SpaceX’s Public-Market Case Now Runs Through AI Compute

Gavin Baker described the SpaceX IPO as well executed, crediting Goldman Sachs and Morgan Stanley after the stock produced what John Coogan characterized as an almost exactly 20% first-day pop. But Baker pushed against the idea that the public-market case for SpaceX depends mainly on outcomes five or ten years away.

His view is that quarters still matter — “a marathon is twenty six one mile runs,” he said — but that public markets are more patient than many venture investors assume. He pointed to Tesla over the prior five or six years and Amazon during the AWS and retail-infrastructure buildout as examples of public investors tolerating heavy investment cycles when they understand the larger economic logic.

For SpaceX, Baker said two near-term variables matter most over the next year. The first is how quickly the company can bring terrestrial compute online. He argued that SpaceX appears able to monetize gigawatts at a higher rate than anyone else, and cited Jensen Huang’s statement that SpaceX brings data centers online faster than anyone else. In Baker’s framing, the AI infrastructure ecosystem has a shared incentive to get landed power into the hands of operators who can energize it, because “everybody starts making more money when the GPUs are energized.”

The key figure he gave was $50 billion of revenue per gigawatt on the Google deal.

That makes the pace of gigawatt deployment central. “If you can energize two, three, four gigawatts in the next year, that’s a lot of revenue,” Baker said, while acknowledging that prices could move up or down. Jordi Hays summarized the implication as “all eyes on Colossus 3, 4, 5, 6, 7, 8, 9, 10,” at least in the short term.

The second variable Baker named was Cursor. He referred to a prior discussion about Composer 2.5, saying that after three weeks of reinforcement learning and supervised fine-tuning on Colossus 2, it had become “kind of Pareto dominant.” The question, in Baker’s view, is what happens when that process is applied to a bigger and better base model. He also called it interesting that Cursor is in half the Fortune 500.

Baker was clear that he remains excited about Mars, asteroid mining, a city and mass drivers on the moon, and other long-range possibilities. He said there is a “decent chance” some of that happens in his lifetime. But he did not want the investable story reduced to distant science fiction. SpaceX, in his framing, has much more tangible near-term drivers: gigawatts, revenue per gigawatt, speed of energizing data centers, and the application layer that can turn compute into valuable tokens.

SpaceX’s compute business creates an allocation problem: the same capacity can serve external customers, internal model training, internal inference, or applications. Anthropic and Google deals create revenue, but they also consume capacity that SpaceX might eventually want for itself.

Baker’s answer was that the willingness to sell or allocate compute externally may itself signal confidence. One interpretation, he said, is that the SpaceX team is highly confident in its ability to bring gigawatts online quickly. He also referred to an Altimeter figure that SpaceX had “hoarded 20% of the Rubens,” calling Nvidia’s Rubin “an epic chip” and a more drop-in replacement than Blackwell, which he described as hard to get online. He added that Groq LPUs would be integrated “at some point in the next six, nine months.”

The contracts may also preserve flexibility, though Baker framed this as his recollection rather than a firm contractual claim. He said he did not remember all the details, but believed the Anthropic agreement included an out if SpaceX needed the compute. Coogan added that the Google deal had a similar provision. That does not remove the allocation tension, but it changes the way Baker reads it: SpaceX can monetize capacity while keeping some optionality if its internal demand becomes more valuable.

The larger question is whether “Elon Web Services” eventually needs to become a full enterprise cloud — databases, memory storage, cloud storage, cold storage, compliance, and the ecosystem features that make large customers comfortable — or whether producing tokens is enough. Baker’s answer was deliberately narrow: “being a token factory is more than enough for the next five to 10 years.”

He said the world needs more token factories. If SpaceX eventually needs the surrounding cloud primitives, he said, it will build them. But the immediate scarcity is not an AWS clone. It is energized compute that can produce valuable tokens at scale.

That framing carried into Baker’s view of other hyperscalers. Meta, for example, has enormous resources, cash flow, and data-center experience, but its route into the token-factory business does not have to look like SpaceX’s. Baker said Meta could give 100,000 GPUs to a company like Fireworks and take a revenue split.

He also said he is increasingly looking at enterprise value to net property, plant, and equipment as a valuation metric. His rationale is that “installed atoms on earth” may appreciate in value. He connected this to a “halo trade” associated with high asset value and low obsolescence. Meta’s EV-to-net-PP&E multiple, in his view, suggests the market has deep skepticism about Meta’s ability to monetize its asset base.

Coogan agreed that the skepticism is understandable. He said Meta’s management has offered investors “personal super intelligence,” but not yet an obvious product proof point. It would be different, he argued, if Meta AI were at the top of the app store, if people were commonly saying they use it, or if Instagram had a clear AI-driven killer feature such as shopping. Coogan’s view was that no one doubts Mark Zuckerberg’s ability to scale an AI product from $1 billion to tens of billions of revenue once one exists. The open question is what the API or enterprise business actually becomes.

Baker’s broader read was that the major players are behaving as if AI is entering an endgame earlier than many expected. He said coding tokens are so valuable that OpenAI cutting Codex pricing can be understood as an attempt to ensure enough usage to enter a recursive self-improvement loop. If a company lacks enough coding-token volume, it may not be able to participate in that loop.

Meta, in that context, may be more flexible than its current messaging suggests. Baker noted that Meta has been telling investors it will not monetize GPUs externally, but he pointed to the company’s rapid shift toward operating-expense discipline after Brad Gerstner’s letter to Zuckerberg in 2022. In Baker’s telling, Meta went from saying it would keep investing to becoming sharply focused on OpEx within a very short period. His conclusion: if investors hear Meta say it will not externally monetize GPUs, they should “check back in a few hours.”

He added that stock price matters for these companies not only as a financial signal but as a talent-retention mechanism. If a company’s stock falls below grant levels, engineers become unhappy, and retaining the people needed to win in AI becomes harder. Baker referred to the industry’s escalation from talk of “10x engineers” to “10,000x engineers.” For big-cap technology companies, he said, AI may feel existential, but equity compensation and stock performance still shape behavior.

The SpaceX IPO also tested assumptions about retail investors. Some commentary, Coogan said, treated the IPO as “only” 2x oversubscribed by retail compared with smaller IPOs that were 10x oversubscribed, missing the much larger absolute scale. He argued that retail may matter more than ever over the next year as hyperscalers turn cash-flow negative and capital needs expand.

Baker rejected “retail” as a pejorative category. “Stupid is as stupid does,” he said, arguing that retail-favorite indices have risen substantially in 2024 and 2025. His claim was not that retail investors are always right, but that they have likely outperformed the overwhelming majority of professional money managers across private equity, venture, and public equities over the relevant period.

In the specific case of SpaceX, Baker said he was not sure the first-day demand was primarily retail. He emphasized uncertainty about the source of his information, but said his understanding was that after the IPO there may have been more institutional buying than retail buying, and that some retail allocations may have gone to people flipping the stock. The point was not a definitive flow analysis; it was that the simple “retail drove it” story may be incomplete.

The more underappreciated fact, in his view, was employee participation. Baker said more than 10,000 SpaceX employees bought stock in the IPO. That matters for supply analysis. Investors often focus on lockups, SPVs, and potential secondary supply coming to market, but Baker argued that SpaceX employees and longtime investors have had repeated chances to sell every six months for the past decade. If they wanted liquidity, they could already have taken it in previous offerings.

That history changes how he thinks about supply and demand. There may be SPVs that unwind, and some stock may come to market, but the assumption that a huge amount of pent-up supply must immediately appear is too crude. Employees own a large amount, they have had opportunities to sell, and many still chose to buy.

On retail more broadly, Baker said retail investors have been a more powerful force over the last three and a half years than at any other time in his career, including the 2000 bubble. He stressed that he does not think the current environment resembles the 2000 bubble. But retail matters, and dismissing it has become costly.

“In the token path” is not a binary category in Baker’s framework. Cloudflare may deliver tokens at the edge, and AI agents may increasingly interact with websites through distributed infrastructure, but that does not make every CDN a core token-production business.

Baker said every CDN has some version of this business, pointing to Akamai’s $1.8 billion deal with Anthropic. If a company has many points of presence and can deliver low-latency tokens to high-value users, it can command a premium. One lesson he drew from Cerebras is that people will pay for speed, and he said the prices being commanded by these infrastructure companies support that.

But Baker did not equate edge delivery with the core of token production. His gut estimate was that CDNs deliver less than 1% of tokens consumed on earth, and perhaps less than 10 basis points. Most tokens, Hays noted, are generated internally, with the user receiving only the result. Baker said CDNs have “a path towards being in the token path,” and part of their business is there today, but they are trying to move more of their business into that flow.

Edge compute may matter more in use cases where latency is the product. Coogan cited Matthew Prince discussing co-locating Nvidia servers at the edge for voice models: a coding model might work in Virginia for 20 minutes or an hour and return a result, but a conversational model needs speed. Hays also noted a startup Nvidia had invested in that would pay people to put a small GPU server outside their home; Baker compared the broad idea of local infrastructure to the Tesla Powerwall, which stores energy locally rather than compute.

The investor search for token-path exposure has also moved upstream into increasingly obscure semiconductor-supply-chain bottlenecks. Hays described investors hunting for the final ingredient, from specialized materials used in chipmaking to sand that becomes silicon. Baker said the “bottleneck bros” trade may be nearing its end.

He referred to investor enthusiasm around Ajinomoto, the Japanese company tied to a semiconductor material, and said many investors bought because they expected price increases. When the company said it would not raise prices because that was the wrong thing for it, Baker almost posted “welcome to Japan” to the bottleneck crowd. His larger point was that asking Claude to identify the next bottleneck was “the game for the last year.” The next game, he said, is identifying what has enduring franchise value after the bottlenecks.

Land, for Baker, is not obviously in the token path — with one exception he joked about as “beachfront property.” But the land question led him to orbital compute, where he sees a real economic argument once Starship becomes reusable.

Baker’s orbital-compute math was explicitly back-of-the-envelope. He said a terrestrial gigawatt costs about $60 billion to bring online, with roughly $25 billion of that in power and cooling. In space, he argued, that power-and-cooling component is not needed. He compared the remaining terrestrial IT equipment cost — GPUs, CPUs, switches, memory, and storage — with launch cost. Once Starship is reusable, Baker estimated launch cost for a gigawatt in space at about $5 billion, implying that a gigawatt in orbit could cost roughly half as much as a gigawatt on Earth.

He also said the terrestrial power-and-cooling component feels “reasonably inflationary.” That makes the orbital comparison more compelling in his view if Starship reusability works at the necessary scale.

Sovereign AI, in Baker’s view, will be widely pursued but mostly not at the frontier. Countries will want national AI strategies at least for defense, but power shells and GPUs are not enough if a country lacks frontier talent and massive usage.

For essentially every country other than the United States and China — and Baker added that he thinks China will fall further behind — he expects sovereign AI to look like localized adaptation rather than independent frontier development. That means using one of the leading providers to perform reinforcement learning on a country’s language, culture, and values; adding a system prompt; doing supervised fine-tuning; and running the result on the best available open-source model.

Countries would then run those systems in their own data centers so they feel secure about defense, intelligence, and possibly policing agents operating continuously. Baker expects a significant buildout for that kind of sovereign AI. But “sovereign AI at the frontier,” he said, is not something he sees.

On China, Baker’s view was that the country has made a “terrible mistake” by not taking what he characterized as the administration’s apparent willingness to let China buy H200s or B30s. He said China seems to believe its internal chips are good enough, and his judgment was blunt: “they’re not.”

Baker argued that Chinese labs’ strength in distillation may be masking the hardware problem. He said someone told him it took only 160,000 reasoning traces from o1 and o3 to get the original DeepSeek. He described Chinese labs as highly clever at industrial-scale distillation, running through multiple APIs, using many endpoints, and building what he compared to the iPhone farms seen in China — but for distilling models across every available API.

That advantage, in his view, depends on frontier models leaking enough useful traces or outputs to distill from. If labs stop releasing frontier models in ways that expose those signals, the distillation path weakens. Baker said Mythos is a sign of things to come in that respect. Coogan agreed that labs appear to be improving at locking down reasoning traces and espionage, and said that during the o1 period the labs may not have fully appreciated how feasible and threatening distillation was.

Baker has seen rare companies early. He said he owned 15% of Nvidia when it was below a $2 billion market cap and 10% of Tesla when it was below a $2 billion market cap. SpaceX, Nvidia, and Tesla have each been special to him in different ways. But SpaceX is different because of the scale of what it may become and the physical experience of what it already does.

Baker said there is a chance SpaceX becomes one of the most important and iconic companies of all time. It already is, in his view. It may even become the most important. His argument was not limited to investor returns. He described SpaceX and Musk’s broader companies as touching decarbonization, low-cost internet access for low-income countries, schools, and hospitals, and technologies that help blind people see and interact with the world. He also highlighted the SpaceX story’s blue-collar wealth creation, saying many blue-collar workers have become very wealthy.

The investment implications link back to the infrastructure argument. If someone opposes data centers on Earth for environmental reasons, Baker said, SpaceX may be part of the solution through orbital compute. If reusable launch lowers the cost of putting compute in orbit, the company’s space capability becomes relevant not just to exploration but to AI infrastructure scarcity.

The most concrete advice Baker gave was to see a rocket launch in person. He said there is “nothing like” it and recommended taking children. People do not need special SpaceX tickets, he said; they can go to a beach at Boca Chica and be as close or closer than the special viewing area, or find hotels near Vandenberg. The experience is more visceral than people expect: the sound, the blast of hot air, and the emotional reaction. “A lot of people cry,” he said.

That physical, inspirational quality is why he doubts another company will deliver a similar visceral experience. If SpaceX puts a city on the moon and then Mars, and becomes what he called “the British East India Company of the solar system,” Baker said it will probably be the most important company of his lifetime and maybe of all time. He acknowledged that much hard engineering remains, but the possibility is what makes the company feel singular.

Baker’s experience owning Nvidia and Tesla below $2 billion led to a narrower question: where, if anywhere, the next sub-$2 billion opportunities are. Baker said the more interesting activity today is not necessarily in public companies. He is seeing the knock-on effects of AI in biology and materials science, and pointed out that Neuralink was below $2 billion for a long time. The companies he is most interested in may not be “in theme” as pure AI companies, but are accelerated by AI.

In venture markets, he said, there are “loads” of incredible startups under $2 billion, and he speaks with them several times a week. Public small-cap land, by contrast, had not been especially active until AI and the bottleneck trade arrived.

That public small-cap AI trade worries him. Baker said it feels strange that a social-media account with a large following can post a thesis on a $35 million market-cap company and send the stock up 1,000%. He clarified that he was not referring to one specific account and said some of the people involved are smart. Still, he said there appears to be “a lot of shady stuff” happening.

Coogan connected the moment to the end of the dot-com cycle, when euphoric behavior eventually produced consequences. Baker agreed that after that period there were civil penalties, bankruptcies, and other market perturbations, even though the era also produced enduring generational companies. Hays noted that many current accounts are anonymous and may be operating through made-up emails and VPNs.

Baker said the current setup is different from sell-side research that goes too far but still sits inside a compliance apparatus. Social-media-driven microcap promotion is “uncharted territory.” His advice was correspondingly plain: stay safe, and do your own research.