The episode opens in medias res with a direct question — 'Are we in an AI bubble?' — and Gavin Baker's equally direct answer: no. To make his case, he reaches for the most powerful historical analogy available: the 2000 internet and telecom bubble, which he lived through as a tech investor. His defining image is dark fiber — the vast majority of physically installed but unused cable that represented the era's catastrophic overbuilding. Then comes his punchline: today, there are no dark GPUs ^{[1] — Gavin Baker "In the 2000 bubble, 97% of all fiber laid in America was dark — unused and worthless. Today, every GPU is running hot. Baker's ROIC test co…" 03:38 Play} . It's a tight, memorable frame that sets the intellectual tone for everything that follows.

A narrator bridges the cold open to the live event context, identifying this conversation as part of a16z's Runtime conference on AI infrastructure and the future of computing. The emcee then steps in with a provocative setup: AI is supposedly the biggest trend in the world, so why is there so little evidence of it in the broader economy — and as Andrej Karpathy provocatively asked, are agents just ghosts? Gavin Baker is introduced as Atreides Management's CIO and the person many turn to when they want an honest read on major AI news developments. David George, general partner at a16z, joins him on stage.

David George opens with staggering numbers: $1 trillion of data centers in the US, plans to add $3–4 trillion over five years, a data center buildout that already dwarfs the entire US interstate highway system in inflation-adjusted dollars, and a 150x increase in Google's token processing over just 17 months ^{[2] — Gavin Baker "10-point ROIC increase for big CapEx spenders: The biggest public GPU spenders have seen approximately a 10-point increase in return on inv…" 04:45 Play} . The scale sounds terrifying, but Baker is unfazed ^{[1] — Gavin Baker "In the 2000 bubble, 97% of all fiber laid in America was dark — unused and worthless. Today, every GPU is running hot. Baker's ROIC test co…" 03:38 Play} . He walks through two key comparisons: Cisco peaked at 150–180x trailing earnings in 2000 versus NVIDIA's current ~40x multiple, and unlike dark fiber — where 97% of installed cable sat unused at the bubble's peak — every GPU is running at capacity. The most rigorous test, Baker argues, is ROIC: the biggest public GPU spenders have seen roughly a 10-point increase in their returns since ramping up CapEx ^{[2] — Gavin Baker "10-point ROIC increase for big CapEx spenders: The biggest public GPU spenders have seen approximately a 10-point increase in return on inv…" 04:45 Play} . The ROI on AI infrastructure spending, he concludes, has been unambiguously positive — even if future Blackwell spending will be a fair test.

The conversation turns to round-tripping — the practice, echoing internet-era accounting games, where NVIDIA invests in companies like OpenAI that then use those funds to buy NVIDIA chips. Baker is blunt: it's objectively happening, but at a very small scale and for rational strategic reasons ^{[1] — Gavin Baker "Round-tripping — where a chip company invests in a customer who then buys chips — is objectively happening but at a small scale. Baker argu…" 07:47 Play} . The real driver isn't financing need — it's competitive dynamics. NVIDIA's most dangerous rival isn't AMD, Broadcom, or Intel; it's Google, which owns both the TPU chip and the DeepMind research lab and is arguably the leading AI company today, having taken 15–20 points of traffic share in just 2–3 months through Gemini. Anthropic, meanwhile, is effectively captive to Google TPUs and Amazon Trainium. That leaves OpenAI and xAI as the independent labs, and if Google is subsidizing its AI ecosystem, NVIDIA must respond in kind. Baker relays Jensen Huang's view that these investments are expected to be good investments — a rational strategic bet, not bubble-era excess.

David George presses Baker on which AI model companies will win, and Baker's response is a master class in epistemic humility. If ChatGPT is the Netscape Navigator of AI, then at this equivalent moment in the internet boom, Google had not been founded, Zuckerberg was in middle school, and Travis Kalanick was in kindergarten ^{[1] — Gavin Baker "At the equivalent moment in the internet era, Google hadn't been founded, Zuckerberg was in middle school, and Kalanick was in kindergarten…" 10:37 Play} . Making high-conviction application-layer bets this early is premature. Baker then introduces a subtle but important distinction: AI might prove to be a sustaining innovation for today's biggest tech companies rather than a disruptive one for new entrants, because the three raw ingredients — unique data, capital for compute, and distribution — are all already concentrated in the Magnificent Seven. The failure mode, he warns, isn't disruption from below; it's IBM-style irrelevance from poor execution. ChatGPT was, as Baker puts it, 'Pearl Harbor for Google' — and we're only now starting to see their response.

The economics of frontier AI models are fundamentally different from the software businesses that preceded them. Baker points to the SaaS playbook of 2021–2022 — companies routinely running at 80–90% gross margins — and explains why that benchmark is structurally unachievable for AI labs. The culprit is compute intensity: scaling laws and the growing importance of test-time compute mean AI models are inherently more expensive to run. Richard Sutton's 'Bitter Lesson' — the observation that methods leveraging raw computation consistently outperform those encoding human knowledge — is a structural anchor for these costs. The silver lining, Baker notes, is that lower gross margins don't preclude great businesses: if opex can be kept low, the math can still work. But investors and founders expecting SaaS-era margins from frontier AI are looking at the wrong benchmark entirely.

In early 2024, Baker publicly predicted all application SaaS might be worth zero. He's moderating that view — some application SaaS companies, especially those serving fragmented SMB markets, could be big winners. But the competitive threat remains acute and the window is narrowing. He draws an analogy to retailers who saw Amazon's thin margins and decided they didn't want to compete — a mistake that took 25 years to fully reveal itself. For SaaS companies today, the equivalent mistake is protecting existing gross margin structures to preserve stock prices ^{[2] — Gavin Baker "It is definitionally impossible, given what we just discussed, to succeed in AI without gross margin pressure. And I do not know why they h…" 15:35 Play} . Baker is direct: it's definitionally impossible to succeed in AI without gross margin pressure, and Microsoft and Adobe both prove that a software company can navigate a margin-compressing transition and still produce a decade of strong returns. The most urgent example is Cursor, which has accumulated 1 trillion coding tokens — a data flywheel that incumbent public coding companies have inexplicably failed to challenge ^{[1] — Gavin Baker "Baker once predicted all application SaaS could go to zero. He's walked that back, but the urgency remains: Cursor has accumulated 1 trilli…" 14:43 Play} . The window to compete, he warns, is closing.

The conversation pivots to how consumer internet business models will be reshaped by AI. Google's original model — capturing search intent and directing users to third-party sites — is already being disrupted, though Baker notes even today's AI browsers still have rough edges. He then issues a sharp warning to AI-native companies launching browsers: Chrome has roughly 5 billion users, and Google has been deliberately restraining itself due to ongoing antitrust litigation. The implication is stark — Google has been letting OpenAI and Anthropic take first-mover risk in the browser space, accumulating learnings from a safe distance, before likely entering with a superior product and overwhelming distribution. It's a pattern worth fearing.

Baker identifies a pivotal inflection point in AI economics: the emergence of reasoning models. Before this, a frontier model without unique data and internet-scale distribution was — in his memorable phrase — the fastest depreciating asset in history. The problem was that the consumer internet flywheel (good product → large users → improved algorithm → better product) wasn't available to AI models. Reasoning changed that because reinforcement learning during post-training means a large, active user base now directly improves the model ^{[1] — Gavin Baker "Pre-reasoning, a frontier AI model without unique data and internet-scale distribution was the fastest depreciating asset in history. Reaso…" 20:30 Play} . The flywheel isn't fully spinning yet, Baker admits — but you can squint and see it beginning. This insight fundamentally changes the competitive dynamics for Anthropic, xAI, and OpenAI, giving large user bases a compounding strategic value they previously lacked. Baker also fires a short, pointed salvo at GPT-5 scaling law skeptics: GPT-5 is a smaller, economical model, not a frontier capability push — citing it as evidence scaling laws are dead is simply wrong.

Baker delivers his most detailed technical analysis in the chip segment, mapping a competitive landscape that most investors misread. NVIDIA's real enemy isn't AMD — it's Google's TPU, the only credible alternative for AI training and possibly the best inference alternative today. Broadcom is playing a fascinating enabling role, offering hyperscalers like Meta an Ethernet-based fabric that theoretically competes with NVIDIA's NVLink stack, alongside custom ASIC development services ^{[1] — Gavin Baker "NVIDIA's real competitor isn't AMD — it's Google's TPU. Broadcom is playing a clever game, offering hyperscalers custom Ethernet-based fabr…" 23:00 Play} . The catch: custom silicon is brutally hard. Google took three generations to get the TPU right. Baker predicts that within three years, most high-profile custom ASIC programs will be canceled, especially if Google begins selling TPUs externally — a development widely rumored on social media. He notes that Amazon's Annapurna silicon team is the most talented at any hyperscaler, and Trainium 3 will be significantly better than Trainium 2, but Google ultimately controls the TPU and can pull it away from Broadcom at any time. AMD will always exist as the market's necessary second source.

David George presses Baker on the harder question beyond obvious use cases like customer support: how does AI actually displace human services at scale, and what business models emerge? Baker's framework is simple — humans are paid for outcomes, AI will increasingly be paid for outcomes too. Customer service is the easy first example: the data is textual, LLMs are strong at text, and the reward signal (happy customer, first call resolution) is easily verifiable with RL. The harder, more interesting question is consumer business models ^{[1] — Gavin Baker "Google never built a marketplace because advertisers systematically overpay, and that inefficiency was Google's profit. AI agents will sque…" 27:05 Play} . Baker's vision: personalized AI agents that know your preferences and actively negotiate on your behalf — booking hotels by soliciting competing offers rather than passively directing you to a website. This eliminates the systematic inefficiency that made Google's advertising model so lucrative: advertisers have always paid more than they should because the customer conversion happened post-click. AI closes that loop, compressing margins but creating more genuine value. He closes with Elon Musk's tweet that 'work will become optional' — a scenario Baker says is not wildly implausible given the technology's trajectory.

The conversation's final substantive topic is robotics, and Baker brings the same directness he applies to chips and software. The question of whether humanoid or non-humanoid robots will dominate is settled, in his view, in favor of humanoids — primarily because they can learn from existing human video data and accept direct human demonstration through motion capture. The verification problem is also cleanly solved: did the glass go in the dishwasher correctly? That binary reward signal makes RL training straightforward ^{[1] — Gavin Baker "The humanoid vs. non-humanoid robot debate is settled. Humanoids can learn from YouTube videos and human demonstrations, and the verificati…" 29:56 Play} . Baker points to Tesla Optimus — a video of 50 robots performing 50 different tasks — as compelling evidence. As for competitive dynamics, it will mirror the EV story: Tesla versus Chinese manufacturers, with the race defined by who can scale production and lower costs fastest.

David George brings the fireside chat to a close, thanking Gavin Baker for a characteristically thoughtful and wide-ranging conversation. The episode ends with the standard a16z podcast outro — an invitation to like, subscribe, leave a rating, and find the show on YouTube, Apple Podcasts, Spotify, X at @a16z, and Substack at a16z.substack.com, followed by the standard legal disclaimer noting that content is for informational purposes only and does not constitute investment advice.