(Preview) SpaceX Hype and the Elon Bargain, Nvidia and the Neoclouds, Q&A on Dropbox, Google, Ferrari Luce Backlash

Summary

• SpaceX’s proposed $2 trillion valuation does not pencil against $18.6 billion of revenue, a $4.9 billion loss, and growth slowing from 35% to 33%. The cited figures say adding xAI—and thus X—tipped the company from a small profit to a massive loss via $5.1 billion of AI R&D, which went toward a fifth-place model whose entire founding team recently left. Yet Ben Thompson supports the IPO because Elon Musk repeatedly converts belief into the capital required to pursue otherwise impossible infrastructure: “He creates movements to fund infrastructure instead of infrastructure building to a movement.” The bargain is that hype may eventually “backwards justify” the valuation.
• Tesla demonstrated why conventional valuation objections alone are inadequate for analyzing Musk companies. Roughly 300,000 people put down $1,000 for the Model 3 within about 12 hours, while later Tesla stock issuances perversely drove the share price higher despite dilution. Skeptics could be fundamentally right while holders still became rich: Musk is “the master of retail” and memes.
• The SpaceX IPO gives public investors a genuine venture-style opportunity—including the possibility of losing everything. Thompson points to Tesla, Starlink, routinely landed rocket boosters, and remarkably capable driver assistance as evidence that Musk has earned some credibility, while stressing that vision-only full autonomy remains unresolved and Starship is the S-1’s central risk. “You wanna be a VC? Here’s your chance.”
• Terrestrial compute remains cheaper and has numerous escape hatches before orbital data centers become necessary. Stranded natural gas can replace flaring, old Bitcoin facilities are already becoming GPU sites, and offshore infrastructure could host compute; the immediate constraint is that proper turbines may take “seven or eight years” to obtain. Thompson accepts that simply manufacturing more turbines might prove easier than space compute. Andrew also cites a $45 billion, three-year Anthropic compute contract for SpaceX, though either side can exit with 90 days’ notice; the AI-infrastructure business is already real and should grow.
• Orbital compute is ultimately a marginal-capacity thesis, not a claim that space will quickly undercut Earth. Solar power is free, but cooling a hypothetical satellite-scale unit of, say, 100 kilowatts is difficult—the International Space Station is only about 70 kilowatts—and the whole model depends on Starship working at scale. Earth facilities may retain better margins while SpaceX supplies the marginal next unit of compute and tries to “make it up in volume.”
• SpaceX’s launch monopoly may erode precisely because its success proves reusable rockets are possible. Blue Origin has landed suborbital boosters, and China has strong incentives to follow, so Thompson sees xAI’s potential model and tooling network effects as more durable than infrastructure alone. If multiple companies eventually launch reusable rockets and deploy orbital compute, he considers the SpaceX thesis vindicated even if SpaceX loses exclusivity.
• The preview leaves NVIDIA’s neocloud economics as an explicit unanswered risk. NVIDIA’s ACIE segment covers AI, cloud, industrial, and enterprise buyers outside the major hyperscalers—Google, Microsoft, Amazon, and Facebook—including CoreWeave and GPU “wildcatters.” The open question is whether those customers survive normalized supply while paying NVIDIA’s huge gross margins on its bundle against hyperscalers using cheaper internal ASICs.

Deep dive

1. Musk turns speculative belief into productive capital

• Sharp reads Thompson’s deliberately hostile arithmetic: a $2 trillion SpaceX valuation against $18.6 billion in revenue, $4.9 billion in losses, and growth slowing from 35% to 33%. The cited figures say adding xAI—and thus X—tipped the company from a small profit to a massive loss via $5.1 billion of AI R&D, which went toward a fifth-place model whose entire founding team had recently left the company.

• Thompson’s framework begins with an inversion: traditional mass movements required years of organization before producing spectacle, while social media can summon crowds overnight without supporting infrastructure. Musk has “nuclearized” the Jobsian reality-distortion field, using those crowds to finance the infrastructure behind the promise.

• The Model 3 supplied the canonical example: about 300,000 people placed $1,000 deposits within roughly 12 hours. Later, Tesla could issue dilutive shares and watch its price rise as investors interpreted the capital raise as fuel for achieving the dream.

• Sharp’s pushback is empirical: his money-manager friend had told him for eight years that Tesla’s fundamentals did not make sense, yet believers were rewarded and Tesla succeeded where a new car company had not in roughly a century, in Sharp’s framing. Whatever the valuation, “clearly something he’s doing is working.”

2. The IPO sells venture risk directly to retail

• Thompson is not offering a blanket defense of Musk, but the physical achievements matter: boosters land on ships multiple times weekly, Starlink exists, and Tesla’s driver assistance can reliably travel door to door without steering-wheel intervention. Whether vision-only systems reach full autonomy remains open.

• Starship is the load-bearing caveat. SpaceX caught the vehicle with “chopsticks” in 2024, but roughly two years later development remained slower, harder, and more expensive than expected; the S-1’s risk factors begin with multiple pages on Starship because the broader capacity thesis depends on it.

• Supporting the IPO is therefore not the same as recommending the stock. “You wanna be a VC? Here’s your chance,” Thompson says—and venture investing means possibly losing everything. He might buy a few shares because “we might pull it off,” but repeatedly stresses buyer beware.

3. Earth has plenty of ugly compute capacity left

• A listener’s challenge lands: if power and zoning are the bottlenecks, offshore platforms, stranded-energy sites, and other terrestrial workarounds should beat disposable orbital hardware. Thompson calls this “totally valid pushback” and concedes that SpaceX is partly searching for the largest nail available to its rocket-shaped hammer.

• West Texas already has natural gas that producers simply flare because they cannot move it. Turning that gas into AI compute could be environmentally preferable to burning it uselessly, while former Bitcoin facilities are being gutted and repopulated with GPUs by a new generation of compute “wildcatters.”

• The hidden bottleneck is generation equipment: converting gas into electricity requires turbines that Ben estimates may take seven or eight years to procure. Siemens, GE, and possibly Mitsubishi Heavy possess valuable blade expertise; less-efficient alternatives such as Bloom Energy fuel cells are being pursued because demand exists now.

• Thompson nevertheless writes the listener’s rebuttal for him: “Wouldn’t it be easier to make more turbines than to have data centers in space?” His answer is candid: “Yes, that might be the case.”

4. Orbital compute wins only at the margin

• Space offers one fundamental advantage—“free power”—with less conversion machinery than extracting gas, burning it, and spinning generators. The proposed unit is not a warehouse in orbit but roughly Starlink-satellite scale, albeit with a far harder cooling problem.

• Thompson benchmarks that problem against the International Space Station at roughly 70 kilowatts; a hypothetical compute unit might reach 100 kilowatts. Keeping one side shaded and the other sunlit might help, but deploying enough radiating surface remains an engineering requirement, not a solved detail.

• Musk’s characteristic pattern is enormous upfront difficulty followed by a simple, repeatable endpoint. Starlink on aircraft now looks obvious, but required rockets, reusable rockets, and the constellation itself; terrestrial GPU sites are easier to start yet retain continuous fuel, turbine, and maintenance complexity.

• Earth data centers will probably remain cheaper and extremely profitable. The orbital thesis is that commodity pricing depends on the cost of the next unit: if demand becomes effectively infinite and SpaceX alone can add marginal compute, it can “make it up in volume”—a phrase Thompson says should always trigger suspicion.

5. Competition shifts the prospective moat toward xAI

• SpaceX’s lead is not necessarily permanent. Blue Origin has landed boosters in suborbital flights, China is working on reusable launch, and sufficiently large incentives solve problems—especially once a pioneer demonstrates that the problem is solvable.

• Thompson therefore sees logic in continued xAI investment. His tentative framing is that xAI’s new team is “basically Cursor”: a coding-focused business with useful data, enterprise growth, heavy compute needs, and no viable future if its ultimate model is to send requests to Anthropic, OpenAI, or another provider.

• Sharp adds that SpaceX has a $45 billion, three-year compute contract with Anthropic, although either party can exit with 90 days’ notice. Anthropic currently needs the capacity, making the AI-infrastructure business real and likely to grow even before orbital compute arrives.

• Software tooling and model network effects could outlast a rocket monopoly. Thompson’s broader bet is that agentic inference may make compute look radically different in 10 or 20 years; if several companies ultimately deploy reusable rockets and space compute, “it’s a wonderful world. It’s a world I want to live in,” so he is not worried about SpaceX losing exclusivity.

6. NVIDIA’s neocloud moat remains the preview’s cliffhanger

• Thompson defines ACIE as AI, Cloud, Industrial, and Enterprise—effectively NVIDIA sales outside Google, Microsoft, Amazon, and Facebook. It includes CoreWeave, governments, corporations purchasing their own GPUs, and the “wildcatters” converting stranded facilities.

• The unanswered challenge is whether ACIE’s strength reflects scarcity rather than durability. Neoclouds must buy NVIDIA’s high-margin bundle, while hyperscalers can deploy internal ASICs and avoid the NVIDIA tax; the preview ends before Thompson answers what happens when accelerator supply and demand normalize.