CHAPTERS
AI bubble question, investor context, and why comparisons to 2000 matter
The conversation opens with the “are we in an AI bubble?” provocation and frames the discussion through Gavin Baker’s experience investing through the 2000 telecom/internet bubble. The hosts set expectations: evaluate AI build-out with evidence (usage, returns, and market structure) rather than hype.
- •Taboo question: whether AI is a bubble and where the economic evidence shows up
- •Baker’s credibility lens: lived experience of the 2000 bubble
- •Promise to compare today’s AI capex cycle vs. telecom/internet era dynamics
- •Early framing around measuring real adoption vs. “ghost” narratives about agents
The data-center spending surge vs. token-usage growth
David George lays out the scale of AI infrastructure expansion—trillions in planned data-center investment—alongside rapid growth in actual AI usage (tokens processed). This sets up the central tension: scary capex headlines versus signs of genuine demand.
- •~$1T existing US data centers; $3–4T more projected over five years
- •Recent build-out rivals the (inflation-adjusted) interstate highway system in dollars
- •OpenAI’s massive committed deal volume as a signal of capital intensity
- •Google reports ~150x token processing growth in ~17 months as a demand proxy
Lessons from 2000: “dark fiber” vs. “no dark GPUs”
Baker argues the core feature of the 2000 telecom bubble was overbuilt, unused capacity—dark fiber. He contrasts that with today’s AI compute environment where GPUs are heavily utilized, even stressed, implying demand is real rather than speculative overbuild.
- •2000 bubble as largely a telecom/infra bubble, not just “dot-com”
- •Dark fiber concept: fiber laid but not “lit” (no optics/switches/routers)
- •Peak stat: ~97% of laid fiber was dark
- •Today’s contrast: “no dark GPUs”; compute is constrained and heavily used
ROI and valuation checks: why Baker says this cycle isn’t a bubble (yet)
The discussion shifts to financial reality checks: valuations and returns on capital. Baker claims the largest GPU buyers have seen meaningful ROIC improvement since ramping capex, suggesting investment is paying off so far—even if future spend (e.g., Blackwell) is debated.
- •Valuation contrast: Cisco peak multiples vs. Nvidia’s lower multiple today (per Baker)
- •GPUs as revenue-generating assets rather than idle capacity
- •ROIC lens: biggest public spenders show roughly a 10-point ROIC increase after capex ramp
- •Open question: whether the next wave of spend sustains returns, but “so far, positive”
Hyperscaler balance sheets and the “win at all costs” infrastructure race
George emphasizes that the primary spenders are exceptionally strong companies with massive free cash flow and cash reserves. The conversation frames the AI capex race as existential for some incumbents, with Google and Meta willing to spend aggressively to avoid losing.
- •Capex funded by companies generating enormous free cash flow and holding large cash balances
- •High cost to build: tens of billions to equip ~1GW full-stack data center capacity
- •Google/Meta “existential” framing; reported internal sentiment like “don’t lose the race”
- •Expectation of near-term financial pressure (FCF dips) tied back to ROIC discussion
Round-tripping deals: real, but (so far) limited risk
They address fears about “round-tripping” (vendors financing customers who then buy from the vendor), a notorious feature of prior bubbles. Baker acknowledges it happens because money is fungible, but argues the scale is small relative to the overall market.
- •Round-tripping defined and why it spooks investors (historical precedent)
- •Money fungibility makes restrictions hard to enforce in practice
- •Claim: it is happening, but at small scale today
- •Emphasis on monitoring rather than treating it as systemic
Big Tech’s “right to win” vs. execution risk (Google’s wake-up call)
The talk explores how advantages in data, distribution, talent, and capital give Big Tech strong positioning, but not guaranteed outcomes. Baker describes ChatGPT as a “Pearl Harbor” moment for Google—an external shock forcing faster execution.
- •Big Tech advantages: distribution, data, compute dollars, and talent
- •“Right to win” framing—yet execution determines whether it’s sustaining innovation
- •Google’s cautious product posture (and regulatory scrutiny) vs. need to respond quickly
- •Historical cautionary note: incumbents can still fail if they don’t execute (IBM analogy)
AI infrastructure economics: lower gross margins, scaling laws, and why that’s okay
They discuss how AI’s compute intensity structurally lowers gross margins versus classic SaaS. Baker argues lower gross margins don’t preclude great businesses; they reflect real usage and the realities of scaling laws and test-time compute.
- •SaaS-era benchmark: 80–90% gross margins (2021–2022) vs. AI’s compute-heavy reality
- •Scaling laws and “Bitter Lesson” imply continued reliance on compute
- •Frontier labs unlikely to reach SaaS-like gross margins soon; Opex may be lower though
- •Business quality should be judged by overall economics, not gross margin nostalgia
Application layer reset: SaaS isn’t dead, but must accept margin pressure
Baker revisits his earlier pessimism about application SaaS and offers a more nuanced view: winners can emerge, especially serving fragmented SMBs. He warns SaaS leaders against clinging to legacy margin structures, arguing margin compression may signal successful AI adoption.
- •Updated view: application SaaS can produce big winners, especially in SMB/fragmented markets
- •Warning against repeating retailers’ mistake of dismissing low-margin businesses (Amazon analogy)
- •AI success “definitionally” creates gross margin pressure; resist fear-driven underinvestment
- •Cloud transition analogy: Microsoft/Adobe proved margin shifts can still create great outcomes
A practical signal: lower gross margins can indicate real AI product usage
George and Baker describe an emerging investor heuristic: very high gross margins may indicate AI features aren’t truly being used at scale. They argue companies should communicate margin strategy clearly and leverage profitable legacy cash flows to fund AI products aggressively.
- •Investor pattern: “AI company” + very high gross margins can be a red flag on usage
- •Preference for larger revenue at lower margins versus small revenue at SaaS-like margins
- •Legacy software companies can run AI offerings at break-even using existing profits
- •Examples of competitive urgency (e.g., coding tools) and the risk of data/moat flywheels forming
Consumer AI and distribution: browsers, Chrome’s gravity, and platform power
The discussion turns to consumer market structure and the battle for distribution. Baker suggests AI-native browsers may be vulnerable if Google leverages Chrome’s massive user base, and he cautions against betting against incumbents with entrenched distribution.
- •Shift from search “referral” model toward AI interfaces completing tasks end-to-end
- •AI-native browsers vs. Chrome’s multi-billion-user distribution advantage
- •Google’s strategic caution (and legal scrutiny) may delay but not prevent entry
- •General thesis: large user bases and distribution remain decisive
Reasoning models revive the consumer flywheel and reshape frontier lab economics
Baker argues reasoning and RL-based post-training make user scale more valuable, re-enabling the classic consumer internet flywheel: more users → better model → better product → more users. This shifts the outlook for frontier labs that may lack proprietary data but can build strong distribution.
- •Pre-reasoning view: frontier models without unique data/distribution are rapidly depreciating assets
- •Reasoning + RL post-training increases the value of large user bases for improvement loops
- •Consumer flywheel analogy: product improves via user interaction and feedback
- •Mentions of leading labs and competitive dynamics; skepticism about “GPT-5 ends scaling laws” claims
Chips and AI infrastructure competition: Nvidia vs. Google TPU, plus Broadcom/AMD
They outline a multi-layer infrastructure battle: Nvidia as a systems/data-center company versus Google’s TPU stack, with Broadcom and AMD collaborating to offer alternative fabrics and ASIC pathways. Baker predicts many custom ASIC efforts may be canceled within a few years, especially if TPUs become broadly available.
- •Nvidia evolution: from semis → CUDA software → rack-level systems → data-center architecture
- •Core rivalry framed as Nvidia vs. Google TPU
- •Broadcom’s pitch: Ethernet-based open-standard fabric + custom ASICs; AMD as plug-in fallback
- •Prediction: multiple high-profile ASIC programs get canceled within ~3 years; TPU externalization would accelerate this
Business model shift: paying for outcomes, affiliate economics, and services displacement
They explore how AI enables outcome-based pricing, especially where results are measurable (e.g., customer support resolution). Baker extends the idea to consumer purchasing via AI agents, predicting affiliate/marketplace-like economics that compress today’s advertising inefficiencies.
- •Outcome pricing is natural for AI because performance can be verified (e.g., resolution, satisfaction)
- •AI may augment and replace human work, pushing payment models toward outcomes
- •Consumer agents could negotiate purchases (e.g., travel bookings), capturing affiliate fees
- •Prediction: ad-market inefficiencies shrink as AI closes loops and optimizes transactions
Robotics and humanoids: why Optimus (and China) define the near-term race
The conversation ends with a forward-looking view on robotics, where Baker argues humanoids are increasingly favored because they can learn from human demonstrations and existing video data. He highlights Tesla’s Optimus progress and expects competition to mirror the auto market dynamic: Tesla versus Chinese manufacturers.
- •Robotics framed as “very real,” with a competitive landscape like the car market
- •Humanoids vs. non-humanoids: Baker suggests the debate is resolving in favor of humanoids
- •Learning advantage: training from video/demonstrations; humans can teach via wearable capture
- •Optimus progress cited as impressing roboticists; rapid task learning/verification as key
