SpaceX Becomes World's Fourth-Largest AI Compute Provider in Just 30 Days
Source: BG2Pod | Published: 2026-06-11T23:15:21Z
SpaceX landed major compute deals with Anthropic and Google back-to-back, now outearning every major AI company on a per-gigawatt operating profit basis — with Colossus Phase 1 posting a staggering 55% IRR.
SpaceX Became the Fourth-Largest AI Compute Provider in 30 Days — This May Be the Most Underrated Story Before the IPO
The Ground Data Center Business Is Sexier Than Rockets
SpaceX's IPO is priced at $135 per share, implying a $1.77 trillion market cap. Wall Street and Goldman Sachs project $160 billion in revenue by 2028. But investor Gavin Baker argues that most people are looking at the wrong thing — not rocket launches, but two key variables: how fast ground data centers come online, and how X.AI's models perform on the Pareto curve in coding.
Clark Tang ran an analysis: X.AI's cloud computing agreement with Google generates higher operating profit per gigawatt than any deal from Anthropic, Meta, Google, or OpenAI. The Anthropic agreement ranks near the top as well. Altimeter's Freda calculated that Colossus Phase 1 delivers a 55% internal rate of return — borrow at 6% to 8%, deploy into a 55%-return asset. The math speaks for itself.
Jensen Huang has said that Elon took a data center from zero to operational in just 122 days — the fastest in the world. Speed itself is cost — every additional day means another day paying electricians and plumbers.
30 Days: From Zero to the Fourth-Largest Hyperscale Compute Provider
Six months ago, almost nobody had SpaceX on the AI compute chessboard. The prevailing model was simple: Starlink does connectivity, X.AI does models. Then within weeks, SpaceX signed major contracts with both Anthropic and Google, and an entirely new business category — reselling compute at high margins — materialized out of nowhere.
After the Google deal, SpaceX ranks roughly as the world's fourth-largest hyperscale compute provider, ahead of Oracle. Brad Gerstner put it bluntly: 30 days ago we weren't a hyperscale AI player. Now we're number four.
A wave of NeoCloud companies is emerging — an estimated 50 are raising money in Silicon Valley right now. But Gavin Baker doesn't believe data centers are a commodity business. His analogy: Elon redesigned rockets from first principles to make them reusable, then redesigned electric vehicles from first principles, and now he's applied the same approach to data centers. Gavin even told the SpaceX team: "You probably don't realize how differentiated what you've built is — it's so obvious to you that you can't see it, but to everyone else it's a revelation."
Rocket Reusability Is the Foundation of Everything
Andrew Fox calls reusability the bedrock of SpaceX's entire commercial empire. The old rocket industry was like flying to California and having the plane explode after you land. SpaceX's goal is to make both stages of Starship — not just the booster — fly 30 to 50 times before needing refurbishment. Each additional flight further amortizes the per-launch cost. Eventually, costs converge toward the price of fuel alone.
The company plans to attempt second-stage recovery later this year and achieve second-stage reuse next year. On launch cadence, last year saw roughly 160 to 165 launches; the next few years should reach the hundreds, and within three years after that, thousands — meaning two to three launches per day.
This is extraordinarily hard. Gavin Baker emphasized the point: he's watched Elon pull off many difficult things, but rapid two-stage reusability remains a formidable engineering challenge.
Starlink: 0.3% of the Global Telecom Market
Brad Gerstner can't hold a stable phone call on Sand Hill Road in Silicon Valley — the heart of the tech industry, where mobile connectivity is still abysmal after twenty years. He says once Starlink Mobile launches, everyone will switch.
Fox points out that Starlink's broadband business is still in its earliest stages, with less than 1% global household penetration. If rapid Starship reusability is achieved, the terminal count could scale to hundreds of millions. Leaked bank models show communications revenue (primarily Starlink direct-to-cell) growing from roughly $10 billion to $50 billion by 2028. Fifty billion sounds massive, but it represents only 0.3% of the global telecom market.
Gavin Baker travels with a Starlink unit everywhere — he's a hardcore gamer. He says no matter where in the world you are, Starlink delivers the fastest connection with the lowest latency. Once rapid reusability is achieved, the cost per gigabyte delivered will also be the lowest. Better, faster, cheaper — that formula has never failed.
Space Data Centers: One-Fifth the Cost Per Gigawatt vs. Ground
Clark Tang ran the numbers. Elon recently disclosed AI satellite specs: each Starship launch can carry about 5 megawatts of compute capacity with a 100-metric-ton payload. Working backward, the capital expenditure to put one gigawatt of compute in orbit is roughly $5 billion. On the ground, factoring in generators, transformers, shells, power, and cooling, one gigawatt costs $20 to $25 billion.
Brad Gerstner broke it down further: total ground cost for one gigawatt is about $60 billion, of which $35 billion goes to GPUs and chips, and $25 billion covers land, shells, power, and cooling. In space, land is free ("there's no land in space, but there's plenty of space"), power comes from solar, and cooling is handled by the vacuum itself. So total cost for one gigawatt in orbit comes to roughly $30 billion, and that $5 billion launch cost will keep declining as reusability improves.
They do acknowledge that GPUs burn out and lasers fail — that happens in ground data centers too. As long as the failure rate in space isn't "astronomically" higher, the economics work.
But the investors unanimously agree: space data centers are not a prerequisite for buying the IPO. The contract prices for ground compute — Anthropic at $2.2 to $2.3 billion per gigawatt per year, Google at $5 billion per gigawatt per year — already far exceed the market's implied assumption of $1.4 billion. Space is a massive call option.
The Cursor Deal May Be the Most Overlooked Variable
Six months ago, X.AI was performing well in the model race but wasn't leading. Then they struck a partnership with Cursor and secured an acquisition option — SpaceX has the right to acquire Cursor for $60 billion by year-end, with a $10 billion breakup fee if the option isn't exercised. Cursor has a team of roughly 300 people and explosive revenue growth; Altimeter internally projects Cursor could hit $10 billion in revenue this year. More critically, Cursor and Anthropic each possess more proprietary coding data than exists on the entire public internet.
The Cursor team used Kimi K2.5 as a base model, injected their proprietary data, ran reinforcement learning and supervised fine-tuning, and trained on the Colossus 2 cluster for three weeks. The resulting Composer 2.5 model was Pareto-optimal as recently as 12 days ago. Now Grok 4.5 — 1.5 trillion parameters — is in training, with Cursor's data being injected at the pre-training stage, not just during reinforcement learning.
Gavin Baker considers this the lowest-attention, highest-upside part of the entire SpaceX IPO story. The logic is straightforward: Anthropic has already proven that once you reach the Pareto frontier, revenue can explode. Replit founder Amjad Masad has articulated what he calls a "bitter lesson adjacent" thesis — coding may be the fastest path to AGI and ASI, because a model that excels at coding can write code to do anything.
Fable 5 and Long-Running Agents: We Don't Know How Smart These Models Really Are
Anthropic released Fable 5 (essentially Mythos with classifiers and safety guardrails for cybersecurity and biochemistry). Karpathy called it state-of-the-art on every benchmark, but what's truly remarkable is its long-running task capability. OpenAI's Noam Brown reposted a tweet making a profound point: snapshot benchmarks are becoming irrelevant. The x-axis should be time, or token count, or compute — because if you let a frontier model run long enough, most problems can be solved.
Gavin Baker took the idea to its logical extreme: imagine a driver who never gets distracted, never tires, never takes calls, never drinks — FSD's advantage is obvious. Now imagine an Einstein who doesn't need to eat, sleep, or rest, never ages, and thinks continuously about fundamental physics for an entire year. Nobody has let Mythos run continuously for a year. We may never know how smart each generation of models truly is, because before we can fully evaluate one, the next generation has already arrived.
This observation points directly to one conclusion: if you were already bullish on compute demand, you should be even more bullish now.
Frontier Models Capture 90% of Revenue; Open Source Handles 80% of Tokens
At the start of the year, market consensus was: open-source models are catching up to the frontier on cheap tokens, frontier models may be converging, and users won't pay for expensive tokens. Six months later, reality looks like the exact opposite — frontier tokens capture the overwhelming majority of revenue.
But Gavin Baker argues both things can be true simultaneously: frontier models continue to capture most of the economic value, while open-source models consume the majority of global token volume. Harvey ran reinforcement learning and fine-tuning on open-source models using their proprietary legal data, paired with a model router, and achieved better results than Opus 4.7 or 4.8 at lower cost — yet they're still consuming massive amounts of Opus tokens.
A key corollary: the better open source gets, the better it is for compute providers. If frontier models capture a smaller share of profits, customers end up spending more on compute instead.
Clark Tang added a perspective from Asia: in Silicon Valley, the dominant belief favors closed-source and cloud; in Asia, the preference is to match different models to different workloads and avoid overpaying. Next year could be decisive.
Nvidia Can Join the Frontier Anytime — and It's Holding a Nuclear Option
On the competitive landscape around ASICs and Nvidia, the group reached a consensus: on paper, Nvidia holds only about 30% share of OpenAI's gigawatt-scale orders, but in actual deployment that number is likely much higher. The reason: in a power-constrained world, Nvidia produces more tokens per watt — which directly translates to more revenue.
Gavin Baker raised an interesting game-theory scenario: if all of Nvidia's customers are going to compete with it, why wouldn't Nvidia compete with its customers? Nvidia already has strong open-source models (Nemotron 3 is excellent on compute efficiency); it's deliberately releasing only small models to avoid stepping on Anthropic, OpenAI, and Google. But if the landscape shifts, Nvidia could join the frontier faster than anyone expects and become one of the world's largest cloud computing companies.
He summed up the power of that trump card in one line: "Oh, what a cute little ASIC you built. Would you like open source to join the frontier too?"
$1.5 Trillion in Capex vs. $300 Billion in Inference Revenue: Does the Math Work?
Morgan Stanley raised its 2027 industry-wide capex forecast from $950 billion to $1.1 trillion, but that doesn't include SpaceX, CoreWeave, and others. Brad Gerstner thinks the real number could approach $1.5 trillion. Meanwhile, combined AI lab revenue for 2027 is projected at roughly $300 billion.
Gavin Baker pushed back: what's the gross margin on that $300 billion? Brad said 50%; Gavin thinks it's 60% to 70%. And he believes $300 billion is too low — inference revenue will be well past $200 billion by the end of this year. Dario Amodei previously said 2028 revenue would be in the "low hundreds of billions" (i.e., $300 to $400 billion), with trillion-dollar revenue before 2030.
Another overlooked fact: roughly 35% of capex goes toward training next-generation models and doesn't directly generate revenue. The return on the remaining 65% is climbing rapidly — at the start of the year, peak monetization per gigawatt was $20 billion; now it's pushed to $30 to $40 billion. Fixed costs stay the same, but revenue per gigawatt is rising, and the marginal gains are almost pure profit.
Alex from Whale Rock offered a data point: fewer than 0.2% of people globally are using AI in an agentic manner. Gavin himself isn't a technical person, but he's consuming 500 CPU cores and 5 GPUs around the clock. If you extrapolate that usage intensity to even a tiny fraction of the population, the compute shortage could persist for a very long time.
The Market Needs to Catch Its Breath, but It's Hard to Be Truly Bearish
Semiconductors surged this year, but internet stocks are down 16% and software is down 8%. Brad Gerstner says if someone had told him at the start of the year what would happen — conflict with Iran, oil back to $100, CPI above 4% — he would have said don't think twice, stay far away from the market. But AI revenue beat expectations and held everything together.
Gavin Baker used a running metaphor: in 2022 the market was running downhill, storing energy. Over the past two months, semiconductors in particular have scrambled up a steep cliff face. A lot of stocks need a breather — the question is whether they'll hang on a rope at the top of the cliff, or need to climb back down a stretch.
He also noted that AI consumption has shown seasonal slowdowns every summer for the past three years, because college students are heavy users and usage drops during break. Silicon Valley data from the past two weeks suggests some token consumption may be shifting toward cheaper open-source models.
Altimeter's approach has been to dial positions from "large" down to "medium-small." Not all-in or all-out — just reassessing risk-reward at current prices. But when Brad thinks about what Noam Brown said, and about what Fable 5 can do, he says it's hard to be truly bearish.
Three Companies, Four to Five Years, One Trillion Dollars in New Revenue
Brad Gerstner revisited a number: over the past seven years, the Mag 7 added $1 trillion in revenue, corresponding to $17 trillion in market cap gains. The current projection is that SpaceX, Anthropic, and OpenAI will add another trillion dollars in revenue within four to five years — not seven companies, but three; not seven years, but half the time.
Ten percent of global GDP is $10 trillion. He believes AI will reshape 5% to 15% of global GDP, and that conviction is beyond question in his mind. Coming back to the SpaceX IPO itself, his conclusion: there has never been an IPO of this scale in history, and no company has ever been added to indices this quickly. Elon holds roughly 50% of shares with a 365-day lockup. Over the past decade, SpaceX employees and investors have had a liquidity window every six months — anyone who wanted to sell has already sold. This is a situation without precedent.