Vault
research

The Mask That Eats What It Hides — Memory, Power, and the Cyclical 20%

Created

The Mask That Eats What It Hides — Memory, Power, and the Cyclical 20%

Builds-on: cyclical-20-and-the-ai-capex-mask Related: compute-as-collateral-the-residual-value-wrap, hormuz-to-ai-repricing-causal-chain, ai-circular-financing-and-banking-exposure-audit, ai-survival-theater-and-the-bubble, the-data-center-convergence, ai-infrastructure-endgame-indicators, mechanism-vs-narrative-method

cyclical-20-and-the-ai-capex-mask treated AI capex the way 2007 treated housing: a line item large enough to hold up headline GDP while the cyclical slice underneath contracts. The analytical procedure that follows from that framing is subtraction. Strip out the masking line item, read what's left.

That procedure has a defect, and the memory market is where it shows. AI capex is not sitting on top of the cyclical 20% as a neutral cover. It is bidding away that slice's physical inputs and raising its household cost base at the same time. Some of the weakness you find after subtracting AI capex was produced by AI capex. The mask is eating what it hides.

This is not a rhetorical point. It changes what the counterfactual means, it changes which indicators are clean, and it adds a component-cost channel that runs straight back into the residual-value question in compute-as-collateral-the-residual-value-wrap.

Channel one: memory, and the durable goods it prices out

Durable goods consumption — vehicles, appliances, furniture, electronics — is the first named component of the cyclical 20%. It is now in a supply-driven price shock caused directly by the buildout.

HBM production for AI accelerators consumes roughly three times the wafer capacity per gigabyte of conventional DRAM. Samsung, SK Hynix and Micron control north of 95% of global DRAM and have reallocated capacity toward it. The result:

And the demand response, which is the part that lands in the national accounts: PC shipments are projected to contract 11.3% in 2026 and smartphone shipments 12.9%.

That is a double-digit volume contraction in consumer electronics, inside durable goods consumption, inside the cyclical 20% — caused by the same buildout that is holding up the equipment-investment line that makes the cyclical 20% look fine. An analyst subtracting AI capex from GDP and finding weak consumer electronics would code that as independent evidence of a consumer-led downturn. It is partly the mask's own shadow.

On the "cartel" question

The Gamers Nexus framing is coordination. The legal record is an allegation: 17 plaintiffs filed a class action on June 25, 2026 in N.D. Cal. against all three, alleging coordinated supply restriction, with a cumulative price increase near 697% from Q3 2024 through Q1 2026. The complaint's sharpest claim is that the HBM pivot was a pretext — that DDR3/DDR4 output was curtailed further than HBM demand required, and that in January 2026 all three simultaneously began vetting customers and policing order books. There is precedent: Samsung and Hynix pleaded guilty to DOJ DRAM price-fixing for 1999–2002 and paid $300M and $185M. There is also precedent for these suits failing — this is the third such civil action in two decades.

The useful analytical move is to notice that the conclusion does not depend on the conspiracy. A three-firm oligopoly with 95% share, facing a demand shock in a structurally higher-margin adjacent product, restricts commodity supply as a matter of ordinary profit maximization. No phone call is required. That matters because it means the price shock persists on the same trajectory whether the plaintiffs win or lose — which is not true of a genuine cartel, where enforcement is the remedy. Treat "cartel" as a description of market structure, not of conduct, and the forecast is unchanged and better founded.

Channel two: power, and the household budget

(The ratepayer channel itself is well-covered in the vault — the-data-center-convergence, ai-infrastructure-endgame-indicators, the-government-put-question-2026-conditions-and-models. What is added here is reading it as a transfer out of the cyclical 20% specifically, rather than as an energy-sector or political story.)

The second channel raises the non-discretionary share of household spending, which mechanically squeezes the discretionary share the cyclical 20% is made of.

PJM's capacity auction went from about $29/MW-day to $270/MW-day for 2025–26 and $329.17/MW-day for 2026–27 — roughly a tenfold move in two years. IEEFA and PJM's own market monitor attribute 63% of the 2025/26 increase to data center demand, about $9.3B recovered from customers in rates. From June 2026 the region collectively pays an additional $1.4B in capacity costs. Typical residential bills went up 4.9% on July 1, 2026; NRDC projects roughly $70/month above pre-surge levels by 2028; western Maryland and Ohio households are looking at $16–18/month from the capacity line alone.

Twenty-three states have already legislated or regulated on who pays. That is the tell that this has stopped being an energy-sector story and become a distributive one — which is the ai-infrastructure-endgame-indicators "ratepayer socialization" archetype arriving on schedule, in the open, with a number attached.

For the cyclical-20 framework the mechanism is simple: an involuntary $40–70/month increase in a non-discretionary category comes out of the same household budget that would otherwise replace a laptop or a refrigerator. It shows up in the data as weak durables demand and weak consumer sentiment — UMCSENT at 53.3 in cyclical-20-and-the-ai-capex-mask — attributed to the cycle. Part of it is a transfer to the buildout.

Channel three: the costs that aren't being paid

The third channel is not a price at all, which is why it doesn't appear in any series.

EPA's January 15, 2026 NSPS for stationary combustion turbines cut Title V major-source permitting obligations for "low use" turbines across small, medium and large subcategories, pushing them to state non-major programs, and created a "temporary" subcategory for turbines up to 850 MMBtu/hour operating under 24 months. A July 27 guidance memo exempted roughly 60 islanded off-grid AI data center power plants from Acid Rain Program caps. A proposed rule cutting public participation from minor-source air permits — the class that covers the trailer-sized diesel gensets — had comments due August 21, 2026.

Whatever one thinks of the policy, the accounting consequence is unambiguous: measured AI capex understates the real resource draw by the value of the compliance and externality costs that were waived. The buildout is larger than the capex line says, and the difference is being carried by people who did not sign a contract. Gamers Nexus calls this gaslighting; the analytically neutral version is that a portion of the buildout's cost has been moved off every ledger that a macro analyst reads.

Channel four: the loop back into residual value

This is the part that connects to compute-as-collateral-the-residual-value-wrap and it is the most consequential.

HBM is not a component of an AI accelerator. It is close to a majority of one. HBM3E is roughly 45% of the B200's bill of materials; memory is over 50% of a Blackwell Ultra B300; and the memory share of the next-generation Vera Rubin platform rises from 53% to 62% of BOM. Nvidia is reportedly weighing cuts of up to 81% to next-generation HBM capacity because of what memory now costs.

So when Nvidia writes a six-year take-or-pay at roughly $2.36/hr and a residual-value guarantee on the liquidation price of a GPU stack, the majority of what it is guaranteeing is the resale value of memory content bought at the top of the most violent price spike in DRAM history.

Memory is the most cyclical commodity in semiconductors. Peak-to-trough in the last two completed cycles: ASPs fell more than 55% from the 2018 peak through 2019–20, and DDR4 8Gb spot fell from over $5 in 2022 to $1.63 by January 2025, about 67%. The current level sits above every prior peak. The historical base rate for what happens next is not ambiguous.

Nvidia's residual-value guarantee is, in substance, a short position on the memory cycle, written at a cycle high, with a six-year tenor. That is a sharper statement of the wrong-way risk than the demand-side version. It does not require AI demand to disappear. It only requires memory to do what memory has done at the end of every cycle since the 1990s — at which point the replacement cost of the collateral falls, the liquidation value of the installed stack falls with it, and the shortfall the guarantee covers opens up while Nvidia's own margins are compressing on the same input.

flowchart TB
  A[AI capex] --> B[HBM absorbs<br/>3x wafer per GB]
  B --> C[Commodity DRAM<br/>and NAND restricted]
  C --> D[Consumer electronics<br/>prices spike]
  D --> E[PC -11.3%, phones -12.9%<br/>durable goods contract]
  A --> F[Data center load]
  F --> G[PJM capacity<br/>29 to 329 per MW-day]
  G --> H[Household bills up<br/>non-discretionary share rises]
  H --> E
  B --> I[Memory is 45-62%<br/>of accelerator BOM]
  I --> J[Residual value guarantee<br/>is mostly a memory bet]
  C --> K[Memory cycle turns]
  K --> J
  E --> L[The cyclical 20%<br/>the capex was masking]

Two arrows arrive at the cyclical 20% from the buildout, and the same supply restriction that drives them also sets up the collateral bet. One variable, three exits.

What this does to the method

cyclical-20-and-the-ai-capex-mask proposed a clean procedure: strip the masking line item, read the residual, compare to 2007 housing. The housing analogy is where the defect enters. Residential investment in 2005–06 was additive to the rest of the cyclical slice — the housing boom pulled appliance sales, furniture, truck demand and construction employment up with it. Subtracting housing gave you a conservative read on the underlying economy, because housing's spillovers were positive.

AI capex spills over negatively into its own slice. Subtracting it does not give you a counterfactual economy; it gives you an economy that has absorbed AI capex's input competition and cost pass-through without receiving its GDP contribution. That is strictly worse than the true counterfactual.

The correction runs in both directions and is worth stating precisely, because it is uncomfortable:

Which is to say: the mask thesis survives, the timing gets murkier, and the shape of the other side gets sharper. The clean tests in the original doc — equipment investment ex-information-processing, Class 8 orders, Cass freight — remain the right ones precisely because they are the least memory- and power-exposed lines in the cyclical 20%. Consumer electronics and durables ex-autos are now contaminated indicators and should be demoted.

On the clown car

The Gamers Nexus fourth point — executives and officials making AGI and singularity claims to hold stock momentum ahead of IPOs — is the ai-survival-theater-and-the-bubble thesis in different vocabulary, and it does not need re-litigating here. The narrow observation worth carrying is the timing correlation: hyperbolic capability claims cluster ahead of financing events, not ahead of product releases. That makes them a sentiment instrument rather than an information one, and it means they should be read as a funding-cycle indicator. Escalating rhetoric is evidence about the difficulty of the next raise, not about model capability.

Watchlist

Sources

Backlinks