Skip to content

AI Market · Forecast desk · Live · hardware

40%

NVIDIA publicly confirms a Rubin Ultra memory configuration at or below 512GB, or an 8-high HBM4E stack option, in official specifications or an earnings disclosure by March 31, 2027.

Probability history

0%25%50%75%100%08-1509-06deadline

Called at 37% on August 15, 2026. Repriced 1 time; now 40%. 206 days to deadline.

How the number was built

Final method: stacked-fixed-weights (resolved n=57 < 80). Each rung is an independent estimate; the final is a stacked combination.

Stated (issue date)37%
Reference class (n=29)50%
Recalibrated52%
Feature model45%
Council32%
Final40%

Drivers

  • Reference class qtype:disclosure runs 0.50 (n=29) vs stated 0.37.
  • Hierarchical calibration moves stated 0.37 to 0.52 (lens hardware, kappa_stated=24.0, kappa_lens=200.0).
  • Feature model at 0.45: conjuncts=1, named disclosure=True, horizon=228d.
  • Council of 5 at 0.32.

Cautions

  • all 5 votes came from one vendor (openai); extremization disabled, persona diversity only

Council

5 personas on one model (openai/gpt-5.6-sol); other vendors were unavailable this run, so read this as one model's persona-averaged view, aggregated by log odds mean without extremization.

31%

grid and power economist · openai/gpt-5.6-sol

The deadline likely captures GTC 2027, creating a plausible venue for Rubin Ultra specifications, but detailed memory disclosure could still await a later launch datasheet. The ≤512GB outcome also faces hardware risk: Ultra-class designs may aggregate multiple GPU dies and HBM stacks into a package whose advertised memory exceeds 512GB even if a per-die figure does not. Under the ambiguity rule, unclear die-versus-package labeling would not qualify. An explicit walk-back is a weak alternative because NVIDIA commonly replaces roadmap specifications without characterizing the change as a retraction. The supplied public-series data shows strong NVIDIA demand but does not resolve the memory architecture or disclosure timing.

34%

semiconductor supply chain analyst · openai/gpt-5.6-sol

Rubin Ultra’s previously indicated higher capacity is architecturally plausible for a multi-die, many-stack HBM4E package, so a reduction to 512GB or less is not the base case. HBM4E 16-high yield, thermals, CoWoS-L integration, and package availability create meaningful downgrade risk, but NVIDIA more often preserves headline specifications and shifts timing or volume than explicitly walks back a roadmap figure. GTC 2027 provides a natural disclosure window, yet NVIDIA could discuss system performance without publishing a qualifying memory configuration. The strict source requirement and silence-as-miss rule lower the probability below the hardware-led reference rate.

32%

capital markets credit analyst · openai/gpt-5.6-sol

GTC in March is a natural venue for updated Rubin Ultra specifications, so there is a meaningful chance NVIDIA addresses memory before the deadline. However, reaching ≤512GB would be a substantial reduction from the previously discussed higher configuration. NVIDIA is more likely to preserve the roadmap target, revise other specifications, or defer exact capacity than explicitly acknowledge a large downgrade. The strict source requirement and silence-as-miss rule further reduce the probability. Hardware forecasts resolve relatively well, but the longer-horizon disclosure reference class is weaker.

31%

superforecaster statistician · openai/gpt-5.6-sol

Base rate: 0.40 for >90-day questions, cross-checked against 0.50 for disclosures and 0.615 for hardware. I lower it because resolution requires a specific ≤512GB Rubin Ultra statement or explicit walk-back in NVIDIA GTC, IR, or a datasheet by a hard date; leaks, partner materials, vague platform-memory figures, and silence all fail. A likely GTC window before the deadline helps, but the digest provides no Rubin-specific evidence. The forecaster’s low-confidence calibration has only one observation, so it warrants no adjustment; broader results show overconfidence rather than grounds to raise the estimate.

32%

geopolitics regulation analyst · openai/gpt-5.6-sol

A March GTC creates a plausible disclosure window before the deadline, but Rubin Ultra is positioned as a high-memory 2027 product, making a configuration above 512GB more likely than one at or below it. NVIDIA also rarely issues explicit walk-backs; it more often replaces roadmap slides without characterizing the change. Even if capacity has been reduced, architecture-level GTC material may omit final memory specifications, while a definitive datasheet could arrive after March. The strict source requirement and silence-as-miss rule push this below the hardware and disclosure base rates.

Evidence the engine used

direct · undated

Best of NVIDIA GTC: AI Breakthroughs and Recommended Sessions | NVIDIA GTC

Cited by grid-and-power-economist: The deadline likely captures GTC 2027, creating a plausible venue for Rubin Ultra specifications, but detailed memory disclosure could still await a later launch datasheet. The ≤512GB outcome also faces hardware risk: Ultra-class designs may aggregate multiple GPU

direct · undated

NVIDIA Corporation - Home

Cited by grid-and-power-economist: The deadline likely captures GTC 2027, creating a plausible venue for Rubin Ultra specifications, but detailed memory disclosure could still await a later launch datasheet. The ≤512GB outcome also faces hardware risk: Ultra-class designs may aggregate multiple GPU