Key Takeaways
- • Ken Griffin, speaking at Citadel's Future of Global Markets conference, said AI is "almost certainly a bubble," while remaining bullish long term; his concern is that companies are not yet converting AI spending into productivity.
- • Chamath Palihapitiya estimates that 1% of companies account for 90% of AI token spending, and that 80% of that spend comes from unprofitable VC-backed startups, a figure Ramp chief economist Ara Kharazian reportedly called "pretty much correct."
- • Palo Alto Networks CEO Nikesh Arora says the PHP 2 Shell exploits show attack cadence accelerating, with "10 days is too long" as his view on patch timing, and that technical debt and scarce compute are the core constraints for enterprises.
- • The episode's France segment, including Polymarket odds on a state of emergency and a Moody's or S&P downgrade, is framed by the hosts as opinion rather than established fact.
Summary
Griffin's warning that AI is "almost certainly a bubble" anchors the episode, according to the All-In hosts' recap of his Citadel remarks. He argued bubbles are normal alongside radical economic transformation. Calacanis noted market reactions cited on the show: data-center builders down about 10%, Meta down 6% after earnings, AMD down 5% and SpaceX down 5%, though all have trended up over time.
Palihapitiya offered a quantitative version of the concern, tweeting a back-of-envelope estimate that 1% of companies represent 90% of AI spending and that 80% of that comes from unprofitable VC-backed firms. He described a loop in which VCs raise from pension funds, LPs, foundations and family offices, startups spend on growth including on other companies' AI, and then raise again. Ramp gave the show an unpublished chart of top-1% AI spender allocations, showing baseline spend on closed frontier labs and rising spend on companies running open-source models on their own GPUs. Palihapitiya tied that shift to data sovereignty and cited a large US bank CIO who reportedly cut AWS and GCP spend and built six independent data centers. Calacanis countered that startups are the most nimble buyers and that a startup stampede signals the right technology direction.
Arora, CEO of Palo Alto Networks, favors open-weight models for auditability and chain-of-thought interrogation, but says compute supply is the binding constraint. He argued building proprietary data centers is "kind of an empty threat" given chip lead times, and expects a hybrid model with bare metal from hyperscalers. On the PHP 2 Shell exploits, found in August 2026 by AI-assisted researchers and rated 9.2 out of 10 by the PHP security team, Arora said attacks will come faster than every 10 days and attributed the exposure to deferred upgrades. Reported impacts included an AWS-reported jump in hacking activity, millions of attempts seen by Cloudflare, and government site takedowns in several European countries. Arora said frontier labs' models can find and chain vulnerabilities, and he rejected the idea that defenders will out-compute attackers today.
The episode also covered agent liability. Friedberg and Arora debated whether liability sits with the agent provider or the party granting agency, and Calacanis raised Obliteration AI, a company he said his fund incubated that supplies unconstrained models, which Palihapitiya and Arora criticized. The France segment, which the hosts treated largely as opinion, cited reports of more than 500,000 protesters, over 3,000 damaged businesses, and French 10-year yields above 5% for the first time since 2002.
Palihapitiya offered a quantitative version of the concern, tweeting a back-of-envelope estimate that 1% of companies represent 90% of AI spending and that 80% of that comes from unprofitable VC-backed firms. He described a loop in which VCs raise from pension funds, LPs, foundations and family offices, startups spend on growth including on other companies' AI, and then raise again. Ramp gave the show an unpublished chart of top-1% AI spender allocations, showing baseline spend on closed frontier labs and rising spend on companies running open-source models on their own GPUs. Palihapitiya tied that shift to data sovereignty and cited a large US bank CIO who reportedly cut AWS and GCP spend and built six independent data centers. Calacanis countered that startups are the most nimble buyers and that a startup stampede signals the right technology direction.
Arora, CEO of Palo Alto Networks, favors open-weight models for auditability and chain-of-thought interrogation, but says compute supply is the binding constraint. He argued building proprietary data centers is "kind of an empty threat" given chip lead times, and expects a hybrid model with bare metal from hyperscalers. On the PHP 2 Shell exploits, found in August 2026 by AI-assisted researchers and rated 9.2 out of 10 by the PHP security team, Arora said attacks will come faster than every 10 days and attributed the exposure to deferred upgrades. Reported impacts included an AWS-reported jump in hacking activity, millions of attempts seen by Cloudflare, and government site takedowns in several European countries. Arora said frontier labs' models can find and chain vulnerabilities, and he rejected the idea that defenders will out-compute attackers today.
The episode also covered agent liability. Friedberg and Arora debated whether liability sits with the agent provider or the party granting agency, and Calacanis raised Obliteration AI, a company he said his fund incubated that supplies unconstrained models, which Palihapitiya and Arora criticized. The France segment, which the hosts treated largely as opinion, cited reports of more than 500,000 protesters, over 3,000 damaged businesses, and French 10-year yields above 5% for the first time since 2002.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.