Observed Signal · Dec 19, 2025 · Technical Release · Source: Aakash Gupta Product Growth · Impact: 4/5 · Sentiment: Positive
Amazon’s Trainium3: Rack‑Level Parity and Trainium4 Plan
This newsletter deep-dive explains why Amazon’s quietly developed custom silicon and infrastructure may be an undervalued AI strategic asset. AWS launched Trainium3 (first 3nm AWS ML accelerator) and a liquid-cooled Trn3 UltraServer rack that packs 144 Trainium3 chips to deliver ~0.36 ExaFLOPS FP8 (rising to 1.4 ExaFLOPS with structured sparsity). Trainium3 chips are specified at ~2.5 petaFLOPS FP8 with 144 GB HBM3E and ~4.9 TB/s bandwidth; AWS claims 30–40% better price-performance vs comparable Nvidia instances and lower power draw (~1,000W vs ~1,400W). Anthropic and several production customers use Trainium. AWS acknowledges software and ecosystem gaps vs CUDA and plans to open-source PyTorch backend/compilers. AWS also announced Trainium4, expected late 2026, which will support NVLink Fusion to enable mixed racks of Trainium, Graviton and Nvidia GPUs and promise large FP8/FP4 and memory improvements. The piece frames Annapurna Labs’ 2015 acquisition as foundational to this capability.
AWS (a major cloud provider) shipped a 3nm training accelerator with rack-level parity to Nvidia and announced Trainium4 with NVLink Fusion support — a material infrastructure development that can change cost and multi-vendor deployment choices for large-scale AI training and inference.
Track Amazon Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- AWS announced Trainium3: a 3nm ML accelerator with ~2.5 petaFLOPS FP8 per chip and 144GB HBM3E (4.9 TB/s).
- Trn3 UltraServer racks pack 144 Trainium3 chips delivering ~0.36 ExaFLOPS FP8, rising to ~1.4 ExaFLOPS with 16:4 structured sparsity.
- AWS claims Trainium instances offer 30–40% better price-performance than comparable Nvidia GPU instances; reported instance-level costs cited ~ $1/hour vs ~$3/hour for H100s (long-term contracts down to ~$0.50/hour).
- Anthropic trains on Trainium (Project Rainier uses hundreds of thousands of Trainium chips); Trainium3 is reported in production use by JetBrains, Figma, Cursor, Replit and Bridgewater.
- AWS announced Trainium4 (targeted late 2026) will support Nvidia NVLink Fusion for mixed Trainium/Nvidia racks and promise ~3x FP8, ~6x FP4, more memory bandwidth and HBM4 stacks.
Connected Companies & Entities
8 Entities mappedRelated Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Inside AWS Trainium Lab: Amazon's AI Chip Push
TechCrunch toured AWS’s Austin chip lab where Amazon develops its Trainium AI chips. The article describes Trainium’s evolution from training-focused silicon to inference-optimized processors, the deployment of roughly 1.4 million Trainium chips across three generations, and Amazon’s commitments under a recent OpenAI deal that include supplying 2 gigawatts of Trainium capacity. It highlights Trainium3 (a 3nm chip produced by TSMC) and related system innovations — Neuron switches and Trn3 UltraServers — that AWS says cut inference cost for comparable workloads. The piece notes major customers (Anthropic, OpenAI), AWS partnerships (Cerebras Systems), PyTorch support to ease migration from Nvidia GPUs, and operational details from the lab’s engineers about chip “bring-up,” sleds, and liquid cooling.
AWS Exploring Sale of Trainium AI Chips to Third Parties
Amazon Web Services is in early talks to sell its in-house AI chip Trainium to other companies for data-center use, AWS AI chief Peter DeSantis told Bloomberg. Amazon says discussions are preliminary and DeSantis declined to name potential buyers. The idea follows Amazon CEO Andy Jassy’s shareholder letter estimating that a standalone chips business selling to AWS and third parties could have an annual run rate of about $50 billion. AWS has historically prioritized serving its own cloud customers and says current Trainium capacity — and capacity for the next-generation Trainium4 — has sold out; Trainium4 is not expected to be available for over a year. AWS would likely need surplus manufacturing via partners such as TSMC to sell chips externally. AWS spokesperson Doron Aronson also confirmed the company may sell racks of its chips to third parties in the future.
Uber Expands AWS Use, Trials Amazon Trainium3 Chips
Uber is expanding its AWS cloud contract to run more ride‑sharing features on Amazon-designed chips. The company will increase use of AWS Graviton (an ARM-based, low-power server CPU) and begin a trial of Trainium3, AWS’s in-house AI accelerator positioned as an alternative to Nvidia. The move follows Uber’s 2023 multi-year cloud deals with Google and Oracle to move infrastructure off its own data centers. Amazon has highlighted that its custom chips have attracted major customers — including Anthropic, OpenAI and Apple — and CEO Andy Jassy has said Trainium is already a multibillion-dollar business.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.