# AWS Trainium 2

> Source: https://aiwiki.ai/wiki/aws_trainium2
> Updated: 2026-10-15
> Fact-checked: 2026-10-15
> Categories: AI Companies, AI Compute, AI Hardware, AI Infrastructure, Data Centers
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "AWS Trainium 2." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/aws_trainium2
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

AWS Trainium 2 (also written as Trainium2 and abbreviated Trn2) is the second generation of [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services)' custom [machine learning](https://aiwiki.ai/wiki/machine_learning) training accelerator, developed by Amazon's in-house chip design team Annapurna Labs. Amazon announced Trainium 2 at AWS re:Invent on November 28, 2023, and made Trn2 instances generally available at AWS re:Invent 2024 on December 3, 2024, in the US East (Ohio) region.[36][1][5] Each chip integrates eight NeuronCore-v3 compute cores, 96 GiB of [high-bandwidth memory](https://aiwiki.ai/wiki/high_bandwidth_memory) at 2.9 TB/s, and 1,299 TFLOPS of dense FP8 compute (667 TFLOPS of dense BF16), and Amazon states that Trn2 is four times faster than the first generation, with four times the memory bandwidth and three times the memory capacity of the original [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium).[2][1] The chip powers Amazon EC2 Trn2 instances (16 chips, 20.8 dense FP8 petaflops) and the Trn2 UltraServer (64 chips, 83.2 dense FP8 petaflops), which joins four servers into one NeuronLink domain with 6 TiB of shared accelerator memory.[1][39] In his 2025 letter to shareholders, published in April 2026, Amazon CEO Andy Jassy wrote that Trainium 2 "had about 30% better price-performance than comparable GPUs, and has largely sold out."[27] AWS deployed Trainium 2 at massive scale through [Project Rainier](https://aiwiki.ai/wiki/aws_project_rainier), a cluster of nearly 500,000 Trainium 2 chips activated in October 2025 in partnership with [Anthropic](https://aiwiki.ai/wiki/anthropic) for training and serving the Claude family of models.[6][7] By March 2026, Claude was running on more than one million Trainium 2 chips, a figure Anthropic confirmed in April 2026.[23][12] [OpenAI](https://aiwiki.ai/wiki/openai) committed in February 2026 to about 2 GW of Trainium capacity, but on Trainium3 and Trainium4 rather than Trainium 2, while [Apple](https://aiwiki.ai/wiki/apple) said in December 2024 that it was evaluating Trainium 2 for pretraining.[32][33] The successor, [Trainium 3](https://aiwiki.ai/wiki/aws_trainium_3), reached general availability in Trn3 UltraServers in December 2025, and the April 2026 [expansion of the Anthropic-Amazon agreement](https://aiwiki.ai/wiki/anthropic_amazon_trainium) covers Trainium 2 through Trainium 4.[13][12]

## What is AWS Trainium 2 and how does it differ from the first generation?

Amazon announced the original [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium) accelerator at re:Invent in December 2020 as a purpose-built [AI chip](https://aiwiki.ai/wiki/ai_chip) for [deep learning](https://aiwiki.ai/wiki/deep_learning) training on AWS, and the EC2 Trn1 instances built on it became generally available in October 2022.[37][38] The first generation, which powered the EC2 Trn1 and Trn1n instances, used two large NeuronCore-v2 compute cores per chip and shipped with 32 GiB of device memory at 820 GiB/s.[43] Trn1 was designed before [large language model](https://aiwiki.ai/wiki/large_language_model)s dominated training workloads; Trainium 2 tripled the memory per chip and raised chip-to-chip interconnect bandwidth 3.3 times, from 384 GB/s to 1,280 GB/s.[2]

Amazon announced Trainium 2 at AWS re:Invent on November 28, 2023, positioning it for training foundation models and other [generative AI](https://aiwiki.ai/wiki/generative_ai) workloads. AWS said at the time that the chip would deliver up to four times faster training, three times the memory capacity and up to twice the energy efficiency of the first generation, and that customers could scale to 100,000 chips in next-generation EC2 UltraClusters.[36] The revised architecture increased the NeuronCore count from two to eight per chip, tripled device memory to 96 GiB, raised memory bandwidth from about 0.8 TB/s to 2.9 TB/s, and introduced NeuronLink-v3.[2] The chip also added 16 dedicated Collective Communication Cores (CC-Cores), up from six in the first generation, to handle the all-reduce and all-gather operations that dominate large-scale [distributed training](https://aiwiki.ai/wiki/distributed_training) without competing with the [tensor core](https://aiwiki.ai/wiki/tensor_core)s for execution resources.[2]

AWS made Trn2 instances generally available in the US East (Ohio) region on December 3, 2024, at the same AWS re:Invent event where the company also previewed Trainium 3 and disclosed Project Rainier details.[1][5][9] By March 2026, 1.4 million Trainium chips were deployed across all three generations, according to TechCrunch, and Claude alone ran on more than one million Trainium 2 chips.[23]

## How is the Trainium 2 chip designed?

### NeuronCore-v3

Each Trainium 2 chip contains eight NeuronCore-v3 compute cores.[2] As in NeuronCore-v2, each core has four functional engines:[41]

- A tensor engine built around a 128x128 [systolic array](https://aiwiki.ai/wiki/systolic_array) for [matrix multiplication](https://aiwiki.ai/wiki/matrix_multiplication), convolution, and transposition. This engine handles the bulk of the floating-point work in transformer training.[10][41]
- A vector engine for operations in which each output element depends on several input elements, such as layer normalization and pooling.[41]
- A scalar engine for element-wise operations, in which each output element depends on one input element.[41]
- A GPSIMD (general-purpose single instruction, multiple data) engine of eight programmable 512-bit vector processors, which lets users write custom C++ operators that run directly inside the NeuronCore without routing through the host CPU.[2][41]

NeuronCore-v3 supports dynamic shapes and control flow through ISA extensions, a capability the first generation also had through NeuronCore-v2 extensions.[2][43] This matters for workloads like [reinforcement learning from human feedback (RLHF)](https://aiwiki.ai/wiki/rlhf) and [sequence-to-sequence task](https://aiwiki.ai/wiki/sequence-to-sequence_task)s where batch dimensions vary at runtime. Trainium 2 additionally supports user-programmable rounding modes (round-nearest-even or stochastic rounding), giving researchers control over numerical behavior in reduced-precision formats.[2]

The chip provides the following aggregate compute across all eight cores:[2]

| Precision | Dense TFLOPS | Sparse TFLOPS |
|---|---|---|
| FP8 | 1,299 | 2,563 |
| BF16 / FP16 / TF32 | 667 | 2,563 |
| FP32 | 181 | -- |

Note: the Neuron documentation lists a single sparse figure, 2,563 TFLOPS, for FP8, FP16, BF16 and TF32.[2] AWS's product page describes the sparsity support as "4x sparsity (16:4)",[3] and launch coverage applied that 4x multiplier to the dense FP8 rating, which gives a higher sparse FP8 figure than the documentation's table (see the UltraServer table below).[39] The NKI architecture guide rates the tensor engine of each NeuronCore-v3 at 158 FP8, 79 BF16/FP16/TF32 and 20 FP32 dense TFLOPS, with FP8 matrix multiplication running at double the BF16 rate.[10]

### Logical NeuronCore configuration

Starting with Trainium 2, AWS Neuron supports Logical NeuronCore Configuration (LNC), which combines the compute and memory resources of physical NeuronCores into a single logical NeuronCore.[2] LNC can be set to 1 or 2. The default on Trainium 2, LNC=2, pairs two physical cores, so each chip presents four logical NeuronCores and a trn2.48xlarge presents 64; LNC=1 exposes all eight physical cores separately.[42] This feature is useful for workloads that benefit from a larger scratchpad or higher sustained memory bandwidth per logical compute unit, at the cost of reduced parallelism. LNC gives operators a tuning knob that did not exist on Trainium 1.

### Collective Communication Cores

Trainium 2 includes 16 dedicated CC-Cores per chip, compared to six in Trainium 1.[2] These cores handle collective operations (AllReduce, AllGather, ReduceScatter, AllToAll) independently of the tensor cores, so gradient synchronization across chips proceeds in parallel with the forward and backward passes.[2][4] On NVIDIA and AMD GPUs, communication operations run on the same streaming multiprocessors as compute, according to SemiAnalysis, so the two can contend for resources in a [GPU cluster](https://aiwiki.ai/wiki/gpu_cluster).[4] The dedicated CC-Core design is one of the architectural choices AWS made to optimize the chip for distributed training throughput rather than single-chip peak compute.[4]

## What are the Trainium 2 specifications?

### Per-chip specifications

| Specification | Value |
|---|---|
| Architecture | NeuronCore-v3 |
| NeuronCores per chip | 8 |
| FP8 dense compute | 1,299 TFLOPS |
| BF16 dense compute | 667 TFLOPS |
| FP32 compute | 181 TFLOPS |
| HBM type | HBM3 per AWS;[3] SemiAnalysis reports HBM3e stacks run at HBM3 speed[4] |
| HBM capacity | 96 GiB (four 24 GB stacks)[39] |
| HBM bandwidth | 2.9 TB/s |
| DMA bandwidth | 3.5 TB/s |
| Scratchpad (SBUF) | 224 MiB |
| CC-Cores | 16 |
| NeuronLink-v3 bandwidth | 1.28 TB/s per chip |
| Process node | TSMC N5 (5 nm class)[15] |
| Thermal design power | About 500 W (SemiAnalysis estimate)[4] |

### Trn2 instance specifications

| Specification | trn2.3xlarge | trn2.48xlarge |
|---|---|---|
| Trainium 2 chips | 1 | 16 |
| NeuronCores | 8 | 128 |
| Accelerator memory (HBM) | 96 GB | 1.5 TB |
| HBM bandwidth | 2.9 TB/s | 46 TB/s |
| FP8 dense compute | 1.3 PFLOPs | 20.8 PFLOPs |
| System vCPUs | 12 | 192 |
| System memory | 128 GB | 2 TB |
| Instance storage | 1x 470 GB NVMe | 4x 1.92 TB NVMe |
| Network bandwidth (EFA v3) | 0.2 Tbps | 3.2 Tbps |
| EBS bandwidth | 5 Gbps | 80 Gbps |

### Trn2 UltraServer specifications

| Specification | Trn2 UltraServer |
|---|---|
| Trainium 2 chips | 64 |
| NeuronCores | 512 |
| Accelerator memory | 6 TiB |
| HBM bandwidth | 185 TB/s |
| FP8 dense compute | 83.2 PFLOPs |
| FP8 sparse compute | 332.8 PFLOPs with the 4x sparsity mode, as reported at launch;[39] about 164 PFLOPs at the Neuron documentation's per-chip sparse rate[2] |
| Network bandwidth (EFA v3) | 12.8 Tbps |
| Constituent instances | 4 x trn2u.48xlarge |
| Status | Launched in preview in December 2024; still listed as preview on AWS's product page in October 2026[1][3] |

## How much memory and bandwidth does Trainium 2 have?

Each Trainium 2 chip carries 96 GiB of HBM in four 24 GB stacks, with an aggregate bandwidth of 2.9 TB/s.[2][39] AWS's product page calls the memory HBM3; SemiAnalysis reported that the stacks are [HBM3e](https://aiwiki.ai/wiki/hbm3e) parts that firmware runs at HBM3 speeds.[3][4] Relative to the first-generation chip (32 GiB at about 0.8 TB/s), this is three times the capacity and, by the Neuron documentation's count, 3.6 times the bandwidth; AWS's launch materials round the bandwidth gain to four times.[2][1]

Physically, the chip is a multi-chiplet design in which two compute chiplets share access to all four HBM stacks.[4] The layout creates non-uniform memory access (NUMA) characteristics: a compute chiplet accesses adjacent HBM stacks at full bandwidth but incurs a penalty when accessing stacks attached to the other chiplet.[4] This is similar to the NUMA behavior observed in AMD's Instinct MI300X, and SemiAnalysis noted that NUMA-aware programming may be needed to reach peak performance.[4]

Trainium 2 also provides 224 MiB of on-chip scratchpad memory (SBUF), 4.7 times larger than Trainium 1's 48 MiB.[2] The SBUF is partitioned among NeuronCores, 28 MiB per core, and acts as a software-managed tile buffer for the tensor engine, avoiding redundant HBM loads during matrix operations.[10]

DMA bandwidth reaches 3.5 TB/s with inline memory compression and decompression.[2]

The 16-chip Trn2 instance aggregates to 1.5 TiB of HBM at 46 TB/s combined bandwidth.[1][3] The Neuron documentation lists memory pooling across up to 64 chips, compared with 16 on the first generation, which matches the UltraServer's 64-chip domain and its 6 TiB of HBM; AWS pitches the UltraServer for training and inference "at the trillion parameter level and beyond."[2][1]

## Trn2 instances

Amazon Web Services offers two Trn2 instance sizes in the EC2 family:

The trn2.3xlarge is the entry-level instance with a single Trainium 2 chip, 12 vCPUs, 128 GB of system memory, and 96 GB of accelerator memory.[3] With one chip, it suits inference, fine-tuning of smaller models, and other workloads that do not require multi-chip connectivity. Network bandwidth is 200 Gbps.[3]

The trn2.48xlarge is the full-instance configuration with all 16 Trainium 2 chips enabled, 192 vCPUs, 2 TiB of system memory, 1.5 TiB of HBM, and 3.2 Tbps of Elastic Fabric Adapter (EFA) v3 network bandwidth.[3] The 16 chips are connected in a 4x4 two-dimensional torus via NeuronLink-v3 at 1 TB/s of chip-to-chip bandwidth.[1][4] This instance is the building block for both standalone large-model training and for composition into UltraServer systems.

A third variant, the trn2u.48xlarge, has the same compute and memory specs as the trn2.48xlarge but includes UltraServer-capable hardware to support NeuronLink connections across server boundaries.[1][3] Four trn2u.48xlarge instances combine to form one Trn2 UltraServer.[1]

At launch, Trn2 instances were reserved through Amazon EC2 Capacity Blocks for ML, which allows customers to book blocks of up to 64 instances for periods up to six months, with instant start times. Capacity Blocks can be booked up to eight weeks in advance, giving training teams predictable access without long-term commitment.[1]

On the software side, AWS's launch post pointed developers to the AWS Deep Learning AMIs and said the Neuron SDK integrates natively with [JAX](https://aiwiki.ai/wiki/jax), [PyTorch](https://aiwiki.ai/wiki/pytorch) and libraries such as [Hugging Face](https://aiwiki.ai/wiki/hugging_face), [PyTorch Lightning](https://aiwiki.ai/wiki/pytorch_lightning) and NeMo.[1]

### Where are Trn2 instances available?

Trn2 instances launched in US East (Ohio), with AWS saying other regions would follow.[1][9] In October 2026, AWS's Capacity Blocks for ML pricing page listed trn2.48xlarge capacity in US East (Ohio) and Asia Pacific (Hyderabad).[16] The April 2026 Anthropic agreement included an expansion of Claude inference in Asia and Europe, without saying which Trainium generation would serve it.[11][12] TechCrunch reported in March 2026 that the biggest share of Trainium 2 chips was deployed in Project Rainier, whose Indiana campus The New York Times described as Amazon's biggest data center.[23][30]

## What is the Trn2 UltraServer?

The Trn2 UltraServer connects four trn2u.48xlarge instances using NeuronLink-v3 to form a single scale-up domain with 64 Trainium 2 chips.[1] SemiAnalysis estimated that the system spans two racks of about 24 kW each.[4] AWS describes the UltraServer's 6 TiB of HBM as shared accelerator memory across the 64 chips.[1][3] AWS specifies up to 83.2 dense FP8 petaflops of compute, 185 TB/s of total memory bandwidth, and 12.8 Tbps of EFA v3 networking per UltraServer.[1][39]

Keeping more traffic on NeuronLink instead of EFA is meant to shorten the time large models spend on gradient synchronization and tensor-parallel communication, which can be a significant fraction of step time for large language models.

The UltraServer extends the chip topology from a 4x4 two-dimensional torus (within a single instance) to a 4x4x4 three-dimensional torus (across the four instances).[4] The Z-axis connections between instances use OSFP-XD active electrical cables and provide 64 GB/s of point-to-point bandwidth per chip pair, while X and Y axis connections within a server run at 128 GB/s per connection.[4] AWS deliberately avoided optical transceivers for these links, using passive and active copper cables instead.[4] SemiAnalysis put the reliability advantage of copper at 10 to 100 times better mean time to failure than optics, with far less link flapping, which matters in a cluster with thousands of interconnects.[4]

Multiple UltraServers can be connected into an UltraCluster. AWS's 2023 announcement said customers could scale to 100,000 Trainium 2 chips in next-generation EC2 UltraClusters, and the December 2024 launch post described scale-out training across tens of thousands of Trainium chips on a single petabit-scale, non-blocking network.[36][1] Trn2 UltraClusters have access to Amazon FSx for Lustre high-performance storage for checkpoints and training data.[1]

Trn2 UltraServers launched in preview in December 2024, and AWS's Trn2 product page still described them as available in preview in October 2026.[1][5][3] Project Rainier was built as an EC2 UltraCluster of Trn2 UltraServers.[6]

## How does NeuronLink networking work?

NeuronLink-v3 is the scale-up interconnect that connects Trainium 2 chips within and across servers. Unlike NVIDIA NVLink, which uses a switch topology enabling all-to-all communication between GPUs in a node, NeuronLink uses direct point-to-point connections arranged in torus configurations.[4] Each chip connects to up to six immediate neighbors at 640 GB/s of unidirectional bandwidth: four inside its server, forming the 4x4 2D torus, and two in other servers in UltraServer mode, extending it to the 4x4x4 3D torus.[4]

The torus topology determines which collective communication algorithms run efficiently. AWS's NeuronX Collective Communication Library provides AllReduce, AllGather, ReduceScatter and other collectives purpose-built for the 2D and 3D torus; SemiAnalysis's December 2024 analysis initially reported that it lacked All-to-All support, which matters for Mixture of Experts expert parallelism, and later noted that AWS said All-to-All was supported.[4] The library is conceptually similar to NVIDIA's NCCL but tuned for Trainium 2's topology rather than NVLink's all-to-all connectivity.

The aggregate NeuronLink-v3 bandwidth per chip is 1.28 TB/s, which is the sum across all active ports.[2] This figure captures the total bidirectional chip-to-chip capacity rather than a single-port measure.[4]

For scale-out communication between UltraServer systems, Trn2 instances use Elastic Fabric Adapter v3, which provides 3.2 Tbps of network bandwidth per instance.[3] EFA is used for data-parallel gradient synchronization across racks, while NeuronLink handles tensor-parallel and pipeline-parallel communication within a scale-up domain.

## What is the Neuron SDK?

All Trainium 2 instances run the AWS Neuron SDK, a software stack that includes a graph compiler, runtime, libraries and developer tools.[18] Neuron releases have shipped every one to three months: version 2.28.0 came out in February 2026 and 2.32.0 in August 2026.[21]

### Framework support

Neuron supports PyTorch and JAX, and integrates with vLLM for inference serving.[18] Through 2025, PyTorch ran on Trainium 2 through torch-neuronx, which is built on PyTorch/XLA's lazy-tensor tracing. AWS introduced TorchNeuron, a native PyTorch backend with eager execution and torch.compile, as a private beta in December 2025, and announced that PyTorch 2.9 would be the last version based on PyTorch/XLA, with native support through TorchNeuron starting with PyTorch 2.10.[20][44] JAX support arrived as a beta in September 2024 through a PJRT plugin that presents Trainium and Inferentia to JAX as pluggable devices.[46]

Neuron SDK 2.26.0, released in September 2025, added support for PyTorch 2.8 and JAX 0.6.2, enhanced inference on Trn2, and added expert parallelism (beta) for Mixture of Experts models.[19] Neuron SDK 2.27.0, announced in December 2025 after the Trainium 3 launch at re:Invent, added support for Trainium 3 UltraServers and introduced several private betas.[20]

For inference serving, Neuron integrates with [vLLM](https://aiwiki.ai/wiki/vllm) and includes the NxD Inference library, and Hugging Face supports Trainium through its Optimum Neuron library.[18][3] Lightning AI offers native Trainium support across Lightning Studios and its open-source tools.[3]

### Neuron Kernel Interface (NKI)

Neuron Kernel Interface (NKI), introduced as a beta with Neuron 2.20 in September 2024, gives performance engineers direct access to the NeuronCore-v3 instruction set architecture, memory allocation, and execution scheduling.[47][10] NKI uses a tile-based programming model with a Triton-like interface (Triton is used for CUDA kernel development) and allows developers to write custom operators in Python that compile to native NeuronCore instructions.[3][10] This is useful for fused attention kernels, custom normalization layers, and other operators that would otherwise be decomposed into multiple passes through HBM.

The NKI Library of optimized kernels launched with seven kernels in Neuron 2.27.0 in December 2025, covering attention, MLP, output projection, QKV and RMSNorm-quantization operations, and reached 16 documented kernel APIs with Neuron 2.28.0 in February 2026, which added RoPE, router top-k and [Mixture of Experts (MoE)](https://aiwiki.ai/wiki/mixture_of_experts) kernels among others.[45] AWS publishes the kernels in an open-source GitHub repository.[45]

Neuron 2.27.0 also introduced, in private beta, an open-source NKI compiler built on MLIR, native PyTorch support through TorchNeuron, and Neuron DRA (Dynamic Resource Allocation) for Kubernetes-native resource management.[20] In 2026, Neuron 2.29.0 (April) declared NKI 0.3.0 stable and limited the NxD Inference library to Trn2 and newer hardware, and Neuron 2.31.0 (July) made a new code-generation backend the default on Trn2 and Trn3.[21]

### Compiler and tooling

The Neuron graph compiler turns framework graphs into binaries for NeuronCore execution, and applications built for the first generation can be recompiled for Trn2.[18][1] The open-source, MLIR-based compiler that AWS introduced at the end of 2025 is the NKI compiler, which was in private beta at launch.[20]

Neuron Explorer, a tools suite introduced with Neuron 2.27.0, and the Neuron Profiler, which added system profile grouping for distributed workloads in Neuron 2.26.0, give developers visibility into NeuronCore utilization, memory traffic, and collective communication.[20][19]

## How does Trainium 2 perform?

AWS's published performance figures position Trn2 against prior-generation Trainium and against GPU-based EC2 instances:[1][3]

| Benchmark | Trn2 vs Trn1 | Trn2 vs P5e/P5en (H200) |
|---|---|---|
| Training throughput | 4x faster | 30-40% better price-performance |
| Memory bandwidth | 4x higher | -- |
| Memory capacity | 3x larger | -- |
| Energy efficiency | 3x better (AWS's 2023 announcement projected up to 2x)[36] | -- |

For inference, AWS reported that Trainium 2 delivers three times higher token-generation throughput for Meta's Llama 3.1 405B model on Amazon Bedrock compared to other available offerings by major cloud providers at the time of launch in December 2024.[9]

[Claude 3.5 Haiku](https://aiwiki.ai/wiki/claude_3_5_haiku) running on Trainium 2 with latency optimization delivers up to 60 percent faster inference compared to the non-latency-optimized version of the same model, according to Anthropic's December 2024 announcement.[8] Tom Brown, Anthropic's Chief Compute Officer, described the two launch advances directly: "a new 'latency-optimized mode' for Claude 3.5 Haiku which runs 60% faster on Trainium2 via Amazon Bedrock," and Project Rainier, which he called "a new cluster with hundreds of thousands of Trainium2 chips delivering hundreds of exaflops, which is over five times the size of our previous cluster."[3]

In the [SemiAnalysis](https://aiwiki.ai/wiki/semianalysis) analysis of the Trainium 2 architecture, the chip's arithmetic intensity is 225.9 BF16 FLOP per byte, lower than the [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100), GB200 and Google TPU v6e, which SemiAnalysis put in a range of about 300 to 560 FLOP per byte.[4] SemiAnalysis argued that lower arithmetic intensity may be the right design choice: memory-bandwidth-bound inference and [Mixture of Experts](https://aiwiki.ai/wiki/mixture_of_experts) models, whose grouped matrix multiplications load many more weights per token, favor bandwidth over raw TFLOPS.[4]

When Trainium 3 launched at re:Invent 2025, AWS published internal benchmarks comparing the two generations on OpenAI's [GPT-OSS](https://aiwiki.ai/wiki/gpt_oss) open-weight model. In that testing, Trn3 UltraServers delivered three times higher throughput per chip and four times faster response times than Trn2 UltraServers.[40]

## How does Trainium 2 compare to NVIDIA GPUs?

The following table compares Trainium 2 specifications with NVIDIA's H100 SXM5, H200 SXM, and Blackwell B200 accelerators:

| Specification | Trainium 2 (chip) | NVIDIA H100 SXM5 | NVIDIA H200 SXM | NVIDIA B200 |
|---|---|---|---|---|
| FP8 dense compute | 1,299 TFLOPS | 1,979 TFLOPS (3,958 sparse) | 1,979 TFLOPS (3,958 sparse) | 4,500 TFLOPS (9,000 sparse) |
| BF16 dense compute | 667 TFLOPS | 989 TFLOPS (1,979 sparse) | 989 TFLOPS (1,979 sparse) | 2,250 TFLOPS (4,500 sparse) |
| FP32 compute | 181 TFLOPS | 67 TFLOPS | 67 TFLOPS | 75 TFLOPS |
| HBM capacity | 96 GiB | 80 GB | 141 GB | 180 GB |
| HBM bandwidth | 2.9 TB/s | 3.35 TB/s | 4.8 TB/s | 7.7 TB/s |
| Scale-up interconnect | NeuronLink-v3 (1.28 TB/s) | NVLink 4 (0.9 TB/s) | NVLink 4 (0.9 TB/s) | NVLink 5 (1.8 TB/s) |
| TDP | About 500 W (estimate) | Up to 700 W | Up to 700 W | Up to 1,000 W |

Note: NVIDIA publishes its Tensor Core figures with sparsity; the dense values above are half of those, as NVIDIA's own footnotes state.[48][49][50] H100 and H200 values are for SXM variants.[48][49] B200 compute values are per GPU, derived from NVIDIA's eight-GPU HGX B200 specifications; its memory figures are the shipping 180 GB and 7.7 TB/s rather than the 192 GB and 8 TB/s of early material.[50] Trainium 2 figures come from the AWS Neuron documentation, except the power figure, which is a SemiAnalysis estimate.[2][4]

Trainium 2's raw TFLOPS at FP8 and BF16 are lower than the H100, H200, and B200.[4] AWS argues that the relevant metric for training cost is price-performance rather than peak TFLOPS, claiming 30 to 40 percent better price-performance for Trn2 than for H200-based P5e and P5en instances.[1][52] In his 2025 shareholder letter, Jassy drew an explicit parallel to Amazon's earlier Graviton CPUs displacing general-purpose server chips, writing that Trainium 2 "had about 30% better price-performance than comparable GPUs, and has largely sold out," and that at scale "we expect Trainium will save us tens of billions of capex dollars per year, and provide several hundred basis points of operating margin advantage versus relying on others' chips for inference."[27]

From a software ecosystem standpoint, NVIDIA's CUDA platform has a much larger library of hand-optimized kernels, profiling tools, and community examples than the Neuron SDK. AWS has invested in closing this gap through NKI, the open-source NKI compiler, and TorchNeuron, and through partnerships with framework teams at PyTorch and Hugging Face.[20][3] NVIDIA's larger scale-up domain in [GB200 NVL72](https://aiwiki.ai/wiki/nvidia_gb200_nvl72) systems (72 GPUs sharing NVLink) gives Blackwell-generation clusters more flexibility in tensor-parallel degree, which can reduce inference costs at large model sizes. Trainium 2's answer is the 64-chip Trn2 UltraServer. For Trainium 4, AWS said it is designing the chip to support NVIDIA [NVLink Fusion](https://aiwiki.ai/wiki/nvlink_fusion), so that Trainium 4, Graviton CPUs and EFA can work together in common MGX racks that hold both GPU and Trainium servers.[40][14]

## What is Project Rainier?

The most consequential deployment of Trainium 2 is Project Rainier, an [AI compute](https://aiwiki.ai/wiki/ai_compute) cluster built by AWS for [Anthropic](https://aiwiki.ai/wiki/anthropic) to train and serve the Claude model family.[6][25] Ron Diamant, AWS Distinguished Engineer and Head Architect of Trainium, called it "one of AWS's most ambitious undertakings to date."[6]

AWS detailed Project Rainier at AWS re:Invent in December 2024, describing a cluster with hundreds of thousands of Trainium 2 chips.[8][9] AWS brought Project Rainier online in October 2025, less than a year after it was first announced.[6][22] At activation the cluster contained nearly 500,000 Trainium 2 chips spread across multiple U.S. [data center](https://aiwiki.ai/wiki/data_center)s, including a 1,200-acre campus outside New Carlisle in St. Joseph County, Indiana.[6][7][30] Semafor reported, citing a person familiar with the matter, that Rainier spanned sites in Indiana, Mississippi and Pennsylvania.[31] AWS stated that it had built Trainium 2 infrastructure "70% larger than any other AI computing platform in AWS history," and that the cluster provided more than five times the compute Anthropic had used to train its previous models.[6]

Project Rainier is named after a 14,410-foot (4,392-meter) stratovolcano that can be seen from Seattle on a clear day.[6] AWS is investing $11 billion in the St. Joseph County site, where it broke ground in October 2024.[7] The New York Times reported in June 2025 that Amazon had promised the $11 billion for 16 buildings under a 2023 tax deal, had seven buildings up by early June, planned about 30, and expected the campus to consume 2.2 gigawatts of electricity when complete.[30] AWS CEO Matt Garman told CNBC after the activation that Anthropic was running about 500,000 chips in Indiana.[26]

After activation, Anthropic scaled its Trainium 2 footprint quickly. Amazon had said Claude was expected to run on more than one million Trainium 2 chips by the end of 2025, across direct usage and Amazon Bedrock; TechCrunch reported in March 2026 that Claude ran on over one million Trainium 2 chips, and Anthropic confirmed the figure in April 2026.[22][23][12] This scale-up coincided with sharp growth in Anthropic's run-rate revenue, which the company said had surpassed $30 billion by April 2026, up from about $9 billion at the end of 2025.[12]

### How does Anthropic schedule workloads on Project Rainier?

Semafor reported in October 2025 that the majority of Rainier's chips were used for inference and that Anthropic used a portion of them for training runs in the evening, pausing those runs during the day when inference demand increased.[31] Amazon and Anthropic have described the cluster as used both to train and to serve Claude.[6][12]

### Hardware co-design

Anthropic said in November 2024 that it was working closely with Annapurna Labs on future generations of Trainium accelerators, and Amazon said in April 2026 that Anthropic works with Annapurna on developing and optimizing future Trainium chips.[35][11] The New York Times reported that Annapurna developed Trainium 2 alongside Anthropic engineers.[30] CNBC reported around the Rainier activation that Trainium 3 was also developed with direct input from Anthropic.[26]

### Commercial structure

Amazon CEO Andy Jassy said on Amazon's third-quarter 2025 earnings call, on October 30, 2025, that Trainium 2 was "fully subscribed" and "a multibillion-dollar business that grew 150% quarter-over-quarter."[28][29] Amazon's investment in Anthropic began with $1.25 billion in September 2023 and reached $4 billion in March 2024; a further $4 billion in November 2024 brought the total to $8 billion and made AWS Anthropic's primary cloud and training partner.[34][35] On April 20, 2026, the two companies announced an [expanded agreement](https://aiwiki.ai/wiki/anthropic_amazon_trainium) under which Anthropic committed more than $100 billion over ten years to AWS technologies, securing up to 5 GW of capacity across Graviton and Trainium 2 through Trainium 4. Anthropic said significant new Trainium 2 capacity would come online in the second quarter of 2026 and that nearly 1 GW of Trainium 2 and Trainium 3 capacity would be online by the end of 2026. Amazon invested $5 billion at the time, with up to $20 billion more to follow.[12][11]

## How is Trainium 2 used on Amazon Bedrock?

[Amazon Bedrock](https://aiwiki.ai/wiki/amazon_bedrock) is the fully managed service through which AWS customers access [foundation models](https://aiwiki.ai/wiki/foundation_models) from Anthropic, Meta, Mistral, and other providers via API. Trainium 2 shows up in Bedrock in two ways: a latency-optimized inference option for specific models, and the general inference fleet, most of which runs on Trainium.

### Latency-optimized inference

Bedrock's latency-optimized inference option, launched in preview in December 2024, uses Trainium 2 for the latency-optimized versions of Llama 3.1 405B and Claude 3.5 Haiku.[1][8] Bedrock's documentation lists latency-optimized versions of:[51]

- Claude 3.5 Haiku (Anthropic): up to 60 percent faster than non-optimized inference[8]
- Meta Llama 3.1 405B: three times higher token-generation throughput versus other available cloud offerings[9]
- Meta Llama 3.1 70B
- Amazon Nova Pro

Latency-optimized Claude 3.5 Haiku was initially available in the US East (Ohio) region via cross-region inference.[8] Bedrock's documentation, which still labeled the feature a preview in 2026, lists Claude 3.5 Haiku and the Llama 3.1 models in US East (Ohio) and US West (Oregon) and [Amazon Nova](https://aiwiki.ai/wiki/amazon_nova) Pro in US East (N. Virginia), US East (Ohio) and US West (Oregon), all through cross-Region inference; no newer Claude model appears on the list.[51]

### Infrastructure for model providers

Beyond the latency-optimized option, Trainium 2 runs workloads for Bedrock model providers, including the Anthropic Claude family. On the October 2025 earnings call, Jassy said "the majority of token usage in Amazon Bedrock is already running on Trainium."[29] TechCrunch reported in March 2026 that Trainium 2 handled the majority of Bedrock's inference traffic, and Jassy's April 2026 shareholder letter said Bedrock "runs most of its inference on Trainium."[23][27]

### Pricing on Bedrock

Claude 3.5 Haiku accessed through Bedrock with latency optimization on Trainium 2 was priced at $1.00 per million input tokens and $5.00 per million output tokens in the US East (Ohio) region at launch in December 2024.[8] At the same time, Anthropic cut the standard Claude 3.5 Haiku price to $0.80 per million input tokens and $4.00 per million output tokens across all platforms.[8] Pricing for later Claude versions and other models is published per model on the Bedrock pricing page.

## Who uses Trainium 2?

### Anthropic

Anthropic is the largest disclosed Trainium 2 user: the biggest share of Trainium 2 chips sits in Project Rainier, and Anthropic said in April 2026 that it used over one million Trainium 2 chips to train and serve Claude.[23][12] The company moved from roughly 500,000 chips at Project Rainier activation in October 2025 to more than one million by early 2026, with additional capacity committed under the April 2026 agreement.[7][23][12] Claude is served both through Anthropic's own products and through Amazon Bedrock, where over 100,000 customers ran Claude as of April 2026.[12]

### OpenAI

On February 27, 2026, Amazon and [OpenAI](https://aiwiki.ai/wiki/openai) announced a strategic partnership under which Amazon agreed to invest $50 billion in OpenAI, starting with $15 billion and adding $35 billion once certain conditions were met, and the two companies expanded OpenAI's existing $38 billion AWS agreement by $100 billion over eight years.[32][24] OpenAI committed to consume approximately 2 GW of Trainium capacity, spanning Trainium 3 and Trainium 4 rather than Trainium 2, and AWS became the exclusive third-party cloud distribution provider for OpenAI Frontier.[32]

### Apple

[Apple](https://aiwiki.ai/wiki/apple) does not have a disclosed production deployment of Trainium 2. At AWS re:Invent in December 2024, Apple machine learning and AI director Benoit Dupin said Apple was testing whether Trainium 2 could be used to pretrain [Apple Intelligence](https://aiwiki.ai/wiki/apple_intelligence) and other models, and expected "up to a 50 percent improvement in efficiency with pre-training"; the 40 percent efficiency gain he cited came from Apple's use of AWS's Graviton and Inferentia chips for search services.[33] TechCrunch's March 2026 report on the Trainium lab named Apple among the companies the chip family had won over, citing those public remarks.[23]

### Other customers

Other customers that disclosed Trainium 2 plans or use include:

- poolside, a code-focused AI company that planned to train future models on Trn2 UltraServers, expecting savings of 40 percent compared with EC2 P5 instances[3]
- [Databricks](https://aiwiki.ai/wiki/databricks), which planned to use Trn2 to lower total cost of ownership for its customers by up to 30 percent[3]
- Karakuri, a Japanese AI company that trained KARAKURI LM 8x7B Chat on AWS Trainium and expected Trainium 2 to halve its training time[3]
- NinjaTech AI, which cited Trn2's low latency and competitive pricing for its inference workloads[3]
- Ricoh, which pretrained a 13-billion-parameter LLM in eight days on 4,096 first-generation Trainium chips and said it was looking ahead to Trainium 2[3]

AWS's Trainium customer page also features Decart, which uses Trainium for real-time video models, without naming the chip generation.[17] At the Trainium 3 launch, AWS said customers including Anthropic, Karakuri, SplashMusic and Decart had significantly cut their inference costs with Trainium.[14]

On the October 2025 earnings call, Jassy said Trainium 2 was fully subscribed and that "a small number of very large customers" were using it, with Amazon expecting to accommodate more customers starting with Trainium 3.[29] In April 2026 he wrote that Trainium 2 "has largely sold out."[27]

## Pricing and availability

AWS did not publish an on-demand hourly rate for Trn2 at launch; capacity was reserved through EC2 Capacity Blocks for ML.[1] Dated list prices:

| Instance | Purchase option | Listed price | Date |
|---|---|---|---|
| trn2.48xlarge | EC2 Capacity Blocks for ML, US East (Ohio) and Asia Pacific (Hyderabad) | $35.7608 per instance-hour ($2.235 per chip-hour) | October 2026[16] |

AWS updates Capacity Block prices periodically based on supply and demand; in October 2026 its pricing page said the next update was scheduled for January 2027.[16] Large customers such as Anthropic buy capacity under multi-year agreements rather than through public reservations.[12]

AWS claims Trn2 instances deliver 30 to 40 percent better price-performance than P5e and P5en instances (which use NVIDIA H200 GPUs), citing internal benchmark results.[1][52]

## Power and data center deployment

Trainium 2 chips draw about 500 W each, according to a SemiAnalysis estimate; AWS does not publish a TDP.[4] At that figure, a fully populated trn2.48xlarge with 16 chips draws roughly 8 kW from the accelerator subsystem alone before host CPU, memory, networking, and cooling overhead. SemiAnalysis estimated rack densities of about 27 kW for standard Trn2 server racks and about 24 kW for each of the two racks of a Trn2 UltraServer.[4]

Project Rainier's Indiana campus is expected to consume 2.2 gigawatts of electricity when complete, according to The New York Times.[30] AWS designed the campus for outside-air cooling: its data centers use no water for cooling from October to March and, on an average day from April to September, use cooling water for only a few hours.[6][26] AWS also chose copper rather than optical cabling for NeuronLink, which SemiAnalysis tied to lower failure and link-flapping rates.[4] Trainium 3, at a 3 nm process, was disclosed at re:Invent 2025 as delivering four times better performance per watt than Trn2 UltraServers.[13]

## What is the successor to Trainium 2?

AWS announced Trainium 3 at AWS re:Invent 2024, one year after announcing Trainium 2, establishing an annual chip cadence.[9] Trainium 3 reached general availability in Trn3 UltraServers at AWS re:Invent 2025 on December 2, 2025.[13]

Trainium 3 is built on TSMC's 3 nm class N3P process, compared with the N5 node used for Trainium 2.[14][15] Per-chip specifications as disclosed by AWS include:

- 2.52 PFLOPs of FP8 compute per chip (approximately 1.9x the Trainium 2 dense FP8 figure)[13]
- 144 GB of HBM3e memory (1.5x Trainium 2)[13][15]
- 4.9 TB/s HBM bandwidth (1.7x Trainium 2)[13]

The Trn3 UltraServer scales to 144 chips, delivering 362 FP8 PFLOPs, 20.7 TB of HBM3e, and 706 TB/s of aggregate memory bandwidth.[13] AWS claims Trn3 delivers 4.4 times higher performance and 3.9 times higher memory bandwidth than Trn2 UltraServers, with four times better performance per watt.[13] In his 2025 shareholder letter, Jassy noted that Trainium 3 "just started shipping at the start of 2026 and is 30-40% more price-performant than Trainium2," and was already "nearly fully-subscribed."[27]

At the same re:Invent 2025 event, AWS previewed Trainium 4, saying it is being designed to deliver at least six times the FP4 processing performance, three times the FP8 performance and four times the memory bandwidth of Trainium 3, and to support NVIDIA NVLink Fusion.[40] AWS gave no Trainium 4 timeline at re:Invent.[14] In February 2026, Amazon said Trainium 4 was expected to begin delivery in 2027, and Jassy wrote in April 2026 that it was "still about 18 months from broad availability" and that a significant chunk of it had already been reserved.[32][27] Trainium 2 capacity was still being added in 2026: Anthropic said significant new Trainium 2 capacity would come online in the second quarter of 2026, and its April 2026 commitment covers Trainium 2 alongside the newer generations.[12]

## See also

- [AWS Trainium](https://aiwiki.ai/wiki/aws_trainium)
- [AWS Trainium 3](https://aiwiki.ai/wiki/aws_trainium_3)
- [Anthropic-Amazon Trainium expansion](https://aiwiki.ai/wiki/anthropic_amazon_trainium)
- [Project Rainier](https://aiwiki.ai/wiki/aws_project_rainier)
- [Amazon Bedrock](https://aiwiki.ai/wiki/amazon_bedrock)
- [Anthropic](https://aiwiki.ai/wiki/anthropic)
- [AWS Inferentia](https://aiwiki.ai/wiki/aws_inferentia)
- [AI chips](https://aiwiki.ai/wiki/ai_chip)
- [NVIDIA H200](https://aiwiki.ai/wiki/nvidia_h200)
- [Google TPU](https://aiwiki.ai/wiki/tpu)
- [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services)
- [Apple](https://aiwiki.ai/wiki/apple)
- [OpenAI](https://aiwiki.ai/wiki/openai)

## References

1. AWS News Blog. "Amazon EC2 Trn2 Instances and Trn2 UltraServers for AI/ML training and inference are now available." December 3, 2024. https://aws.amazon.com/blogs/aws/amazon-ec2-trn2-instances-and-trn2-ultraservers-for-aiml-training-and-inference-is-now-available/
2. AWS Neuron Documentation. "Trainium2 Architecture." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/arch/neuron-hardware/trainium2.html
3. Amazon EC2. "Gen AI Compute Instance - Amazon EC2 Trn2 Instances." https://aws.amazon.com/ec2/instance-types/trn2/
4. SemiAnalysis. "Amazon's AI Self Sufficiency | Trainium2 Architecture & Networking." December 3, 2024. https://newsletter.semianalysis.com/p/amazons-ai-self-sufficiency-trainium2-architecture-networking
5. AWS What's New. "Amazon EC2 Trn2 instances are generally available." December 3, 2024. https://aws.amazon.com/about-aws/whats-new/2024/12/amazon-ec2-trn2-instances-available/
6. About Amazon. "AWS activates Project Rainier: One of the world's largest AI compute clusters." October 29, 2025. https://www.aboutamazon.com/news/aws/aws-project-rainier-ai-trainium-chips-compute-cluster
7. Data Center Dynamics. "AWS activates Project Rainier cluster of nearly 500,000 Trainium2 chips." October 30, 2025. https://www.datacenterdynamics.com/en/news/aws-activates-project-rainier-cluster-of-nearly-500000-trainium2-chips/
8. Anthropic. "Claude 3.5 Haiku on AWS Trainium2 and model distillation in Amazon Bedrock." December 3, 2024. https://claude.com/blog/trainium2-and-distillation
9. TechCrunch. "AWS' Trainium2 chips for building LLMs are now generally available, with Trainium3 coming in late 2025." December 3, 2024. https://techcrunch.com/2024/12/03/aws-trainium2-chips-for-building-llms-are-now-generally-available-with-trainium3-coming-in-late-2025/
10. AWS Neuron Documentation. "Trainium2 Architecture Guide for NKI." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/nki/guides/architecture/trainium2_arch.html
11. About Amazon. "Amazon and Anthropic expand strategic collaboration." April 20, 2026. https://www.aboutamazon.com/news/company-news/amazon-invests-additional-5-billion-anthropic-ai
12. Anthropic. "Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute." April 20, 2026. https://www.anthropic.com/news/anthropic-amazon-compute
13. AWS What's New. "Announcing Amazon EC2 Trn3 UltraServers for faster, lower-cost generative AI training." December 2, 2025. https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/
14. TechCrunch. "Amazon releases an impressive new AI chip and teases an Nvidia-friendly roadmap." December 2, 2025. https://techcrunch.com/2025/12/02/amazon-releases-an-impressive-new-ai-chip-and-teases-a-nvidia-friendly-roadmap/
15. SemiAnalysis. "AWS Trainium3 Deep Dive | A Potential Challenger Approaching." December 4, 2025. https://newsletter.semianalysis.com/p/aws-trainium3-deep-dive-a-potential
16. Amazon EC2. "Amazon EC2 Capacity Blocks for ML Pricing." Accessed October 2026. https://aws.amazon.com/ec2/capacityblocks/pricing/
17. AWS Trainium Customers. https://aws.amazon.com/ai/machine-learning/trainium/customers/
18. AWS Neuron SDK GitHub. https://github.com/aws-neuron/aws-neuron-sdk
19. AWS What's New. "Announcing AWS Neuron SDK 2.26.0." September 19, 2025. https://aws.amazon.com/about-aws/whats-new/2025/09/aws-neuron-2-26-announce/
20. AWS What's New. "Announcing AWS Neuron SDK 2.27.0." December 22, 2025. https://aws.amazon.com/about-aws/whats-new/2025/12/announcing-aws-neuron-2-27/
21. AWS Neuron Documentation. "What's New in the AWS Neuron SDK." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/whats-new.html
22. AWS News Blog. "AWS Weekly Roundup: Project Rainier online, Amazon Nova Web Grounding, MCP Proxy for AWS, and more (November 3, 2025)." November 3, 2025. https://aws.amazon.com/blogs/aws/aws-weekly-roundup-project-rainier-online-amazon-nova-amazon-bedrock-and-more-november-3-2025/
23. TechCrunch. "An exclusive tour of Amazon's Trainium lab, the chip that's won over Anthropic, OpenAI, even Apple." March 22, 2026. https://techcrunch.com/2026/03/22/an-exclusive-tour-of-amazons-trainium-lab-the-chip-thats-won-over-anthropic-openai-even-apple/
24. Tom's Hardware. "Amazon invests $50 billion in OpenAI, committing to 2 gigawatts of Trainium silicon." https://www.tomshardware.com/tech-industry/amazon-invests-50-billion-in-openai
25. Constellation Research. "AWS fires up Project Rainier, Trainium2 cluster for Anthropic." October 29, 2025. https://www.constellationr.com/blog-news/insights/aws-fires-project-rainier-trainium2-cluster-anthropic
26. Data Centre Magazine. "AWS: How 500,000 Trainium2 Chips Power Project Rainier." November 3, 2025. https://datacentremagazine.com/news/aws-how-500-000-trainium2-chips-power-project-rainier
27. Amazon. "Amazon CEO Andy Jassy's 2025 Letter to Shareholders." April 9, 2026. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-2025-letter-to-shareholders
28. Amazon. "Amazon.com Announces Third Quarter Results." October 30, 2025. https://ir.aboutamazon.com/news-release/news-release-details/2025/Amazon-com-Announces-Third-Quarter-Results/default.aspx
29. The Motley Fool. "Amazon (AMZN) Q3 2025 Earnings Call Transcript." October 31, 2025. https://www.fool.com/earnings/call-transcripts/2025/10/31/amazon-amzn-q3-2025-earnings-call-transcript/
30. The New York Times. "At Amazon's Biggest Data Center, Everything Is Supersized for A.I." June 24, 2025. https://www.nytimes.com/2025/06/24/technology/amazon-ai-data-centers.html
31. Semafor. "AWS' mega multistate AI data center is powering Anthropic's Claude." October 29, 2025. https://www.semafor.com/article/10/29/2025/aws-massive-multi-state-ai-data-center-is-powering-anthropics-claude
32. Amazon. "OpenAI and Amazon Announce Strategic Partnership." February 27, 2026. https://press.aboutamazon.com/2026/2/openai-and-amazon-announce-strategic-partnership
33. MacRumors. "Apple Uses Amazon's Custom AI Chips for Search Services." December 3, 2024. https://www.macrumors.com/2024/12/03/apple-amazon-ai-chips-search/
34. About Amazon. "Amazon completes $4B Anthropic investment to advance generative AI." March 27, 2024. https://www.aboutamazon.com/news/company-news/amazon-anthropic-ai-investment
35. Anthropic. "Powering the next generation of AI development with AWS." November 22, 2024. https://www.anthropic.com/news/anthropic-amazon-trainium
36. Amazon. "AWS Unveils Next Generation AWS-Designed Chips." November 28, 2023. https://press.aboutamazon.com/2023/11/aws-unveils-next-generation-aws-designed-chips
37. TechCrunch. "AWS launches Trainium, its new custom ML training chip." December 1, 2020. https://techcrunch.com/2020/12/01/aws-launches-trainium-its-new-custom-ml-training-chip/
38. Amazon. "AWS Announces General Availability of Amazon EC2 Trn1 Instances Powered by AWS-Designed Trainium Chips." October 10, 2022. https://press.aboutamazon.com/2022/10/aws-announces-general-availability-of-amazon-ec2-trn1-instances-powered-by-aws-designed-trainium-chips
39. The Register. "Amazon promises 4x faster AI silicon in 2025, turns Trainium2 loose on the net." December 3, 2024. https://www.theregister.com/2024/12/03/amazon_ai_chip/
40. About Amazon. "Trainium3 UltraServer delivers faster AI training at lower cost." December 2, 2025. https://www.aboutamazon.com/news/aws/trainium-3-ultraserver-faster-ai-training-lower-cost
41. AWS Neuron Documentation. "NeuronCore-v3 Architecture." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/arch/neuron-hardware/neuron-core-v3.html
42. AWS Neuron Documentation. "Logical NeuronCore configuration." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/arch/neuron-features/logical-neuroncore-config.html
43. AWS Neuron Documentation. "Trainium Architecture." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/arch/neuron-hardware/trainium.html
44. AWS Neuron Documentation. "Announcing Transition to PyTorch Native Support for AWS Trainium in the Next Neuron Release Supporting PyTorch 2.10." December 16, 2025. https://awsdocs-neuron.readthedocs-hosted.com/en/latest/about-neuron/announcements/neuron2.x/announce-transition-pytorch-trainium.html
45. AWS Neuron Documentation. "Release Notes for Neuron Component: NKI Library." February 26, 2026. https://awsdocs-neuron.readthedocs-hosted.com/en/v2.28.0/release-notes/components/nki-lib.html
46. AWS Neuron Documentation. "Neuron XLA pluggable device (libneuronxla) release notes." https://awsdocs-neuron.readthedocs-hosted.com/en/latest/release-notes/archive/libneuronxla.html
47. AWS Neuron Documentation. "Neuron Kernel Interface (NKI) release notes." https://awsdocs-neuron.readthedocs-hosted.com/en/v2.27.0/nki/nki_rn.html
48. NVIDIA. "NVIDIA H100 Tensor Core GPU." https://www.nvidia.com/en-us/data-center/h100/
49. NVIDIA. "NVIDIA H200 Tensor Core GPU." https://www.nvidia.com/en-us/data-center/h200/
50. NVIDIA. "NVIDIA HGX Platform." https://www.nvidia.com/en-us/data-center/hgx/
51. Amazon Bedrock User Guide. "Optimize model inference for latency." https://docs.aws.amazon.com/bedrock/latest/userguide/latency-optimized-inference.html
52. Amazon EC2. "Amazon EC2 P5 Instances." https://aws.amazon.com/ec2/instance-types/p5/


---

## Live data: Trainium2 rental prices and specs

> Generated from AI Compute Tracker (https://aicomputetracker.com/reference-prices.json), not part of the article text. Quotes collected 2026-10-20. AI Compute Tracker is a companion site owned and operated by the same team as AI Wiki. Data: AI Compute Tracker (aicomputetracker.com), CC BY 4.0; third-party terms apply to data from Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Cloud Infrastructure, GPU cloud pricing pages, Beam, Chip vendor documentation.

- Trainium2 spec strip (reviewed against vendor documents 2026-10-15): Dense BF16 667 TFLOPS; Dense FP8 1,299 TFLOPS; Memory 96 GiB; Bandwidth 2.9 TB/s; TDP 500 W.

More: https://aicomputetracker.com/gpu-rental-prices/trainium2/