THE AI LANDSCAPE 05 / SILICON & SYSTEMS

The silicon
behind the models.

Compare the accelerator programs powering AI: merchant GPUs, cloud chips, and chips designed for a company’s own workloads.

THE BIG PICTURE

AI Chips

Selected coverage. Check each entry for context and evidence.

8 entries

Each entry links to its primary source ↗

AI Chips. Reviewed 2026-09-23. Editorial classifications, not a performance ranking.
Accelerator / designerWorkload focusSystem contextRoute to useMaturity evidence
Vera RubinNVIDIA1 source Training + inferenceAgentic AI systemsRack / pod platformIntegrated compute and networkingSystem and cloud partnersPartner distributionProduction rampNVIDIA: May 31, 2026
Instinct MI455X / MI400AMD1 source Training + inferenceAI accelerator platformHelios systems72 MI455X GPUs per rackSystem and cloud partnersROCm ecosystemIn productionAMD: July 23, 2026
Trainium3AWS1 source Training + inferenceNeuron software stackTrn3 UltraServersScale through EC2 clustersAWS cloudEC2 service accessProduct availableSee region and capacity limits
Ironwood TPUGoogle1 source Training + inferenceSeventh-generation TPUCloud TPU systemsLarge connected podsGoogle CloudCloud TPU accessGenerally availableCurrent Cloud TPU catalog
TPU 8t / TPU 8iGoogle2 sources 8t: training · 8i: inferenceDifferent optimization targetsEighth-generation TPUsSpecialized system designsGoogle Cloud plannedUpcoming customer offeringComing soonCatalog checked Sep 23
Maia 200Microsoft1 source InferenceModel servingAzure infrastructureCustom Microsoft acceleratorAzure-operated systemsService-level accessDeployedMicrosoft: Jan 26, 2026
MTIA 300Meta1 source Training recommendationsMeta-specific workloadsIntegrated networkingBuilt-in network interfacesMeta infrastructureInternal deploymentOperating chip programEngineering report: Aug 24
JalapeñoOpenAI1 source InferenceCustom inference siliconOpenAI compute strategyAlongside partner acceleratorsInternal deployment plannedNot a public cloud SKUDeployment plannedBy end of 2026
Available / deployed / in productionComing soon / deployment plannedRoute to useRead the classification notes ↗

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

How to read this landscape.

Our interpretation of the comparison.

01

Compare systems, not one number

Precision, memory, networking, software, and workload all affect results. This map deliberately avoids a synthetic performance ranking.

02

In production is not everywhere

A chip can be in production while systems, regions, and customer deployments are still rolling out.

03

Custom does not mean isolated

Designers depend on manufacturing and system partners. A custom chip usually complements other accelerators rather than eliminating them.

TRANSPARENT BY DESIGN

Follow the evidence.

Official announcements, product documentation, and model repositories. All linked sources were checked on September 23, 2026. Publication dates and review dates are shown separately.

8primary sources
behind this comparison
Suggest a correction
Scope & methodology

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.

Vera Rubin1 source

NVIDIA / Merchant platforms

Vera Rubin

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.

Instinct MI455X / MI4001 source

AMD / Merchant platforms

Instinct MI455X / MI400

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.

Trainium31 source

AWS / Cloud silicon

Trainium3

Trainium3 is accessed through AWS Trn3 infrastructure. It is a cloud accelerator offering rather than a chip sold as a consumer or workstation component.

Ironwood TPU1 source

Google / Cloud silicon

Ironwood TPU

Ironwood is the generally available generation in the cited Cloud TPU catalog. It is listed separately from the announced eighth-generation pair.

TPU 8t / TPU 8i2 sources

Google / Cloud silicon

TPU 8t / TPU 8i

Google positions TPU 8t for training and TPU 8i for inference and post-training. Its current product catalog still marks both as coming soon.

Maia 2001 source

Microsoft / Cloud silicon

Maia 200

Maia 200 is Microsoft’s custom inference accelerator deployed within Azure. The deployment announcement does not describe a retail hardware offering.

MTIA 3001 source

Meta / Internal programs

MTIA 300

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.

Jalapeño1 source

OpenAI / Internal programs

Jalapeño

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.

PRIMARY SOURCES

Reviewed 2026-09-23 · AIstify Research