Trainium3 Ships Now; Azure Maia 300 Starts Production in 2027

By Grace Kim

Trainium3 is shipping through Amazon EC2 Trn3 UltraServers; Microsoft announced Azure Maia 300 on October 12, 2026, but says production deployments will start in 2027. Only Trainium3 supports a budget decision in the next two quarters.

Azure customers will meet Maia indirectly through GitHub, Microsoft Foundry, Copilot, Teams and Dragon Copilot; they cannot rent Maia 300 this year. No independent lab has compared the chips, so no defensible speed verdict exists.

What do the specifications show?

AWS calls Trainium3 its fourth-generation 3-nanometer AI accelerator, sold for training and inference of frontier-scale, reasoning, mixture-of-experts, long-context and video models. A Trn3 UltraServer provides up to 20.7 TB of high-bandwidth memory (HBM) and 706 TB/s aggregate bandwidth.

Secondary AWS-sourced material reports 144 chips, 144 GB HBM3e, 4.9 TB/s per chip and 362 quadrillion 8-bit operations per second (FP8 PFLOPs) per server; only one cross-check supports it.

Microsoft calls Maia 300 a complete inference system. Its third-generation design has eight HBM4 stacks and 288 GB per accelerator at 18 TB/s. A rack holds 96 accelerators; a two-tier network links 12,288 in a superpod. The process node is undisclosed.

Microsoft says Maia 300 doubles density versus Maia 200, raises bandwidth 2.5 times, adds 33% capacity and roughly doubles HBM bandwidth per watt. A single-source rack breakdown cites 27.6 TB HBM, 1.7 PB/s aggregate bandwidth and 167 kW.

AWS emphasizes proprietary interconnect and UltraCluster scale; Microsoft uses software-defined dataflow for long-context reasoning, chat and agents. Neither vendor has workload-level proof that paper memory becomes lower cost.

What performance claims matter?

Microsoft claims up to 40% more tokens per megawatt than leading GPU systems, up to seven times Maia 200's rack-scale token throughput and more than 2.5 times its performance per watt. The baselines differ, and no independent lab has verified the figures.

AWS cites unaudited results: 30% lower training cost for Amazon Search M5; 50% less training time and cost for SplashMusic; an expected 40% saving for Poolside's Trainium2 plan; and 60% faster Claude 3.5 Haiku inference on Trainium2 in Bedrock. Poolside's is a forecast.

Microsoft's headline metric measures output per unit of grid power; renters measure price per million tokens. The ratios diverge when power is cheap, silicon is expensive or utilization is low.

What can buyers purchase today?

Neither vendor publishes a Trn3 or Maia 300 rate. A single-source EC2 Capacity Blocks reading showed Trn1.32xlarge at $9.532 per hour and Trn2.3xlarge at $2.235, not Trn3. AWS prices AI Factories by location, scale, accelerator and services.

AWS has sold Trn3 since a general-availability announcement dated December 2, 2025, but reviewed pages use both “available” and “preview.” No confirmed Trn3 region list appears in the sources.

Microsoft has no Maia 300 SKU, region list, reservation plan or price; its advertised savings of up to 65% on select compute are not Maia-specific. Maia 200 entered production October 28, 2026, making it the only live baseline.

Does the software change the decision?

AWS's dated developer path is its strongest argument. Neuron 2.33 shipped in October 2026 with native PyTorch and Neuron Observer, which AWS says flags Trainium faults in under one second. TorchNeuron entered public beta October 20 as an open-source PyTorch backend for Trainium2 and Trainium3, supporting eager mode and torch.compile.

AWS says standard models can move to Neuron without full rewrites, but custom CUDA kernels and unusual operators need testing and rewriting. Microsoft has disclosed Maia's architecture but no public SDK, compiler notes or porting guide.

Amazon said Trainium and Graviton revenue commitments each rose more than 20 times year over year; earlier Trainium capacity sold at a premium to first-half 2026. CEO Andy Jassy said Trainium will be offered outside AWS by end-2027, without price, form factor or buyers.

Microsoft says Maia will become the primary inference system for its own products, scaling to multi-gigawatt deployments from 2027. Reflection's Beam was reportedly ported to Trainium3 for SageMaker JumpStart, based on one announcement. Microsoft discloses no Maia share across Copilot, GitHub or Foundry.

AWS still lists Nvidia P6-B200, P6-B300, P6e-GB200 and P6e-GB300 options. CUDA's kernels, profilers, research support and cross-cloud portability favor GPUs for fast-changing research; custom silicon fits stable, highly utilized model families.

How should buyers proceed?

For AWS, get a written quote covering instance size, region, term and exit conditions. Baseline throughput, batch size, precision, utilization and GPU price; inventory custom operators; test eager mode before torch.compile; validate fixed outputs; and measure cost per million tokens including idle time, compilation and engineer hours.

Shadow production traffic for two weeks and retain a GPU kill switch. Azure teams should keep code in standard PyTorch, avoid hard-wired vendor operators and track Foundry endpoint cost. Before signing, ask who patches firmware, what telemetry exports and what happens on reassignment. Microsoft has published no Maia equivalent.

Microsoft Ignite runs November 17-20 in San Francisco and online; AWS re:Invent follows, with AWS CEO Matt Garman scheduled to speak. Watch for a Trn3 price, TorchNeuron general availability, a Maia 300 SKU, an independent benchmark and off-cloud Trainium details. Die Signal's procurement test is blunt: “Treat stage announcements as plans, and treat pricing pages as products.”

For an AWS buyer, pilot Trainium3 on a stable workload and demand a quote. For Azure, use GPU instances and Foundry endpoints now, request a Maia roadmap briefing and revisit direct capacity after 2027. An unreleased chip should not decide a cloud contract.