Skip to content
a16za16z

"Is there an AI bubble?” Gavin Baker and David George

In this conversation from a16z’s Runtime, Gavin Baker, Managing Partner and CIO of Atreides Management, joins David George, General Partner at a16z, to unpack the macro view of AI: the trillion-dollar data center buildout, the new economics of GPUs, and what this boom means for investors, founders, and the global economy. Timestamps: 00:00 Intro 01:02 Are We in an AI Bubble? Setting the Stage with Data 02:41 Lessons from 2000: Dark Fiber vs. “No Dark GPUs” 05:07 ROI on AI: Why This Time Is Different 06:42 Nvidia, Google, and the Race to Win AI Infrastructure 08:36 Round-Tripping Deals and Competitive Pressure 10:58 Google’s TPU Advantage and the Four Leading Labs 12:22 The Application Layer: It’s Still Early 13:46 Big Tech’s “Right to Win” and the Execution Gap 15:24 Margins, Scaling Laws, and the Business Model Shift 17:18 Why Lower Gross Margins Mean Real AI Usage 19:44 SaaS, Software, and the Cloud Transition Analogy 21:12 Consumer AI, Browsers, and the Battle for Distribution 23:26 Reasoning Models and the Data Flywheel Effect 25:02 Chips, TPUs, and Broadcom’s Bet Against Nvidia 27:18 The Future of Business Models: Paying for Outcomes 29:41 Robotics, Humanoids, and Elon’s Optimus Vision Resources: Full Transcript on our Substack: https://a16z.substack.com/p/gavin-baker-and-david-george-on-positional Follow Gavin on X: https://x.com/GavinSBaker Follow Atreides Management on X: https://x.com/atreidesmgmt Follow David on X: https://x.com/DavidGeorge83 Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Resources: Find a16z on X: https://x.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.

David GeorgehostGavin Bakerguest
Oct 30, 202531mWatch on YouTube ↗

CHAPTERS

  1. 0:00 – 3:21

    Are we in an AI bubble? Framing the question with today’s capex and usage stats

    David George opens by contrasting the massive planned data-center spend with rapidly rising AI usage (e.g., token growth). The chapter sets up the core tension: scary infrastructure numbers vs. real-world adoption signals.

    • Trillion-dollar existing US data-center base; multi-trillion expansion plans
    • Infrastructure build-out compared to the US interstate highway system (inflation-adjusted)
    • OpenAI’s large-scale infrastructure commitments cited as bubble-like indicators
    • Google’s reported 150× token processing growth as evidence of real usage
    • Central question posed: does spend imply a bubble or a real platform shift?
  2. 3:21 – 4:52

    Lessons from the 2000 telecom bubble: “dark fiber” vs. “no dark GPUs”

    Gavin Baker argues today doesn’t resemble the 2000 bubble because that era’s defining feature was overbuilt, unused capacity. In AI, he claims the opposite: scarce, heavily utilized compute with no equivalent of idle “dark” infrastructure.

    • 2000 was more telecom than “internet”: capital chased network build-out
    • Dark fiber definition: laid fiber that wasn’t ‘lit’ and thus economically useless
    • At the 2000 peak, ~97% of laid fiber was dark (unused capacity)
    • AI contrast: ‘no dark GPUs’—compute is constrained and actively consumed
    • Operational evidence: training runs stress hardware; utilization is high
  3. 4:52 – 5:32

    Why this cycle may be different: ROI and valuation sanity checks

    Baker points to public-market valuation multiples and, more importantly, return on invested capital for the biggest AI capex spenders. He claims ROIC has risen alongside AI investment, supporting the idea that spending has been productive so far.

    • Valuation comparison: Cisco at bubble peak vs. Nvidia’s lower multiple today
    • Key metric: ROIC of major GPU/AI capex buyers (public companies)
    • Claimed outcome: ~10-point ROIC increase since AI capex ramp
    • Open question: whether ROI stays strong with next-gen spend (e.g., Blackwell)
    • Conclusion: thus far, spending has paid off; not a classic bubble setup
  4. 5:32 – 6:06

    Adoption dynamics: internet’s two-sided build vs. AI’s instant distribution via cloud/APIs

    David contrasts the early internet’s friction (building sites + acquiring users) with AI’s ability to ship instantly through APIs and existing cloud/internet distribution. They argue AI’s adoption curve can be faster and less structurally constrained than past platform shifts.

    • Internet required building a two-sided network (websites and users)
    • AI tools can be activated via API or a single consumer endpoint (e.g., ChatGPT)
    • AI benefits from layers already built: cloud + internet
    • Instant distribution claim: potential to reach enormous user bases quickly
    • Implication: faster adoption reduces “bubble built on hope” risk
  5. 6:06 – 7:01

    Who is funding the build-out: Big Tech balance sheets and ‘win at all costs’ mindset

    They argue the main capex funders are unusually resilient: top companies with huge free cash flow and cash balances. The discussion highlights the existential framing inside Big Tech—spend aggressively to avoid losing the AI platform race.

    • Capex counterparties described as exceptionally strong businesses
    • Collective free cash flow and cash reserves cited as buffers against a ‘pop’
    • Rough cost benchmark: tens of billions to build a 1GW full-stack AI data center
    • Noted possibility of temporary free-cash-flow compression as spend ramps
    • Cultural signal: ‘go bankrupt rather than lose’ mentality inside some hyperscalers
  6. 7:01 – 12:15

    Round-tripping deals: what’s real, what’s overstated, and why competition drives it

    They address concerns that vendors and labs may be indirectly financing purchases of their own infrastructure (money-fungible ‘round-tripping’). Baker acknowledges it happens but argues it’s currently small relative to the overall market and is amplified by competitive pressure.

    • Round-tripping defined as financing arrangements that recycle money back into purchases
    • Acknowledgment: it is ‘objectively happening’ due to fungibility
    • Claim: scale is small relative to total capex and demand
    • Competitive dynamics can encourage aggressive deal structures
    • Implicit comparison to past cycles where such behavior was more systemic
  7. 12:15 – 13:08

    Big Tech’s ‘right to win’—and the execution gap risk

    They discuss why incumbents may have structural advantages (data, distribution, compute, talent) yet still risk failure if execution falters. Baker frames ChatGPT as a wake-up call for Google and warns that platform shifts can punish slow responders.

    • Incumbent advantages: data distribution, compute access, capital, talent
    • Phrase used: Big Tech has ‘every right to win’—if they execute
    • ChatGPT described as ‘Pearl Harbor’ for Google, triggering urgency
    • Existential stakes: failure to execute can lead to an IBM-like fate
    • Market remains open enough that responses and strategy still matter
  8. 13:08 – 14:34

    Margins and scaling laws: why AI economics won’t look like SaaS

    Baker argues frontier AI is structurally more compute-intensive, leading to lower gross margins than classic SaaS. He notes that lower gross margins don’t preclude great businesses, but they change expectations and how value accrues.

    • SaaS gross margins cited (80–90%) as historical benchmark
    • AI inference/training compute makes gross margins structurally lower
    • Scaling laws and ‘The Bitter Lesson’ referenced to justify compute intensity
    • Opex may be lower even if gross margins are lower (different business shape)
    • Expectation reset: frontier labs likely won’t reach SaaS-like margins soon
  9. 14:34 – 17:09

    Application layer reality check: SaaS isn’t dead, but must accept margin pressure

    The conversation shifts to software companies and whether AI ‘kills SaaS.’ Baker revises earlier bearishness, arguing there will be winners—especially in fragmented markets—but only if companies embrace margin compression as part of real AI adoption.

    • Baker: earlier view ‘application SaaS might be a zero’ now more nuanced
    • Potential winners in fragmented SMB markets where customization/data control matter
    • Warning: trying to preserve legacy margins can be a strategic mistake
    • Analogy: retailers misread Amazon margins and ceded long-term advantage
    • Prescription: treat declining gross margins as a success signal when usage grows
  10. 17:09 – 19:20

    ‘Low gross margins = real usage’: public-market messaging and competition examples

    They note a shift in investor interpretation: lower margins can indicate genuine AI costs and adoption. Baker argues incumbents can fund AI products at break-even using profitable legacy businesses, and he questions why more public software companies aren’t competing aggressively.

    • Investor heuristic flips: high AI-company margins can signal low usage
    • Trade-off framed: larger revenue at lower margin can be better than small high-margin revenue
    • Public-market challenge is communication, not economics (cloud transition analogy)
    • Microsoft/Adobe cited as examples navigating margin transitions successfully
    • Competitive example: incumbents should lean in (e.g., coding tools vs. fast-growing challengers)
  11. 19:20 – 21:07

    Consumer AI and distribution battles: browsers, Chrome, and platform leverage

    They debate whether AI-native browsers and chat interfaces will subsume consumer internet distribution. Baker suggests AI browser launches may be vulnerable to Google’s Chrome scale and timing strategy, and he emphasizes the power of existing user bases.

    • Question: do consumer internet companies become components inside chat interfaces?
    • Baker: AI-native browser makers may regret competing against Chrome’s scale
    • Google portrayed as cautious due to regulatory scrutiny and historical lessons
    • Strategy idea: let others prove demand, then integrate at platform level
    • Takeaway: distribution moats (existing users) remain decisive in consumer AI
  12. 21:07 – 23:25

    Reasoning models and the data flywheel: why user bases now matter more

    Baker argues that reasoning and post-training RL change the economics for frontier labs by enabling stronger feedback loops from users. He claims this restores a classic consumer-internet flywheel: usage improves models, which improves product, which drives more usage.

    • Pre-reasoning claim: frontier models without unique data/distribution depreciate fastest
    • Reasoning + RL post-training increases value of large user bases
    • Flywheel described: better product → more users → better algorithm/model → better product
    • Implication: strengthens prospects for leading labs beyond pure pretraining scale
    • Skepticism of ‘GPT-5 ends scaling’ narratives; notes model design trade-offs
  13. 23:25 – 26:18

    Chips and infrastructure competition: Nvidia vs. TPU, and Broadcom/AMD’s ‘open’ bet

    They map the AI silicon landscape as a fight between Nvidia’s integrated systems approach and Google’s TPU ecosystem. Baker highlights Broadcom’s role in Ethernet-based fabrics and custom ASIC programs, while predicting many ASIC efforts will be canceled.

    • Core rivalry framed: Nvidia platform vs. Google TPU platform
    • Nvidia evolution: chips → CUDA software → rack-scale systems → data-center architecture
    • Broadcom pitch: open Ethernet-based fabric + custom ASICs to rival Nvidia’s stack
    • AMD positioned as complementary/backup option in some custom deployments
    • Prediction: multiple high-profile ASIC programs may be canceled within ~3 years
  14. 26:18 – 30:04

    Business model shift: from seats and ads to ‘paying for outcomes’

    They discuss where disruption emerges when platform shifts coincide with business model shifts. Baker expects more AI services to be priced on measurable outcomes (e.g., resolutions), and he forecasts consumer monetization moving toward closed-loop affiliate/outcome economics rather than inefficient ad auctions.

    • Platform shift + business model shift creates startup openings
    • Customer support highlighted as early fit for outcome-based pricing
    • Broader claim: humans are paid for outcomes; AI will increasingly be too
    • Consumer agent example: personalized booking/shopping with affiliate-like fees
    • Ad inefficiency argument: outcomes pricing may compress advertiser overpayment vs. search ads
  15. 30:04 – 31:52

    Robotics and humanoids: Optimus, learning from video, and the Tesla vs. China framing

    They close with a forward-looking discussion on robotics, with Baker bullish on humanoids and Tesla’s Optimus progress. He argues the humanoid form factor wins because it can learn from human demonstrations and existing video, making task generalization and training simpler.

    • Robotics framed as ‘very real’ with major near-term momentum
    • Competitive landscape: Tesla vs. Chinese manufacturers (similar to autos)
    • Humanoid debate: Baker believes it’s effectively settled in favor of humanoids
    • Training advantage: learning from YouTube/demonstrations; humans can teach by example
    • Optimus videos cited as evidence; task success judged by simple outcome verification

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.