Analyst: AI Storage Bills May Account for 60% of Cloud Capex Next Year; Ten-Year Demand Visibility Is a Lie
nashnova research
Anonymous analyst "Mr. P" estimates storage alone could account for 50%–60% of hyperscalers' projected $1.1–1.2 trillion capex next year; he calls their claim of "ten-year demand visibility" a flat-out lie, arguing the real bottleneck has shifted from GPUs to storage, power, and ABF substrates.
Why is "ten-year demand visibility" an empty claim?
"Mr. P," founder of P Equity Research, said in a September 26 podcast: hyperscalers cannot predict conditions two years out, let alone ten. The "ten-year visibility" line is spin.
This means → the cyclical nature of AI hardware spending has not disappeared. Spending will slow eventually; storage's boom-and-bust pattern does not end just because AI arrived.
In plain terms = cloud giants are spending at a furious pace today, but no one can guarantee that pace holds for a decade — every prior hardware super-cycle hit a wall.
How large is the storage bill?
Mr. P estimates hyperscaler capex next year at $1.1–1.2 trillion. Storage alone — HBM, DRAM, and NAND combined — could consume 50%–60%, roughly $500–700 billion.
UBS's estimate is more aggressive: $900 billion on storage next year, exceeding the total capex of all hyperscalers this year.
This means → storage is shifting from a supporting line item to the single largest capex category. Whoever controls storage capacity controls the tempo of AI expansion.
Can long-term contracts lock in supply?
To smooth cyclical swings, memory makers are pushing long-term supply agreements (LTAs) — contracts that lock in volumes and pricing years ahead. Samsung has begun receiving contract requests stretching up to ten years.
But an industry expert who previously worked at both Samsung and AMD warned Mr. P: LTAs may be overhyped — "if demand collapses, these contracts can be quietly cancelled behind closed doors with no penalty."
This reflects a deeper tension: a ten-year contract on paper does not equal ten years of certainty. When a downturn arrives, contractual force may prove far weaker than the market assumes.
Where is the real bottleneck — not GPU dies?
Used H100s still trade at roughly $25,000. B-series GPU hourly rental rates have risen from $5 to $7–8. On the surface, compute remains scarce.
Yet an AMD expert confirmed to Mr. P: raw chip supply is actually adequate. The real chokepoints are data-center power, advanced packaging, and storage.
Nvidia's next-gen Vera Rubin bill of materials shows triple-digit price increases in ABF substrates — a high-end carrier material that holds the chip — along with PCBs and memory. Mr. P expects the ABF shortage to persist into 2028 or even beyond 2030.
On the energy side, gas-turbine backlogs at Mitsubishi, Siemens, and GE Vernova stretch past 2030 — order today, earliest delivery 2030. In plain terms = the constraint is not making chips; it is finding enough electricity and substrates to house and feed them.
Copper vs. optics: which comes first?
Mr. P believes optical interconnects win long-term — "nothing travels faster than light."
But co-packaged optics (CPO) — embedding optical modules directly alongside the chip — still faces yield and thermal issues, and the cost is high. With the storage bill already consuming budgets, hyperscalers must cut spending elsewhere.
He expects copper and optics to coexist through 2027. Near-package optics (NPO) may appear first around 2027; CPO capacity ramp could start in 2028–2029 but at low penetration. CPO dominance likely waits until after 2030.
What are the two checkpoints that test the AI capex narrative?
Checkpoint one: when the storage cycle peaks — whether the current spending growth rate is sustainable will directly test the credibility of "ten-year visibility."
Checkpoint two: whether the ABF substrate bottleneck eases before 2028 — if it does not, compute expansion will be capped by substrates, not chips.
This means → rather than tracking short-term GPU shipment swings, investors should watch these two timelines: the storage-cycle inflection point and the substrate-capacity release schedule. They are the hard constraints that determine whether the AI capex narrative holds together.
市场有风险，内容仅供研究参考，不构成投资建议。