Starship Reached Orbit. Here's the $1B-Revenue-Per-Launch Math | The Brainstorm 152

Sam (0:00)

The Hood Event, Egentic Trading. Are consumers going to be sending their agents into a battlefield to be demolished by a $200 billion spending high-frequency trading Goliath agent? Welcome to The Brainstorm. Today, we're going through OpenAI Dev Day. Are you going ultra fast or are you going ultra ultra fast because you've got $200 billion per gigawatt to throw at it? And of course Starship successfully making it to orbit for the first time and everything that makes for SpaceX.

Brett, you want to kick it off with OpenAI Dev Day. What blew your mind?

Brett (0:37)

Yeah, I mean, there was a lot of announcements at OpenAI Developer Day. Some of them somewhat kind of like chaotic and they're doing live demos, some of which didn't work.

So they have DOTS, which is like a Grokbot competitive product, Grokbot and Muse competitive product.

They announced a newer version, GPT 6.1 Sol, which is actually a very, very useful for cost model and not quite as token consumptive as Astra. They created more of a platform for people to build on top. You can use your token spend at OpenAI for other applications. So if you're paying for a monthly package, that can be a monthly subscription to all kinds of AI applications as developers develop to them, and they announced, there's plenty of other stuff, but Ultra Fast, which is you can pay, get 8x the tokens for 6x the price per second, and so you can pay up to get your answers faster or to provide voice-like experiences. People were actually a little surprised that this product was powered by NVIDIA as opposed to their Ultra Ultra Fast partner, Cerebris, which makes kind of pizza box-sized chips that can run models very, very fast. And as it turns out, at least as reported by SemiAnalysis, the reason why Ultra Ultra Fast, which should be able to go 16x faster, was not announced as publicly available in the developer day is because all of that inventory has been sold out mostly, apparently, to Jane Street, a high-frequency trading firm. And I think this is a nice illustration of, and who are paying apparently $200 billion per gigawatt for that capacity. So compare to, we don't know exactly what the monetization rate is for kind of the Anthropic and OpenAI's off-the-shelf products, but we know that Anthropic is willing to pay $30 billion per gigawatt to rent capacity from SpaceX AI. And so $200 billion is quite a lot. Why would Jane Street do such a thing? It's because when you're in high-frequency trading, having an ultra-intelligent thing that is faster than your competitor is a license to mint money. And so I think there's all kinds of business actions where speed really matters in competitive contests and where companies will pay up for that speed. And so this is in addition to being able to differentiate by the level of intelligence you're accessing, hey, I need the Astra model, I need the Sol model, I need Clod, you know, Opus 5.5. You also, I think the competitive landscape of enterprises will also be willing to pay for speed. And this is an example, an illustration of that.

Sam (3:34)

It'll be super interesting to see what the ROI is on a $200 billion per gigawatt expenditure. Because it's like, it is, right? If it is truly being able to front run all other high frequency trading competition, could be enormous. And then that would be crazy.

But also, you know, how long does that lead last? Is it just a constant battle?

Brett (4:04)

Right? I think high-frequency trading is an example, yeah, is an example where just having the most performant intelligence, like, in action, it's almost the perfect kind of like market for these AI companies because if you have a marginal edge, then suddenly you're printing money. And if you have a marginal disadvantage, then suddenly you're bleeding cash. And so, you know, it really is an arms race between these companies. And for Jane Street, they can go to OpenAI and be like, listen, why don't you just give us all of this capacity? That's actually much more lucrative for them than they're buying like stuff that's publicly available. And so, who knows what the actual price they're paying is. And, you know, they might be willing to be like, we will capitalize, you know, all of your cerebrous build up until X year or up until X dollar amount because they see it gives them this marginal advantage that makes their, you know, ROIC, you know, positive. So.

Nick (5:05)

Doesn't this force OpenAI to continue to have to focus on enterprise instead of consumer? Because I thought, you know, the DOTS announcement was interesting. I think it is more competitive with Grok and not necessarily Muse because it is behind a paywall. But it does, to me, make OpenAI's choice of, do we dedicate compute and fast compute to the consumer? Because, you know, speed will matter for the consumer as well.