How DeepStack settles a question · Part 10

Does AI hardware hold its value?

Does the hardware hold its value?

From Part 10: The Bubble Is in the Denominator. 16 voices on 6 sides, attributed and dated; the DeepStack Call with its history; 6 dated tests that settle it; the results as they come in.

The mechanism

The S&P 500 trades at 19.0 times forward earnings, and the market looks cheaper because profits surged. Those profits depend on AI spending: capital spending becomes supplier revenue and reported profit, so the denominator of the multiple is partly the buildout itself, while about $250 billion of hyperscaler bonds issued through August compete for the savings that set the discount rate, with the 10-year Treasury at 5.31 percent on October 5. Whether the profits are real depends on whether the hardware holds its value: if used-chip prices fall once scarcity fades and server lives prove shorter than the schedules, depreciation grows faster than operating income and the earnings being capitalized belong to no one yet. The Call would change if chip values held, productivity turned higher and lab revenue compounded.

Approved by Marco Cauduro on October 10, 2026.

The DeepStack callSkeptical

The market is not cheap. It is capitalizing AI earnings before anyone knows who owns them.

Cheap is conditional · Conviction High (4/5)

Our Call takes the side “A model, not a price”.

Read the full call

Call history2 versions · last changed
  1. v2Cheap is conditionalConviction High (4/5)Side: A model, not a priceSide named by the owner.
  2. v1Cheap is conditionalConviction High (4/5)First call, as published with the story. The conviction, the actions and the flips were set down for the site on October 7, 2026.

6 voices

A model, not a price

  1. Michael Burry

    CoreWeave and Nvidia's efforts to show stamina in the pricing of their chips are simply at odds with the depreciation realities lenders actually face.

    Source

  2. Michael Burry

    Manufacturers such as Nvidia and hyperscalers such as Oracle and Google Cloud are offering residual value guarantees for chips not because they want to but because lenders require them, given that chip depreciation is a very real problem following…

    Source

  3. Ren

    Architectures change faster than factories can retool, and with NVIDIA targeting up to 30% lower total cost of ownership with 800 VDC, a shift that large rewards whoever redesigns fastest, so today's winning box may not win in five years, which…

    Source

  4. Ren

    The popular Jevons-paradox story that cheaper inference means bulk consumption and justifies simply buying the picks-and-shovels trade does not match Astra's actual pricing, since Astra's tokens got more expensive per unit (versus Sol, 2.5x on…

    Source

  5. Kakashii

    Kakashii calculates that Nvidia's plan to ship over 10 million chips of Blackwell and Rubin by the end of 2026 implies a total required power of roughly 17 to 23 GW, averaging around 20 GW, and questions whether enough data centers are under…

    Source

  6. Kakashii

    Kakashii's long-standing thesis, which they say has gained support and validation over time, is that Blackwell GPUs had to be shipped to warehouses because there were not enough compatible datacenters (with the needed cooling and equipment) to…

    Source

6 voices

Still renting at January prices

  1. David George

    GPUs are already running hot even while the data indicates it's still very early in mature AI adoption and utilization.

    Source

  2. David George

    Since 2023, tech has been the earnings growth story, contributing about 76% of the S&P 500's total earnings growth in 2026 as of late August.

    Source

  3. JunkBondInvestor

    Contracted GPU lease rates and second-hand prices have held up even for chips launched 3 to 6 years ago, and compute is now about 60% of hyperscaler capex, so the scarce, short-lived asset is a growing share of what is being financed.

    Source

  4. JunkBondInvestor

    The entire AI buildout has been constrained by scarcity of power, GPUs, transformers, interconnection, skilled labor and ready-to-use capacity, and this scarcity has propped up the economics of everything financed against it by holding up lease…

    Source

  5. Tae Kim

    Tae Kim believes that if OpenAI were public, its stock would have gapped up 30% or more on hitting $70 billion in ARR, a blowout versus expectations on the level of Nvidia in May 2023, and that the figure proves his August report that OpenAI's ARR…

    Source

  6. Tae Kim

    Axios reported, and the Financial Times later confirmed the same day, that OpenAI's annual recurring revenue is nearing $70 billion, with enterprise sales more than doubled since July according to sources familiar with the financials.

    Source

Where they cross

Who answers whom.

The specific points where one voice meets another: the same evidence read two ways, or the same mechanism with a different sign.

Alternative explanations

Four other readings.

What would have to be true for the Call to be wrong, as the other side argues it.

  1. The hardware earns because it is used: GPUs are already running hot while the data says adoption and utilization are still early, so the capacity is not idle collateral (the case David George makes).
  2. Old chips still rent at a price: contracted lease rates and second-hand prices have held for chips launched three to six years ago, and compute is about 60 percent of hyperscaler capex (the case JunkBondInvestor makes).
  3. The checks are arriving: OpenAI's recurring revenue is nearing $70 billion, with enterprise sales more than doubled since July, so the earnings being capitalized have an owner (the case Tae Kim makes).
  4. The rise is earnings, not multiples: tech's forward multiple is up about 14 percent since October 2022, against 250 percent in the run to March 2000, so this is not 1999 (the case Ren makes).

What settles it

The dated tests.

The tests the story set, each with its specification where the owner has written one and the piece of the argument it verifies; on the Docket, resolved on the Results page.

  1. Microsoft and Alphabet report October 28, after Meta on October 21, and Amazon follows on October 29; depreciation, server lives and 29.5% growth will test the earnings denominator.

    Part 10: The Bubble Is in the Denominator Pendingp10-t01Pre-specifiedChecks the depreciation claim

    Question
    Did the combined depreciation and amortization of Microsoft, Alphabet, Meta and Amazon for the quarter ended September 30, 2026 grow faster from the year-earlier quarter than their combined operating income?
    Source
    SEC EDGAR, the four companies' earnings releases (Form 8-K) and quarterly reports (Form 10-Q)
    Rule
    Sum depreciation and amortization, and sum operating income, for the four companies for the September quarter of 2026 and of 2025, from each release or 10-Q. The answer is yes if the year-over-year growth rate of depreciation is above the rate of operating income; no if it is equal or below. A change to server useful lives disclosed in the same filings is DeepStack's reading into the Result, not the answer.
    If the source is late, revised or silent
    If any of the four has not reported by November 7, 2026, or restates the year-earlier quarter, the test is inconclusive, and the Result says what was read and what was missing.

    The Call, before the dateExpects yes. Gives the yes answer 75%.

    75%

    0 · no100 · yes

  2. Anthropic filing or listing reveals audited revenue, losses and concentration, testing the $400B revenue milestone.

    Part 10: The Bubble Is in the Denominator Pendingp10-t02 · monthPre-specifiedChecks the lab-revenue claim

    Question
    Did Anthropic file a registration statement (Form S-1 or equivalent) with audited financial statements on SEC EDGAR, or begin trading, by October 31, 2026?
    Source
    SEC EDGAR, Anthropic filings
    Rule
    The answer is yes if a registration statement for Anthropic carrying audited revenue, net loss and customer-concentration disclosures is public on EDGAR by October 31, 2026, or its shares begin trading by that date; no otherwise. The audited figures against the $400 billion run-rate line Romero named for the end of 2027, and the two-customer share Silverio cited, are DeepStack's reading into the Result.
    If the source is late, revised or silent
    A confidential draft filing reported in the press but not public on EDGAR, or a filing without audited statements, is inconclusive, and the Result says what was read and what was missing.
  3. At 5.5% 10-year

    A 5.5% Treasury yield tests Howell’s turning-point history against TSCS’s warning that the long end will keep rising.

    Part 10: The Bubble Is in the Denominator Pendingp10-t03Open questionChecks the discount-rate claim

    Specification pending

  4. Next used-GPU print

    Three-year-old chip rents and resale prices show whether scarcity or durable economics supports hardware values.

    Part 10: The Bubble Is in the Denominator Pendingp10-t04Open questionChecks the hardware-value claim

    Specification pending

  5. Productivity data

    A rise in total factor productivity would show the rest of the economy can write the checks; a flat reading leaves bondholders exposed.

    Part 10: The Bubble Is in the Denominator Pendingp10-t05Open questionChecks the productivity claim

    Specification pending

  6. The record holds or breaks: whether the Thrift Savings Plan's C Fund, which holds the S&P 500 with dividends reinvested, closes above its October 6 high of $126.18 on any day through July 12, 2027.

    Part 10: The Bubble Is in the Denominator Pendingp10-t06Pre-specified

    Question
    Does the C Fund of the Thrift Savings Plan, the federal employees' S&P 500 index fund, close above $126.1817, its highest share price through October 12, 2026 (set on October 6, 2026), on any business day from October 13, 2026 through July 12, 2027?
    Source
    Thrift Savings Plan, share price history of the C Fund (daily closing share prices, tsp.gov)
    Rule
    Read the C Fund's daily share prices from October 13, 2026 through July 12, 2027 in the plan's share price history. The answer is yes if any of them is above $126.1817, the highest close through October 12, 2026 (October 6; the plan posted no price on October 12, a federal holiday); no if every one is at or below it. The fund holds the stocks of the S&P 500 and reinvests their dividends, so its share price is the record the test reads.
    If the source is late, revised or silent
    If the plan has not posted the prices through July 12, 2027 by July 19, 2027, the test is inconclusive, and the Result says what was read, what was missing and the day of the next reading. A price the plan corrects is read as corrected.

    The Call, before the dateEither answer fits the Call. Gives the yes answer 88%.

    88%

    0 · no100 · yes

What changed

The record of Part 10.

The Part’s publication, every change to the Call, every result that came in and the next dated test, oldest first.

  1. PublishedPart 10 published; the Call: Cheap is conditional.
  2. Call updatedThe Call: Cheap is conditional. The market is not cheap. It is capitalizing AI earnings before anyone knows who owns them. Our Call takes the side “A model, not a price”. by Marco Cauduro
  3. Next testNext: By Oct 31, 2026, Anthropic filing or listing reveals audited revenue, losses and concentration, testing the $400B revenue milestone. Pre-specified
Sources (9)
  1. FactSet - S&P 500 earnings multiples and estimates
  2. CNBC - 10-year Treasury yield
  3. a16z - technology profit growth
  4. Prof G Markets - AI and earnings
  5. Aurelion Research - Nvidia and infrastructure
  6. Apollo - hyperscaler credit and cash flow
  7. Axios - OpenAI annual revenue
  8. PYMNTS - Anthropic listing expectations
  9. BLS - productivity release

Market data are as of the dates cited. Voices are paraphrased from public writing and attributed by name.

Analysis and opinion, not investment advice. Voices are paraphrased from public writing and attributed by name; DeepStack shows no score, ranking or accuracy for any person. Every question, one page each.