most would agree that when the bubble eventually pops, Nvidia and OpenAI are poised to experience some of the largest falls from grace. With the amount of hubris shown by both companies, I wonder which one will deserve it more.”

So, Ben, lots of different directions we could go here, but I'll just repeat Aaron's question: Are we already at the point in the bubble where we're asking for bailouts? What do you think?

1. OpenAI Wants a Backstop

No, I think that is a mischaracterization of what was discussed here, which is not to excuse the overall request. It is a little bit of saying the quiet part out loud. So, again, Sarah Frier, the CFO, was out with some sort of clarification: “We're not looking for federal support.” I didn't know what that word meant. Blah, blah, blah.

What I believe she was referring to is financing. We discussed this in the context of Nvidia—or I wrote about it in the context of Nvidia. I can't remember if we talked about it on the podcast—with some of the deals they've done with other companies, like some of the neoclouds. They've guaranteed to buy a certain amount of capacity through a certain number of years. The goal from Nvidia in that case is not to actually buy that capacity.

The goal is to allow the neocloud to get a lower rate of financing to buy Nvidia chips, with the assumption that the neocloud is going to rent those chips to someone else. But the risk that a financier would have is, “What if you don't find customers? If you don't find customers, you're going to go bankrupt, and we're going to lose all our money.”

Nvidia is saying, “No, they have a guaranteed customer, which is us. We will rent the capacity if no one else rents it.”

And that should allow the neocloud to have a lower interest rate on its financing because it has a guaranteed source of revenue.

Mm-hmm. I think that's what she was referring to. What OpenAI is doing—and this is what I said this week, and I wrote about it—is that, obviously, we're in a bubble. The marker of shifting into a bubble is when debt starts entering the equation.

All these deals, to the extent they happen, are getting too big for traditional VC. It's not going to just be equity; there's going to be debt. Debt is a very good match, at least in theory, with upfront infrastructure investments that pay off over time, because you need to build it, and then you have the low marginal cost of serving it in the long run. You can pay off your debt going forward.

The reason VC companies don't take on debt is because the uncertainty is very high. Their interest rates would be sky-high. It would consume their free cash flow. So, they give away equity, which means they don't have to make interest payments and they don't have to pay it back, but they do lose part of the upside. They have to share the upside in the long run. That's just the fundamental difference between debt and equity.

In theory, once you're established and you have predictable cash flows, debt is better because you're not giving away your upside. All you're doing is shifting time. Debt is a way to pay for the stuff now with revenue that you're going to earn in the future by virtue of buying the stuff now.

The issue is, if you're going to take on a bunch of debt, the interest rate really matters. What I suspect this reference was to was that OpenAI would love for there to be an understanding by whoever is issuing debt to them that, no matter what, the U.S. was going to be good for it—

The U.S. was going to be good for it because—

That's right. (laughter) I mean, it was part of a discussion of banks and government. I think, obviously, it was an error to say this. It was an error, even if you wanted to say it, not to clarify what you were talking about, which is financing.

That's the thing. I honestly feel for Sarah Friar because the word “backstop” conjures the federal government bailing out big banks or something like that, which is not really what she was describing.

I don't feel too much sympathy for her because what she was doing, I think—and I sort of said this line before—was saying the quiet part out loud.

Yeah. Well, some kind of public-private partnership does feel inevitable somewhere along the line here.

Well, what OpenAI, I think, is doing with cutting all these deals with anyone and everyone is tying everyone into its web so that no one can afford for OpenAI not to succeed.

Right? There's systemic risk if OpenAI goes, right?

They're trying to make themselves a systemic risk.

And thus have guaranteed loans.

Yeah, right. And so—I'm sure they've talked about this internally—if we make a deal with Oracle, we make a deal with Microsoft, we make a deal with Google, we make a deal with Amazon, and all the like, if we become so essential that we're taking these crazy risks, it's in everyone's interest that we succeed.

It goes back to that story: If you owe the bank $1,000, the bank owns you; but if you owe the bank $1 billion, you own the bank. I think that phrase originated at a time when $1,000 and $1 million were more meaningful. We definitely need to update the numbers.

And a trillion dollars, and OpenAI may in fact owe a trillion dollars somewhere.

That's my overall read on what they're doing to the system as a whole. Obviously implicit in that is, once you get too big to fail, it is valid to draw the line back to banks. So, I feel bad for her in that she shouldn't have said the quiet part out loud, and it sucks.

There's a way to use less loaded terminology that doesn't turn this into a national news story this week.

Maybe it should be a national news story, because what they are doing is—we are all getting signed up for this whether we know it or not.

It was one of those things where it's definitely an error. I can feel bad for anyone making an error, but in my position as an analyst, I'm not going to say, “Oh, she misspoke. That's not what she meant.” I think that's exactly what she meant.

And so, again, it's not a bailout. What's the line? What's the difference? What I think it's driving for is that they want cheaper financing. The best way to get cheaper financing is basically to have the U.S. government as a backstop. That's literally what she—

2. OpenAI Should Go Public

I mean, let's ask a more fundamental question, because I found it interesting as more evidence that paying for the compute necessary to serve AI to massive audiences may be impossible if you don't already have monopoly profits to fund what you're doing. That's been a question with OpenAI all along.

On the other hand, I understand why people would recoil at this idea. But if you accept the premise that leading-edge AI is essential to national security over the next several decades—which is definitely debatable—but developing leading-edge AI may only be possible while serving a massive consumer business, then in that scenario, I don't think it's that crazy to have the government de-risk some of the lending in order to advance the future of the ecosystem. Do you?

Yeah, no, I'm not a fan of that. Where do you draw the line? “Oh, this one just happens to be so expensive that it has to be done.”

That discussion came in the context of Sarah Frier saying that an IPO is not coming soon.

Mm-hmm. If you want a massive amount of funding—

Do an IPO and issue a bunch of stock. Make your case to shareholders about what they're buying into.

And that comes with certain responsibilities.

That comes with a decrease in equity because you're literally selling equity, and it comes with all these reporting requirements. If you're going to be so systemically important, then I think you should be a public company, and you should have the transparency that goes with it. I actually think the context of these comments makes it worse.

She said specifically, “An IPO.”

I'm going to double down. I completely agree with you. I don't feel bad for her. It was dumb, but I appreciate her saying the quiet part out loud, and I think it's useful for all of us to be aware of the implications of what OpenAI is doing.

Yeah. Well, fair enough. If I look at it and say, “All right, you're a bit of a softie. You're the guy who was Rudy Gobert-sympathetic—”

Touching all the mics, and then he got COVID first.

Which, by the way, was, in retrospect, very appropriate. I would say everyone was going to get it anyway. But, yes, you are a well-known softie for anyone on the receiving end of an internet pile-on.

I tend to be sympathetic to pile-ons, that is.

And, yeah, we'll see what happens in terms of government involvement with OpenAI. It's all part of the bubbly atmosphere we inhabit here. One question I had on the rundown: What's the more appropriate top of the AI bubble—Sam Altman getting indignant when he was asked how OpenAI's $13 billion in revenue can support $1.4 trillion of spending commitments?

Or President Trump saying last week, “Everybody wants AI because it’s the new internet. It’s the new everything. It’s one of the biggest things anyone’s ever seen. So everyone wants it. Yeah, I mean, the only problem is if you don’t get it.” Which of those do you think is the more appropriate epigraph on the bubble here?

I don’t think we’re even close to the top of the bubble, so I think neither of those are relevant. [Laughter] The popping of the bubble is probably going to come when, if you just follow the traditional so-called trough of disillusionment or whatever, something is hyped and everyone gets on board with it, and then everyone is like, “Oh, this doesn’t work. It sucks.” Then the bottom falls out.

What actually happens is it just wasn’t ready. It has to be developed, and it gets better. Suddenly, you look up and, holy cow, we’ve vastly exceeded what we thought we were going to get before.

Certainly, if there were some sort of Microsoft report showing a huge miss because it actually turns out all these companies that signed up for Copilot all failed, or something like that. The reality is—all these companies are saying demand vastly exceeds supply. Actually, this is something I’m very curious about.

That’s why they’re justified in increasing their capex: They can’t fulfill the needs. Again, I’ve been a big defender of Google in particular. I have to take my “where I was right about Google” points when I can grab them, because there are times I’ve been wrong. But in January of this year, their stock had a big drop because they announced big capex increases, and the cloud numbers were a little disappointing. I said, actually, if you look at the cloud numbers, it’s super clear they just don’t have enough capacity.

That’s why they missed. And it’s why you should not just be happy they’re spending more; you should wish they had spent more previously. That has very much come to bear. You could tell they brought a lot of capacity online this year: huge jumps in Google Cloud margins and revenue. It’s all looking fantastic, and by and large, that seems to be the case.

What’s confusing is that OpenAI signs a deal with AWS—a relatively small deal, $38 billion—but the difference is that a $300 billion deal with Oracle is for the future, for data centers that still need to be built. This $38 billion deal with Amazon is for access to GPUs right now. Why does Amazon have GPUs to give OpenAI right now if they’re telling investors on earnings calls that demand exceeds supply?

Yeah. Well, I think the theory is interesting. We’ve seen a real shift in talking about power on these calls, which everyone’s been talking about. Both Microsoft and Amazon said, actually, no, we have plenty of chips; we’ve been power-constrained, which they haven’t said so clearly previously.

We knew this was coming; we didn’t know it was happening today.

I think AWS probably brought on a new data center. They talk about how they’re about to bring on a certain amount of power, and they let OpenAI jump the line because it’s a huge boost for their stock. OpenAI is paying a pretty steep price for these GPUs.

That’s right. We have OpenAI. I think OpenAI paying that steep price is indicative of OpenAI needing chips now, as opposed to waiting for—

The pessimistic take is that OpenAI is gangbusters and no one else is. Actually, this is really just an OpenAI story. I think there is an open question here. The story I told about letting them jump the line is probably the case. I don’t think they’ve had GPUs just sitting there unused and OpenAI has taken them. I think they probably literally just brought capacity online that should have gone to people in line, and they let OpenAI cut the line because OpenAI is paying more. Amazon’s stock goes up 10% when they sign the deal.

Exactly the now-routine OpenAI press release bump. Congrats to the folks at Amazon.

It is important to note what makes this distinct. It’s not just a press release bump. OpenAI is going to be running on Amazon’s servers immediately, which makes this deal significant. I suspect that’s what happened with Google as well. No one’s announced the details of what’s going on with Nvidia and Google, but they did sign a deal. It is for Nvidia GPUs; it’s not for TPUs.

That’s a sign that they need capacity now. They’re bursting at the seams, and that is the OpenAI bull case: Everyone’s falling over backward to give them supply, and they still don’t have enough.

And they need it. Yeah, no, exactly. Well, as far as the OpenAI bull case, did you have any response to Sam Altman’s answer to Brad Gersonner over the weekend—the indignant response?

I don’t know. I actually should watch it. I did not watch it. I guess I probably should watch it.

Wow, that’s great. Good job touching grass. It went viral for a couple of days there, but you missed it. I like it.

Yeah. I don’t know.

My take is this: Sam’s answer was fine. The problem with his answer was his tone. The questions about OpenAI’s spending commitments are obviously fair, and he could acknowledge that the questions are fair and then just make his case.

Speaking as a podcast host myself, Gerstner was just being a good host and teeing up a softball for Sam Altman to address. He addressed it in an annoyed way that I think rankled a lot of people. But I think the questions are fair, and Sam’s answers are also fair. It’s the same case he made to you in the strategy interview a couple of weeks ago: They’re spending because they’re betting that they’re going to need the compute, and the bigger risk is that they don’t have the compute once the market opportunity materializes.

Yeah, they are operating, even though they’re a very large company with a ton of revenue, as a startup, and I think appropriately so. This is an all-in bet by them. They could stop now and have a nice business, but what’s the fun in that? They want to be the AI for the world.

I thought that’s what we celebrated: the audaciousness, going for the moon, not being content to hit a streetlight.

Altman’s background is in startups. Google, Microsoft, and all these big companies—even Meta—do have constraints that a company like OpenAI doesn’t. Maybe this is the defense for why they don’t want to go public. They don’t want to answer to shareholders every quarter.

Well, they are losing a lot of money right now, despite all the success. They’re losing a lot of money every quarter, and so they need to expand.

The reason it would be fascinating to the public is to get visibility into whether they’re losing money on actual servicing. They’re spending a lot of money on R&D and things of that nature, but when you actually calculate the cost of serving their customers relative to what they’re making from those customers, what are the profit margins there? That’s the real question.

Given their growth rate, you can assume that they’re profitable on a unit-cost basis—where they’re making X amount of money on their customers—and that’s fine. I have no objection to how much money they’re losing.

But we don’t know that, right? The $20-a-month plan was pulled out of thin air. Are they actually making money on those $20-a-month plans or not?

Yes. Well, maybe one day we’ll get an OpenAI IPO, which I’m cheering for strictly for the sake of takes here on the podcast.

Oh, yeah. I mean, public companies are much better than private ones. Let’s have some fun. Indeed. All right. Why—