BG2 Podcast: SpaceX Emerges as Fourth-Largest Hyperscaler in AI Compute Within 30 Days
Brad Gerstner hosts Gavin Baker, Andrew Fox, and Clark Tang ahead of SpaceX IPO, June 12, 2026
SpaceX transformed itself from a space and communications company into the fourth-largest AI hyperscaler in just 30 days, according to Gavin Baker of Atreides Management speaking on Brad Gerstner's BG2 podcast. The revelation came as the company prepares for its $1.77 trillion IPO at $135 per share, with Goldman Sachs and Wall Street Journal reporting expectations of $160 billion in revenue by 2028, up from approximately $18 billion last year.
Unprecedented Monetization of AI Compute Infrastructure
The most striking development discussed was how SpaceX's AI compute business emerged as a major revenue driver that virtually no one had modeled just weeks ago. Clark Tang of Altimeter demonstrated that SpaceX's deals with Google and Anthropic for cloud computing generate more operating profit per gigawatt than any other provider, including Meta, Google's own infrastructure, and OpenAI. The Google deal in particular commands approximately $50 billion in annual monetization per gigawatt, while the Anthropic deal generates $22 to $23 billion per gigawatt.
Baker emphasized the significance of this development, noting that "nobody had AI compute in the SpaceX model" until recently. He pointed to analysis by Altimeter's Freda showing a 55% IRR on the Colossus 1 data center. "If you can borrow money at six, seven, eight percent and invest in something with a 55% IRR, that math maths," Baker said.
Jensen Huang told Gerstner that Elon Musk brings data centers online faster than anyone, with a 122-day deployment for 100,000 GPUs compared to the typical three-year planning cycle plus one year for installation. "Speed is literally cost because every day you're paying electricians and plumbers," Baker explained. This execution advantage has positioned SpaceX ahead of numerous competitors in the NeoCloud space, surpassing established players like Oracle in total AI compute capacity.
The Cursor Acquisition Changes the Frontier Model Race
Gerstner argued that the market is underappreciating what the Cursor acquisition means for SpaceX's xAI business. The 700 to 800-person company was projected to exit this year at up to $10 billion in revenue and possessed proprietary coding data exceeding what exists on the public internet. Baker noted that both Cursor and Anthropic have more tokens of proprietary coding data than are available publicly, which has proven crucial for model performance.
The combination appears to be working. SpaceX's Composer 2.5 model, trained using the Cursor data on the Kimmy K 2.5 base model with three weeks of reinforcement learning on the Colossus 2 cluster, achieved Pareto dominance on coding benchmarks just 12 days before the podcast. "What I think is so impressive is that Composer 2 was Pareto dominant at the lowest level of intelligence with very little training," Baker said, adding that the proprietary Cursor data, when trained with chinchilla optimal or beyond chinchilla optimal reinforcement learning, suggests "xAI and SpaceX has a shot of being a real player in coding."
Baker believes this is the most underrated variable in the SpaceX story. "If I had to say what the one piece that's being lost in the story, I think they've dramatically advanced their capability when it comes to building a frontier model," he said. The integration gives SpaceX a unique position where they can pull all their monetized compute in-house to train models and then run them, creating a powerful flywheel effect.
Orbital Data Centers Offer Five-Fold Cost Advantage
Andrew Fox laid out the economics for orbital compute, which depends critically on achieving rapid two-stage reusability with Starship. At current projections with reusable rockets, the cost to put a gigawatt of compute capacity in space would be approximately $5 billion versus $20 to $25 billion for the terrestrial infrastructure alone, not including the GPUs and silicon.
Fox explained that with Starship's 100 metric ton payload capacity and satellites designed for five megawatts each, the math works out to about $5 billion per gigawatt of orbital capacity. Combined with the roughly $35 billion for GPUs and silicon, total orbital deployment costs would be around $40 billion per gigawatt versus $60 billion terrestrially. The key is that power, cooling, and physical space are essentially free in orbit.
Gerstner noted that the terrestrial costs are likely inflationary while orbital costs could be deflationary over time, creating an expanding advantage. However, Baker cautioned that rapid reusability "is a really hard thing" even though "I've watched Elon do many hard things and this is a really hard thing, so I think it's reasonable to think that they're going to succeed." The company plans to attempt bringing back the Starship second stage later this year and making it reusable by next year.
Core Business Remains Strong Foundation
Fox emphasized that rapid reusability is foundational to everything else SpaceX aims to accomplish. The company is moving from approximately 160 to 165 launches last year to the high hundreds in coming years and potentially thousands of launches within three years. "The company's aspirations thousands of launches, you're launching, you're doing two or three launches a day," Fox said.
On Starlink, Fox noted that broadband penetration remains under 1% of global households despite hundreds of millions of terminals being a realistic target if rapid reusability is achieved. The revenue models show connectivity revenue including Starlink direct-to-cell growing from roughly $10 billion to $50 billion by 2028. Baker framed this as just 0.3% penetration of the global telecom market and observed, "I travel with Starlink. I'm a big video gamer and very consistently wherever I am in the world, Starlink is the best connection, fastest, lowest latency."
Gerstner pointed out that the implied AI business monetization rate in the leaked $160 billion 2028 revenue figure is about $14 billion per gigawatt per year, yet SpaceX just signed deals at $23 billion and $50 billion per gigawatt. "I think you can invest behind the AI business terrestrially and still be excited about it," Fox said, suggesting orbital compute represents additional upside rather than a requirement to justify the valuation.
The Frontier Model Race Accelerates
The discussion of Anthropic's Fable 5 launch and the broader model landscape revealed a fundamental shift in how intelligence should be measured. Gerstner highlighted a tweet from Noam Brown suggesting snapshot benchmarks are no longer relevant because the x-axis needs to be time, tokens, or compute rather than single-point measurements. Models can now solve most problems if given sufficient runtime.
Baker called this insight "so profound" and used an analogy to explain its significance. "Imagine Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn't have to eat. He doesn't have to sleep. He doesn't have to relax. He doesn't drink and never gets old, never has diminution of intelligence. And he thought for one year. I mean we might already have solved a lot of these intractable problems." He added that "we do not know how smart these models are and we may never know how smart each generation of models actually is or was because we don't have time to appropriately evaluate their intelligence before the next model comes out."
Tang provided hands-on examples of Fable 5's capabilities, including dumping seven financial models into the system and requesting a master view with contradictory assumption analysis, as well as analyzing three years of notes to identify highest-signal sources. "The model is able to reason through all of our assumptions," he said, adding that multi-agent orchestration represents just the beginning of these capabilities.
Open Source Versus Frontier Debate Resolves Toward Frontier
Gerstner challenged the consensus view from earlier this year that open-source models and cheap tokens would close the gap on frontier models. The evidence from the first six months of 2026 points decisively in the opposite direction. Baker noted that empirically, "90% of the revenue, probably more than 90%" accrues to frontier models despite open source potentially representing 80% of tokens consumed.
Tang observed a stark geographic divide in beliefs, with Silicon Valley heavily favoring closed-source cloud approaches while Asia expects model-to-workload optimization to dominate. He suggested the next year will be "the most indicative of which way this falls," but noted that closed source captured value this year because "the models actually get the intention and actually carry through the work" with the first truly useful agents.
Baker raised the possibility that Nvidia could disrupt the entire landscape by bringing open source to the frontier whenever it chooses. "If all of his customers are going to compete with him, then why not compete with his customers?" Baker asked, noting that Nvidia already has excellent open-source models like Nemotron 3.1 but has been careful not to compete directly with Anthropic, OpenAI, and Google. The implication is that ASICs from various hyperscalers create an incentive for Nvidia to strengthen its own position through superior open-source offerings.
Capital Expenditure Acceleration Justified by Revenue Growth
Morgan Stanley recently raised its 2027 AI capex forecast from $950 billion to $1.1 trillion, though Gerstner believes the actual number including SpaceX, CoreWeave, and other players will approach $1.5 trillion. This compares to approximately $300 billion in AI lab inference revenue projected for 2027, raising questions about returns on invested capital.
Baker argued the math works at 50% to 70% gross margins on that revenue base, but more importantly believes the $300 billion figure is too conservative. "I think we end this year well over $200 billion in inference revenue. Well over," he said. The group noted that Dario Amodei predicted revenues reaching the low hundreds of billions by 2028 and said "it's hard for me to see that there won't be trillions of dollars in revenue before 2030."
Tang highlighted that monetization rates per gigawatt have been increasing from roughly $20 billion early this year to $30 to $40 billion currently, representing pure margin flow-through on a heavy fixed-cost base. "The revenue might actually outstrip our fixed cost base by a significant amount," he said, explaining why labs are accelerating spending. Fox added that less than 0.2% of people on Earth are using AI in an agentic way, suggesting massive headroom for demand growth.
Gerstner put the scale in perspective by noting that in the last seven years, the Mag 7 added one trillion in revenue generating $17 trillion in market cap. The forecast now calls for adding another trillion in revenue across just three companies, SpaceX, Anthropic, and OpenAI, in half the time. "We are going to have bumps in the road, but we're going to higher highs because the size of the prize, this is going to transform 5, 10, 15% of global GDP," he said.
Portfolio Management in a Seasonal Market
Both Gerstner and Baker discussed dialing back from large to medium-small positions after the substantial run in semiconductor stocks, though both emphasized these are relative adjustments rather than bearish calls. Baker used a runner analogy, describing how the market climbed out of the 2022 downturn with substantial energy but has now "run up a very steep hill" particularly in semiconductors, with many stocks having "gone straight up a cliff."
Gerstner noted that despite war with Iran, oil at $100, rising CPI, internet down 16%, and software down 8%, the market has performed well because "the world underestimated AI revenues and underestimated the amount of compute that was going to be needed." However, he flagged concerns about heading into a seasonally weak period, noting that AI has shown seasonal patterns for three summers as college students reduce usage.
Baker acknowledged multiple near-term risks but concluded, "When I think about what Gnome Brown said and when I see the capabilities of Fable, it's just hard for me to get too bearish." He emphasized always assuming "a bullet is coming for me" with head on a swivel, but the fundamental trajectory remains compelling.
On the SpaceX IPO specifically, Baker noted unprecedented dynamics including Musk's 50% ownership with a 365-day lockup and the fact that employees and investors have had liquidity every six months for roughly a decade through secondary markets. "If you're a SpaceX employee or former employee and you wanted to sell, you've had whatever that is close to 20 chances," he said, suggesting much less pent-up selling pressure than typical IPOs despite a chart showing average maximum drawdowns over 50% for comparable large technology IPOs.
Gerstner characterized SpaceX as "a must buy, a must own, set it and forget it" for institutional investors wanting "a real bet on both the space and the AI future," while acknowledging the need to manage position sizing around volatility. The group's assessment is that breaking down each business line from first principles, Starlink connectivity growth looks achievable, terrestrial AI compute monetization looks achievable, and the model business post-Cursor acquisition represents potential upside that the market is underappreciating.