The Wind Must Be Deep Enough: A Roundtable Two Days Before the Bell, and the Phrase 'The Math Maths'

Photorealistic painterly cover: a pre-dawn view from a coastal ridge, where long low data-center halls, cooling banks and transformer yards run to the horizon, still sunk in cold blue shadow; at the far end of that same plain a single rocket on its tower is the one object catching the first gold light

If the accumulated wind is not deep,
it has no strength to bear up great wings.
—— Zhuangzi, “Free and Easy Wandering” (4th c. BC)

What this episode is about

This BG2 episode is a four-way roundtable: host Brad Gerstner, his Altimeter partner Clark Tang, plus Gavin Baker and Andrew Fox from Atreides.

The timing is unusual — two days before the SpaceX IPO, with all four men at the table holding shares. They say so up front.

Three threads: how to break the deal into variables you can estimate separately; what Fable 5 and this new class of long-running models actually mean; and a check on capex and the market.

Original episode: BG2Pod, “The SpaceX IPO, Fable 5, AI Capex Update & Market Check w/ Gavin Baker, Andrew Fox & Clark Tang” (2026-06-11)

The notes I took

They break the company into four pieces: launch, Starlink, AI compute sold to others, and their own model. The ordering is explicit — launch is the foundation, because only once reusability drives cost per kilogram down do the other three make economic sense.

“Rapid reusability” is the phrase that recurs most. Booster recovery is done; the second stage is only being attempted this year, with a re-flight targeted for next. Why it’s the dividing line: amortising the cost of a vehicle over thirty to fifty flights is a completely different cost structure from amortising it over one. Fox’s analogy — the old rocket industry was like boarding a plane to California and having the plane explode after you got off.

The Starlink story is penetration. Under 1% of global households, plus direct-to-cell. The measuring stick they use is deliberately crude: even a multi-fold increase in revenue is a fraction of a percent of the global telecom market — that isn’t proof they’ll get there, it’s a statement that the ceiling isn’t the binding constraint yet. Gavin adds the counterweight himself: that requires rapid reusability, and rapid reusability is genuinely hard.

The real surprise is compute. In thirty days the company went from not being on the list of AI clouds to being the fourth largest, by selling compute to other people. Their comparison: the monetisation rate implied by widely reported revenue forecasts is roughly $14bn per gigawatt per year, while the two contracts just signed land at $22–23bn and $50bn. Gavin’s verdict is blunt: if you borrow at 6–8% and invest in something returning 55%, the math maths.

Clark’s observation is that the market keeps treating data centers as commodities. None of the four agrees. The same person re-engineered a rocket from first principles and an electric car from first principles, so he probably re-engineered the data center the same way — which is where 122 days comes from. And in a supply-constrained world, speed is money for everyone upstream: the sooner GPUs are energised, the sooner suppliers of land, power and turbines get paid, which tells you who gets to the front of their queue.

The orbital data center arithmetic is laid out cleanly. On the ground, a gigawatt-scale site costs roughly $20–25bn once you exclude the silicon itself — land, shell, power, cooling. Their space version comes to around $5bn, conditional on second-stage reusability driving cost per kilogram from ~$1,500 down below $250, and on satellite and chip failure rates not being wild. But they volunteer the important caveat: orbital compute is not a precondition for the investment case, it’s an option on top of it. The terrestrial pieces already carry the numbers that are circulating.

On Fable 5, Gavin’s takeaway isn’t the scores. He points at Noam Brown’s observation that snapshot benchmarks are losing meaning — the x-axis should become time, or compute, or tokens, because a frontier model given long enough will solve most of what you throw at it. Then comes the line I think is worth keeping: we don’t actually know how smart these models are, because nobody has ever run one continuously for a year — and we won’t have time to find out before the next one ships.

He follows it with a thought experiment: an intellect that never eats, sleeps, gets distracted or ages, thinking about one problem continuously for a year.

One detail carries more information than its own content: the Pareto curve they use to compare model efficiency went out of date twice within twelve days of being made.

On open source versus frontier, he offers a “two things can both be true” structure. Frontier models capture the overwhelming majority of the revenue — his number is over 90% — while open-source models may account for the overwhelming majority of tokens consumed. The failed inference of the past few years was translating volume directly into value. He adds a second-order point: if open source does well, that’s bullish for compute suppliers, because margin the frontier labs don’t capture flows to the hardware end instead.

Marking capex to reality. The 2027 consensus estimate has been raised toward $1.1tn, and the table thinks the real figure, including what isn’t being counted, is closer to $1.5tn. The denominator against it is roughly $300bn of combined AI lab revenue. Their answer comes in three parts: gross margins may be 60–70%; a third or so of the spend is training the next model and was never meant to generate current revenue; and they think $300bn is low.

Another underrated denominator is users. The estimate cited on the show: fewer than 0.2% of people on Earth are using AI agentically.

The most conservative passage is the market check. Semis ripped this year, but internet is down 16% and software down 8% — enormous dispersion. Gavin’s metaphor is a runner: the market spent the last two months not climbing a hill but going straight up a cliff, and the runner is tired. The only question is whether it rests at the top or has to come back down a stretch first. He also says something uncomfortable for anyone doing research for a living: the game of finding the next bottleneck is over. The host’s own response has been to dial exposure from large down to medium-small.

On post-IPO volatility they offer no answer. The statistic on the table is that across twenty large listings the average maximum drawdown exceeded 50%. Gavin’s reply is that this situation is unprecedented — employees and early investors have had roughly twenty liquidity windows over the past decade, so whoever still holds has chosen to hold. But he says plainly that this is a guess, that he doesn’t know what happens short term, and that the only advice is to work through each variable yourself.

One last number. Because of those new contracts, the company’s revenue multiple went from roughly 100x to roughly 39x in a month.

What I took away

1. When the denominator can halve in a month, the multiple isn’t an anchor.

100x to 39x, with no change in price — what changed was the denominator. That exposes the limits of the tool rather completely: a multiple is a numerator over a denominator, and when the denominator is being rewritten quickly, the multiple carries almost no information.

Which is why the useful move in this episode isn’t a view on the headline valuation but a decomposition into four pieces, each with its own estimate and its own probability. My own discipline runs the same way: first ask what assumptions hold up the denominator, then ask whether those assumptions are falsifiable, and only then look at the multiple.

2. “The math maths” is a conditional, not a conclusion.

The phrase recurs throughout, and every instance has a set of premises hanging off it. The honest version is to write it as a receipt: it works because X, Y and Z hold — then ask which one reality breaks first.

The episode supplies its own falsification list: whether second-stage reusability lands, on-orbit failure rates for satellites and chips, land and power procurement, and the termination terms in those long contracts. All of those are observable and have dates attached. “2028 revenue,” by contrast, isn’t something you can check an answer against — it’s the output of everything above it.

3. The bottleneck didn’t disappear. It moved somewhere a screen can’t reach.

Gavin says the bottleneck-hunting game is over. But the episode’s own content contradicts the literal reading: the discussion migrates from chips to power, land, transformers and gas turbines, and then to who can stand a site up in 122 days.

That difference matters enormously. The earlier items can be screened with capacity and financial data; the last one is execution capability, and it appears in no column of any dataset. So the correct update isn’t “stop looking for bottlenecks” — it’s “this round’s bottleneck is becoming unscreenable,” which is precisely the category a quantitative framework misses wholesale.

4. The death of snapshot evaluation is both a methodological warning and a trap.

Put time on the x-axis: that warning transfers directly to investing. Judging something that only unfolds over time by a single quarter’s EPS or a single day’s signal will systematically understate it.

But the reverse matters just as much — a claim that “you’d have to run it a year to see how good it is” is unfalsifiable on any reasonable horizon. Unfalsifiable things can’t be reasons to hold; they can only be things to watch. You have to hold both halves at once.

5. How to use an insider’s words: check their facts, recompute their probabilities.

All four are shareholders, two days before the listing. They disclose it, and they repeatedly say do your own work and make your own decision. Taleb’s skin-in-the-game principle says people who bear consequences are more credible — but the principle assumes they genuinely bear the cost of being wrong, and at certain moments the payoff from speaking and the cost of being wrong aren’t symmetric.

So I split it: take their facts (cost structures, contract terms, engineering timelines) and go verify them; take their probabilities and recompute them myself. The most honest passage in the episode is the host describing his own habit — the future is a distribution of unknown probabilities, so don’t tell me “it could happen,” give me the distribution: 20% or 30%?

6. Two things can be true at once, and many arguments are just volume mistaken for value.

Frontier takes 90% of the revenue; open source may take 80% of the tokens. That sentence pattern is worth memorising on its own, because it’s the antidote to a whole family of bad inferences: market share isn’t profit share, usage isn’t value capture, penetration isn’t profitability.

And the corollary — that open source doing well is bullish for compute suppliers — is where this kind of discussion earns its keep. It isn’t a restatement of a position; it’s a second-order claim you can go and test.

7. The most conservative lines come from the most bullish episode.

All four are close to unanimous on long-term direction. But when the conversation turns to how they’re actually positioned, one has moved exposure from large to medium-small and the other says the runner needs a rest.

There’s no contradiction there. Directional judgement and position sizing are two separate decisions, and my own most frequent error is translating “I think this trend is right” straight into “I should own more of it.” This episode keeps them cleanly apart, which I think is its most useful lesson.

What this episode is about

The notes I took

What I took away

Further reading