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

At a Glance

Gavin Baker (CIO of Atreides Management) delves into AI chip competition, scaling laws, and the evolving economics of AI in this podcast. Core thesis: Google, once the "lowest-cost token producer" thanks to its TPU, will lose this advantage with the deployment of Blackwell chips, reshaping the strategic landscape of the entire AI ecosystem.

Topic Sections

1. Scaling Laws Intact, but "Reasoning" Fills the Hardware Gap

Gavin Baker argues that pre-training scaling laws remain valid, but AI progress in 2024-2025 stems not from pre-training, but from two new scaling laws: Reinforcement Learning from Verification Rewards (RLVR) and Test Time Compute.

• Historical Context: The release of Gemini 3 was the first significant validation of pre-training scaling laws since the Hopper chip, with results "unambiguously" confirming the law still holds. However, after reaching the upper limit of 200,000 GPUs in a Hopper cluster, the industry must wait for the next-generation chip (Blackwell) to continue pre-training. Baker likens this process to ancient Egyptians precisely measuring the sun's trajectory without understanding orbital mechanics.
• Mechanism Breakdown: Without the emergence of "Reasoning" models, there would have been no AI progress between mid-2024 and the release of Gemini 3. It was precisely these two new laws—RLVR and Test Time Compute—that enabled AI to leap from an 8% to a 95% ARC AGI score during the hardware upgrade "gap."
• Data Chain: The ARC AGI score rose from 0% to 8% over four years, then skyrocketed to 95% within three months (after the first reasoning model was released).
• Extrapolation: These three scaling laws have a "multiplier effect." Once Blackwell models (built on a better base model) begin applying RLVR and Test Time Compute, their capabilities will far exceed current levels.

2. Google's "Low-Cost" Advantage to Shift, Altering Strategic Dynamics

Baker points out that Google, with its TPU, was once the "lowest-cost token producer" and used aggressive pricing to "suck the economic oxygen out of the AI ecosystem." The deployment of Blackwell chips will end this, forcing Google to reassess its strategy.

• Mechanism Breakdown: Baker believes AI is the first area in tech history where a "low-cost producer" can gain a significant advantage (unlike Apple, Microsoft, or Nvidia). Google leveraged this by running its AI business at a negative 30% profit margin, aiming to cripple competitors' ability to raise capital.
• Data Chain: Blackwell's rack weight has increased from Hopper's ~1,000 lbs to 3,000 lbs, and power consumption from ~30 kW (equivalent to 30 US homes) to 130 kW (130 US homes). This complexity has caused deployment delays, during which Google has been training using TPU V6/V7, which Baker likens to the "F-4 Phantom," while Blackwell is the "F-35."
• Competitive Landscape: Baker argues that once Blackwell models (expected to be first launched by xAI in early 2026) are deployed for training and inference, their cost advantage will surpass that of TPUs. At that point, if Google continues its low-price strategy, it will materially impact its own profits. Furthermore, Google's partnership model with Broadcom (which takes a 50-55% gross margin) creates a cost disadvantage, prompting Google to bring in MediaTek as a counterbalance.
• Extrapolation: Baker judges that Blackwell's deployment will force a strategic recalculation for all players. xAI will benefit first due to its fastest build-out, while OpenAI, burdened by paying compute middleman fees, becomes a "high-cost token producer" facing a "Code Red."

3. AI's "Usefulness" Inflection Point: From Intelligence to Utility to Scientific Breakthrough

Baker believes AI development is shifting from pursuing "smarter" models to pursuing "more useful" ones, with "usefulness" hinging on reliability and long context windows.

• Mechanism Breakdown: Baker divides AI's value curve into three stages: Intelligence → Usefulness → Scientific Breakthrough. Currently, for non-expert users, the intelligence gap between top models is barely perceptible, making the next key phase enabling AI to reliably complete complex tasks.
• Data & Cases: Baker uses the example of Gemini 3 successfully booking a restaurant for him to illustrate AI's ability to execute "actions." He further notes that AI can automate any "verifiable" task, such as sales (closed deal or not), customer support (escalated complaint or not), and accounting (balanced books or not). This is underpinned by the RLVR mechanism.
• Extrapolation: Baker predicts that by the end of 2026, AI will be very good at performing the two core corporate functions of "sales" and "customer support." Looking further ahead, long context windows (e.g., containing all of a company's Slack messages and emails) are key to overcoming AI's current limitations. He also presents a "bear case": if edge AI (e.g., running a distilled model on a phone) provides a "good enough" 115 IQ level at 30-60 tokens/second, it could weaken demand for large cloud-based models.

4. Investment Perspective: SaaS's "Fatal Mistake" and AI's "Iron Law"

Baker strongly criticizes SaaS companies for trying to maintain high gross margins by refusing to embrace lower-margin AI businesses, calling it a repeat of the mistake physical retailers made by ignoring e-commerce.

• Historical Analogy: Baker compares SaaS companies' hesitation towards AI to physical retailers ignoring e-commerce. The latter saw e-commerce as a low-margin business, only to be overtaken by Amazon. Today, Amazon's North American retail business has profit margins exceeding many traditional retailers.
• Mechanism Breakdown: AI's nature is to "recalculate the answer every time," resulting in inherently lower gross margins (around 40%). However, AI companies, with very few employees, can generate cash flow earlier. Baker argues that SaaS companies, with their cash cow businesses and customer data, are naturally advantaged to run AI agent businesses. They should offer AI agent services at 10-20% gross margins, leveraging their data advantage, rather than letting other AI agents access their data via API and eventually replace them.
• Extrapolation: Baker believes that, except for Microsoft, nearly all application-layer SaaS companies are making this "existential" mistake. He calls for a new type of "constructive" investor to push these companies to change their strategies.

Position Moves

Position	Analyst Stance	Key Data
Nvidia	Bullish	Blackwell rack weighs 3,000 lbs, consumes 130 kW; GB300 is a "great chip," seamlessly replaceable with GB200; Rubin will widen the gap with ASICs.
Google	Neutral with Risk Warning	TPU was once the "lowest-cost token producer"; TPU V6/V7 performance is like the "F-4 Phantom"; Partnership with Broadcom results in 50-55% gross margin cost.
xAI	Bullish	Will be first to launch a Blackwell model; Processes 1.35 trillion tokens on OpenRouter, ahead of Google (800-900 billion) and Anthropic (700 billion); "Fusing" with SpaceX and Tesla.
OpenAI	Risk Warning	Is a "high-cost token producer"; Faces a "Code Red" and $1.4 trillion in spending commitments; Its Stargate project aims to solve the cost problem.
Anthropic	Bullish	Burn rate is much lower than OpenAI's, with faster growth; Benefits from TPU and Tranium via relationships with Google and Amazon; Recent $5 billion deal with Nvidia shows strategic flexibility.
Meta	Risk Warning	Failed to build a frontier model; Zuckerberg's "highly confident" prediction early this year proved "as wrong as it could be"; Relies on Chinese open-source models for a "lifeline."
C.H. Robinson	Bullish	Stock rose ~20% due to AI-driven productivity gains; AI reduced quote time from 15-45 minutes to seconds, increasing quote rate from 60% to 100%.
SpaceX	Bullish	Its Starship is the only tool to economically launch data centers into space; Starlink's "Direct to Cell" capability is key.
Broadcom	Neutral	Takes a 50-55% gross margin on the Google TPU project; Its SerDes technology is worth ~$10-15 billion/year but is not irreplaceable.
Tesla	Bullish	Optimus robot will combine with xAI's intelligence module and Tesla Vision; Tesla is xAI's "built-in customer."
Microsoft	Bullish	Is the only SaaS company correctly executing an AI agent strategy; Distributing Copilot via GitHub has made it a massive business.
Intel	Neutral	New CEO Lip-Bu Tan is benefiting from predecessor Patrick Gelsinger's strategy; Its vacant fabs will eventually be filled.

Memorable Judgments

1. "AI can automate anything that can be verified." — Gavin Baker. Support: This is the core of RLVR (Reinforcement Learning from Verification Rewards), which turns verifiable outcomes like "deal closed" or "books balanced" into training signals for AI, enabling it to learn through "win or lose" like AlphaGo.

2. "SaaS companies are making the exact same mistake physical retailers made by ignoring e-commerce." — Gavin Baker. Support: Clinging to 70-90% gross margins, SaaS companies refuse to accept the ~40% margins of AI businesses, giving AI-native companies an opening. Baker argues SaaS companies should leverage their cash cows and customer data to offer AI agent services at 10-20% margins, or risk being replaced.

3. "Google's TPU cost advantage is temporary; Blackwell will change everything." — Gavin Baker. Support: Google was the "lowest-cost token producer" and used this strategy to "suck the economic oxygen out of the AI ecosystem." But with Blackwell's deployment, the cost advantage shifts to the Nvidia ecosystem, forcing Google to reassess its aggressive pricing strategy.

4. "Reasoning models saved AI; they filled the 18-month hardware gap from Hopper to Blackwell." — Gavin Baker. Support: Without the two new scaling laws of RLVR and Test Time Compute, AI would have made zero progress from mid-2024 to late 2025, which would have been disastrous for the market.

5. "Data centers in space are, from first principles, superior in every way to data centers on Earth." — Gavin Baker. Support: Space offers 24/7 solar power that is 30% more intense (no batteries needed), near-free cooling (using the vacuum of space), and laser communication between satellites in a vacuum (faster than fiber optics).

6. "AI is the first area in tech where a 'low-cost producer' can gain a significant advantage." — Gavin Baker. Support: The success of Apple, Microsoft, and Nvidia was not due to low cost, but AI token production has commodity-like attributes, making cost a key competitive factor.

7. "The Blackwell chip has geopolitical leverage; it will widen the gap between US frontier labs and Chinese open-source models." — Gavin Baker. Support: China is restricted from using Blackwell, and DeepSeek has acknowledged in its technical papers that "insufficient compute" is its main obstacle to competing with US labs. Baker believes China's strategy regarding rare earths is "a terrible mistake."

8. "Whatever AI needs, it gets." — Gavin Baker. Support: Baker observes that whenever AI development hits a bottleneck (e.g., power), public opinion and solutions quickly pivot (e.g., changing attitudes toward nuclear energy, the emergence of space data center concepts), as if a force is driving AI forward.