All-In Podcast: Gerstner Says Anthropic and OpenAI Could Dwarf SpaceX's $2 Trillion IPO, While Chamath's CTO Reveals Token Costs Are Doubling Every 45 Days for Just a 5% Productivity Gain
Besties discuss the AI IPO pipeline, a sobering enterprise ROI data point, a widening frontier-model duopoly, and China's move to wall off its open-source models
On episode 280 of the All-In Podcast, hosts Jason Calacanis, Chamath Palihapitiya, David Sacks and Brad Gerstner (Freeberg was on vacation) used the SpaceX IPO as a template for what may be the two largest tech listings in history, while surfacing a pointed data point on artificial intelligence monetization that cuts against the industry's default optimism.
A CTO's warning on token economics
The most consequential disclosure came from Chamath Palihapitiya, who runs several portfolio companies through his venture platform 8090 and shared a direct conversation with his own chief technology officer. "Our token costs are doubling every 45 days," Palihapitiya said, recounting the exchange. When he asked what the downstream productivity gain had been, the answer was "maybe 5% max." His CTO's explanation, in Palihapitiya's telling, was blunt: "What we're finding out is that you need to use a lot more tokens to get to this next iteration of improvement because we've effectively already asymptoted." Palihapitiya said he does not know how many companies will hit this wall, but predicted "everybody in the next three or four years will for sure go through it," and argued that founders eyeing an IPO should get out before "that starts to seep into the water table."
The remark reframes the panel's later, more bullish commentary on enterprise AI spending. Sacks and Gerstner both pushed back on the idea that this signals broad trouble, arguing the market is still in an early, experimental phase where near-term ROI scrutiny is limited. But Palihapitiya's framing — that a "reckoning" on token spend versus earnings-per-share lift is coming — sets up a debate that public-market investors in Anthropic and OpenAI will need to underwrite before either files an S-1.
SpaceX becomes the playbook, and the price tags being floated are enormous
Gerstner, who was an investor in the SpaceX offering, called it "textbook." SpaceX raised $75 billion at a $1.75 trillion valuation and now trades around $150 a share, roughly a $2 trillion market cap, making it the seventh-largest company in the world on what Gerstner pegged at roughly $35 billion of forward revenue. He credited Elon Musk, Gwynne Shotwell and the SpaceX team with pioneering the mechanics other frontier companies will now copy: staged lockup releases tied to milestones, early index inclusion, and pricing discipline designed to avoid the 50% peak-to-trough drawdowns common in the first six months of trading.
Anthropic, which confidentially filed on June 1, is rumored — per prior guest Gavin Baker — to be on pace to exceed $100 billion in revenue by the end of 2026 while turning meaningfully profitable; Baker has said the company could trade at $3 trillion if it went public today. Polymarket currently prices a 65% chance of an Anthropic IPO this year, though on light volume. Gerstner said Altimeter "would be a buyer at scale and at size" in both Anthropic and OpenAI at today's operating metrics, while cautioning that a $3 trillion entry price would not be a "get-rich-quick" trade: "I don't expect that they're going to be priced in a way where you're going to get a 50 to 100% durable bounce out of the IPOs. If so, that would mean they were probably mispriced right into the IPO." He does expect both companies to compound revenue at "well over 30%" annually for years, a growth rate he noted is unprecedented at this scale.
On OpenAI, Gerstner said the company has "gotten its swagger and mojo back," pointing to a fresh model cycle, GPT-6 rumors circulating for a within-30-days release, and reported run-rate revenue near $70 billion — still below Anthropic's rumored trajectory but double SpaceX's revenue base. He said OpenAI likely lags Anthropic to market given the complexity of its corporate restructuring, but that both are IPO-ready "when it's time."
The frontier-model duopoly is widening, not narrowing
A recurring thread was the gap between token-level commoditization and revenue concentration. Sacks cited data showing open-source models' share of enterprise spend fell from 19% to 11% year over year, even as total token usage across the industry exploded. "Anyone who's saying that these closed models are gonna lose or are somehow losing, you're just not seeing it in the data," he said, attributing this to a "spirit is willing but the flesh is weak" dynamic: enterprises want to diversify away from frontier labs for cost and data-sovereignty reasons but lack the technical middleware — intelligent routing, portable memory and context — to actually do it.
Gerstner went further, arguing the market is coalescing into a duopoly measured by revenue, with Anthropic at roughly $60 billion-plus ARR and OpenAI in the $40 billion range and "I don't know if anybody else even registers." He raised the possibility that this gap widens rather than converges: "As it becomes recursive you actually extend the lead because the smarter your model gets, the more revenue you get, the more compute you can buy... I think there's a chance that over the course of the next two to three years... the distance between the frontier and everybody else doesn't converge. It actually extends." Palihapitiya was less certain, noting that heavy AI-native customers such as Lovable and ElevenLabs — both multi-hundred-million-revenue businesses spending tens of millions with frontier labs — are simultaneously building proprietary models to reduce dependence on those same vendors.
Real-world routing examples reinforced both sides of the debate. Uber's engineering organization, per CTO comments cited on the podcast, has built 200 "agentic skills" and routes over 70% of pull requests to local or cloud agents, while DoorDash CTO Andy Fang disclosed publicly that the company now delegates lower-complexity code review to Moonshot's Kimi 2.6 model while reserving Anthropic's frontier model for the hardest tasks — a benchmark DoorDash has released publicly. Databricks' Ali Ghodsi reported that switching the "harness" around an open model like GLM-5.2, independent of the underlying model, cut costs by roughly 2x.
China may restrict its own open-source models
Citing Reuters reporting, the panel discussed indications that Chinese regulators met with Alibaba, ByteDance and Z.ai (maker of GLM-5.2) to discuss limiting overseas access to China's leading AI models, and are reportedly moving to classify AI research leaks as a national security offense. Sacks called the reporting "probably a little bit overstated," noting ByteDance's flagship model has always been closed and that the pattern — stay open until you catch the frontier, then close the model to capture value — mirrors OpenAI's own transition under Sam Altman. Gerstner added that GLM-5.2 carries "watermarks from mythos all over it," alleging distillation from U.S. models, and said Washington is likely to act against distillation practices regardless of China's own posture. Both agreed a Chinese regulatory clampdown would harm China's AI ecosystem more than it would help competitively against the U.S.
Energy, not chips, may be the real constraint
Palihapitiya flagged an internal analysis showing the U.S. faces an energy shortfall equivalent to roughly three entire states the size of California by 2050, based on projected load growth from data centers plus ordinary consumer demand. He separately noted Taiwan's reliance on liquefied natural gas reserves that could be exhausted within two to three weeks under a blockade scenario, tying the AI buildout's physical constraints directly to geopolitical risk in the Taiwan Strait.
Trump Accounts launch with over $1 billion funded in 24 hours
Gerstner, who has spent four years building the "Invest America Act" into law, detailed the July 4 launch of Trump Accounts — $1,000 seeded federal investment accounts for every child, invested in the S&P 500, funded further by family contributions up to $5,000 per year and employer contributions up to $2,500 tax-free. The app became the number-one download in the App Store, with more than 1.5 million accounts created and over $1 billion deposited in the first 24 hours. Michael and Susan Dell committed $250 for each of 25 million lower- and middle-income children, a pledge exceeding $6 billion; SpaceX president Gwynne Shotwell contributed $350 million in SpaceX shares targeted by zip code and age; Micron committed $250 million, up to $1,000 per employee's child. Gerstner said the platform could raise $100 billion in philanthropic capital in its first 12 months and eventually scale to more than 100 million accounts within a decade, calling it "the largest direct philanthropic platform in the history of the country." He noted the accounts function as a de facto Roth IRA created at birth — something previously unavailable to anyone without earned income — and pointed to a wave of CPA commentary describing the tax mechanics as best-in-class estate and retirement planning.