"Is there an AI bubble?” Gavin Baker and David George
Summary
Gavin Baker’s answer is no: today’s AI buildout does not resemble the 2000 bubble by either valuation or utilization. Cisco peaked around 150–180 times trailing earnings versus roughly 40 times for NVIDIA; 97% of peak-era fiber was dark, while “there are no dark GPUs” and training clusters are pushing chips until they melt. The largest public GPU buyers have gained roughly 10 points of ROIC since ramping capex—though whether that persists through Blackwell spending remains “an interesting and open debate,” and Baker personally thinks it will.
The infrastructure bill is enormous, but its buyers have an unusually deep balance-sheet buffer. David George framed roughly $1 trillion of existing US data centers, another $3–4 trillion planned over five years, and estimated more than $1 trillion of OpenAI commitments; against that, the major spenders generate about $300 billion of annual free cash flow and hold $500 billion of cash. At $40–50 billion per NVIDIA-powered gigawatt, George sees “an $800 billion buffer growing $300 billion every year,” even if near-term buildout creates some mismatch.
The feared round-tripping is real but, in Baker’s view, small and strategically rational. NVIDIA funding OpenAI while OpenAI buys NVIDIA chips looks circular because “money is fungible,” but Baker says the real driver is competition with Google’s TPU, DeepMind and Gemini—not weak underlying demand. Baker estimates Gemini had taken roughly 15–20 points of traffic share in two or three months and suspects Google may already have more AI traffic than OpenAI or Anthropic on an actual-traffic basis; the cited share gain did not include AI Overviews.
AI could reinforce much of the Mag 7, but execution failure remains existential. Incumbents possess the essential inputs—data, distribution, compute, capital and talent—so AI might be a sustaining innovation if they execute; otherwise, “IBM might be a good fate.” David George called ChatGPT “Pearl Harbor for Google,” while Baker cautions that frontier labs will structurally carry lower gross margins than SaaS because scaling laws and test-time compute keep the products compute-intensive.
Application SaaS is not necessarily dead, but winning requires embracing margin compression. Baker has softened his early-2024 view that all application SaaS “might be a zero,” especially for vendors serving fragmented SMB customers; his warning is that protecting 80–90% gross margins can sacrifice the AI opportunity. The operative choice is “10 bucks of revenue with 90% gross margins or 50 bucks of revenue with 60%,” while incumbents can subsidize break-even AI products before leaders such as Cursor accumulate enough tokens to make catching up difficult.
Distribution and reasoning have made consumer AI less hostile to durable platforms. AI-browser launches could let Google watch the pioneers for three to six months before responding through Chrome’s roughly 5 billion users. Reasoning and RL can turn a large user base into the classic product-data flywheel: users improve the algorithm, which improves the product. David George says GPT-5 is not evidence that scaling laws ended because it was “a smaller model” designed to run more economically, not to maximize capability.
Outcome pricing is the likely business-model shift, with robotics as the physical extension of the same logic. Customer support can charge per resolved task because success supplies a verified reward; personal agents may collect affiliate fees for completed purchases, squeezing the advertiser overpayment that made Google search so lucrative. Baker calls robotics “very real,” expects Tesla versus China, and thinks the humanoid debate is effectively over because robots can learn from video or human demonstrations and receive clean task-level feedback.
Deep dive
1. Utilization—not the capex headline—is Baker’s bubble test
George opened with the intimidating ledger: roughly $1 trillion of US data centers, another $3–4 trillion planned over five years, and three years of construction already exceeding the inflation-adjusted cost of the interstate highway system. He also cited Google’s 150-fold increase in tokens processed over 17 months and estimated that OpenAI alone had more than $1 trillion of committed deals.
Baker’s 2000 comparison turns on both price and use. Cisco reached roughly 150–180 times trailing earnings versus NVIDIA near 40 times. Dark fiber was fiber laid but not lit, useless without the optics, switches and routers needed to activate it; 97% of installed fiber was dark at the telecom bubble’s peak. Today, “there are no dark GPUs”—technical papers instead describe GPUs melting during training runs.
His cleanest economic test is the return on invested capital of the largest public GPU buyers: since capex accelerated, their ROICs have risen by roughly 10 points. “There’s no debate that thus far the ROI on AI has been really positive”; whether that continues through the quantum of Blackwell spending is explicitly an open debate, though Baker personally thinks it will.
George said those buyers collectively generate around $300 billion of annual free cash flow and hold $500 billion in cash. The discussion acknowledged some near-term mismatch as construction peaks, while George said Larry Page had apparently indicated he would rather “go bankrupt than lose” the race.
2. Circular financing is a side effect of the NVIDIA–Google war
Baker concedes the accounting optics: round-tripping is “objectively happening,” and restrictions cannot eliminate circularity because “money is fungible.” His qualifier is scale—it remains small—and motive: NVIDIA is responding to Google, which funds labs and supplies them with TPUs.
NVIDIA’s most important competitor is therefore “not AMD, not Broadcom, not Marvell” or Intel—it is Google. The TPU may be the only serious training alternative today and perhaps the best inference alternative; Google also owns DeepMind and Gemini, whose traffic share Baker estimates had risen 15–20 points in two or three months. With Anthropic tied to Google and Amazon infrastructure, NVIDIA has strategic reasons to respond, while xAI and OpenAI remain at the forefront.
The hardware contest now spans whole systems. NVIDIA progressed from chips to CUDA, rack-scale systems, networking and data-center architecture; Broadcom counters with open Ethernet fabrics, custom ASICs and AMD as a fallback. Baker expects a bunch of high-profile ASIC programs to be canceled within three years, while Trainium 3 “will probably be a much better chip” than Trainium 2 and AMD remains the necessary second source.
3. Distribution may let incumbents own the model transition
Baker’s restraint is historical: at the equivalent point after Netscape, Google did not exist, Mark Zuckerberg was in middle school and Travis Kalanick was in kindergarten. George contrasted the internet’s need to build both websites and users with AI tools that can be exposed through an API or ChatGPT and distributed to a billion people immediately.
Unlike the internet’s disruption of incumbents, AI might be sustaining because today’s giants already possess data, distribution, compute, dollars and talent. Baker says they have every right to win provided they execute; George called ChatGPT “Pearl Harbor for Google,” while Baker says failure could leave an incumbent with IBM as the good outcome.
Frontier labs should not be modeled like 2021 SaaS. Scaling laws, the “Bitter Lesson” and test-time compute make AI structurally more compute-intensive, so gross margins should remain below cloud-era software margins even if lower operating expenses still produce excellent businesses.
Consumer distribution compounds the advantage: Chrome has roughly 5 billion users, so AI-browser pioneers may regret giving Google time to watch and then respond. Reasoning also makes frontier models less like “the fastest-depreciating asset in history”: RL can turn users into a product-improvement flywheel. George calls Chinese open-source models “a godsend” for American challengers trying to catch the four leading labs.
4. SaaS winners must treat lower margins as proof of adoption
Baker has revised his early-2024 belief that application SaaS might all “be a zero.” Large winners remain plausible, particularly among companies serving fragmented SMB customers, but vendors cannot preserve legacy economics while meaningfully adopting compute-heavy AI.
His cautionary analogy is retail’s response to Amazon: incumbents rejected the business because its margins looked unattractive, only to watch Amazon build healthy margins over 25 years. Software already has an existence proof in Microsoft’s move from perpetual on-premise licenses to lower-margin cloud delivery—followed by “a pretty good stock for 10 years.”
Baker argues that lower gross margins should be “a badge of honor”; George adds that an alleged AI company still posting 82% may simply have little usage. George’s arithmetic captures the choice—$10 of revenue at 90% gross margin versus $50 at 60% is “not that complicated,” even if public-market communication is.
Legacy vendors can fund AI products at break-even from profitable installed businesses. Baker gives public coding companies only “a chance” against Cursor, which already has a trillion coding tokens, but says attaching an aggressive product everywhere is still worth attempting; Figma’s willingness to guide toward lower AI margins showed investors can accept the trade.
5. AI monetization moves from seats and clicks to completed outcomes
Customer support is the clearest starting point: abundant text data suits LLMs, while customer satisfaction or first-call resolution provides a verifiable reward. Because humans are fundamentally paid for outcomes, AI that augments or replaces their work should increasingly be priced the same way.
Baker imagines a personalized Grok soliciting hotels for the best room and price, then probably collecting an affiliate fee when it closes the booking. That may degrade platform economics: Google favored advertising because merchants systematically overestimate their ability to retain customers acquired through Google and consequently overpay for acquisition; an outcome-based agent squeezes out that inefficiency.
The long-range claims remain deliberately loose but consequential. Baker finds Elon Musk’s idea that work could become optional “not wildly implausible,” and rejects treating Karpathy as a skeptic for putting AGI 10 years away: “Are you kidding? Insane. Ten years. Sign me up.” In robotics, he expects Tesla versus China and favors humanoids because observation, human demonstration and binary task feedback make training tractable.