Memory-demand durability — does algorithmic efficiency cap the memory super-cycle?

memory-demand-durability · conviction low · status open · horizon 2026-2028 (adoption is multi-quarter) · as of 2026-08-03

The memory re-rating (SK Hynix +257% rev, HBM/DRAM 10-30x) prices long-context AI needing linearly-growing KV cache. Kimi K3 proved hybrid-linear (constant-state) attention works at frontier scale with near-full recall and ~73% memory saving — the net-new fact that makes the load-bearing premise contestable. This thesis holds the DEBATE, not a naked short: the variant is that the market underweights algorithmic efficiency; the falsifiable gate is frontier-lab ADOPTION, and Jevon's + deployment inefficiency are real offsets.

Rests on filed figures, not on modelled shares. 23 premises (11 field, 10 entity, 1 signal, 1 edge); no derived cell is involved, so the undisclosed supply weights that put a range on other pages in this bank cannot move this one.

Exhibits

Exhibit 1Relative performance, indexed to 100How the names in this thesis have traded against SOXX.

Series available as data/memory-demand-durability.csv

Exhibit 2What the conviction is actually made ofEach premise and the number it composes to. A conjunction of plausible premises is far weaker than any of them.

Memory is taking share of the bill of materials, which is cost-side evidence the demand thesis previously lacked

52% if the 2 gates are independent, 65% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: Ai Capex Cycle — signal 2026-08-04 at 0.65 — Wells Fargo NVL72 BOM: memory (HBM+LPDDR) rises from 26.6% to 30.8% of rack cost between GB300 and Vera Rubin, growing +128% against compute's +116%.

The long-context leg of the memory thesis is architecture-dependent, not time-dependent

14% if the 3 gates are independent, 30% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: KV cache bytes per token, by attention architecture at 0.30 — Computed from published configs: GQA at 252 KB/token against MLA at 34 KB/token, a 7.3x spread. Held at 0.80 because the arithmetic is trivially check

Hybrid-linear (constant-state) attention breaks the linear-KV-cache premise the memory super-cycle rests on

† 3 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

One gating premise, so the conclusion is exactly as strong as it. The figure is an ordering device, not a calibrated probability — see how the numbers are made.

Weakest link: Hybrid Linear Attention (constant-state KV) at 0.85 — Hybrid-linear attention exists and is implemented, not a paper proposal. The doubt is adoption breadth, scored separately.

Kimi K3 is the net-new fact: hybrid-linear now works at frontier scale, so industry adoption is credible, not hypothetical

† 1 premise marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

85% if the 2 gates are independent, 85% if they move together. They are claims about one industry, so the truth is between and nobody can say where. Treat this as an ordering device rather than a calibrated probability — the ranking of premises is the information, not the level.

Weakest link: Kimi K3 (Moonshot) at 0.85 — Kimi K3 ships and is documented. Existence is not the doubt.

Therefore memory-demand durability is a scored bet on ADOPTION, not a settled given — and the tradable claim is measurable

† 11 premises marked supporting — shown and arguable, but the conclusion does not depend on them, so they are not multiplied into the composed figure. Citing a filed figure should not cost conviction.

No gating premise. Every premise here is supporting evidence, so this conclusion states no necessary condition — it is asserted from cited data rather than derived from a claim that could fail. Read it as a summary, not as a falsifiable call.

Weakest link: Frontier hybrid-linear (constant-state) adoption — value single-lab at 0.60 — SINGLE-LAB is the load-bearing premise and the weakest. One lab shipping constant-state attention at frontier quality is proof of possibility, not of

The variant

Consensus

Long-context/agentic AI drives structurally rising memory demand; the KV-cache-per-token requirement is a given, so HBM/DRAM shortage persists — validated by SK Hynix Q2 records, $33B 2026 capex, HBM4 mass production, and KLA raising 2026 WFE to ~$150B. Any efficiency gain is eaten by Jevon's paradox.

Variant

The consensus overweights a premise that just became contestable. Hybrid-linear attention (KDA) removes the linear-KV mechanism at ~73% memory saving, and Kimi K3 proved it frontier-viable AND open-sourced it — so adoption is cheap and likely (as MLA diffused post-DeepSeek-V3). Jevon's is NOT guaranteed to outrun the saving, because denser scaling laws mean smaller models per unit intelligence. So memory-demand durability is lower than a +257% tape implies — a small demand-growth deceleration could trigger game-theoretic profit-taking in a crowded trade.

Differentiator

The desk does not take a naked short (the bull spine has overwhelming current confirmation). It SEPARATES the scoreable from the narrative: adoption count and HBM-guidance trajectory are tracked, falsifiable measurements; the 'game-theoretic crash' mechanism is flagged unfalsifiable-in-advance and carries no weight. The edge is holding a scored, updating probability on a premise the momentum crowd treats as settled.

Falsifiers

Hybrid-linear attention (KDA-class) reaches >20% of served frontier-model tokens by end-2027 WITHOUT HBM bit demand growth decelerating below trend. That would show the memory saving is real but Jevons absorbed it, which is the consensus leg this thesis contests.
HBM bit shipments grow <10% year-on-year in any 2027 quarter while accelerator units still grow >20%. That is the decoupling the thesis predicts; its absence through 2027 falsifies the mechanism.
No frontier lab ships a production model using linear/hybrid attention at scale by end-2027 (model cards, published architectures). The thesis rests on cheap diffusion of an open-sourced technique; two years without production adoption refutes 'adoption is cheap and likely'.
Memory contract pricing holds or rises through 2027 alongside falling memory intensity per token — pricing power would have survived the architectural change, which is the outcome the thesis says is contestable.
SCOPE: 000660.KS/MU/005930.KS outperform the semis benchmark over the horizon while memory-per-token intensity falls >30% — the architectural read was right and produced no tradeable consequence.

Open questions

TWO PACKETS, OPPOSITE SIGNS, AND THEY ARE NOT ACTUALLY IN CONFLICT. The KDA packet argues architecture cuts memory BIT VOLUME ~73%. The Huawei tau-scaling packet argues logic-memory re-fusion (HBM, hybrid bonding, 3D-stacked SRAM) raises memory vendors' STRUCTURAL POWER and the value captured per bit. Both can hold simultaneously: fewer, more tightly integrated, higher-value memory units. This thesis currently models ONE blended demand series, which cannot represent that and will produce incoherent forecasts if both effects are real. The desk already owns the right distinction — it is the same volume-versus-rent split the ABF and neocloud theses established, where proximity to scarcity turned out not to be ownership of rent. PROPOSED, NOT APPLIED: split into bits demanded per unit of intelligence delivered, and value captured per bit.
The Huawei paper asserts >70% of AI system cost sits in storage and >80% of cluster energy goes to data movement. Both are unverified single-source figures, but if either is close, architectural cuts to per-token memory intensity have far more cost leverage than a bit-volume view implies — which strengthens this thesis's variant rather than the consensus. Neither figure is used as a premise anywhere; obtaining an auditable version is a research task.

Reasoning chain

Memory is taking share of the bill of materials, which is cost-side evidence the demand thesis previously lacked VALID

premises

Ai Capex Cycle — signal 2026-08-040.65 moderate
26.6%30.8%128%116%
Wells Fargo NVL72 BOM: memory (HBM+LPDDR) rises from 26.6% to 30.8% of rack cost between GB300 and Vera Rubin, growing +128% against compute's +116%.
why
Memory is the fastest-growing major line except storage.
HBM40.80 strong
HBM4 is the generation carrying that increase, and it is in the model.

Every prior leg of this thesis argued from DEMAND — tokens, context length, cache pressure. This is the first evidence from the COST side, and it is independent: a bill of materials does not care why memory is needed. Memory growing faster than compute inside the same package means the accelerator vendor is choosing to spend marginal package cost on memory rather than on logic, which is a revealed preference from the party with the best information about the bottleneck. HELD AT 0.65 because it is a sell-side BOM estimate rather than a disclosed teardown, and this desk refused a sell-side memory TAM this month for failing an arithmetic check. AMD points the same way from a different direction: MI455X ships 432GB HBM4 per GPU against Vera Rubin's 288GB, a 1.5x capacity lead chosen as a competitive axis.

The long-context leg of the memory thesis is architecture-dependent, not time-dependent VALID

premises

AI total addressable market (revenue, forward)0.60 moderate
THE DENOMINATOR CHECK
70%1.18 tn1.3 tn100%1.69 tn
Memory is an INPUT to AI compute, so a memory TAM has to fit inside the AI TAM it serves with room left for everything else.
why
AI-attributable memory is ~70% of DRAM, so ~$1.18tn in 2028 against an interpolated ~$1.1-1.3tn AI TAM that year — roughly 90-100% of the revenue pool, which is not survivable for an input. Scope is what makes or breaks this check: the $1.69tn DRAM+NAND total covers phones, PCs and autos and cannot be set against an AI-only denominator. Held at 0.60 because the AI TAM itself is a supplied buy-side build with no public link, so this bounds the memory forecast only as well as the denominator is known.
KV cache bytes per token, by attention architecture0.30 weak
252 KB34 KB7.3x
Computed from published configs: GQA at 252 KB/token against MLA at 34 KB/token, a 7.3x spread.
why
Held at 0.80 because the arithmetic is trivially checkable but production serving quantises, evicts and shares prefixes, so the figure is an upper bound.
Long-context inference (KV-cache growth)0.75 strong
Context lengths rising, held in the model.

LARGELY REFUTED. Compression does not stop context growth from driving HBM demand, because the workload grew faster than the compression did. Kimi K3 ships the most aggressive compression available — Kimi Delta Attention linear layers plus MLA at a 3:1 ratio — and still exhausts a B300 node at EIGHT concurrent users: 3.25M tokens of cache budget against a workload of 142k input tokens per turn over a median 65 turns, with hit rate collapsing from a theoretical 95% to under 10% past that point (SemiAnalysis, free portion, 2026-08). THE ARITHMETIC: compression bought 7.3x; the agentic workload took ~71x on context per turn versus a chat turn. Compression lost by an order of magnitude. The general form of the error is holding two terms and multiplying neither — tokens-per-agentic-task and kv-bytes-per-token were both in the model and nothing joined them, the same way tokens and megawatts sat unconnected. THE SURVIVING CLAIM IS NARROW: the leg is contingent on attention architecture rather than on the calendar, and at 1M context a single GQA request wants more HBM than an eight-accelerator node holds while the MLA equivalent fits. WHAT WOULD SETTLE IT is the SHARE OF SERVED TOKENS by attention scheme, which nobody publishes and which is estimable from open-weight download counts and provider model catalogues.

Hybrid-linear (constant-state) attention breaks the linear-KV-cache premise the memory super-cycle rests on VALID

premises

Hybrid Linear Attention (constant-state KV)0.85 strong
Hybrid-linear attention exists and is implemented, not a paper proposal.
why
The doubt is adoption breadth, scored separately.
Hybrid Linear Attention (constant-state KV) — kv cache scaling constant0.80 strong
Constant-state scaling is the architectural property that breaks the linear-KV assumption the memory super-cycle rests on.
why
Held at 0.80: constant in the attention layer does not mean constant end-to-end once weights and activations are counted.
Hybrid Linear Attention (constant-state KV) — memory saving vs full attention 0.730.70 moderate
73%
A 73% saving against full attention is a large number resting on a single implementation's reported figure rather than on a measured serving comparison.
Memory (KV/state) per token — value linear-with-context0.75 strong
The desk's own held coefficient is linear-with-context, which is exactly what constant-state attention contradicts — so this premise is what makes the conclusion a live disagreement.

The bull spine assumes memory/token grows with context (linear KV cache). A constant-size recurrent state removes that growth at ~73% saving, so the mechanism the demand curve is priced on is not a law — it is an architecture choice that can change.

Kimi K3 is the net-new fact: hybrid-linear now works at frontier scale, so industry adoption is credible, not hypothetical VALID

premises

Kimi K3 (Moonshot)0.85 strong
Kimi K3 ships and is documented.
why
Existence is not the doubt.
Kimi K3 (Moonshot) — frontier proof confirmed0.75 strong
Frontier-quality at constant-state attention is the net-new fact.
why
Held at 0.75 because 'frontier' is a comparative judgement against a moving benchmark, not a measured threshold.
Hybrid Linear Attention (constant-state KV) supplies Kimi K3 (Moonshot)1.00 strong
The architecture-to-model link is documented in the model's own technical report.

Hybrid-linear was dismissed as sub-frontier until K3 delivered near-full recall + 2.5x train efficiency at 1T+ scale and open-sourced it. Cheap-to-copy frontier proof is exactly how MLA diffused after DeepSeek V3 — so the demand-dampening is a live adoption vector, which is what makes the premise contestable NOW rather than someday.

Therefore memory-demand durability is a scored bet on ADOPTION, not a settled given — and the tradable claim is measurable VALID

premises

Memory Super-Cycle0.80 strong
CONTEXT, not a necessary condition.
why
Names the theme this conclusion argues with; the conclusion does not fail if the theme is renamed.
SK Hynix0.85 strong
CONTEXT.
why
The name most exposed to the outcome, with primary-sourced financials.
Hybrid Linear Attention (constant-state KV)0.80 strong
CONTEXT here — the architecture is scored on its own merits in the conclusion above; this cites it as the subject of the adoption bet rather than re-arguing it.
Frontier hybrid-linear (constant-state) adoption — value single-lab0.60 moderate
SINGLE-LAB is the load-bearing premise and the weakest.
why
One lab shipping constant-state attention at frontier quality is proof of possibility, not of adoption — and the thesis is a bet on adoption. It is why this conclusion is scored low rather than refuted.
HBM supply tightness — value severe0.80 strong
CONTEXT.
why
Tightness is what makes the question worth asking; it is not a condition of the adoption claim.
China HBM market entry — value in-development0.65 moderate
CONTEXT.
why
A second supply source in development bears on durability but does not gate the adoption bet, and 'in-development' is a status with no date attached.
Incentive × Capacity — the indigenization / margin-compression generator0.70 moderate
CONTEXT.
why
The prior that a firm with both incentive and capacity to enter eventually does.
Tokens consumed per agentic task — value rising-fast0.80 strong
7.3x71x
Rising fast is the term that beat compression: 7.3x compression against ~71x context growth per agentic turn.
why
This premise is why compression losing is the base case.
Aggregate inference token demand — value rising-fast0.85 strong
Token demand growth is the least-contested input in the model.
Inference cost per token — value falling-fast0.75 strong
CONTEXT.
why
Falling unit cost is the mechanism by which demand keeps rising; it does not gate whether hybrid-linear attention gets adopted.
Semis/tech sector valuation (FMP sector P/E) — value elevated0.70 moderate
CONTEXT.
why
Elevated valuation sets what a re-rating costs if the thesis is wrong. It is a risk statement, not a condition.

The variant resolves on falsifiable measurements (frontier-lab hybrid-linear adoption count; HBM demand-guidance trajectory; memory-per-token trend) that the momentum crowd is not tracking. The desk holds a probability on those, fades the 'premise is settled' consensus, and explicitly excludes the unfalsifiable game-theoretic-crash mechanism from the weighted case.

Sources

Kimi K3: 69 KDA + 24 full-attention layers, constant per-head state, ~73% KV/state saving vs all-MLA at 128K-1M context, near-full recall, 2.5x training efficiency vs K2; architecture + FlashKDA kernel open-sourced — link acc 2026-07-29
Bull-spine confirmation: SK Hynix Q2 FY26 record rev +257% YoY, 76% op margin, HBM4 mass production, ~$33B 2026 capex; KLA hikes 2026 WFE to low-$150B (from $140B+), advanced packaging +70% Y/Y — link acc 2026-07-29

Write-up

Pre-filled skeleton: memory-demand-durability.md