Venture capitalist and SpaceX board member Steve Jurvetson joins the Moonshots podcast to explain why his own partners voted down SpaceX and what the record Anthropic IPO means. He talks with Peter Diamandis, Salim Ismail, Dave Blundin, Alexander Wissner-Gross and Emad Mostaque, recorded the day after Anthropic began trading on the Nasdaq. The episode moves from the 2008 SpaceX deal to Elon Musk’s compute buildout, the coming memory bottleneck, Starlink, Terafab, and Voyager 1 reaching one light-day from Earth. Watch the full episode on YouTube.
TLDW
Jurvetson recounts how his firm, Draper Fisher Jurvetson, formally rejected SpaceX in 2008 and only said yes after he added deal terms that neither he nor Musk thought mattered. The panel reads the Anthropic IPO as the moment the public markets started deciding how big the next AI model gets, and debates who survives among Anthropic, OpenAI and Google. The second half covers Musk’s vertical integration, why memory rather than GPUs may be the next bottleneck, Starlink as a carrier for machine traffic, Terafab, and why Jurvetson thinks mission-driven companies beat ones that chase profit. It closes on Voyager 1, which passed one light-day from Earth on November 18, 2026.
Thoughts
“Maximize shareholder value” sounds like a clear goal, and Jurvetson’s best argument is that it is not one until you say over what time scale. A CEO trying to maximize revenue or shareholder return has to decide whether that means this quarter, this year or this decade, and every short-term decision trades against the long term: “You can always mortgage your future to make more money now.” A company whose goal is a mission, such as making humanity multiplanetary or speeding up the shift to sustainable energy, never has that fight, so the whole organization pulls the same way as it grows. He backs it with a Tesla earnings call where Musk said that as long as he was CEO, profits were not the priority, and with Musk’s pay package tied to a million people on Mars. Profit then becomes the byproduct of being good enough to lead an industry transition, because, in Jurvetson’s phrase, “you can’t suck and lead.” That reframes a worry raised earlier in the episode, that quarterly reporting will slow Anthropic down. The quarterly cycle only bites a company whose real goal is the stock price.
The most original framing of the Anthropic IPO is that it turned a stock offering into the budget process for the next frontier model. Jurvetson points out that the Apollo program had a budget range set by Congress, while AI has no natural ceiling on how much you could spend on the next neural net. So the question of how much of the future society mortgages to build something “20, 50, 100 IQ points higher” is now settled by whether an IPO is oversubscribed. That flows into GPUs, which sets the size of the next model. Whatever you think of the outcome, that is a strange way for a civilization to make one of its biggest decisions.
The part investors should pay most attention to is Jurvetson’s history of memory chips. His father ran the biggest memory fab in the world in the 1980s and watched it go from the top of the industry to bankrupt in about a year as Japanese capacity came online. DRAM has always been boom and bust because fabs take years to build while demand grew slowly. His point is that exponential AI demand breaks that cycle in the same way Tesla’s battery forecasts forced the first Gigafactory, and it is why Dave Blundin notes that Musk’s first target for Terafab is high-bandwidth memory, not GPUs. The cautious habits of established chipmakers came from a world of slow demand, and that world may be over. If Emad Mostaque is right that models will get far more efficient within two or three years, though, the same boom-bust risk comes straight back.
Blundin raises the nightmare case for anyone building giant GPU clusters: AI-assisted research finds a way to train without backpropagation, and hardware with a three-year life becomes obsolete in six months. Jurvetson’s answer is Jevons paradox. Nobody has found an upper bound on demand for computation, and if inference gets cheap, continuous learning in every business would absorb the surplus. The better half of his answer is about where compute could go. Older, less tightly connected clusters can run inference anywhere there is stranded energy, such as a geothermal site or a remote solar farm, with Starlink carrying the traffic. That is a more plausible future than a single winner owning all the compute.
Jurvetson’s line about seeing dead companies is played for laughs, but it holds a useful test for founders. NASA’s SLS and ULA’s Vulcan, plus the roughly 200 venture-backed small launch startups still pitching every week, all fail the same check: even if they hit every objective, they land at tens of thousands of dollars per kilogram while Starship aims far lower. As one host put it, why build a product that is a failure even if it succeeds? It is the same instinct that made Jurvetson leave internet investing when he saw the fourth pantyhose-on-the-web company get funded.
Key Takeaways
- Draper Fisher Jurvetson formally voted down its first SpaceX investment in 2008. It took two weeks, and added terms (anti-dilution, a 2x minimum at IPO, a 2x liquidation preference and a 2x redemption right), to turn two partners from no to yes.
- Those partners could not see SpaceX ever being worth more than two or three times the entry price, because they valued it against United Launch Alliance, Boeing and Lockheed. Musk accepted the terms without blinking because every outcome was under his control.
- Jurvetson’s investing rule since the dot-com era: back things that are “unlike anything I’ve seen before, yet adjacent to where I’ve been.”
- Anthropic and SpaceX each raised on the order of the largest IPOs ever. Jurvetson says the two together were about three times Saudi Aramco’s, and Wall Street’s machinery is not yet built to digest offerings that size.
- Because AI spending has no natural ceiling, IPO demand now helps decide how big the next frontier model will be.
- Jurvetson would put Google’s Gemini ahead of OpenAI among the likely top three labs. His reasons are Google’s pace of releases, its own TPU chips, and its willingness to integrate vertically.
- Alexander Wissner-Gross ranks Anthropic first, OpenAI second and Google a close third, and calls OpenAI’s safety explanation for delaying its IPO “safety ate my IPO.”
- Peter Diamandis said Musk described SpaceX adding about 500 megawatts of AI compute a month at roughly 60% of the industry’s cost, and expecting to pass a gigawatt a month by the end of 2027.
- Jurvetson credits SpaceX’s speed to hiring from outside traditional industries, refusing to accept mediocrity, and building in-house what suppliers cannot do well, including what he calls the largest printed circuit board facility in America.
- Memory and data movement, not GPUs, are becoming the bottleneck. Optical interconnects and vertically stacked analog chips could pack compute more densely than ever.
- Signals between satellites travel through a vacuum, which Jurvetson puts at about 30% faster than light through fiber, and on a straighter path. That gives Starlink a latency advantage for traffic between machines and orbital data centers.
- Netflix is Jurvetson’s lone example of a company that led its own disruption, and he argues only a founder with a large stake could survive the stock drop that came with it.
- Mission-driven companies avoid the time-scale fights that come with chasing profit, and they attract the best people.
- Jurvetson does not expect humans to reach Alpha Centauri. He expects data packages, robots and AIs to go instead, and says one Anthropic co-founder has shown interest in personally funding a permanent moon base.
Chapters
2:37 The SpaceX Bet His Partners Voted Down
In late 2008, Tesla was in crisis and Musk was borrowing against everything to keep it alive. Jurvetson offered to invest in SpaceX, and the company never even made a pitch deck. His partnership turned it down, the only time he remembers that happening, and one senior partner copied the whole firm on a reply warning against letting personal passion cloud investment judgment. Jurvetson argues that rockets are simply cool and passion should guide investing, and describes the rule he formed after seeing yet another pantyhose-on-the-web company get funded.
12:06 Anthropic’s IPO, Market Digestion and Peak Fear
The hosts react to Anthropic’s first day on the Nasdaq, where it closed about 17% above its offering price before falling the next day. Jurvetson compares it to Facebook’s IPO, when Wall Street could not absorb that much stock in a day, and says big funds need weeks to study a breakthrough before they buy. Wissner-Gross pairs the listing with New York’s ban on horse-drawn carriages the same day. The talk then turns to recent cyberattacks and outages. Jurvetson calls this a period of maximum vulnerability that should improve as AI hardens software.
25:49 IPO-ing the Singularity and Who Survives
Asked whether the size of the IPO means AI is mature, Jurvetson describes a singularity as “perpetual future shock,” with each year’s capital flows and winners’ returns growing more extreme. He contrasts Apollo’s congressional budget with AI’s lack of any spending ceiling, and worries that new wealth will flow into unproductive nonprofits. On which labs survive, he stresses culture and recruiting, takes a shot at Meta, and puts Google ahead of OpenAI.
38:49 Sam Altman, Gradual Handover and AI on the Board
A DealBook clip of Sam Altman warning against a gradual handover of the future to AI starts a debate. Ismail calls the handover inevitable for decisions like monetary policy, and Jurvetson expects analyst, legal and accounting roles to go to AI agents. The hosts doubt Altman’s safety explanation for delaying OpenAI’s IPO, and Blundin notes how often minority-shareholder CEOs are pushed out after going public. Mostaque suggests companies would do better with an AI on the board, and Diamandis says every startup should have one.
51:52 Gemini 4 Neon and the Frontier Race
Google released Gemini 4 Neon, its fourth Gemini model in seven weeks. Wissner-Gross places it just below the cost-performance frontier set by Anthropic and OpenAI. Jurvetson, whose neighbor is Sergey Brin, says what matters is the pace of improvement, and that Brin is more engaged than at any time since his early twenties. Mostaque points to the TPUs Google is quietly deploying and its new bundled AI plan, and explains “small model smell” as the point where smaller models fall apart on complex work.
1:06:58 Musk’s Compute Machine and Vertical Integration
Diamandis runs through Musk’s recent comments at DealBook and elsewhere: SpaceX adding about 500 megawatts of compute a month, and Anthropic’s filing disclosing large compute agreements with SpaceX. Speaking only in generalities as a board member, Jurvetson credits the talent, the refusal to accept mediocrity, and bringing software skills into industries that had not changed in decades. Ismail adds that strong companies bring their biggest outside dependencies in-house, as Tesla did with batteries and SpaceX with engines.
1:15:14 The Next Bottleneck: Training, Memory and Chips
Blundin lays out Musk’s dream (control the compute while models become easy to copy) and his nightmare (training without backpropagation makes today’s GPU clusters obsolete). Jurvetson, who is investing in continuous learning at the edge, argues demand for compute has no visible ceiling. The group agrees the bottleneck is moving to memory, data movement and optical interconnects. Mostaque warns that this is bearish for Nvidia within two or three years.
1:28:54 Starlink, Telecom and Orbital Data Centers
Musk says AT&T and Verizon will survive because bandwidth demand will explode, while Starlink goes after traffic between AI agents and robots. Jurvetson calls the target of carrying half of all internet traffic audacious and highlights the speed-of-light advantage in a vacuum. Wissner-Gross asks about laser links from orbit to the ground. Mostaque argues that on-device models acting as “bit translators” could cut the bandwidth the internet needs.
1:35:33 Terafab, the Fear of Hardware and Mission Over Money
Musk has reconfirmed Terafab, a planned 100-million-square-foot chip factory. Jurvetson says his board meetings feel like a startup’s, only at a trillion-dollar scale. He recalls Michael Dell advising Apple in the 1990s to shut down and return the cash to shareholders. Blundin and Wissner-Gross argue that 20 years of “hardware is hard” left founders afraid of physical things. Jurvetson, who has known Musk for 30 years, insists the mission, not money, drives him.
1:50:52 Dead Companies, Optimus and Voyager 1
Jurvetson calls SLS a dead company that does not know it yet. On Optimus, he argues each job will get its own form of machine, so tunnels will still be dug by boring machines. He is building a house facing the Starbase launch pad. The episode ends on Voyager 1 passing one light-day from Earth after 49 years. Jurvetson notes how empty space is past the Oort cloud, and one host recalls JPL showing him the spacecraft at 19 and a half light-hours.
Notable Quotes
“I try to invest in things that are unlike anything I’ve seen before, yet adjacent to where I’ve been.”
Steve Jurvetson, on the rule he adopted after leaving internet investing
“Rockets are inherently cool. There’s no getting around that.”
Steve Jurvetson, answering the partner who warned against letting passion cloud investment judgment
“Mandating horseless carriages on the one hand and publicly listing humanless minds on the other. I think yesterday was a key marker in history.”
Alexander Wissner-Gross, on the Anthropic IPO and New York’s carriage ban
“AI has infinite scalability in terms of the amount you could invest in the size of the next neural net.”
Steve Jurvetson, contrasting AI with the Apollo program’s budget
“The money that is made at the company level and as the leader of the company is the byproduct of meeting the mission.”
Steve Jurvetson, on what drives Elon Musk
“Why would you invest money and effort into a product that if it succeeds in all of its possible objectives is a failure?”
A Moonshots host, on legacy rockets and small launch startups
“We are birthing a symbolic immortality through probes and technologies that will be nonhuman in any sense of the word.”
Steve Jurvetson, on why humans will not travel to other star systems
For the full conversation, including the Gemini benchmarks and Musk’s Optimus timeline, watch the full episode here.
Related Reading
- Steve Jurvetson (Wikipedia) background on his career from Hotmail to SpaceX, Tesla and Future Ventures.
- Future Ventures Jurvetson’s current fund and the deep-tech companies it backs.
- Jevons paradox (Wikipedia) the idea behind Jurvetson’s argument that cheaper compute will raise demand rather than cut it.
- NASA Voyager 1 mission page the official record of the spacecraft that just passed one light-day from Earth.
- Exponential Organizations 2.0 by Salim Ismail, Peter Diamandis and Michael Malone, the book Ismail cites on purpose and culture as the CEO’s main job.