How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
Trainium2 is AWS’s second training chip, announced 28 November 2023 and on general sale from 3 December 2024 1 2. It became the anchor chip for Anthropic’s Project Rainier, one of the largest AI compute clusters publicly disclosed.
Deep diveThe technical detail
By April 2026, AWS said Trainium2 had “largely sold out,” with roughly 30% better price-performance than comparable GPUs, according to CEO Andy Jassy’s shareholder letter 3. By February 2026, AWS reported 1.4 million Trainium2 chips delivered and “fully subscribed” 4 — Amazon’s clearest public demand signal for any chip on this site.
2.Launch and history
SimpleStart here
Trainium2 is the chip behind Project Rainier — the massive Trainium2 cluster AWS built specifically for Anthropic, its largest disclosed AI-chip customer.
Deep diveThe technical detail
AWS activated Project Rainier in October 2025 with nearly 500,000 Trainium2 chips, which AWS described as “70% larger than any other AI computing platform in AWS history”; Amazon said Anthropic would have more than a million Trainium2 chips in the cluster by the end of 2025 5. By February 2026, AWS said all 1.4 million Trainium2 chips it had delivered were fully subscribed 4 — Rainier alone accounts for the large majority of that total.
3.What’s inside it
SimpleStart here
Each Trainium2 chip has eight NeuronCores — twice the original Trainium’s four — and adds “structured sparsity,” a technique that skips unnecessary zero-value math for up to four times the speed 6.
Deep diveThe technical detail
AWS’s Neuron architecture documentation confirms the eight-NeuronCore layout, building on the same four-engine (tensor/vector/scalar/GpSimd) design used in the original Trainium and Inferentia2 (see those pages on this site) 7. The 4× sparse-versus-dense compute figure (1.3 PFLOPS dense, 5.2 PFLOPS sparse) is AWS’s own published number 6; AWS’s own Neuron documentation marks its detailed structured-sparsity API section as pending a future release, not yet part of its public reference, so this page cites the compute figure to AWS’s product announcement rather than the architecture docs. AWS has not published a die size or transistor count for Trainium2.
4.Full spec table
SimpleStart here
1.3 PFLOPS of FP8 compute per chip (5.2 PFLOPS with sparsity), 96 GiB of HBM memory, and up to 83 PFLOPS in a full 64-chip UltraServer — AWS’s own figures.
Deep diveThe technical detail
| Spec | Amazon Trainium2 |
|---|
| AI compute (FP8, dense) | 1.3 PFLOPS 6 |
|---|
| AI compute (FP8, sparse) | 5.2 PFLOPS 6 |
|---|
| Memory | 96 GiB HBM 6 |
|---|
| Memory bandwidth | 2.9 TB/s (AWS product page) or 3 TB/s (AWS Neuron docs) |
|---|
| NeuronCores | 8 6 7 |
|---|
| Process node | Not published by AWS |
|---|
| System | Chips | AI compute (FP8, dense) | Memory |
|---|
| Trn2 instance | 16 Trainium2 | 20.8 PFLOPS | 1.5 TiB |
| Trn2 UltraServer | 64 Trainium2 | 83 PFLOPS | 6 TiB |
AWS’s product page and its Neuron technical documentation give slightly different memory-bandwidth figures for the same chip (2.9 vs. 3 TB/s); this page states both rather than picking one 6 7.
5.Where it’s made
SimpleStart here
AWS has not named a foundry for Trainium2. Press reports name TSMC 3.
Deep diveThe technical detail
This is a press attribution, not an AWS statement — as with every Amazon chip on this site, AWS itself has never named a foundry. Press coverage of Amazon’s chip supply chain (see the Trainium3 page on this site for the fuller sourcing) names TSMC for Trainium2 specifically, distinguishing it from the original Trainium and both Inferentia chips, where this site found no equivalent foundry reporting at all.
6.Which systems use it
SimpleStart here
The core of Project Rainier, Anthropic’s dedicated Trainium2 cluster — nearly 500,000 chips at activation in October 2025, growing past a million by the end of that year 5.
Deep diveThe technical detail
Beyond Anthropic, AWS names Databricks, poolside, Itaú Unibanco, Ricoh, Hugging Face and Datadog among named Trn2 customers on its own instance page 8. Not every account has been positive: press coverage cites an internal Amazon document (dated July 2025) in which Cohere found Trainium1 and Trainium2 underperforming NVIDIA’s H100 with access “extremely limited,” and Stability AI called Trainium2 “less competitive” on latency and cost; Amazon responded that Trainium2 was fully subscribed 9 — a press report citing a leaked document, not independently verified by this site.
7.Official pricing
SimpleStart here
AWS does not publish an on-demand price for Trainium2 — the only published route is an advance-booked Capacity Block 8 10:
| Machine | Price/hr | Per chip | Where |
|---|
| trn2.48xlarge (16 chips) | $35.7608 | $2.235 | US East (Ohio) |
| trn2.3xlarge (1 chip) | $2.235 | $2.235 | Melbourne, São Paulo |
Deep diveThe technical detail
For comparison, AWS lists an NVIDIA H100 at $5.19 per GPU-hour on the same Capacity Block page (see this site’s NVIDIA guide) — Trainium2’s $2.235 is less than half that, though the chips are not equal in speed 10. AWS claims Trn2 has 30 to 40% better price-performance than its GPU-based P5e and P5en instances 8. AWS says Capacity Block prices are next updated in October 2026 10.
9.What came before, what came next
SimpleStart here
Came before: the original Trainium, covered on its own page on this site.
Came after: Trainium3, covered on its own page on this site.
Deep diveThe technical detail
Trainium2 was announced in November 2023 and shipped in December 2024 — about 13 months, far faster than the original Trainium’s 22-month gap (see that page on this site). Trainium3, announced the same day Trainium2 shipped, kept a similar pace, reaching general availability about a year later 1 2.
10.Hidden in plain sight
SimpleStart here
By February 2026, every one of the 1.4 million Trainium2 chips AWS had delivered was already spoken for.
Deep diveThe technical detail
AWS’s own fourth-quarter 2025 results state that 1.4 million Trainium2 chips had landed and were “fully subscribed” 4 — a sold-out figure this site found for no other AI chip on the site, from any vendor. That demand sits alongside the Business Insider report of named startups finding the same chip underperforming or hard to access in mid-2025 9 — both are real, sourced, and not necessarily contradictory: overall allocation can be fully subscribed while individual customers still describe friction getting access or matching workloads to it.