Invest Like the Best with Patrick O'Shaughnessy

Gavin Baker - Nvidia v. Google, Scaling Laws, and the Economics of AI - [Invest Like the Best, EP.451]

December 9, 2025

Key Takeaways Copied to clipboard!

  • The confirmation that pre-training scaling laws remain intact, evidenced by Gemini 3, is crucial because progress in 2024/2025 was otherwise expected to stall due to the delay in Blackwell chip deployment. 
  • Google currently holds a temporary advantage as the lowest-cost producer of AI tokens using its TPUs, a position that allows it to rationally pursue a strategy of 'sucking the economic oxygen out of the AI ecosystem' to pressure competitors. 
  • Data centers in space are presented as the next major technological shift, offering superior, lower-cost inputs (solar power without batteries and free cooling via radiators) compared to terrestrial data centers, contingent on the economic viability of launch costs provided by vehicles like Starship. 
  • The convergence of Elon Musk's companies (Tesla, SpaceX, XAI) creates synergistic competitive advantages, with XAI's intelligence module powering Tesla's Optimus using data centers in space provided by SpaceX. 
  • The iron law of capital cycles suggests compute gluts will eventually follow current shortages, but power constraints and Taiwan Semi's caution act as natural governors slowing down potential overbuilds. 
  • SaaS companies risk obsolescence by clinging to high gross margin structures (80%+) and failing to embrace the lower gross margins (potentially 40%) inherent to AI-native agent businesses, mirroring the mistake brick-and-mortar retailers made with e-commerce. 

Segments

Processing New AI Updates
Copied to clipboard!
(00:04:58)
  • Key Takeaway: Evaluating new frontier models requires using the highest-paid tiers, not free versions, to accurately gauge adult-level capabilities.
  • Summary: To process new AI releases like Gemini 3, one must use the highest-tier paid versions, as free tiers offer an inadequate comparison to a fully capable adult model. The cutting edge of AI development is often revealed through intense public debate and technical papers shared on platforms like X by a small group of leading researchers. Following key figures like Andrej Karpathy and listening to podcasts featuring researchers from the top four labs (OpenAI, Gemini, Anthropic, XAI) provides critical signal.
Gemini 3 and Pre-Training Laws
Copied to clipboard!
(00:08:21)
  • Key Takeaway: Gemini 3 confirmed the empirical observation that pre-training scaling laws remain intact, despite a lack of fundamental understanding of why they work.
  • Summary: The progress seen since late 2023 was primarily driven by post-training scaling laws: reinforcement learning with verified rewards (RLVR) and test-time compute. Gemini 3’s confirmation of pre-training scaling laws means that future models will benefit from applying these new post-training laws to much better base models. The complexity of deploying Blackwell chips delayed the continuation of pre-training scaling until reasoning models bridged the gap.
Google vs. NVIDIA Hardware Race
Copied to clipboard!
(00:11:14)
  • Key Takeaway: Blackwell’s delayed deployment created an 18-month AI progress bridge, giving Google a temporary pre-training advantage due to their advanced TPU V6/V7 chips.
  • Summary: Blackwell deployment was severely complicated by the transition from air-cooled to liquid-cooled racks requiring massive power and structural upgrades, slowing its scaled adoption. Reasoning models saved the AI timeline by enabling progress while the industry waited for Blackwell infrastructure to stabilize. Google’s current advantage as the lowest-cost producer using TPUs will shift once Blackwells are fully deployed for inference, changing the economic calculus for all players.
Future Hardware and Cost Dynamics
Copied to clipboard!
(00:17:07)
  • Key Takeaway: The introduction of the GB300, which is drop-in compatible with existing liquid-cooled racks, will enable companies vertically integrated with NVIDIA hardware to become the new low-cost token producers, forcing Google to reconsider its margin strategy.
  • Summary: The economics of success in AI are shifting toward being the low-cost producer of tokens, a dynamic previously irrelevant in tech investing. Google’s strategy of running AI at negative margins to capture economic oxygen will change when NVIDIA customers deploy GB300s. Furthermore, Google’s reliance on Broadcom for backend semiconductor design (paying a 50%+ margin) makes bringing TPU development fully in-house an inevitable economic move, potentially impacting Broadcom.
ASIC Development Learning Curve
Copied to clipboard!
(00:21:48)
  • Key Takeaway: Developing competitive custom accelerators (ASICs) requires at least three generations of learning by doing, encompassing networking, software, and system integration beyond just the chip design.
  • Summary: The difficulty of creating a successful ASIC extends beyond the chip itself to include the NIC, CPU, scale-up/scale-out protocols, and supporting software. Even Amazon’s highly innovative ASIC team took several generations (Tranium 3/4) to reach competitive parity with established GPU ecosystems. This complexity explains why many internal ASIC efforts lag behind the rapid acceleration seen in GPU development.
Near-Term AI Utility Unlocks
Copied to clipboard!
(00:24:17)
  • Key Takeaway: The reduction in per-token cost enabled by Blackwell and future chips will allow models to think for longer, leading to near-term utility in personal assistance (booking travel) and significant automation in sales and customer support.
  • Summary: Increased compute efficiency will allow models to handle longer context windows and task lengths, moving the focus from pure intelligence to practical usefulness. Functions with verifiable outcomes, like accounting reconciliation or sales conversion, are prime candidates for early, high-ROI automation. Fortune 500 companies are beginning to report quantitative AI-driven uplift, exemplified by C.H. Robinson’s seconds-long quoting time versus previous 15-45 minute manual processes.
Frontier Lab Dynamics and Competition
Copied to clipboard!
(00:40:55)
  • Key Takeaway: Reasoning models initiated a powerful flywheel where user interaction generates verifiable rewards that can be fed back to improve the model, creating significant barriers to entry for labs without the latest internal checkpoints.
  • Summary: The ability to leverage user feedback as verifiable rewards (RLVR) has created an increasing returns-to-scale dynamic for the top four labs (OpenAI, Gemini, Anthropic, XAI). Meta, Microsoft, and Amazon have struggled to match the performance of these leaders, highlighting the difficulty of mastering GPU cluster coherence, experimental taste, and post-training optimization. China’s decision to avoid Blackwell chips risks blowing out the performance gap with US frontier labs, as Blackwell is essential for maintaining competitive checkpoints.
Musk Company Convergence
Copied to clipboard!
(00:55:36)
  • Key Takeaway: XAI, Tesla, and SpaceX are converging to create mutual competitive advantages.
  • Summary: XAI will provide the intelligence module for Tesla’s Optimus, which uses Tesla Vision for perception. SpaceX will power much of this AI infrastructure with data centers in space. This convergence provides built-in customer bases and technological leverage for each entity.
Compute Shortages and Gluts
Copied to clipboard!
(00:57:10)
  • Key Takeaway: AI’s fundamental compute consumption differs from traditional software, but historical capital cycles suggest gluts will follow current shortages.
  • Summary: Companies like Mark Chen’s could consume ten times more compute if available, indicating a massive shortage persists. However, the iron law of capital cycles dictates that gluts follow shortages, though AI’s per-use compute cost changes monetization models (e.g., ads).
Semiconductor Inventory Cycles
Copied to clipboard!
(00:58:15)
  • Key Takeaway: Taiwan Semi’s caution regarding overbuilding acts as a governor against historical semiconductor capacity cycles.
  • Summary: Historically, semiconductor inventory cycles were driven by customer buffer inventories equaling lead times. Taiwan Semi’s current paranoia about overbuilding, stemming from past experiences, is preventing capacity expansion despite shortages, potentially allowing Intel’s new fabs to fill up.
Power as Compute Governor
Copied to clipboard!
(01:00:13)
  • Key Takeaway: Power constraints favor the most advanced compute players by making compute price irrelevant relative to tokens per watt.
  • Summary: When watts are the constraint, achieving higher tokens per watt translates directly into higher revenue, making the Total Cost of Ownership (TCO) of the ASIC irrelevant. The immediate solutions for power generation are natural gas and solar, as building nuclear power plants quickly in America is too difficult due to regulations.
AI Native Entrepreneurs’ Edge
Copied to clipboard!
(01:03:05)
  • Key Takeaway: Young AI-native entrepreneurs are highly polished faster due to leveraging AI for pitch refinement and operational advice.
  • Summary: Young CEOs are becoming polished rapidly because they use AI tools to simulate difficult conversations, like pitching investors or handling HR issues. This AI productivity is evident across VC portfolios, accelerating the competence of the youngest generation of technologists.
Semiconductor VC Ecosystem
Copied to clipboard!
(01:04:44)
  • Key Takeaway: The resurgence of semiconductor venture capital is crucial for enabling the annual cadence of advanced chip development.
  • Summary: Semiconductor venture founders are typically older (around 50), but NVIDIA’s success has reignited this VC sector. This funding supports the thousands of necessary component suppliers whose acceleration is required for public giants like NVIDIA to maintain their annual chip release cadence.
SaaS Margin Preservation Mistake
Copied to clipboard!
(01:08:28)
  • Key Takeaway: SaaS companies preserving 80%+ gross margins are guaranteed to fail against AI natives operating at 40% margins.
  • Summary: SaaS companies are making the same mistake as brick-and-mortar retailers facing e-commerce: refusing to accept lower margin structures dictated by transformative technology. AI requires recomputing answers, leading to lower gross margins, but existing SaaS firms have the cash flow advantage to transition their platforms using agents.
Rolling Bubbles and Quantum
Copied to clipboard!
(01:14:03)
  • Key Takeaway: Nuclear fusion/SMRs and quantum computing represent current rolling bubbles where public investment vehicles are not the leading innovators.
  • Summary: The post-2020 era has seen rolling bubbles, currently including speculative plays in nuclear and quantum. The true leaders in quantum, such as Google and IBM, are not the publicly traded quantum companies that are currently seeing speculative investment.
AI’s Unstoppable Growth
Copied to clipboard!
(01:15:38)
  • Key Takeaway: Technology, or the ’technium,’ seems to accelerate solutions whenever a bottleneck threatens to slow AI’s growth.
  • Summary: Public opinion on nuclear power shifted rapidly just as AI needed more power, and concepts like data centers in space emerged when terrestrial power became constrained. Whatever AI needs to advance, whether it’s compute (Rubin, Blackwell) or power, seems to materialize.
Gavin Baker’s Origin Story
Copied to clipboard!
(01:16:52)
  • Key Takeaway: Gavin Baker’s passion for investing stems from a childhood love of history intersecting with a competitive drive for skill-based games.
  • Summary: Baker’s initial plan was to be a ski bum and wildlife photographer until a single internship at DLJ exposed him to research reports, which he found fascinating. He conceptualized investing as a game of skill and chance where an edge is gained by combining deep historical knowledge with accurate current event analysis.