Open scale-up fabrics run in two hyperscalers by 2031

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

Reference class: consortium interconnect standards progressing from roadmap to explicitly documented production at multiple hyperscalers within five years. I assign 0.80 to at least one qualifying disclosure, aided by the supplied Oracle/Helios order, but discount it because an order is not production and company statements may omit the 64-accelerator domain or protocol. Conditional on one, I assign 0.67 to a second: Microsoft and Meta provide plausible adoption paths, while Amazon’s proprietary stack and NVLink alternatives constrain breadth. Applying 0.97 for continued standards publication gives about 0.52. The strict requirement for two company-authored, production-specific disclosures makes 0.75 too high.

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

Reference class: new multi-vendor datacenter interconnects typically require 2–4 years from specification and switch sampling to disclosed production deployment. Helios makes Oracle the likeliest first qualifier, but an order or product offering is not necessarily a company-authored disclosure of a production 64-accelerator domain. I assign Oracle roughly 0.68 cumulative qualification probability, Microsoft 0.45, Meta 0.42, and Amazon 0.25 by 2031. Compounding those hazards suggests about 0.60 for at least two, reduced for correlated delays, continued NVLink/proprietary adoption, and strict disclosure wording. Consortium cessation is a small additional miss risk. Thus 0.75 appears optimistic.

enterprise-cio · openai/gpt-5.6-sol

Reference class: hyperscale networking-standard transitions, typically taking 3–6 years from specification to qualified production, with public architectural disclosure later and less complete. I assign roughly 0.8 to a viable UALink/UEC/OCP ecosystem by 2029, then 0.6 conditional probability that at least two named companies deploy and disclose every required fact by 2031. The supplied Helios order improves Oracle’s odds, but an order is not production evidence. A second company, 64-accelerator threshold, open protocol, and company-authored disclosure are conjunctive requirements. Correlated delays and proprietary or NVLink alternatives make 0.75 too high.

credit-analyst · openai/gpt-5.6-sol

I use the reference class of hyperscaler adoption and public documentation of new data-center fabrics during the first four years after switch silicon reaches GA. Conditional on a credible 2027 ecosystem, I assign cumulative qualifying-disclosure probabilities of roughly 0.65 for Oracle, 0.40 for Microsoft, 0.35 for Meta, and 0.25 for Amazon. Correlated ecosystem failure and the strict production, 64-accelerator, and company-authored requirements reduce the at-least-two result. Oracle/Helios is a strong first path, but an order is not a qualifying production disclosure. The forecaster’s 0.75 underweights conjunctive disclosure risk and continued proprietary or NVLink deployment.

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

I use the reference class of open interconnect standards moving from specification to disclosed hyperscaler production, typically a four-to-seven-year path with uneven adoption. Helios materially improves the chance of one qualifier, especially Oracle, but an order or vendor announcement is not the required company-authored statement that a production domain has at least 64 accelerators on the named fabric. A second qualifier is conjunctive: it requires product readiness, deployment, production status, sufficient domain size, and unusually specific disclosure. Proprietary-stack incentives and NVLink competition narrow the pool, while the absent switch-GA and deployment indicators provide no confirmation. The 2029 falsifier also removes late-start paths. My estimate is therefore below 0.75.

superforecaster-statistician · openai/gpt-5.6-sol

Base rate: 0.35 for a new accelerator-interconnect standard obtaining company-authored production disclosures from at least two of four major cloud operators within five years. The supplied Helios/Oracle ordering signal raises the outlook because a 72-GPU design could clear the technical threshold, but vendor-side evidence is not the required Oracle disclosure. No indicator verifies even one qualifier. Using roughly 0.25 annual hazard for a first qualifying disclosure and 0.16 for a second after ecosystem validation gives about 0.44 before discounting correlated ecosystem failure and disclosure timing. The two-company conjunction, 64-accelerator threshold, primary-source requirement, and exclusion of proprietary fabrics make 0.75 overconfident.