Ten open forecasts
Each one is a statement that will turn out true or false, shown with the model’s latest recorded odds.
Google open-sources the native TorchTPU backend used by vLLM or SGLang.85%
still open · check by
last looked
first recorded · live
as written in the record
Google has publicly open-sourced the native TorchTPU backend used by vLLM or SGLang
Cerebras Systems goes public, with its shares trading on a public exchange.80%
still open · check by
last looked
first recorded · before the live record began
as written in the record
Cerebras Systems completes an initial public offering (its shares begin trading on a public exchange), per SemiAnalysis / credible reporting
AMD ships an MI450X-class rack that links at least 64 GPUs as one machine.80%
still open · check by
last looked
first recorded · before the live record began
The odds were revised: 60% on → 72% on → 80% on .
as written in the record
AMD has launched / shipped a rack-scale AI system (MI450X-class, CDNA-Next) with a single scale-up domain of ≥64 GPUs
A top AI lab or cloud giant trains, not just runs, a huge model on GB200 NVL72 at production scale.72%checked 1 time · evidence not yet decisive · next check 26 Oct 2026
unresolved · evidence not yet decisive
Condition not established: "Mega-scale model training on GB200 NVL72 at production scale" is measured at the cluster: a training run, pre- or post-training, for a frontier-model programme on a GB200 NVL72 cluster of at least 10,000 GPUs. Per-job GPU counts are rarely reported; cluster scale is.
last looked
first recorded · before the live record began
as written in the record
At least one frontier lab or hyperscaler is running mega-scale model TRAINING (not just inference) on GB200 NVL72 at production scale, per SemiAnalysis / credible reporting
Nvidia ships the Groq-IP LPX (LP30) rack, built for the decode side of inference.70%
still open · check by
first recorded · before the live record began
as written in the record
Nvidia has shipped / made generally available the Groq-IP LPX (LP30) decode-specialized inference rack, per SemiAnalysis / credible reporting
Cerebras CS-4 hits at least 3,500 output tokens a second per user on GPT-5.6 Sol in a public test.65%
still open · check by
last looked
first recorded · live
as written in the record
Cerebras CS-4 demonstrates ≥3,500 output tokens/sec/user on GPT-5.6 Sol in a public Cerebras, SemiAnalysis, or InferenceX benchmark
Nvidia launches a Rubin Ultra Kyber rack linking 144 GPU packages, 576 dies, as one NVLink domain.62%
still open · check by
last looked
first recorded · before the live record began
as written in the record
Nvidia has launched a Rubin Ultra–generation product built on the Kyber rack architecture (NVL576-class: 144 GPU packages / 576 compute dies in a single NVLink domain)
Google's TPU v8 line includes a MediaTek co-designed model, Zebrafish, in production.62%
still open · check by
last looked
first recorded · before the live record began
as written in the record
Google's TPU v8 generation includes a MediaTek-co-designed SKU (Zebrafish / TPU 8X) in production, per SemiAnalysis / credible reporting
Nvidia ships the Rubin CPX accelerator, built for the prefill side of inference.60%
still open · check by
last looked
first recorded · before the live record began
The odds were revised: 78% on → 60% on .
as written in the record
Nvidia has shipped / made generally available the Rubin CPX prefill-specialized accelerator (in a VR NVL144 CPX or standalone VR CPX rack), per SemiAnalysis / credible reporting
AMD ships an MI500-generation rack linking at least 128 GPU packages as one machine.45%
still open · check by
first recorded · before the live record began
as written in the record
AMD has launched / shipped an MI500-generation rack-scale AI system with a single scale-up domain of ≥128 GPU packages (UAL128/UAL256-class)
Checked against SemiAnalysis's reporting, company filings, or credible press.
How forecasts are checked