Article — "Huang Calls Memory Shortage Structural as Nvidia Trims HBM on Rubin Ultra to Cope": At GTC earlier this year, Nvidia held up a Rubin Ultra compute tray glittering with 1TB of HBM4E across sixteen stacks. On a test bench today, engineers are reportedly validating the same flagship at 192GB—reportedly on last-generation HBM4.
That is, by the article's own reckoning, roughly a five-fold retreat by the article's own reckoning, and it is reportedly not a rumor about pricing but hardware under testing.
Last week in this column, we covered Huang reportedly describing the HBM shortage as "structural"—a multi-year condition, not a spike. This week the framing acquires a body: the shortage is now shipping as reduced-spec hardware, and doing so inside the company's most advanced product.
According to The Information, Nvidia has tested at least three lower-memory Rubin Ultra variants—some at 192GB, some at 256GB, several using fewer than the announced sixteen stacks. The most telling detail is the generation: HBM4, not the HBM4E Huang promised on stage.
This matters because HBM4E was the point. Its customizable base logic die—Micron reportedly building it at TSMC, tuned per customer—was, in this article's reading, reportedly the feature that justified the "E." Stepping back to HBM4 is not a smaller number on a spec sheet. It is abandoning a capability.
The cause is blunt. Samsung, SK hynix, and Micron have sold through HBM capacity into 2027, and HBM4E's complexity has strained even that. SK hynix's CEO has reportedly called 2027 the "worst year" for the shortage, with constraints running toward 2030. Nvidia may not have chosen to ship less memory; per The Information's reporting, it is testing lower-memory designs in part because it may not be able to secure enough advanced memory chips.
Note the direction of the humiliation. The 192GB and 256GB figures reportedly fall below the 288GB that the base Rubin GPU—already in mass production—carries today. The Ultra, the halo part, is being tested with less memory than the tier beneath it.
Some of that may be explained by the reported cancellation of the quad-die Ultra in favor of a dual-die design, which Nvidia has yet to confirm; fewer compute chiplets need less memory to feed them. But even granting the dual-die math, the capacities land lower than anyone modeling the roadmap expected.
Here is where the story turns from supply to demand.
A GPU with less memory does not reduce the memory a workload requires. A frontier model needs the parameters and the KV cache it needs; that number is a property of the model, not the chip. If the chip can no longer hold it, the workload must be split.
Splitting means more GPUs. The Information's reporting says it plainly: while lower memory could affect performance, "AI companies running large models on the reduced-memory versions may need to deploy more GPUs than they otherwise would."
Read that as a conservation law. Memory scarcity inside the package does not vanish—it is displaced outward, reappearing as GPU count, then rack count, then power draw, then floor space in a datacenter that already can't get enough of any of those.
Memory scarcity inside the package does not vanish—it is displaced outward, reappearing as GPU count, then rack count, then power draw, then floor space.
The customer absorbs the shortage Nvidia could not. What could not be bought as HBM gets bought as silicon, interconnect, and megawatts.
And this is where "structural" quietly changes jobs. As a description of supply, it explains why the tray shrank. As a mechanism, it manufactures demand: a memory-starved flagship forces buyers to purchase more units to run the same model, and every extra unit carries Nvidia's full margin stack—compute die, networking, the platform.
The token-throughput roadmap Nvidia sells—the promise that each generation delivers more inference per dollar, culminating in a projection of 40x token demand by 2030—was pitched on denser, more capable chips. A five-fold memory cut on the flagship inflates the unit count needed to hit those throughput targets. The roadmap the compromise threatens is the same roadmap the compromise now props up, one incremental GPU at a time.
None of this requires bad faith. It requires only that the constraint be real and that Nvidia sit at the single chokepoint where it converts into orders.
Which brings us to the quietest line in the reporting, and the most revealing.
One Nvidia customer told The Information that per-GPU memory "isn't a top concern," valuing the relationship with Nvidia over the long term.
Sit with that. The defining specification of an accelerator—how much model you can hold per chip—is being waved off by the people who write the checks. Not because it doesn't affect their cost, but because the alternative to accepting Nvidia's terms is not being in line at all.
That sentence tells you the transaction has changed underneath the product. The buyer is no longer purchasing a spec. The buyer is purchasing an allocation—a place in a queue for scarce silicon—and treating the memory configuration as whatever Nvidia decides to ship.
When a customer stops evaluating the flagship and starts protecting its relationship with the vendor, the flagship has stopped being a competitive artifact and become an entitlement. Specs are for markets where you can walk. In an allocation, you take the tray you're handed.
Huang's own public posture fits the frame. Asked about the cuts in Korea, he reportedly said Nvidia would keep using HBM heavily but had to be more "wise" and "clever" with it while fighting for more supply—and, separately, that scarcity everywhere is reportedly "fantastic for us." Both statements are true at once. That is precisely the problem.
There is a defensible engineering reading here, and it deserves airing: a dual-die part with less memory can be manufactured in higher volume, so Nvidia produces more GPUs from the same constrained HBM pool and the crunch eases. Retail traders reading the same news landed there, guessing Micron benefits either way.
But "more GPUs from the same memory" and "customers need more GPUs per model" are not competing stories. They are two ends of the same pipe. Nvidia makes more units because each unit does less, and customers buy more units because each unit does less. The shortage is preserved and monetized in a single motion.
The tell is that the retreat happened to the flagship. Cutting memory on a mid-tier part is triage. Cutting it on the Ultra—the product whose entire reason to exist is to hold the biggest models on the fewest chips—inverts the value proposition and dares the market to object.
The market, per that one customer, is not objecting.
So the image to keep is not the 1TB tray from the GTC stage. It is the 192GB sample on the bench, and the buyer who shrugged at it.
A year of "structural" language has produced a concrete object: a flagship that holds less, sold to customers who will need more of it, in a market where asking about the spec has been replaced by asking whether you're on the list. The shortage did not get solved. It got shipped—and, quietly, sold twice.