Twelve open forecasts
Each one is a statement that will turn out true or false, shown with the model’s latest recorded odds.
Anthropic uses AWS Trainium chips in production to train or run a Claude model.97%
still open · check by
last looked
first recorded · before the live record began
The odds were revised: 70% on → 86% on → 90% on → 95% on → 97% on .
as written in the record
Anthropic uses AWS Trainium chips in production to train or serve a Claude model (per SemiAnalysis / credible reporting)
AWS open-sources a native PyTorch backend for Trainium with Aten eager mode and torch.compile.85%
still open · check by
first recorded · before the live record began
as written in the record
AWS has open-sourced a native PyTorch backend for Trainium (Aten eager-mode + torch.compile), per SemiAnalysis / AWS
Microsoft's Maia still ships far fewer AI chips than Google TPU, AWS Trainium and Meta MTIA.82%
still open · check by
first recorded · before the live record began
as written in the record
Microsoft remains last among the four in-house-ASIC hyperscalers — its Maia line still ships materially fewer AI-accelerator units than Google TPU, AWS Trainium, AND Meta MTIA (per SemiAnalysis Accelerator Model)
OpenAI's Jalapeño serves production inference on hardware past the engineering-sample stage.78%
still open · check by
last looked
first recorded · live
as written in the record
OpenAI Jalapeño is serving production inference tokens on non-engineering-sample hardware, per OpenAI, SemiAnalysis, or credible reporting
Google puts TPUs into production with a major customer beyond Anthropic: Meta, xAI, SSI or OpenAI.75%
still open · check by
first recorded · before the live record began
as written in the record
Google has deployed or sold TPUs (via GCP rental or direct sale) to at least one major external AI customer beyond Anthropic — Meta, xAI, SSI, or OpenAI — in production, per SemiAnalysis / credible reporting
Arm ships its own full Phoenix datacenter CPU, with Meta as a customer.70%
still open · check by
first recorded · before the live record began
as written in the record
Arm Holdings (ARM) has shipped / made generally available its own full datacenter CPU (Phoenix) with Meta as a customer (per SemiAnalysis / credible reporting)
SemiAnalysis reports Anthropic as the biggest TPU user in 2029, ahead of Google DeepMind.70%
still open · check by
last looked
first recorded · live
as written in the record
SemiAnalysis reports Anthropic as the largest TPU user in calendar 2029, ahead of Google DeepMind
Qualcomm ships its SD2-generation datacenter CPU for revenue.60%
still open · check by
first recorded · before the live record began
as written in the record
Qualcomm (QCOM) ships its SD2-generation datacenter CPU for revenue (per SemiAnalysis / credible reporting)
Anthropic runs a Claude model in production on Cerebras WSE hardware.55%
still open · check by
last looked
first recorded · before the live record began
as written in the record
Anthropic is serving a Claude model in production on Cerebras WSE hardware (via AWS Bedrock prefill/decode disaggregation or directly), per SemiAnalysis / credible reporting
AMD or a hyperscaler announces or ships a prefill inference chip with GDDR-class memory, not HBM.55%
still open · check by
last looked
first recorded · before the live record began
as written in the record
At least one non-Nvidia AI-accelerator program (AMD or a hyperscaler ASIC) has announced or shipped a prefill-specialized inference accelerator that omits HBM in favor of GDDR-class memory, per SemiAnalysis / credible reporting
AWS announces or ships a Trainium4 that uses Nvidia's NVLink Fusion to link chips.45%
still open · check by
first recorded · before the live record began
as written in the record
AWS has announced or shipped a Trainium4 accelerator with a scale-up track using Nvidia's NVLink (NVLink Fusion) interconnect, per SemiAnalysis / credible reporting
Meta's Rivos-derived Phoebe accelerator finishes its chip design.45%
still open · check by
first recorded · live
as written in the record
Meta's Rivos-derived Phoebe accelerator has taped out (per SemiAnalysis / credible reporting)
Checked against SemiAnalysis's reporting, company filings, or credible press.
How forecasts are checked