AWS and Cerebras announced a multi-year collaboration on March 13, 2026, to bring Cerebras CS-3 systems into AWS data centers and offer Cerebras-powered inference through Amazon Bedrock. Their “5×” figure describes expected high-speed token capacity in a particular hardware configuration—not a promise that every model or request will respond five times faster. The proposed Trainium 3–Cerebras service is also distinct from the partnership announcement: public materials reviewed through August 16, 2026, did not establish that the full disaggregated service was generally available to all AWS customers.
What AWS and Cerebras announced
The companies described a strategic, multi-year infrastructure collaboration, not a disclosed acquisition or a publicly priced customer contract. Cerebras CS-3 systems are planned for deployment in AWS data centers, with access intended through Amazon Bedrock. The announcement described support for leading open-source large language models and Amazon Nova models, but it did not provide a definitive model, Region, or pricing list.
A separate part of the collaboration is a disaggregated inference design that pairs AWS Trainium 3 with Cerebras CS-3. AWS is described as the first cloud provider for Cerebras’ disaggregated approach. The companies have not disclosed a deal value, minimum purchase commitment, exclusivity arrangement, or revenue split. See the AWS announcement and Cerebras announcement.
How the Trainium–Cerebras design works
The proposed system assigns the two main phases of large language model inference to different hardware. In prefill, the model processes the prompt and builds its key-value cache; this phase is often compute-intensive. In decode, the model generates output tokens sequentially, repeatedly accessing model state.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
- Trainium 3 handles prefill: It processes the input prompt and creates the state needed for generation.
- AWS Elastic Fabric Adapter carries data between systems: The networking layer connects the prefill and decode stages.
- Cerebras CS-3 handles decode: It generates the response tokens.
This division is intended to use each system for a different stage of serving. It is the companies’ described architecture, not an independently validated production benchmark. Cerebras explains the design in its disaggregated inference overview.
What the “5×” claim means—and what it does not
Cerebras describes the expected gain as roughly five times more high-speed token capacity in the same hardware footprint, or a throughput advantage for the disaggregated design compared with an aggregated arrangement. That is primarily a claim about capacity and token throughput under particular architecture and workload assumptions. It is not evidence of a universal fivefold reduction in response time.
Rank #2
- High-Performance AI Processing: The MX3 is designed to handle the most demanding AI computer vision workloads, delivering exceptional performance and efficiency.
- Flexible Integration: The MX3 can be easily integrated into your existing systems via its M.2 M-key form factor and support for Linux operating systems.
- Energy Efficient: The MX3 is designed to provide high performance while minimizing power consumption.
- Comprehensive Software Development Kit (SDK): The MX3 is supported by a comprehensive SDK that simplifies development and deployment.
- Hardware compatability: The MX3 is compatible with the PCI-SIG M.2 M-key 2280 Specification. It can be used with the Raspberry Pi 5 with a M-key 2280 HAT.
These measures answer different questions:
- Time to first token (TTFT): How long a user waits before generation starts.
- Inter-token latency: How long the user waits between successive output tokens.
- Tokens per second for one request: The generation rate experienced by an individual request.
- Aggregate tokens per second: Total output across many concurrent requests.
- Capacity per hardware footprint: How many sessions or how much throughput can fit within a physical and power envelope.
- Tokens per second per watt: An energy-efficiency measure, not a direct measure of user-perceived speed.
A system can increase aggregate throughput substantially while making little difference to a single low-concurrency request. Conversely, a workload dominated by token generation may benefit more than one dominated by prompt processing. The public announcement does not provide a complete independent benchmark methodology or establish that the stated 5× result applies to every model and workload.
Who could benefit most
The design is most relevant where fast, sustained generation and concurrent demand matter. Potential fits include:
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
- Interactive coding assistants and agent loops that make multiple sequential model calls.
- Voice and conversational applications where pauses are noticeable.
- Customer-service systems with many simultaneous sessions.
- Real-time search and retrieval-augmented generation, when model generation is a significant part of the delay.
- High-volume applications that need greater aggregate output capacity.
It may be less compelling for offline batch summarization, prompt-heavy tasks where prefill dominates, short responses where application or network overhead is the bottleneck, and small workloads that cannot keep the hardware busy. Unsupported models and strict single-Region requirements can also limit fit.
Availability, models, and access paths
The announcement of a collaboration is not the same as a generally available service. Cerebras’ Q1 2026 materials described the joint Trainium 3/CS-3 strategy as planned and treated Bedrock availability as a future milestone. Public information reviewed through August 16, 2026, did not establish broad general availability for the full disaggregated service. That does not rule out limited access or subsequent changes; customers should verify the live offering rather than infer availability from the announcement.
Rank #4
- 48GB AI graphics accelerator
The intended AWS customer path is Amazon Bedrock, not necessarily a self-service EC2 instance that a customer can provision directly. Cerebras also operates its own direct inference service, which is separate from Bedrock. Model availability, endpoint and API compatibility, Region, lifecycle status, and account access can differ. AWS maintains the current Bedrock model catalog and endpoint availability information; do not assume every model named in a broad announcement is currently selectable in every Region.
For workloads with residency constraints, check the routing mode as well as the model’s listed Region. Bedrock documents In-Region, geographic cross-Region, and global cross-Region routing, which have different capacity and data-residency implications. Review the current Region compatibility and routing documentation before deployment.
Best Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
How to evaluate it for a real application
Do not choose a serving option on a headline throughput figure alone. Compare it with the model, prompts, concurrency, and service conditions your application actually uses.
- Match the workload: Use representative prompt lengths, output lengths, model features, tool calls, and concurrency.
- Measure streaming behavior: Record TTFT and inter-token latency separately, alongside total completion time.
- Measure the tails: Compare P50, P95, and P99 latency, not only averages, and track errors, timeouts, and throttling.
- Check quality and functionality: Test answer quality and tool-calling correctness on the same task set.
- Calculate complete economics: Include input and output tokens, service tier or capacity commitments, networking, storage or retrieval, observability, and migration and engineering work.
- Confirm operating constraints: Verify model and version, Region and routing, quotas, data residency, endpoint and API support, and lifecycle status.
Bedrock offers Standard, Flex, Priority, and Reserved inference tiers. Their suitability and price depend on the model and current terms; AWS describes Flex for workloads that can tolerate longer processing, Priority as a higher-priority option with a price premium, and Reserved as capacity commitments. Check the service-tier documentation and current Bedrock pricing for the selected model. A higher throughput figure alone does not establish lower cost per completed task.
How it compares with other options
| Option | Best suited to | Main advantage | Main trade-off |
|---|---|---|---|
| Bedrock with the Trainium–Cerebras design | AWS-native, latency-sensitive applications, if the service and required model are available | Managed access paired with specialized hardware for different inference phases | Availability, pricing, supported models, and workload performance need verification |
| Other models through Bedrock | Teams seeking managed APIs, AWS governance, and model choice | Common AWS access and a changing model catalog | Performance, pricing, and Region coverage vary by model and tier |
| Cerebras Inference Cloud | Developers evaluating direct access to Cerebras inference | A separate route to Cerebras’ hosted service | Its model catalog, Regions, quotas, pricing, and enterprise controls may differ from Bedrock |
| Amazon SageMaker AI | Teams deploying custom models or needing more endpoint control | Greater deployment customization | More infrastructure and MLOps responsibility than a managed model API |
| Self-managed GPU infrastructure | Teams needing CUDA compatibility, custom kernels, or low-level deployment control | Broad tooling and flexibility | Customers manage serving, drivers, utilization, scaling, and capacity |
For a managed model API, start with Amazon Bedrock. For direct Cerebras access, check Cerebras’ service page for current models and terms. Teams that need custom deployments can compare SageMaker AI with the AWS Bedrock-versus-SageMaker guide; teams seeking self-managed GPU options can review AWS EC2 P5 instances.
What remains to be established
The public material cited here does not establish a general-availability date, initial Regions, exact Bedrock model IDs, per-token prices, capacity limits, or support for fine-tuned and custom models. It also does not provide an independent, apples-to-apples benchmark showing the 5× result at realistic concurrency across representative prompts and models. AWS says most Bedrock inference runs on Trainium and has described Trainium 3’s expected price-performance relative to Trainium 2; those are AWS statements, not independent validation. More about AWS’s chip strategy appears in its Q1 2026 discussion of its chips business.
Free tools Windows power users keep installed
One-click scans. No signup required.
For investors and procurement teams, rollout speed and customer adoption matter alongside the architecture. Cerebras’ Q1 2026 materials and SEC filing discuss strategic customers, data-center capacity needs, and customer concentration risks. Those disclosures are relevant execution considerations, not proof that the proposed service will meet a particular performance or cost target.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
