Ep16. Nuclear Update, AI Fast & Furious, State of VC | BG2 w/ Bill Gurley & Brad Gerstner

Episode Summary

Executive Summary: The conversation argues that a U.S. nuclear revival is underway, driven by hyperscalers’ AI power needs, public acceptance, carbon-offset economics, and bipartisan policy support. It then pivots to AI demand, concluding infrastructure spending is still accelerating and inference will dominate costs. The discussion ends with venture capital’s structural shift toward late-stage quasi-public financings, reduced IPO pressure, and concerns that excess capital can weaken discipline and innovation.

Main Topics: U.S. nuclear renaissance (Priority: 5/5): The speakers frame recent deals and policy momentum—Microsoft/Constellation at Three Mile Island, Oracle and Amazon’s nuclear-related moves, and financial-sector interest—as evidence that nuclear power is back in favor. Hyperscalers as new nuclear customers (Priority: 5/5): They argue that AI companies and hyperscalers are a much better demand base for SMRs and reactor restarts than traditional utility-only sales, because they are more innovation-friendly and willing to share risk. AI infrastructure demand and inference growth (Priority: 5/5): The speakers contend that AI demand remains ahead of supply, with model advances like Strawberry/01 preview increasing inference intensity and pushing future spending toward inference rather than training. OpenAI’s consumer moat and network effects (Priority: 4/5): They discuss why ChatGPT appears to be pulling away in consumer AI, citing advanced voice, memory, product velocity, and possible data/network-effect flywheels, while noting uncertainty about the strength of those effects. Valuation and margin quality in AI (Priority: 4/5): The discussion compares OpenAI’s rumored valuation and revenue trajectory to Google and Meta, while warning that gross margins may be much lower because compute, training, and inference act like a permanent tax. Venture capital market structure changes (Priority: 5/5): They debate how bigger funds, late-stage rounds, founder-friendly behavior, and widespread secondary liquidity are reshaping venture, reducing IPO pressure and changing incentives for founders and LPs. Capital discipline vs. overfeeding startups (Priority: 4/5): A central concern is whether too much capital leads to bloated organizations, lower focus, slower innovation, and weaker returns—illustrated by the “force-feeding geese” metaphor.

Key Arguments: Nuclear is moving from politically taboo to broadly acceptable because AI/data centers need massive baseload power and policymakers now see it as climate and national security infrastructure. Selling nuclear to hyperscalers may be superior to selling only to utilities because hyperscalers are more open to innovation and can underwrite large projects with long planning horizons. Carbon offsets are an important economic driver: if companies switch to nuclear-powered energy, they can reduce offset purchases and improve the project math. AI infrastructure demand is still rising; announcements from Microsoft, BlackRock, Oracle, Amazon, and Middle East players suggest enthusiasm is increasing rather than peaking. Inference is becoming the key cost center: reasoning models may require many passes, making future compute demand far larger than single-shot prompting. OpenAI’s consumer lead may be reinforced by advanced voice, memory, and fast product iteration, but the true durability of network effects is still uncertain. OpenAI-like businesses may look like Google/Meta on the consumer side and AWS on the enterprise side, but margins could be far lower because compute remains a variable cost input. The venture industry has shifted toward quasi-public, late-stage capital markets, which lowers IPO incentives and changes the role of LPs, GPs, and founders. Excess capital can be destructive: it may encourage secondary sales, unnecessary hiring, extra projects, and weaker pressure to reach profitability. Public-market discipline and private-market incentives are increasingly misaligned; companies may stay private longer because secondary liquidity reduces the need to IPO.

Data Points: Carbon offset market (2020 estimate): $2 billion - Morgan Stanley estimate cited for total market size in 2020 Carbon offset market (2030 estimate): $100 billion - Morgan Stanley estimate for large hyperscalers’ future spending OpenAI rumored valuation: $150 billion - Bloomberg-reported fundraising discussion OpenAI rumored revenue: $4–5 billion - Discussed as current annual revenue run rate OpenAI weekly active users: ~200 million - Widely reported usage figure mentioned in the conversation Time to 100 million users: ChatGPT far faster than Facebook, Instagram, or YouTube - Comparison made to show consumer adoption speed Mag 7 forward P/E: 31x - Current multiple cited after recent rerating Mag 7 forward P/E 10-year average: 25x - Historical comparison used to frame valuation January 2023 Mag 7 forward P/E: 21x - Point-in-time low during hard-landing fears OpenAI valuation at $5B revenue: ~15x forward revenue - Implied multiple if current revenue trajectory continues Google IPO valuation multiple: ~10x forward revenue - Historical benchmark cited for comparison Meta investment valuation multiple (2007): ~50x revenue - Historical benchmark cited for comparison Meta IPO valuation multiple (2012): ~13x revenue - Historical benchmark cited for comparison Venture capital market size (pre-COVID): ~$300 billion - Speaker describes current venture levels as roughly back to pre-COVID norms Venture capital exit volume peak (2021): ~$700 billion - ZERP-era one-time high Current venture exit volume: ~$100 billion/year - Current level discussed as well below 2021 highs Historical average IPO count: ~70–80 per year - Used to contrast with today’s historically low IPO count Current IPO count: Single digits - Described as historically anemic Example burn rate: $20 million per month - Used to illustrate how excess capital can distort company behavior Inference cost decline: >90% in 18 months - Claim that inference has already become much cheaper rapidly Potential future inference cost decline: Another 90% over several years - Sonny’s forecast mentioned in passing NVIDIA GPU forecast: ~6 million GPUs next year vs. bearish 4.5 million - Supply/demand debate for AI infrastructure Data center fund size: $30–100 billion - BlackRock/Microsoft fund mentioned for financing data centers Fed rate cut: 50 basis points - Major macro event referenced at the end of the discussion

Pivotal Quotes: "Are we overfeeding these startups?" — Host: Opening metaphor comparing startup capital deployment to force-feeding geese for foie gras "If the hyperscalers become part of the customer set for the nuclear startups, that may be like 10x better than selling just to utilities alone." — Bill: Argument that AI buyers could unlock a much stronger market for nuclear startups "One of the reasons this happened was that there was an irrational public response to the negative risk of these solutions." — Bill: Explaining why nuclear decommissionments and public fear delayed a potential renaissance

Implications: AI demand is likely to keep driving power, compute, and infrastructure investment, with nuclear benefiting most. But investors should expect more capital intensity, lower margins, and weaker IPO pressure as private markets stay crowded and public-market discipline remains absent.