TMTB: BG2 Pod w/Gavin Baker on SpaceX (SPCX) Key quotes
Good BG2 Pod on Spacex w/ Gavin. Link here
#1 — SpaceX = “A Must Buy, Must Own” for Institutional Investors
Gerstner: “I think we’re all pretty AGI-pilled. And if you’re AGI-pilled, that means we got to build a lot more compute than the world thinks and that these models are going to be a lot more valuable than people think. You combine that with their core business — I don’t know another entrepreneur or another business that’s a better bet on the future than SpaceX. And so I think for most institutional investors, it’s a must buy, a must own. It’s set it and forget it, in order to have a real bet on both the space and the AI future.”
“When I look at the bull-bear case on the IPO, the bears are looking at last year’s revenue, say it was $18 billion, and the forecast from the banks of $160 billion three years from now, and they’re saying not many companies in the history of the world have basically 8x’d their revenue over three to four years. When you break it down as an analyst, first principles, part by part — Starlink looks totally doable. AI compute terrestrially looks totally doable. The model itself after the acquisition of Cursor looks to me like it could be an upside. Three years from now, there’s a decent chance everybody’s like, ‘Oh my god, that was super obvious.’”
“It was a hundred times trailing revenue. Well, after the deals they signed, I think it’s 39 times. They added $29 billion in a month. Have you ever seen that happen? Never.”
#2 — “Elon Web Services”: #4 Hyperscaler in 30 Days, Best-in-Class Monetization
Baker: “Clark’s analysis shows that XAI’s deal with Google for cloud computing generates more operating profit per gigawatt than Anthropic, than Meta, than Google, than OpenAI. Freda calculated a 55% IRR on Colossus 1. If you can borrow money at 6, 7, 8% and invest in something with a 55% IRR — I’m not the most sophisticated thinker, but that math maths.”
“Are they the number four or number five hyperscaler today? After the Google deal it will be number four. In 30 days, we went from not being an AI hyperscaler to being number four. And we passed a lot of companies, including Oracle.”
Fox: “The implied monetization rate on that [$160 billion] number is something like $14 billion per gigawatt per year for the AI business. They just signed Anthropic at 22 to 23. They just signed Google at 50. So I think you can invest behind the AI business terrestrially and still be excited about it.”
Gerstner: “Jensen said Elon is an N of one. What they achieved is singular — 100,000 GPUs, easily the fastest supercomputer on the planet. A supercomputer would normally take three years to plan and one year to get working. We’re talking about 19 days.”
#3 — Data Centers Are Not Commodities: First-Principles Design
Baker: “There is a belief that these data centers are commodities. I do not share that belief. In the same way that Elon was able to re-engineer a rocket from first principles and make it reusable, he engineered an electric car from first principles — I think he looked at data center design from first principles and designed something fundamentally different.”
“I did actually ask the team — I said, ‘Hey guys, maybe be a little less public about things that are very obvious to you about how to design a data center, but are revelations to other people. What you’re doing is maybe more differentiated than you perhaps realize.’ And that’s how he was able to do it in 122 days. Speed is literally cost — every day you’re paying electricians and plumbers.”
Tang: “There’s only maybe two or three players now that can actually reliably engineer behind-the-meter data centers. If you’re GE Vernova and you only have a certain number of gas combustion engines — you can sell them to XAI or one of these startup NeoClouds. Who are you going to sell them to? Everyone starts making more money when the GPUs get energized and sold faster. Speed is money for all of the suppliers.”
#4 — Orbital Compute Economics: $5B/GW vs. $20-25B Terrestrial
Fox: “When you back into the numbers, you get to something like five megawatts of capacity per Starship launch. The math you get to is about $5 billion per gigawatt of capex to put these in space. For comparison, terrestrially — the switch gears, the generators, the transformers, the shell, getting the power — that today is about $20 to 25 billion per gigawatt. So we’re talking about a 5x reduction in cost on half of your bill of materials.”
“Two-stage reusability — the cost per kg comes down significantly. We’re talking about going from $1,500 per kg on Falcon to $250, something lower. The more you can reuse the rocket, the more that price comes down, because you’re depreciating the cost of the launch, and eventually you asymptote to the cost of the fuel.”
Gerstner: “It costs $60 billion to put a gigawatt on the ground today — call it $35B the GPUs and silicon, $25B the land, shell, power, and cooling. I would hypothesize those elements are probably going to be inflationary. Space, power, cooling are effectively free in space. You’re talking about putting a gigawatt into space for $30 billion with lower operating costs, versus $60 billion that’s inflationary. As long as the reliability and maintenance is not dramatically lower, the math maths.”
Fox: “I don’t think orbital compute is necessary for the IPO valuation, but it’s certainly important.”
#5 — The Cursor Acquisition: The Most Underappreciated Lever
Baker: “My understanding is that Cursor and Anthropic have more tokens of proprietary coding data than anyone else — and each have more tokens of proprietary coding data than exist on the public internet. Cursor used Kimi K2.5, used their own private data, did some RL, some supervised fine-tuning, and got a really good model. Then they spent three weeks in the Colossus 2 cluster and got a model that 12 days ago was Pareto dominant with Composer 2.5. It suggests the Cursor data is very valuable for coding, and that XAI/SpaceX has a shot at being a real player in coding.”
“Right now the Grok 4.3 1.5 trillion parameter model is training. One would hypothesize based on scaling laws that will be a better base model, and then the Cursor data is being injected into the pre-training process, not just reinforcement learning. That is going to be a very important data point. And once you are at multiple places on that Pareto curve, if you have compute, you can scale really rapidly.”
Gerstner: “I think the thing that’s getting lost is they’ve dramatically advanced their capability when it comes to building a frontier model. Michael and the team at Cursor — this is an extraordinary team that he just downloaded right into SpaceX. If there’s an upside surprise, this is the place that’s getting the least amount of attention and could have the biggest upside surprise.”
Tang: “From a lot of my conversations, it looks like they’ve secured maybe up to 20% of Vera Rubin capacity — especially in the early days when these chips are very scarce.”
#6 — Fable 5 / Mythos and the Noam Brown Insight: “We Don’t Know How Smart These Models Are”
Baker: “Nobody has run Mythos for a year continuously, and we may never know how smart each generation of models actually is or was — because we don’t have time to appropriately evaluate their intelligence before the next model comes out. This is a profound statement.”
“Imagine Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn’t have to eat, doesn’t have to sleep, never gets old, never has diminution of intelligence — and he thought for one year. We might already have solved a lot of these intractable problems. My takeaway was: however bullish I was on compute before, I’m just a lot more bullish.”
Tang: “We’ve just been locked at our desks hammering Claude, because it’s fascinating the things we can now do with Fable 5 that we just couldn’t do with Opus 4.8 a day before. It’s really, really good at multi-agent orchestration now. I threw in seven of our models and said, ‘Create a master view of my beliefs given all these assumptions — TSMC capacity — and produce me a report.’ The model reasoned through all of our assumptions: ‘Actually, if you believe this, what are the contradictions?’ I’ve also dumped all my notes into it and it reasoned across three years of notes — here are your ideas that were consistent, here are the sources that were highest signal to what actually played out. And we’ve just blown through our limit.”
Gerstner: “They gave examples in the release — a 50 million line Ruby codebase at Stripe refactored in a day versus many weeks with many people. If you believe this to be true about long-running agents, we’re going to produce and consume more tokens as far as the eye can see.”
#7 — Frontier Captures 90% of Value; Open Source Carries 80% of Tokens
Baker: “Two things can be true. The majority of economic value may continue to accrue to the frontier — and man, has it ever accrued to the frontier thus far. And the majority of tokens consumed in the world may be open source — and they are today. I think this current state is likely to persist.”
“Harvey used their own proprietary legal data to do reinforcement learning and supervised fine-tuning with Fireworks on an open-source model, then used a router, and they got better outcomes than Opus 4.7 or 4.8 at a lower cost. I think that is the future. They were still consuming a lot of Opus, but a majority of the tokens were probably their own open-source model.”
Gerstner: “The consensus going into this year was that open-source models, cheap tokens, were catching up on the frontier, that these models were beginning to asymptote, that people wouldn’t pay for premium tokens. The evidence six months in is just the opposite — frontier tokens are capturing the vast majority of all the revenues. What folks concluded — that frontier models won’t accrue most of the revenue — has been decisively wrong. It’s 90%, probably more.”
“Think of JP Morgan — back of the house stuff, customer service, they may use an open-source model. But the really high-value stuff — coding — they don’t want to write second-tier code. You don’t need Albert Einstein to book you a trip. You don’t need Albert Einstein to do KYC.”
Baker: “Something very important on open source: there’s this belief that it’s bearish for AI. It’s actually — maybe bearish for the frontier models — but really bullish for compute and hardware. If the frontier models are capturing less of the margin, then you’re going to spend more on compute. The better open source does, the better it is for compute providers.”
#8 — Nvidia: “That’s a Cute ASIC” — The Open Source Counter-Move
Baker: “’Wow, that’s a cute ASIC you’ve built there. That is so cute. How would you like open source to join the frontier? How do you like them apples?’ I do think Nvidia is highly likely to be the world’s dominant provider of open source AI. Right now open source is whatever, six months behind the frontier — we might see it creep closer and closer.”
“If all of his customers are going to compete with him, then why not compete with his customers? He has his own models that are really, really good — Nemotron 3.1 was really cool from a compute-efficiency perspective, and he’s always careful to release small models so as to not tread on Anthropic, OpenAI, Google’s toes. But that is a choice he is making. If the economics change, I think Nvidia can join the frontier and become one of the world’s largest cloud computing companies much faster than people think.”
“You might not have the revenue or the margins to fund that ASIC [if open source closes the gap].”
Tang: “Everyone assumed Nvidia was going to lose share dramatically. Actually, if you look at the last few years, they’ve maintained their share very handsomely — and if you account for the fact that Anthropic was not really using Nvidia, they probably actually gained share in ‘25-’26. For ASICs, the argument now is more and more will look custom to the actual workload — MediaTek with their V8T versus Broadcom’s V8i for TPUs was a big topic. There’s a lot more nuance now versus one year ago when it was a Broadcom-or-Nvidia battle.”
Baker: “People are indexing to the OpenAI gigawatts — Nvidia has 10, Broadcom has 10, AMD has six with warrants, Cerebras has a gigawatt. I’ll be very surprised if that’s where they land. As long as we’re in a watt-constrained world, if you can get more tokens per watt — which is literally revenue — with Nvidia than alternatives, you may save some money building your factory with another chip, but you’re going to have less revenue. Meta and Microsoft have been probably disappointing [on ASICs]. You know who made a good ASIC? OpenAI — Jalapeño. They made a great chip. Unfortunately it needs to run at a much lower temperature than Nvidia GPUs, which means more money on cooling.”
#9 — The Capex Math: $1.5T Spend vs. $300B+ Inference Revenue — “The Math Maths”