Gavin Baker, chief investment officer of Atreides Management, argues the AI buildout avoids the credit collapse that broke the late 1990s fiber optic telecom buildout, because Big Tech cloud operator spending is cash flow funded and Taiwan
Gavin Baker argues the current AI infrastructure cycle is being paid for in cash, not borrowed against future revenue — a distinction most bullish pitches skip, and the central reason he believes this buildout won't collapse the way the 1999 fiber boom did.
The 1999-2000 fiber buildout is the comparison Baker keeps returning to. Telecom operators loaded their balance sheets with debt to lay capacity ahead of demand, and when traffic failed to materialize at the assumed pace, the debt matured into cascading defaults. WorldCom did not fail because fiber stopped being useful. It failed because the funding structure presumed a growth rate that did not show up. Baker argued there is no comparable refinancing wall in front of today's hyperscalers, which removes the credit-event failure mode that broke the 1999 cycle.
Baker's broader framework, laid out across the same conversation, compares the two cycles along four axes: valuation, supply elasticity, the funding structure just described, and downstream content economics. The valuation axis is the loudest. At the time of the recording, Baker said the technology sector was trading at a discount to consumer staples — a configuration he called historically rare, which he reads as dot-com scar tissue still distorting multiples two decades after the last bust. (Note: recent market data showing technology and consumer staples multiples near parity in early 2026 suggests the gap Baker cited has partially closed, consistent with his reading that the discount was an anomaly rather than a structural feature.)
The supply-elasticity axis is where Baker says the bullish case is most underpriced. Fiber was a low-friction commodity to lay: any operator with a balance sheet could trench more conduit, and many did, until supply ran ahead of traffic. AI compute is constrained by the leading-edge production capacity at TSMC, the Taiwan-based foundry that fabricates Nvidia's most advanced processors, and by the power grid's ability to deliver megawatts to data center sites.
Independent data supports the supply-constraint picture. TSMC's advanced-node utilization reached 94% capacity as of mid-2026 — the highest since the 2021-2022 pandemic-era shortage — with 3-nanometer wafer starts booked through Q2 2027, according to MorrowReport. The company controls roughly 54% of global foundry capacity below 7 nanometers, the threshold where AI accelerators operate. Competing foundries lack the technical scale to absorb demand migration in the near term. Meanwhile, TSMC's CoWoS advanced packaging capacity — critical for AI chip stacks including HBM memory integration — was scaling from approximately 35,000 wafers per month in late 2024 toward a projected 130,000 by the end of 2026, per industry reports, with a persistent supply-demand gap that analysts pegged at roughly 20% as recently as early 2026. A $20 billion Arizona and Taiwan expansion plan announced in March 2026 will not reach meaningful production until late 2027 at the earliest. Baker's view, as independently corroborated by these capacity figures, is that TSMC's leading-edge capacity expansion will lag hyperscaler demand through the relevant window, so the AI buildout cannot overbuild the way fiber did in 1999. The bottleneck does not stop the cycle. It throttles it, which by Baker's reasoning is exactly what makes the cycle durable.
That throttle also decides where the value lands. If supply is structurally constrained, the producers of leading-edge chips and the platforms that aggregate end-user demand capture most of the surplus, while the application layer gets commoditized. Baker's working example is generative AI in AAA game development, where Baker estimated — using language he characterized as directional rather than precise — that costs have fallen roughly 90% once studios begin integrating the tools. (This figure is Baker's own estimate and has not been independently verified; it should be read as an illustrative data point from a single source, not a confirmed industry benchmark.) Cheaper content floods platforms, platforms take the rent, and application-layer companies converge on price. Investors who extrapolate SaaS-style multiples onto AI application names may be modeling against the wrong cycle.
Baker's framework does not rule out a downturn. It argues for a specific one: a slower repricing of upstream suppliers if hyperscaler capex eventually outruns the cash flow funding it, with platforms and chip vendors caught in the middle. Investors waiting for a 1999-style default cascade are waiting for the wrong failure mode. If the AI capex cycle breaks, the transmission path looks more like the App Store's economics than WorldCom's bankruptcy court.