A 10x boost in AI compute could double model capabilities, accelerating breakthroughs in physics and math.
The episode opens at Morgan Stanley's 2025 TMT conference in San Francisco, featuring a panel of three analysts: Stephen Bird (global head of thematics and sustainability), Josh Baer (software), and Lindsay Tyler (TMT credit). Host Michelle Weaver frames the discussion around AI adoption dynamics. Bird opens by highlighting takeaways from Morgan Stanley's fifth semi-annual AI mapping survey, which tracks how companies across sectors are exposed to the AI theme. He notes a growing number of firms are quantifying specific AI benefits, which he expects will become industry table stakes. Bird flags a striking narrative shift: investor sentiment has rapidly flipped from focusing on AI benefits to fixating on disruption and deflationary risks. Baer argues that AI is fundamentally software and should be understood as a TAM expander following historical patterns of value accrual. He cites CIO survey data showing incumbent enterprise software vendors are increasingly selected as AI deployment partners. His estimate places the generative AI incremental TAM opportunity for enterprise software at $400 billion by 2028, derived from automation scope, labor cost analysis, and typical software vendor capture rates. Baer adds that software companies are deploying AI tools internally to accelerate R&D and cut operating costs, while also monetizing AI through new SKUs — standalone offerings, advanced suites, and embedded platform capabilities. Bird then addresses LLM trajectory, citing 10x compute increases at major labs and the implications of scaling laws: a 10x training compute increase yields roughly a doubling of model capabilities. He views this as the conference's most important signal, with supporting evidence from recent LLM-assisted breakthroughs in physics and mathematics. Bird notes that Sam Altman (speaking at the event) acknowledged both the impressive capability gains and the risks of misalignment. On geopolitics, Bird argues that Chinese labs have talent, infrastructure, and energy, but lack chips — and that if U.S. compute scaling continues, it could trigger a diplomatic chain reaction around technology transfer and rare earth trade leverage. The conversation turns to U.S. power infrastructure. Bird presents a bottom-up power model starting at the chip level, yielding a U.S. data center power requirement of 74 gigawatts by 2028. He breaks down the supply: roughly 10 gigawatts already built or under construction, 15 gigawatts of incremental grid access, and more than 40 gigawatts requiring unconventional off-grid solutions. Solutions cited include Bitcoin site repurposing (10-15 gigawatts), Bloom Energy fuel cells, gas turbines, and nuclear co-location. Net of all these, the U.S. is still likely 10-20% short. Labor is identified as an underappreciated bottleneck, particularly the shortage of electricians required for data center construction. Bird calculates that Bitcoin mining sites converting to hyperscaler AI data center leases have seen enterprise value jump from $1-2 per watt to $10-18 per watt, with hyperscaler payments continuing to increase due to power scarcity. He notes that not all Bitcoin-related equities have fully re-rated, implying remaining upside.
