Starlink Economics Anchor ARK’s Case for SpaceX’s AI Upside

The core of ? brett-winton’s SpaceX case is not that every science-fiction-sounding ambition arrives on schedule. It is that investors can underwrite current and near-term Starlink economics while retaining upside from a more speculative AI infrastructure business, including the possibility of orbital data centers.

Winton said “SpaceX AI” is the largest position in ARK’s venture fund, an interval fund he described as open to ordinary investors rather than only wealthy investors. He said the fund is designed to give everyday investors access to AI-related private-market exposures, naming OpenAI, Anthropic, and SpaceX among its top positions. The portfolio graphic displayed alongside the discussion listed SpaceX, OpenAI, Anthropic, Tenstorrent, Kalshi, and Replit.

Winton framed the investment as a continuation of ARK’s long-running view that reusable rockets are a disruptive technology. From ARK’s inception, he said, the firm had treated reusable launch as a breakthrough because lower launch costs could unlock markets that were previously inaccessible. His cost progression was explicit: from $10,000 to $1,000 per unit of launch cost, and, in ARK’s view, to “much less than $100” with Starship.

That launch-cost claim is the base layer of the argument. Winton said it supports the Starlink business first, and AI second. The orbital data-center thesis matters, but he argued that investors do not need the AI business to work “that well” to get a strong return if Starlink scales.

The speculative AI layer begins with a simple premise raised by Caroline Hyde: AI needs data centers, and if SpaceX can put massive payloads into orbit, then data centers in space become part of the investment pitch.

Winton translated that premise into a revenue model. In his account, “Elon and Co.” are saying SpaceX will be able to launch “tens of gigawatts per year” by the late 2020s. A gigawatt, in Winton’s simplified model, can be rented at current market rates over the long term for roughly $15 billion. Tens of gigawatts at that rate implies, in his example, about $300 billion in incremental revenue in a year.

He then contrasted that long-term rental-rate estimate with what he said SpaceX has recently rented capacity for. Winton said that, on a blended average basis, the company has “just rented capacity to Google and Anthropic for just a little under $40 billion per gigawatt.” He presented that figure as support for the idea that serving as an infrastructure-as-a-service provider could be “wildly lucrative,” because, in his view, SpaceX can build assets efficiently on the ground and could build them efficiently once launching assets into space.

The distinction matters because Winton did not make the case depend entirely on SpaceX building a frontier AI model. Renting compute capacity to others could itself be a major business. The higher-margin version would be competing directly with Anthropic and OpenAI, but Winton described that as a harder path and not the only investable one.

The launch-cadence risk is the obvious stress point. Ed Ludlow noted that Musk had posted that one million tons to orbit would be possible in roughly five years. Starship, on paper, can carry 150 metric tons; reaching that mass would require 1,333 launches a year, every year, for five years. Ludlow asked whether Winton would concede that this was a long shot.

Winton did not dispute that Musk’s targets are extremely ambitious. His defense was that Musk uses such targets to make his workforce effective at delivering “world-changing technologies.” More importantly, he argued that ARK’s revenue expectations do not require assuming a thousand-plus launches a year.

His alternative scenario was simpler: instead of launching Falcon 9 rockets, imagine launching Starship rockets at the same cadence, filled with Starlink satellites. Winton said that alone would produce, after two years of launching, “a little under $200 billion in revenue.” On that basis, he argued, SpaceX can reach $300 billion to $400 billion of revenue by 2030, mostly from Starlink, without assuming the most aggressive launch cadence implied by Musk’s million-tons claim.

Winton also rejected the idea that near-term price-to-sales concerns are the right way to assess the company. He characterized some criticism as the view that SpaceX looks expensive in the immediate term and that retail investors do not understand what they are buying. His response was that the function of capital markets is to raise capital and deploy it into high-return businesses.

Starlink, in his description, is such a business. He said a rocket loaded with satellites, plus customer acquisition, could cost about $500 million, and should generate more than $1 billion per year over a five-year satellite life. He called that a “six month cash on cash return” for five-year-life satellites.

The counterweight to the growth model is that SpaceX, as described by Hyde, had a net loss of $4.28 billion on revenue of $4.69 billion in the first quarter. The question is when the company becomes profitable, whether it needs to, and whether ARK is comfortable with losses if revenue rises fast enough.

Winton said ARK’s model has profitability arriving “very quickly” as Starlink satellites launch. The key uncertainty, in his telling, is not Starlink but how much capital SpaceX directs toward terrestrial data centers to catch up with frontier AI labs through Grok.

Morgan Stanley revenue expectations presented in the source, with the Wall Street Journal cited on screen, put SpaceX at $160 billion in 2028, $330 billion in 2030, and $3.4 trillion in 2040. A separate Bloomberg graphic showed the expected IPO allocation: at least 20% for retail investors, less than 10% for foreign investors, and the remainder for institutional investors, friends, and family.

Winton’s answer separated three businesses. Starlink can be strong on its own. Renting compute to third parties can be an “amazing business” without SpaceX owning the best AI model. Competing directly with Anthropic and OpenAI could produce higher margins, but Winton said that probably does not happen until SpaceX is launching multiple gigawatts into orbit.

The AI story contains a second disconnect. Ed Ludlow pointed to a $26.5 trillion total addressable market figure in the prospectus for enterprise AI, while the nearer-term business looked more like a ready-to-go neocloud: rent out capacity, generate cash, and use that to fund the more ambitious projects. The question was whether Grok actually needs to beat Anthropic and OpenAI.

Winton said that outcome is possible and, in ARK’s model, more likely. He estimated that reaching the frontier by 2030 would require between 30 and 50 gigawatt-years of compute and training. Because xAI started later and is building its own infrastructure, he said it is “very difficult” for it to catch up in the amount of research-and-development compute applied to its model.

His caveat was Tesla. Winton said he assumes — and described as reasonable, or more likely than not — that after SpaceX lists and its shares are floating, there could be a merger between the entities. Hyde immediately noted that he was making that assumption. Winton said such a combination would make “a ton of strategic sense.”

The reason, in his view, is capital allocation. ARK’s Tesla model assumes cash flow from robotaxi that Tesla cannot effectively deploy inside Tesla alone. Winton said he would not want a special dividend out to shareholders; he would want continued reinvestment in growth. Optimus, Tesla’s robot business, is not capital constrained in his framing, but capability constrained. SpaceX, by contrast, can use additional capital to accelerate growth.