All podcasts / BG2 Pod / Summary

All things AI w @altcap @sama & @satyanadella. A Halloween Special. 🎃🔥BG2 w/ Brad Gerstner

2025-10-31 - 74 min - source - Read full transcript
Brad Gerstner (host)Sam AltmanSatya Nadella

Key insights

Microsoft holds 27% of OpenAI on a fully diluted basis after investing roughly $134 billion since 2019, and the restructuring simultaneously created the OpenAI Foundation, a nonprofit capitalized at $130 billion.
Nadella frames the nonprofit as the bigger story than Microsoft's stake itself, comparing it to the Gates Foundation in scale and noting the first $25 billion is earmarked for health and AI security/resilience.
openai-microsoft-deal
The new deal keeps OpenAI's stateless APIs exclusive to Azure through 2030 and a 15% Microsoft revenue share running until 2032, but both terms end early if an expert panel verifies OpenAI has reached AGI.
Gerstner presses both men on whether this creates a financial incentive to fight over the AGI determination; Altman calls the process itself "a good thing to do" while downplaying near-term stakes, and Nadella notes intelligence will keep improving regardless of the label.
openai-microsoft-deal
Microsoft gets royalty-free access to OpenAI's model IP for seven more years, which Nadella describes as effectively "having a frontier model for free" to embed across GitHub, Microsoft 365, and Copilot.
He says this lets Microsoft take the model's weights, add its own data, and post-train it into products, making the IP access a bigger long-term value driver than the equity stake itself.
openai-microsoft-deal
Altman argues compute demand isn't a fixed quantity but a price-elasticity curve: a 100x drop in cost per unit of intelligence would push usage up by far more than 100x because currently uneconomic use cases become viable.
He compares it to energy demand, which can't be discussed without specifying a price point, and predicts new willingness to pay emerges at much higher intelligence levels (curing disease, discovering physics) that don't exist yet as demand categories.
compute-scaling-economics
Nadella says the actual bottleneck isn't chip supply but power and site buildout speed, Microsoft has chips sitting in inventory it can't plug in because of insufficient power infrastructure.
He frames the hyperscaler's job as running an efficient "token factory" with maximum utilization via a fungible fleet spread across workloads, geographies, and hardware generations, since power constraints, not GPU scarcity, are the near-term ceiling.
compute-scaling-economics
Altman flags a real bubble risk in AI infrastructure: cost per unit of intelligence has been falling roughly 40x per year, and a sudden cheap-energy breakthrough could strand a lot of current buildout commitments.
He explicitly compares this to prior tech infrastructure cycles (citing the dot-com telecom buildout) where some participants get badly burned even as the underlying technology goes on to create far more value than initially projected.
compute-scaling-economics
Azure grew 39% in the quarter on a $93 billion run rate versus GCP's 32% and AWS's roughly 20%, but Nadella says Azure could have grown 41-42% with more available compute, and the $400 billion RPO backlog has only about a 2-year average duration.
He says Microsoft is deliberately shaping demand, sometimes declining requests like OpenAI's for a dedicated multi-gigawatt training facility, to protect long-term fungibility over a single customer's short-term ask.
compute-scaling-economics
The Microsoft-OpenAI contract has a formal AGI-verification clause: if OpenAI's board claims AGI has been reached, an expert panel adjudicates the claim, and that determination changes both companies' exclusivity and revenue-share obligations.
Nadella publicly maintains (per his own earnings call comments) that nobody is close to AGI and describes current capability as "spiky and jagged," while Altman is more bullish on timelines, though both downplay near-term tension over invoking the clause.
agi-definition
Altman calls Colorado's AI Act, effective February 2026, unworkable, saying he doesn't know how OpenAI is supposed to comply with it, and both men argue for federal preemption over a 50-state regulatory patchwork.
Gerstner notes federal preemption was included in the "big beautiful bill" but was killed at the last moment by Senator Blackburn; Nadella adds that a fragmented approach disproportionately burdens startups that can't afford to litigate across 50 jurisdictions the way Microsoft and OpenAI can.
ai-regulation
Nadella's core thesis: SaaS architecture is decoupling because an emerging "agent tier" is replacing the old tightly-coupled data/logic/UI business-logic tier.
He argues low-ARPU, high-usage products like Microsoft 365 are better positioned than high-ARPU, low-usage ones, because the constant usage already feeds the data graph needed to ground agent requests, giving Microsoft 365 Copilot pricing power despite low per-seat revenue historically.
agentic-software-disruption
Nadella splits AI value capture into the "token factory" (raw compute throughput and utilization, the hyperscaler's job) and the "agent factory" (the application layer that decides how to spend tokens toward a specific business outcome).
He uses GitHub Copilot's auto-mode, which selects which model to use per task based on an eval and feedback loop rather than round-robin, as the example of agent-factory value creation distinct from raw compute provisioning.
agentic-software-disruption
Nadella predicts consumer search economics face the same disruption as SaaS: chat-based interactions cost far more per query in GPU cycles than a search index amortized over billions of fixed-cost queries.
He says monetization is shifting toward subscription and possibly agentic commerce rather than the old ad-unit model, and that he personally now uses search mainly for narrow navigational queries while shifting commerce-related search to Copilot.
agentic-software-disruption

Companies

Techniques and frameworks

Summary

Brad Gerstner sits down with Sam Altman and Satya Nadella for a joint interview breaking down the newly restructured Microsoft-OpenAI partnership and its implications for compute, regulation, and the software industry. The two executives walk through the mechanics of the deal: Microsoft's 27% fully-diluted ownership stake built on roughly $134 billion invested since 2019, a new OpenAI Foundation nonprofit capitalized at $130 billion, stateless API exclusivity on Azure through 2030, a 15% revenue share to Microsoft, and royalty-free IP access for Microsoft through 2032. All of these commercial terms are tied to a formal AGI-verification clause, if an expert panel determines OpenAI has reached AGI, exclusivity and revenue-sharing terms end early, which both men downplay as a near-term flashpoint despite Nadella publicly maintaining nobody is close to AGI and Altman sounding more bullish on timelines.

A large chunk of the conversation is about compute economics. Altman frames total compute demand as a price-elasticity curve rather than a fixed number: falling cost per unit of intelligence unlocks usage that doesn't pencil out today, similar to how energy demand can't be discussed without a price point attached. He also acknowledges real bubble risk, citing roughly 40x annual declines in cost per unit of intelligence and warning that a sudden cheap-energy breakthrough could strand current infrastructure commitments the way past tech buildouts have burned overextended players. Nadella counters that Microsoft's actual near-term constraint isn't chip supply but power and site buildout speed, describing chips sitting in inventory that can't be plugged in, and lays out his "fungible fleet" philosophy: build compute that can shift across workloads, geographies, and hardware generations to keep utilization high rather than over-committing to any single customer's ask (including OpenAI's requests for dedicated training capacity).

On regulation, Altman is blunt that he doesn't know how OpenAI is supposed to comply with Colorado's incoming AI Act, and both men push for federal preemption over what they call a dangerous 50-state regulatory patchwork, a preemption clause that was reportedly killed at the last minute in the "big beautiful bill" by Senator Blackburn. They frame the patchwork as especially punishing for startups that lack the legal resources Microsoft and OpenAI have to navigate 50 separate compliance regimes.

The back half pivots to Nadella's thesis on software disruption. He argues SaaS architecture is decoupling as an "agent tier" replaces the old tightly-coupled data/logic/UI stack, and that low-ARPU, high-usage products like Microsoft 365 are better positioned than high-ARPU, low-usage ones because constant usage already feeds the data graph needed to ground agent requests. He splits AI value capture into a "token factory" (raw compute throughput, the hyperscaler's job) and an "agent factory" (the application layer deciding how to spend tokens toward a business outcome), using GitHub Copilot's model-routing auto-mode as the example of agent-factory value. He extends the same disruption logic to consumer search, predicting monetization shifts away from the fixed-cost ad-unit model toward subscription or agentic commerce as chat interactions cost far more per query than an amortized search index. The conversation closes on IPO speculation (Altman confirms no fixed date but assumes it happens eventually), AI's effect on hiring (Nadella expects headcount growth but with far more leverage per employee), and a shared bullish read on US reindustrialization tied to the AI capex buildout.

Notable Quotes

"If the price of compute per like unit of intelligence... fell by a factor of a 100 tomorrow, you would see usage go up by much more than 100." - Sam Altman

"I don't know how we're supposed to comply with that California, sorry, Colorado law." - Sam Altman

"Nothing is a commodity at scale." - Satya Nadella

"It's kind of like having a frontier model for free... if you're an MSFT shareholder." - Satya Nadella

"The reality is agents are the new seats." - Satya Nadella