A 10-zettaFLOP/s cluster is operational by 2028

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

The reference class is annual growth in Epoch’s leading operational cluster. The supplied series rose roughly 3x annually from 2020–2025, but reaching 1e22 requires another 36.6x from July 2025 by year-end 2028. That is near the volatile historical rate but about 2–3x above Epoch’s stated trend for the date. Announced multi-gigawatt capacity is weak evidence without completed power delivery, accelerator procurement, commissioning, and Epoch classification as one operational site. Schedule slippage and the narrow 2029 verification window further reduce the probability. The GBM prior overweights noisy historical jumps and omits these execution and resolution constraints.

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

My reference class is Epoch’s 2020–25 frontier series and phased hyperscale accelerator buildouts. From 2.73e20 in July 2025, the threshold requires 36.6x, roughly an eight-month doubling, by end-2028. The historical series is lumpy enough to permit that, and NVIDIA revenue growth supports accelerator availability. However, Epoch’s extrapolation implies only about 5e21 by end-2028, requiring a roughly 2x positive deviation. A multi-million-accelerator single site must also clear HBM, packaging, networking, power and commissioning constraints; large projects commonly enter service in phases and slip by quarters. Epoch must then classify and publish it by March 2029. These conjunctive risks put me below 0.50 and the GBM prior.

enterprise-cio · openai/gpt-5.6-sol

Using historical frontier-cluster scaling and Epoch’s extrapolation as the reference class, the July 2025 maximum must rise 36.6-fold in 3.4 years—about 2.84x annually. Epoch’s stated path implies only about 5e21 by end-2028, so qualification requires roughly a twofold positive deviation plus timely commissioning. Multi-million-accelerator procurement, power, networking, and acceptance create substantial schedule risk; announcements are weaker evidence than an operational listing. I estimate a 0.36 cumulative crossing hazard through 2028, then discount for the conjunctive single-site, operational-date, Epoch-classification, and dataset-continuity requirements, yielding 0.32. The structural prior likely overweights unconstrained upside.

credit-analyst · openai/gpt-5.6-sol

Reference class: historical frontier-cluster scaling and first-of-kind multi-gigawatt compute-project delivery. The latest supplied maximum, 2.73e20 FLOP/s, requires a 36.6-fold increase, while Epoch’s extrapolation implies only about 5e21 by year-end 2028. I use roughly a 0.10 crossing hazard through 2027 and 0.25 during 2028; compounding and discounting for commissioning delays, single-site classification, dataset lag, and discontinuation risk yields 0.30. Strong accelerator revenue supports supply but not timely power, networking, or operational status. The GBM prior likely overweights volatile historical jumps and ignores the resolution’s classification constraints.

geopolitics-policy-analyst · openai/gpt-5.6-sol

Using frontier-cluster scaling in 2020-2025 as the reference class, the observed maximum must rise 36.6-fold from 2.7288e20 in July 2025, requiring about 2.85-fold annual compounding through 2028. That rate is historically achievable, and multi-gigawatt projects plus newer accelerators provide upside. However, Epoch’s central extrapolation reaches only about 3e21 by mid-2028, requiring a sharp late acceleration. Operational commissioning, single-site classification, a qualifying date, and inclusion by March 2029 are conjunctive requirements. I discount the 0.719 structural prior because GBM extrapolates a volatile boom-era series without construction, power, or publication delays.

superforecaster-statistician · openai/gpt-5.6-sol

Base rate: for frontier-compute thresholds whose published trend crossing is 9–18 months after a deadline, I use 0.30 for early arrival. Epoch’s supplied extrapolation reaches only about 3e21 FLOP/s by mid-2028 and 8e21 by mid-2029, so 1e22 by year-end 2028 requires material acceleration. The last observed maximum, 2.73e20 in July 2025, must rise 36.6-fold in 41 months—historically possible but dependent on lumpy deployments. The 0.719 GBM prior overweights volatility and smooths jumps. Announcements are weaker evidence because operationalization, single-site classification, 16-bit estimation, and timely Epoch listing are all required. Dataset discontinuation or column removal adds a small failure tail.