Basis agent relevance engine Palo Alto Networks CEO on All-In Podcast: AI Cyber Panic, Ken Griffin’s Bubble Warning, The Fall of France?
PANWNVDAAMZNMETAAMDGOOGLGOOGMSFTNETCRWV
AI adoption is real but concentrated in VC circularity, while cyber vulnerabilities and France’s fiscal crisis expose unresolved structural risks.
1. AI spending remains dangerously concentrated in unprofitable startups rather than broad enterprise productivity
- Chamath Palihapitiya estimated that 1% of companies generate roughly 90% of AI spending, with approximately 80% of that top-1% spend coming from unprofitable, VC-backed companies.
- He described a circular funding loop: VCs fund startups, startups spend on other AI companies, revenue traction supports another VC round, and valuations rise without proven end-customer productivity.
- Palihapitiya argued AI must diffuse beyond startups before becoming a predictable earnings driver, echoing Ken Griffin’s concern that companies are questioning Anthropic and OpenAI bills versus realized productivity.
- He pushed back on dismissing startup adoption as purely speculative, arguing startups previously rushed to cloud computing, WeWork and underpriced Facebook ads because they identified a real cost or growth advantage.
- Ramp data previewed on the show indicated top AI spenders are increasingly buying their own GPUs and shifting toward open-source models, rather than relying exclusively on closed-model API traffic.
2. Enterprise AI is moving toward open-weight models and owned infrastructure, but compute scarcity limits independence
- Nikesh Arora said Palo Alto Networks prefers open-source/open-weight models because enterprises need explainability, auditability and verifiability, including the ability to interrogate model reasoning.
- A major unnamed US bank reportedly cut AWS and Google Cloud Platform spending and built six independent, ground-up data centers; Palihapitiya presented this as a broader shift toward data sovereignty and private infrastructure.
- Arora said he previously shut down Palo Alto Networks’ data centers 6–8 years ago in favor of hyperscalers, but scarce AI compute has made full dependence less attractive.
- He said a fully independent model is difficult because Palo Alto Networks would need 10–20 megawatts of inference across roughly 50 global locations, while Jensen Huang reportedly cannot provide chips until 2028.
- Arora expects a hybrid model: bare-metal capacity from hyperscalers or adjacent providers, plus compute stitched across locations until enterprise-grade resilience becomes available.