Gather the strongest counterarguments to Burry's thesis from sell-side analysts, hyperscaler executives, and independent bulls.

Michael Burry’s core claim is accounting, not physics: hyperscalers book Nvidia-class servers over roughly five to six years, while he argues the economic replacement cycle is closer to two to three years, so industry depreciation is understated by about $176 billion from 2026 to 2028, with Oracle’s earnings potentially overstated by roughly 27% and Meta’s by about 21% by 2028. The strongest counters do not deny that newer chips are far more efficient. They show that older chips still clear cash costs, still get multi-year contracts, and still sit inside a GAAP framework that keys off expected use by that company, not Nvidia’s product cadence.

1. Residual value and utilization: the hardest empirical counter

The cleanest attack on a two-to-three-year economic life is the market for the chips themselves. A 2020 A100 is still a contracted asset in 2026, not scrap.

CoreWeave disclosed on its Q2 2026 call an A100 contract running into 2029—nine years after the May 2020 launch—and management said prior-generation pricing is at or above levels from years earlier, with an expiring H100 batch rebooked at about 95% of the prior rate. Jensen Huang amplified that the A100 fleet is “mission-capable from 2020 through 2029,” and Nvidia CFO Colette Kress has said A100s shipped six years earlier were still running at full utilization because CUDA keeps the software stack current. Silicon Data’s residual-value estimate for the A100 has held near $5,000 (about $4,956 in one September 2026 reading), built from projected rental income rather than a fire-sale print. Marketplace and index prints in early October 2026 still show A100 hours around $1.50–$1.80 at neoclouds and higher at hyperscalers, while H100 on-demand has been firming—Ornn recorded a 22% monthly rise to about $3.28/hour in September. Nvidia’s own September investor slide, using Silicon Data, plotted A100 residual value at 25% of an assumed $20,000 cost, H100 at 58% of $35,000, and B200 above assumed cost. Bernstein’s Stacy Rasgon (November 2025) put the operating math under that: cash costs on an older GPU are far below rental prices, so even a five-year-old A100 can still earn a comfortable contribution margin (one summary of the note put A100 hourly revenue near $0.93 against cash costs near $0.28, implying roughly a 70% contribution margin), and only something as old as Volta approaches cash breakeven. Lambda’s Matt Rowe has told Bernstein the effective life can stretch to seven or eight years via warranty extensions and redeployment, with H100 utilization above 85% and public pricing unchanged for more than a year at the time of that comment.

What this means for the bear case: a chip can lose frontier-training relevance and still be a positive-NPV asset. Burry’s own reply—that rental demand does not prove the book value is right—is fair, but it shifts the argument from “these chips die in 24–36 months” to “residual value is lower than a six-year straight-line schedule implies.” That is a narrower claim, and the market data has not yet produced the collapse in old-chip rents his original framing required.

2. The waterfall, not a single obsolescence curve

Operators and independent bulls treat useful life as stacked markets, not one curve. Frontier training ages out fast—Hoppers are weak for new frontier jobs once Blackwell is plentiful—but inference, fine-tuning, batch jobs, and smaller models do not need the latest FLOPS-per-watt. Dave Friedman’s framing is the mechanism: if you price only the training curve, you get a one-to-three-year life and conclude the books are cooked; if you stack inference and long-tail workloads on top, the same silicon keeps earning after it leaves the frontier. Power and cooling reinforce that. A100 racks draw far less than liquid-cooled Blackwell racks, so legacy halls that cannot take 100 kW+ racks keep older silicon productive rather than idle. CUDA fungibility is what makes the cascade financeable: the same software stack means a customer can move a workload down a generation without a rewrite, which is why CoreWeave can still clear “attractive” prices on 2020 silicon.

For anyone shorting the depreciation gap, the test is not “does Blackwell obsolete Hopper for training?” It is “does the residual rental curve, after power and colocation, stay above cash cost for years four through six?” Current prints say yes for A100 and H100. That can reverse if inference supply floods or if power becomes the binding constraint everywhere—but it has not reversed yet.

3. Accounting defenses: estimates, not fraud

GAAP and IFRS define useful life as the period the reporting entity expects to use the asset, including residual value, not as Nvidia’s shipping cadence. Straight-line over four to six years is an industry convention, not a mandate; companies can accelerate if consumption of benefits is front-loaded. Extending a life is a change in estimate, applied prospectively, not an error that forces a restatement. Auditors have continued to sign the hyperscaler filings.

The filings also cut against a one-way earnings-management story. Meta’s move to 5.5 years for most servers and network assets, effective 2025, cut 2025 depreciation by about $2.92 billion and lifted net income by about $2.59 billion. Oracle moved servers from five to six years. Amazon went the other way: effective January 1, 2025 it shortened a subset of servers and networking from six years to five, citing the pace of AI hardware, and that change increased 2025 depreciation by about $1.4 billion. Microsoft’s range remains roughly two to six years; Alphabet has held around six. Nvidia’s seven-page memo to analysts, which named Burry, argued customers depreciate over four to six years based on observed longevity, and that two to three years is simply wrong. Ed Yardeni sided with the hyperscalers on the same grounds: older data-center fleets have historically stayed in service, and the cascade story is how management justifies the estimate. The Footnotes Analyst view is similar—different lives on similar hardware are not evidence of misreporting if expected use and residual values differ.

The limit of this defense is real. A legal estimate can still be optimistic. Amazon’s partial shortening shows managements do not all believe six years is conservative for every GPU cohort. The accounting counter defeats “fraud”; it does not prove six years is the right economic life for the newest, most expensive cohort.

4. Eisman: the case is “too academic,” and the real risk sits elsewhere

On the New Money podcast in early October 2026, Steve Eisman—the other Big Short figure—accepted the accounting logic and rejected its weight.

His words, as reported across Business Insider, Yahoo Finance, Benzinga, and IBTimes:

• “With all due respect to Michael, I think his argument is too academic.”
• “If AI succeeds because Anthropic and OpenAI grow like crazy, and the hyperscalers do well, it won’t matter whether the depreciation schedule changed from three or four years to five or six years.”
• “Where I think he is wrong, for the moment, is that there is such demand for chips right now that there is still huge demand for the older chips.”

Where he agrees: stretching lives flatters near-term earnings, and the dollar amounts Burry cites (more than $175 billion of understated depreciation in 2026–2028) are large enough to matter on a standalone P&L. Where he disagrees: current scarcity means older chips are fully utilized, so the economic write-down Burry models is not showing up in rents or contracts. The binding risk, in his view, is concentration, not depreciation. He has called OpenAI and Anthropic the “Achilles’ heel” of the AI trade and said that if OpenAI fails and “the whole chain goes in reverse,” the result is a “massive correction that has nothing to do with the depreciation schedule.” He has also partially hedged his own AI book rather than dismissing the sector. A related fragility point attributed to him—that a large share of hyperscaler AI revenue traces to those two labs—is directionally consistent with public disclosures such as Microsoft’s OpenAI-related revenue, but the precise “70% of AI revenue” figure is his estimate, not a filed number.

Eisman’s argument is strategically coherent and empirically anchored in today’s tight GPU market. It is weaker as a refutation of Burry’s multi-year accounting math, because it is conditional on AI demand continuing to absorb both new and old capacity. He is explicit that the condition can fail.

5. Lab revenue: the condition Eisman needs is, for now, being met—unevenly

Anthropic told prospective investors on October 14, 2026 that Q3 revenue was $16.8 billion, up about 45% from $11.6 billion in Q2 and from roughly $4.7–$4.8 billion in Q1, and that it was profitable on an adjusted basis (one account cites an 11% adjusted operating margin and positive free cash flow; another notes the company was cash-flow positive in the quarter). Full-year 2025 revenue was about $4.6 billion, so the 2026 ramp is an order of magnitude. It is preparing an IPO, with reported valuation talk in the $1.8–$2 trillion range. OpenAI’s picture is messier and slower. It told investors annualized revenue was approaching $50 billion at the end of September, about $20 billion below a ~$70 billion figure that had circulated; the Financial Times and Bloomberg attribute much of the gap to methodology—Anthropic books gross sales through cloud partners, OpenAI books its net share. Bloomberg has also reported an expectation of $70 billion or more in annualized revenue by year-end 2026, while recognized full-year 2026 revenue has been put nearer $35–$36 billion in investor-material summaries. OpenAI remains loss-making; one report put 2025 revenue at $13.1 billion against large operating losses, and materials summarized by the FT have projected deeply negative cumulative free cash flow through 2030 even as revenue scales. IPO timing has slipped toward 2027, with a reported raise of about $30 billion at a $1.2–$1.5 trillion valuation range discussed in October coverage.

These trajectories support the bull path Eisman describes: the labs are growing fast enough that hyperscalers can point to real AI demand, and Anthropic has crossed into adjusted profit. They do not close Burry’s depreciation math, and they sharpen his circularity concern—much of the “demand” that keeps old GPUs rented is demand from two private companies whose own capex is funded by the same ecosystem. Methodology differences mean headline run rates should not be stacked as if they were GAAP revenue.

6. Capex returns: the weakest pure counter, and partly a bear point

Sell-side work on returns does not cleanly refute Burry. Goldman Sachs (Ryan Hammond and colleagues, notes in late September and mid-October 2026) estimates the largest U.S. hyperscalers need on the order of $300 billion in annual AI revenue to break even on average 2026–27 capex, and roughly $417–$526 billion for a 10–20% return, against cloud revenue only about $70 billion above the pre-AI trend as of Q2 2026. The same work has hyperscalers on track for about $800 billion of capex in 2026, with consensus nearer $1.1 trillion in 2027 and Goldman higher still. Company guidance after the July 2026 prints clusters around Amazon ~$220 billion cash capex, Alphabet $195–$205 billion, Meta $130–$145 billion, and Microsoft on the order of $175 billion including leases for calendar 2026—roughly $745 billion for the four, before Oracle. Realized ROIC has already compressed (Goldman has cited a drop from about 40% to 25%). The bull offset is the growth rate, not the level: year-over-year cloud revenue growth for Amazon, Google, Microsoft, and Oracle accelerated from about 25% in 2024 to about 48% in Q2 2026, and managements still describe themselves as capacity-constrained. An independent tally from Exponential View put annualized AI revenue near $276 billion covering more than the cost of capital already in service, while warning that capital committed over the next 12–24 months requires infrastructure revenue to keep compounding at roughly 27–42% a year for the life of the equipment. Hyperscaler forward multiples have also de-rated to multi-year lows, so the equity market is no longer paying peak-bubble multiples for the spend.

Capex ROI is therefore a partial agreement with Burry, not a rebuttal. The schedule debate changes the timing of expense; it does not create the revenue required to earn a return on $800 billion-plus of annual spend. Bulls win this round only if cloud and lab revenue keep compounding faster than depreciation catches up. Bears win it if the $230 billion-scale gap Goldman describes persists once the current capacity comes online and starts depreciating.

Ranking by strength and evidence quality

Rank	Counterargument	Strength vs. Burry	Evidence quality	Why it ranks here
1	Old GPUs still rent, still contract, still clear cash cost	Strongest direct hit on the 2–3 year life	High: contracts, rental indexes, Bernstein unit economics, CFO utilization statements	Market prices are harder to dismiss than narrative. Does not prove six-year book value is exact.
2	Waterfall: training obsolescence ≠ economic death	Strong mechanism	Medium-high: operator testimony, power constraints, price-per-token math on A100 vs H100	Explains why residual value holds. Still a model, not a guarantee if inference oversupply arrives.
3	GAAP useful life is entity-specific; Amazon shortened a subset	Strong vs. “fraud,” moderate vs. “too long”	High on standards and filings; medium on whether 5–6 years is right	Kills the fraud framing. Leaves the estimate open to challenge, which Amazon’s own shortening illustrates.
4	Eisman: too academic while demand is scarce; labs are the real risk	Strong as a ranking of risks, weak as a multi-year accounting rebuttal	Medium: clear quotes, consistent with rental tightness; conditional on AI success	Best statement of why the P&L distortion may not be the trade. Explicitly agrees the accounting point is real.
5	Anthropic/OpenAI revenue ramps	Supports the “if AI succeeds” branch	Medium: multiple outlets on Anthropic’s $16.8B quarter; OpenAI figures conflict by ~$20B on run-rate definition	Shows demand is not fictional. Private, non-GAAP, and circular with hyperscaler capex.
6	Capex will earn its cost of capital	Weak as a counter; partly supports caution	Medium-high on the gap (Goldman), lower on the path to closing it	The $300B breakeven vs ~$70B above-trend is the number bulls have not yet beaten. Growth rates are the only live rebuttal.

Bottom line. The evidence-weighted counter to Burry is not that depreciation schedules are irrelevant. It is that the two-to-three-year economic death he uses to generate the $176 billion gap is not what rental markets, nine-year A100 contracts, or Bernstein’s cash-cost math currently show. GAAP gives companies room to book five to six years if they genuinely expect to use the assets that long, and at least one hyperscaler has shortened lives when it disagreed. Eisman’s October 2026 intervention is the clearest investor-level version of that view: the accounting is real, the scarcity of compute makes it second-order for now, and the trade breaks if OpenAI and Anthropic stop growing—not if Meta’s server life is 5.5 years instead of 3. The open flank in the bull case is returns on the capex itself. Goldman’s roughly $300 billion annual AI-revenue hurdle versus about $70 billion of cloud revenue above trend is a larger, better-documented problem than the depreciation schedule, and it is the one Burry’s broader short does not need the accounting argument to win.

---

Recent Findings Supplement (October 2026)

Steve Eisman's October 2026 rebuttal on the New Money podcast stands as the most prominent and directly targeted counter to Michael Burry's AI depreciation and overinvestment thesis. Eisman calls Burry's focus "too academic," arguing that overwhelming chip demand—including for older generations—renders changes in depreciation schedules (e.g., from 3-4 to 5-6 years) immaterial if Anthropic and OpenAI scale rapidly and hyperscalers thrive. He agrees the accounting concern has some validity but views it as secondary. The real vulnerability, per Eisman, is extreme concentration: OpenAI and Anthropic reportedly account for ~70% of AI-related revenue at Microsoft, Amazon, Alphabet, and Oracle. A failure there triggers a "massive correction" unrelated to depreciation. Eisman has partially hedged his own AI positions while acknowledging he has participated in the trade.[1][2][3]

This directly addresses Burry's core claim (hyperscalers understating depreciation by >$175 billion across 2026-2028 by extending asset lives) while shifting emphasis to demand sustainability and customer concentration.[4]

Empirical data on older GPU useful life and utilization strongly challenges assumptions of rapid obsolescence. CoreWeave disclosed an A100 (launched 2020) rental contract extending through 2029—nine years post-launch—at attractive pricing, with CEO Michael Intrator and CFO Nitin Agrawal noting prior-generation ASPs holding at or above prior levels and renewals providing incremental returns. Occupancy for tracked A100s rose from 61% average in 2024 to 66% in 2025 and ~77-86% in 2026; rental rates stabilized around $0.92/hour after earlier declines. Nvidia's Jensen Huang highlighted this as evidence of durability via CUDA compatibility and fungibility, with residual-value estimates (e.g., Silicon Data) placing six-year-old A100s near $5,000 based on projected rental income. H100s show similar retention patterns in secondary/resale tracking.[5][6][7]

These real-world contracts and utilization trends (post-April 2026 disclosures) provide concrete evidence that economic life exceeds Burry's implied 2-3 year horizon for many workloads, particularly in power-constrained or legacy infrastructure environments.

Anthropic and OpenAI revenue trajectories demonstrate accelerating monetization that supports ongoing capex. Anthropic reported Q3 2026 revenue of $16.8 billion (up ~46% QoQ from $11.5-11.6 billion in Q2), implying an annualized run rate of ~$67 billion; it achieved adjusted profitability for a second straight quarter. Full-year 2025 revenue reached $4.6 billion (roughly 12x prior levels per its IPO prospectus). OpenAI's annualized run rate approached $50 billion by end-September 2026 (below some prior $70 billion circulating estimates, partly due to gross-vs-net partner sales accounting differences), with full-year 2026 revenue projected around $35 billion. Both show explosive growth tied to cloud partnerships, bolstering hyperscaler AI demand.[8][9][10]

This growth trajectory counters narratives of insufficient AI revenue to justify spend, though concentration risks (noted by Eisman) remain.

Hyperscaler executives and sell-side analyses highlight monetization speed and capex coverage. Amazon's Andy Jassy stated the company is "monetizing capacity as fast as we can install it" amid $200 billion 2026 capex guidance (heavily AI-focused), with AWS showing continued strong growth. Broader analyses (e.g., one investment brief) estimate annualized AI/infrastructure revenue at $276 billion covering 182% of in-service AI capital costs (depreciation + OpEx + 15% ROIC target) for deployed assets. Nvidia's investor materials and memos defend 4-6 year lives as realistic, citing retained values (A100 ~25%, H100 ~58% in their framing) beyond accelerated schedules. Meta disclosed a specific $2.9 billion depreciation reduction from extending certain server lives to 5.5 years.[11][12]

These point to returns materializing faster than bearish timelines assume for operational fleets.

Depreciation accounting defenses emphasize economic vs. book life distinctions and actual performance. Companies like Meta, Microsoft, and Alphabet extended useful lives (to 5-6 years in some cases), with executives noting capacity is often pre-sold. Realized rental economics and renewals (as with A100/H200 examples) support longer horizons than strict 2-3 year product-cycle assumptions. Resale/residual trackers show meaningful retained value (e.g., H100s 60-75% after 2-3 years in some data), and power/cooling constraints in legacy data centers sustain demand for lower-power older GPUs.[13][14]

Ranking of counterarguments by strength and evidence quality (post-April 2026 sources prioritized):
- Strongest: Eisman's demand-and-concentration argument (direct, reasoned rebuttal with qualitative agreement on risks; multiple corroborating reports).
- Very strong: A100/CoreWeave utilization/resale data (specific contracts, occupancy trends, residual estimates from earnings calls and trackers).
- Strong: Anthropic/OpenAI revenue figures (recent quarterly disclosures and prospectus data showing scale and growth).
- Moderate-to-strong: Hyperscaler monetization statements and capex-coverage analyses (executive commentary + independent modeling).
- Supporting: Accounting defenses tied to realized economics (less novel but reinforced by new data).

These elements collectively suggest Burry's thesis underweights current demand dynamics and realized asset performance, though concentration and future return validation remain acknowledged risks even among partial bulls like Eisman. Implications for competitors or entrants: Focus on diversified demand sources, proven utilization economics, and power-efficient older hardware plays could differentiate from pure new-GPU capex bets. Data remains fluid with ongoing IPO filings and earnings.