No one counted SpaceX in the AI compute race until it signed two megadeals in a month
In the past six weeks, Musk quietly transformed SpaceX from a rocket and satellite company into the fourth-largest AI compute cloud provider, yet no one has factored this into its $1.77 trillion valuation.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Reselling compute to others is more profitable than using it in-house
Charts show xAI's cloud compute partnerships with Google generate per-kilowatt operating margins exceeding those of Anthropic, Meta, Google, and OpenAI; even its Anthropic partnership margins rank second only to Anthropic's own. The core driver: Musk's data center construction speed (122 days) far outpaces competitors. Speed itself is cost savings—each day of delay means another day paying electricians and plumbers. SpaceX can now resell this speed advantage at premium rates.
— Gavin BakerNobody saw SpaceX becoming an AI cloud provider
Six months ago, everyone's models treated SpaceX as two separate businesses—‘Starlink connectivity plus x.ai models’—with no compute competition component. In the past 30 days, through megadeals reselling compute to Google and Anthropic, SpaceX became the fourth-largest ‘hyperscale cloud provider’, leapfrogging incumbents including Oracle. Few prior forecasts included this pivot.
Data centers are not commodities—they're design problems
Some treat data centers as standardized commodities, but Gavin disagrees. Musk applied first-principles thinking—the same approach used for rockets and EVs—to redesign data centers, which is why SpaceX built one in 122 days. Gavin even cautioned the SpaceX team against publicizing these design details: ‘what's obvious to you is revelation to others’, hinting the methodology gap is understated.
— Gavin BakerOrbital data centers cost one-fifth as much per kilowatt as Earth-based ones
Launching GPUs to space for orbital data centers costs roughly $50 billion per kilowatt in capex, versus $200–250 billion per kilowatt for land, buildings, power, and cooling on Earth—five times cheaper. This requires Starship to first achieve two-stage reusability, then rapid reuse, cutting per-kilogram launch costs from Falcon 9's $1,500 to $250 or lower. The risk: if satellite and GPU on-orbit failure rates spike, the math breaks.
The Cursor acquisition is the most undervalued piece
Six months ago, x.ai was performing well but compute-constrained. After acquiring Cursor, xAI gained proprietary programming data exceeding public-internet code corpora. Fed through Colossus 2 in three weeks, Composer 2.5 briefly reached the Pareto frontier on Cursor's own benchmarks. The larger Grok 4.3 (1.5 trillion parameters) is feeding Cursor data directly into pretraining—not just reinforcement learning. If this path holds, model capability gains may outweigh the compute resale contracts everyone watches.
— Gavin BakerCrash is probable, but Musk can't sell a single share
The past 20 comparable IPOs (Facebook, Twitter, Alibaba, Shopify, et al.) averaged drawdowns exceeding 50%. But this is unprecedented—among the largest IPOs ever, with rapid index inclusion, nobody knows the selling pressure. What's certain: Musk holds 50% and is locked for 365 days. Employees and early investors, however, had a liquidity window every six months for the past decade; selling pressure may not concentrate on day one.
— Gavin BakerNo one knows how smart these models actually are
Fable 5 and Mythos-class models are hard to assess because no one has run Mythos continuously for a full year. Each generation's true intelligence ceiling likely remains unmeasured until the next one displaces it. Noam Brown's tweet calling snapshot benchmarking ‘profound’ reflects that this evaluation method is obsolete. Measurement should run along time, token count, or compute invested—not a single-run score.
— Gavin BakerOpen-source claims 80 percent of tokens; frontier models claim 90 percent of revenue
Two years ago, this group debated whether open-source would catch frontier models. This year's data shows the opposite: frontier models may capture 90 percent-plus of economic value while open-source accounts for 80 percent of token volume. Open-source goes mostly to low-value back-office tasks like support; enterprises won't run ‘second-tier’ models on high-value work like coding. If this divide persists, compute and hardware benefit—frontier margins compress, forcing enterprises to overspend on compute to compensate.
$1.5 trillion capex for $300 billion revenue—does the math work?
Morgan Stanley raised 2027 capex forecasts from $950 billion to $1.1 trillion. Gavin thinks it could hit $1.5 trillion; against that sits roughly $300 billion in AI lab inference revenue next year. Gavin counters that $300 billion itself is underestimated, and training's share of total capex is dropping from near 50 percent to about 35 percent. More striking: over seven years, Mag 7 generated a second trillion in incremental revenue in half the time their first trillion took (20+ years). SpaceX, Anthropic, and OpenAI are now forecast to generate another trillion in four to five years—half again the time.
— Gavin BakerIn their own words · checked verbatim
generates more operating profit per gigawatt um than Anthropic, than Meta, than Google, than OpenAI.
Gavin Baker3:07
we talked about it on all-in podcast, we called it EWS, Elon web services, right?
there is a belief that these data centers are commodities.
Gavin Baker19:19
it's about $5 billion per gigawatt of CapEx to put these in space.
I suspect if there's an upside surprise, if we went around the table, I'd say this is the place that's getting the least amount of attention and could have the biggest upside surprise.
Gavin Baker31:30
nobody has run Mythos for a year continuously. And we may never know how smart each generation of models actually is or was
Gavin Baker44:34
Frontier might be 90% of the economic value.
we're spending 1.5 trillion of CapEx on 300 billion of inference revenue. Does that math math for you?
Brad1:04:48
Figures
| SpaceX IPO valuation | $1.77 trillion ($135 per share) | 2:05 |
| SpaceX revenue forecast (2028) | $160 billion | 2:05 |
| SpaceX data center construction time | 122 days | 4:09 |
| Orbital compute capex per kilowatt | ~$50 billion, vs. $200–250 billion terrestrial | 26:27 |
| Launch cost per kilogram | Falcon 9: $1,500; Starship target: <$250 | 25:26 |
| Cursor projected 2026 revenue | up to $10 billion | 29:30 |
| 2027 AI capex forecast | Morgan Stanley: $1.1 trillion (raised from $950B); Gavin estimate: ~$1.5 trillion | 1:03:47 |
| AI labs inference revenue forecast (2027) | ~$300 billion | 1:04:48 |
| SpaceX valuation multiples | Post-deals: 100x to 39x TTM; added $29 billion in one month | 39:34 |
Glossary
- Pareto frontier
- A curve measuring how much intelligence you get per unit of cost; further left and up is better.
- EWS (Elon Web Services)
- Nickname for SpaceX's practice of reselling compute to clients like Google and Anthropic, mirroring Amazon Web Services' model.
- ASIC
- Application-specific integrated circuit; a chip optimized for a particular AI workload, distinct from Nvidia's general-purpose GPUs.
- two-stage reusability
- Starship's ability to recover and reuse both the booster and second stage, versus single-use expendable rockets.
- token router
- A system that automatically routes each request to the most cost-efficient model (expensive or cheap) for that task.
How to listen
Secondary-market investors tracking SpaceX IPO valuation and AI compute capex logic, plus practitioners who want to understand xAI/Cursor's model strategy trajectory.
After 1:11, the segment on overall market positioning and inflation data reflects personal fund-manager trading details and can be skipped.