I'm sharing 10 reasons why I think he is screwed this time:
1. AI buildout is still very early to call a top. Hyperscaler capex guidance for 2026-2027 keeps getting revised up, not down. We're still in the phase of building the physical infrastructure (data centers, power, memory, networking) the monetization layer (agents, enterprise AI products) is only just starting to scale. Calling a "top" on infrastructure before the application layer has even matured historically been premature (see: fiber buildout in the late 90s eventually got used, just after a shakeout).
2. Use cases for AI and utility will broaden. Today's AI spend is concentrated in chatbots, coding assistants, and search. Enterprise workflow automation, drug discovery, materials science, and defense/government applications are still in early pilot stages. Broader utility = broader revenue surface over time.
3. Robotics and humanoids are coming. Tesla Optimus, Figure, and others are racing toward commercial humanoid deployment. If this materializes even partially, it opens an entirely new demand category for compute, sensors, and precision components one that doesn't exist in today's revenue base yet.
4. Demand for memory and chips will continue to rise through 2035. Memory suppliers (Micron, SanDisk, SK Hynix) have described HBM as sold out through 2026 already, with multi-year supply agreements now including pricing floors that's a structural shift from the historically brutal memory cycle Burry references, not a repeat of it.
5. Trump and key tech executives are bullish on technology. The administration has praised Micron's U.S. investment commitments and framed domestic chip manufacturing as a strategic priority; Jensen Huang, Sanjay Mehrotra, and other CEOs continue to describe demand as capacity-constrained rather than demand-constrained. Policy tailwinds (CHIPS-style support, easier export rules) matter for the sector's ceiling.
6. Consumer and workplace behavior is shifting toward using AI to do actual work. Adoption curves for tools like ChatGPT, Claude, and Copilot inside real workflows (not just novelty use) have moved faster than prior tech adoption cycles (PC, internet, smartphone). That behavioral shift is what eventually converts infrastructure spend into recurring revenue.
7. P/E ratios are still low relative to the growth rates. Despite the run-up, several of these names trade at surprisingly modest forward multiples given their growth: Micron's forward P/E is in the high single digits, Nvidia's is around 19-20x. That's cheap by historical growth-stock standards if the growth continues which is exactly the disputed part.
8. The AI capex is being funded by genuinely profitable, cash-generative companies not debt-fueled dot-com-era startups. Microsoft, Meta, Google, and Amazon are self-funding much of this buildout out of existing free cash flow, which is a fundamentally different risk profile than 1999-2000, when unprofitable companies borrowed and issued equity to build out fiber nobody needed yet.
9. Bear theses like Burry's have historically arrived years before the actual top, not at it. Burry himself was short the market in 2021 with GameStop-style puts and it took a while to play out; being early and being right are not the same thing, and shorts carry theoretically unlimited downside if the timing is off.
10. The bottleneck argument cuts both ways. If demand is genuinely outstripping supply (as memory and GPU shortages suggest), that's usually a sign of a cycle with more room to run, not one that's about to collapse shortages don't typically precede crashes, gluts do.
Burry's core objection isn't that AI is fake it's that the rate of spend has outrun proven revenue, and that circular vendor-financing arrangements ($NVDA funding customers who then buy Nvidia chips) can make demand look more organic than it is.
That's a legitimate structural concern independent of whether AI itself is useful, and it's the same pattern that preceded real corrections in past capex-heavy cycles (telecom in 2000, shale in 2014).
Both theses can't be fully right at the same time, and reasonable people disagree on which cycle this looks more like.
