Rubin Ultra's 192GB Fallback: the Biggest HBM Buyer Is Rationing Itself — While Spot HBM Trades at 4-5x Contract
The clearest demand-side crack in the "permanently sold out" story is that the single largest HBM buyer is redesigning its flagship chip around scarcity. Nvidia is testing stripped-down memory configurations for Rubin Ultra after abandoning confidence in its original 12-Hi HBM4E design:
"Nvidia is testing stripped-down memory configurations for its next AI accelerator, Rubin Ultra, as a global shortage of high-bandwidth memory forces a rethink of plans headed into 20271... Starting in the third quarter of 2026, the company began testing three alternatives alongside that original design: 8-Hi HBM4E, 12-Hi HBM4, and 8-Hi HBM4." — Shattered.io, citing TrendForce (Aug 4, 2026) and The Information via Chosun Ilbo
The front-runner is the memory-cut option: "Research firm SemiAnalysis, cited by Chosun Ilbo on September 16, 2026, described this 8-layer HBM4/192GB setup as the primary configuration currently being shared with customers for evaluation." That is ~33% below Rubin's 288GB and roughly 81% below the ~1TB early design target. TrendForce attributes the reassessment to two overlapping constraints: DRAM supply tight "through 2027," and unresolved yield/validation on 12-Hi HBM4E. No final spec is locked — Nvidia has not commented — which keeps the existing Rubin Ultra spec watch open.
The same rationing shows up in price structure. A 36GB HBM3E module traded at ~$2,100 on the spot market (Sept 4, 2026, per IntuitionLabs) — four to five times the $300-400 long-term-agreement range. Spot at a 4-5x premium to contract means buyers who missed allocation are paying panic prices — historically a late-cycle signature (spot leads contract in both directions). The shortage is also propagating beyond the US ecosystem: Reuters (Sept 10) reports Huawei's Ascend 950DT quoted at >¥250,000 (~$37,255), up 20-50%, as Chinese AI-chip makers pass through HBM scarcity; and gaming GPUs (RTX 60 series, AMD RDNA 5) are reportedly pushed toward 2028 as data-center memory wins every wafer-allocation fight.
Micron's FQ4 call acknowledged the phenomenon obliquely — management fielded a question about "one large customer de-speccing HBM" and deflected to "latent demand," while its prepared remarks concede server content growth is running at "a modestly lower rate... than prior expectations, amid tight memory supply" (prepared remarks).
What it means: this is what a shortage peak looks like from the demand side — not falling prices, but the largest buyers redesigning products to need less of the scarce thing2. It rhymes with the algorithmic threat in DeepSeek Slashes KV Cache Memory Demand by 75%, Signaling Algorithmic Threat to HBM Scarcity (software rationing KV-cache memory) — together they mark the two demand-elasticity channels (configuration and algorithm) that eventually erode "demand is insatiable" narratives. Note the divergence to flag: suppliers are simultaneously asking 3x+ for 2027 HBM4 (see 2027 HBM4 Settlement Round Hardens: Samsung Asks 3x+, TrendForce Sees +121% HBM ASP, Micron Claims Locked — Bernstein Trims the Hike Anyway) while their biggest customer engineers a 33% memory cut into its next flagship. Both can be true in a shortage; the second one is what ends it.
-
An instance of Silicon roadmaps must bend to the physical scarcity of high-bandwidth memory. — The largest HBM buyer is downgrading its flagship roadmap's memory spec — a 192GB fallback roughly 33% below Rubin — to survive a persistent shortage, which is the silicon-bends-to-scarcity law in action. ↩︎
-
An instance of Memory cycles turn on deceleration, never on falling prices. — Nvidia engineering a roughly 33% memory cut into its flagship is demand-side elasticity rationing the shortage before any price decline prints. ↩︎