SpaceX's IPO Isn't About Rockets — It's a $50B-Per-Gigawatt Compute Play

On 18 June 2026, the hosts of TBPN spent a broadcast cycling through four of the most consequential corporate stories of the year — and only one of them actually had anything to do with the company's nominal business. The thesis that threaded the hour together, articulated most clearly by Gavin Baker, chief investment officer at Trades Management, was blunt: SpaceX's IPO performance is a compute trade dressed in a space costume. Investors aren't paying for Mars or asteroid mining. They're paying for gigawatts, and the Google deal implies roughly $50 billion per gigawatt of monetized capacity.

The reframing matters because it sets the price. If SpaceX is a launch-services business, the math is mundane — a few hundred launches a year, fading margins, and a long-dated optionality on a colonisation thesis nobody can underwrite. If SpaceX is an AI-infrastructure business that happens to own the cheapest delivery mechanism for orbital compute, the math is a different species entirely. That is the price the public market has, apparently, decided to pay.

The bottleneck trade is ending — and the new game is franchise value

Baker's argument, laid out on the same broadcast, is that the easy money in AI's supply chain has already been made. The "bottleneck trade" — buying the picks-and-shovels suppliers, the Japanese toilet-parts manufacturers whose share prices ripped when hyperscalers scrambled for capacity — is approaching exhaustion. A Japanese toilet-supplier named Aimoto, Baker noted, recently declined to raise prices on its AI-adjacent components. That kind of pricing restraint is the death knell of a bottleneck. "The bottleneck trade is kind of nearing its end," he said. "The next game is what has enduring franchise value."

SpaceX, in that frame, qualifies. So does anyone who controls a scarce input that AI cannot do without at scale. Compute is the input, and the question of who controls gigawatts of it — at what cost, on what timeline — is the only question that will move capital for the rest of 2026 and most of 2027.

How cheap is too cheap? The orbital compute math

The numbers Baker sketched are striking. Terrestrial gigawatt-scale data centre capacity runs roughly $60 billion once you price in land, power, cooling, and the increasingly punitive cost of grid interconnection. The same gigawatt, delivered to orbit via a fully reusable Starship, drops to roughly $30 billion — and the launch cost itself is the cheapest line item at around $5 billion. The savings come from skipping the cooling and power-distribution infrastructure that a terrestrial build requires, plus the free solar flux above the atmosphere.

That is a structural cost advantage, not a marketing claim. And it is the basis for SpaceX's compute pricing power. The company, Baker argued, is "monetizing gigawatts at roughly two to three times average NeoCloud pricing," with the Google contract serving as the public anchor for the implied multiple. "If you can energize two, three, four gigawatts in the next year — that's a lot of revenue," he added. Public.com, the retail brokerage whose co-founder Leif Abraham appeared on the broadcast, reported that SpaceX accounted for nearly half of all single-stock purchases on its platform on IPO day — 533% more trading volume than Nvidia, the second most-bought name. Retail retention sat at approximately 125% net share retention, meaning customers who got an allocation added to their position rather than flipping it.

The public market, in other words, did not misread the listing. It read it exactly the way Baker is arguing it should be read: as a bet on the cost curve of compute, with rockets as the delivery mechanism.

Fox's $22B Roku bet is a bet that the OS layer matters more than content

While SpaceX repriced itself as compute, Fox Corp announced the largest acquisition in its history — a deal to acquire Roku valued at $22-25 billion on an enterprise basis, depending on whose reporting you trust. The structure runs roughly $160 per share: $96 in cash and approximately 0.99 Fox Class A shares for each Roku share. Fox will fund the cash leg with $12 billion in new debt plus balance-sheet cash, and is targeting $400 million in annual cost synergies.

The strategic logic is unglamorous and probably correct. Roku controls over 20% of the connected-TV operating-system market — a hair behind Samsung's Tizen at 23% — and reaches more than 100 million streaming households globally. Critically, Roku captures over 40% of CTV engagement and watch-time despite its OS share, a structural over-index that makes its ad inventory disproportionately valuable. Combine that with Tubi, the ad-supported streamer Fox bought for roughly $400 million in 2020, and you get a vertically integrated CTV advertising business sitting on top of the largest independent OS layer outside of the smart-TV manufacturers.

The bullish case is that the CTV ad pie keeps expanding even as linear collapses — a dynamic reinforced by the share of US premium video-on-demand signups that are now ad-supported, which has climbed from 39% two years ago to roughly 50%. The bearish case is that Roku is a hardware-dependent business with deteriorating unit economics, and Fox is overpaying for a platform whose moat narrows as smart-TV manufacturers vertically integrate their own OS layers and ad stacks. The market's initial reaction to the deal terms was, by most accounts, cool. But the structural bet — that owning the operating layer matters more than owning the content that runs on top of it — is the same bet Apple, Google, and Amazon have been making for a decade. Fox is simply arriving late to a thesis that has been validated by the trillion-dollar market caps of the companies that originated it.

Anthropic's export-control mess exposes an unsolvable trilemma

If SpaceX and Fox represent the upside of the AI-capex cycle, Anthropic represents the regulatory downside — and the downside is moving fast. According to the broadcast's reconstruction, the Commerce Department issued an export-control directive on 12 June 2026 covering the company's Mythos and Fable 5 model families, barring any foreign national from using them — inside or outside the United States, including Anthropic's own foreign-national employees. The compliance burden, particularly the know-your-customer and access-authorisation workflow, forced Anthropic to suspend both models for all users entirely.

The timeline tells the story. On 4 March, the company received a supply-chain risk letter. On 7 April, it previewed Mythos alongside an internal project called Glasswing. On 9 June, it launched Fable 5. On 12 June, Fable 5 was suspended. Three days from launch to suspension is a measure of how brittle the frontier-model business has become under the weight of overlapping national-security regimes.

The deeper problem is structural. The US wants to protect frontier models from leakage to adversarial jurisdictions. It also wants to hire the best machine-learning researchers in the world, many of whom are not American. And it wants American model providers to ship products globally, generating the revenue and deployment scale that justify the capital costs of training. You cannot do all three simultaneously. Anthropic's suspension is the visible cost of trying.

Baker's broader framing on the same broadcast is that "sovereign AI at the frontier I don't see it" for anyone other than the US and China — a position that includes a pointed critique of Beijing's chip strategy. "China has made a terrible mistake not taking the H200s or P30s. They have this crazy belief that their own internal chips are good enough. They're not," he said. For most other countries, he argued, the sovereign-AI ambition will reduce to a familiar recipe: take the best open-source model, run reinforcement learning and supervised fine-tuning on local language and cultural data, add a system prompt, deploy domestically. That is not a frontier play. It is a localisation play, and the EU's reported effort to restrict Mistral's exports Baker described bluntly as a "meme."

The Applovin paradox and Meta's token-budget 180

The other thread running through the broadcast was the divergence in how AI-spend is being managed inside the platforms. Applovin, which Raphael Vivas discussed on the show, has built a $180 billion valuation on the back of mobile-gaming advertising and a token-maxing posture that Vivas openly endorsed: "We're token maxing, baby." The company is processing more than $12 billion in annual ad spend, growing roughly 70% year-over-year, and currently holds less than 0.01% of its addressable market — a $120 billion annual in-app purchase market that reaches over a billion people per day, weighted toward household decision-makers and female users.

Meta, by contrast, is reportedly executing a sharp reversal. The company has moved from epitomising the token-maxing ethos — internally tracking "billions a year on Claude" — to implementing token budgets, allocation decisions, and supporting tooling scheduled for 2027. The broadcast's hosts cited an unsubstantiated but reportedly informed account of a Meta engineer who burned through $90,000 in AI tokens in a single day and was terminated within three days. Whether or not that specific anecdote holds up, the directional shift is real: even the hyperscalers that can afford to burn are starting to meter.

The deeper pattern

What unifies these four stories is a single underlying question: who owns the scarce inputs of the AI economy, and for how long? SpaceX owns cheap delivery to orbit and is monetising it at a premium. Fox is buying the CTV OS layer before the smart-TV OEMs finish verticalising. Anthropic discovered that frontier models are now subject to national-security regimes that can switch them off in three days. And the platforms themselves are splitting into token-maxers and token-budgeters, with Applovin and Meta on opposite sides of that line.

The bottleneck trade is over. The franchise-value trade is just beginning. The public market, judging by the SpaceX IPO tape, has already made its call.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://www.youtube.com/watch?v=lMd1APMz7zE