Gavin Baker on the SpaceX IPO, the 'token path' investment thesis, and why sovereign AI won't reach the frontier

Gavin Baker on SpaceX, the Token Path, and Sovereign AI

Gavin Baker, CIO and Managing Partner at Atreides Management, frames the SpaceX IPO as clean execution by Goldman Sachs and Morgan Stanley — a roughly 20% pop with none of the volatility that would have signaled mispricing. The more interesting question for him is what drives the stock from here, and he thinks the answer is closer than the Mars colony narrative suggests.

Near-term SpaceX drivers

Two variables matter over the next twelve months. First, how quickly SpaceX can bring terrestrial compute online. Baker says SpaceX is monetizing at $50 billion per gigawatt on the Google deal — a figure he attributes to Altimeter — which he describes as roughly two to three times higher than average neo-cloud pricing. With SpaceX understood to have ordered around 20% of the Rubin GPU allocation, and Rubin being a more drop-in replacement than Blackwell, Baker argues the company is signaling real confidence in its ability to energize data centers fast.

Second, Cursor. Baker says Composer 2.5, after three weeks of RL and supervised fine-tuning on Colossus 2, is already "Pareto dominant." He thinks that result, applied to a larger base model, makes SpaceX's position inside roughly half the Fortune 500 via Cursor a meaningful near-term revenue lever — not a five-year story.

On the supply-demand dynamics post-IPO, Baker notes that more than 10,000 SpaceX employees bought stock on the IPO. Investors and employees who wanted liquidity have had the option to sell every six months for a decade. The people who still hold haven't been looking for an exit.

“There are two variables that are gonna matter a lot over the next year. The first is just how quickly can they bring on terrestrial compute... they're doing 50,000,000,000 a gigawatt on Google deal... I think cursor is another big variable — Composer 2.5 after three weeks of RL and supervised fine tuning Colossus two is kind of Pareto dominant. So what's going to happen when it's applied to a bigger, better base model?”

Token path thesis

Baker's framework is that value accrues to companies sitting in the "token path" — the infrastructure that produces and delivers AI tokens. He's increasingly watching EV-to-net PP&E as a valuation metric, on the thesis that installed physical assets on Earth will appreciate.

On CDNs like Cloudflare and Akamai — Akamai signed what Baker describes as a $1.8 billion deal with Anthropic — he's constructive but precise. CDNs command a meaningful premium for low-latency delivery, and the Cerebras experience showed people will pay for speed. But in terms of total tokens consumed on Earth, Baker estimates CDNs are delivering less than 1%, possibly fewer than 10 basis points. They have a path into the token path; they're not fully in it yet.

At the other extreme, orbital compute. Once Starship is reusable, Baker's math puts the cost of a gigawatt in space at roughly $30 billion versus $60 billion on Earth — the terrestrial figure breaks down as $25 billion for power and cooling (which space eliminates) plus $35 billion for IT equipment. He puts the post-reusability launch cost at around $5 billion. Land on Earth, by that logic, is not a bottleneck worth chasing. The "bottleneck bros" trade — hunting for obscure upstream suppliers — is, in his view, largely played out, with the Japanese materials company Ajinomoto's decision not to raise prices serving as a recent illustration.

Meta and the endgame framing

Baker reads the competitive behavior across big tech as a signal that the AI endgame is arriving faster than most expect. OpenAI cutting Codex pricing, he argues, is because coding tokens are so valuable that labs without enough of them risk missing the reinforcement self-improvement loop entirely.

On Meta specifically, he sees the low EV-to-net PP&E multiple as the market pricing in deep skepticism about Zuckerberg's ability to monetize the asset base — skepticism he considers at least partly warranted given the absence of a breakout consumer AI product. But he points to how quickly Meta pivoted from "we're not cutting OpEx" to mass layoffs in 2022 after Brad Gerstner's letter, and argues the same pivot speed applies here. The company's stated position on not monetizing GPUs externally may not survive contact with a changed market.

The deeper point is that SpaceX's stock price matters for its ability to retain what Baker calls the "10,000x engineer" — compensation pressure from a declining stock is a real constraint on execution for any of these companies, Meta included.

Sovereign AI

Most countries will get a version of sovereign AI, but it won't reach the frontier. Baker's read is that nearly every nation outside the US and China ends up running a fine-tuned open-source model — some RL on language, culture, and values, a system prompt, supervised fine-tuning — inside its own data centers, primarily for defense and intelligence use. The compute buildout for that is still real, but the ambition stops well short of frontier model development.

China is a separate case, and Baker thinks it's falling further behind. The core mistake, in his view, was not securing access to H100s and P30s when the US administration signaled willingness to allow it. Chinese labs are exceptionally skilled at distillation — Baker says it took only 160,000 reasoning traces from o1 and o3 to produce the original DeepSeek, extracted through industrial-scale API pinging across what he describes as the equivalent of iPhone farms running roughly 100,000 endpoints simultaneously. But that advantage disappears if frontier labs stop releasing models openly, and he sees Mythos as an early signal of exactly that tightening.