Compare systems, not one number
Precision, memory, networking, software, and workload all affect results. This map deliberately avoids a synthetic performance ranking.
THE AI LANDSCAPE 05 / SILICON & SYSTEMS
Compare the accelerator programs powering AI: merchant GPUs, cloud chips, and chips designed for a company’s own workloads.
THE BIG PICTURE
Selected coverage. Check each entry for context and evidence.
Try a different name or a broader category.
A manually reviewed snapshot as of September 23, 2026. Colors organize the evidence; they do not score quality. On smaller screens, swipe across the comparison.
CONTEXT MAKES THE DIFFERENCE
Our interpretation of the comparison.
Precision, memory, networking, software, and workload all affect results. This map deliberately avoids a synthetic performance ranking.
A chip can be in production while systems, regions, and customer deployments are still rolling out.
Designers depend on manufacturing and system partners. A custom chip usually complements other accelerators rather than eliminating them.
TRANSPARENT BY DESIGN
Official announcements, product documentation, and model repositories. All linked sources were checked on September 23, 2026. Publication dates and review dates are shown separately.
Selected data-center accelerator families and programs. Workload labels reflect the designer’s stated focus and do not exclude other uses. Maturity follows the most relevant cited product page or announcement; company statements are identified as such. We do not compare vendor benchmark multipliers, fabrication economics, or chip-level power ratings as if they were equivalent end-to-end measurements.
Dates on continuously updated documentation are labeled “Current documentation” unless a specific publication date is available. This page is maintained manually; the review date is not a live-feed timestamp. Company announcements describe the company’s claims and plans, not an independent audit.
NVIDIA / Merchant platforms
NVIDIA states that Vera Rubin systems are ramping into full production with server and supply-chain partners. Individual provider delivery and service availability still require a separate check.
AMD / Merchant platforms
AMD’s July event states that MI400 and Helios are in production. Helios combines MI455X GPUs with EPYC Venice CPUs; the cited milestone supersedes earlier future-launch schedules.
AWS / Cloud silicon
Trainium3 is accessed through AWS Trn3 infrastructure. It is a cloud accelerator offering rather than a chip sold as a consumer or workstation component.
Google / Cloud silicon
Ironwood is the generally available generation in the cited Cloud TPU catalog. It is listed separately from the announced eighth-generation pair.
Google / Cloud silicon
Google positions TPU 8t for training and TPU 8i for inference and post-training. Its current product catalog still marks both as coming soon.
Microsoft / Cloud silicon
Maia 200 is Microsoft’s custom inference accelerator deployed within Azure. The deployment announcement does not describe a retail hardware offering.
Meta / Internal programs
Meta’s MTIA 300 engineering report focuses on recommendation-model training and built-in networking. It should not be treated as a generic claim about all frontier-model training at Meta.
OpenAI / Internal programs
OpenAI’s September 8 update reports chip test results and says deployment is planned to begin by year-end. Tests and a deployment target do not establish a broad production rollout.