On this page · 12 sections
- What Amazon actually said on 29 October 2026
- Trainium is already on-premises: AWS AI Factories
- Trainium3 by the numbers
- Neuron SDK and model support
- Power, space and facilities
- The contract model: what to negotiate
- Worked planning example: sovereign inference live by Q4 2027
- India, EU and US notes
- Decision table: which route fits
- Implementation checklist
- How eCorpIT can help
- References
Summary. On Amazon's third-quarter results call on 29 October 2026, Andy Jassy said customers keep pressing Amazon to offer Trainium beyond AWS and that Amazon intends to do so late in 2027. Amazon has not said what form the offer takes (chips, servers or racks), who can buy, what it costs, or how Neuron software will be supported away from AWS. The idea is older: in April 2026 Jassy called selling racks to third parties quite possible, and in June AWS described talks with outside buyers as early stage. For on-prem and sovereign buyers, the more useful fact is that Trainium already runs in customer buildings. AWS AI Factories, announced on 2 December 2025, place Trn2 and Trn3 next to NVIDIA Blackwell instances in your own or a colocated facility, operated only by AWS staff, about 3 to 6 months after the site is handed over. A Trn3 UltraServer holds up to 144 Trainium3 chips, 20.7 TB of HBM3e and 362 MXFP8 PFLOPs. Our verdict: if you need sovereign Trainium capacity in 2027, scope an AI Factory now and treat outside-AWS Trainium as a 2028 option until Amazon publishes terms.
About this article. This is for CTOs, infrastructure buyers and AI platform leads planning on-premises or sovereign accelerator capacity for 2027. On 7 November 2026 we checked every date, figure and product name against Amazon's results release and call remarks, Jassy's April 2026 shareholder letter, AWS's AI Factories, Trainium and Trn3 pages, and the Neuron documentation. Press-only claims name the outlet; claims we could not open at source are left out. eCorpIT has not tested Trainium hardware, AWS pages change without notice, and no vendor paid for inclusion. This is general information, not procurement or legal advice. Researched with AI assistance and reviewed by our editorial team.
What Amazon actually said on 29 October 2026
Amazon reported results for the quarter ended 30 September 2026 on 29 October. AWS segment sales rose 39% to $46.0 billion, its fastest growth in nearly five years, at a $184 billion annualised run rate. The next day About Amazon published Jassy's remarks on custom chips. Next to the outside-AWS line, he said the chips business is growing at triple-digit rates, revenue commitments for Graviton and Trainium have each risen more than 20 times year over year, and customers using Amazon's custom silicon saved more than $2 billion in the quarter. He added that short-term capacity on prior-generation Trainium instances now sells at a premium to the first half of the year, and that inference makes up well over half of AWS's AI revenue.
Two earlier statements set this up. In his 9 April 2026 shareholder letter, Jassy wrote that selling racks to third parties was quite possible, that Trainium3 started shipping at the start of 2026 and was nearly fully subscribed, and that Trainium4 was about 18 months from broad availability. On 18 June 2026 TechCrunch reported that Peter DeSantis had told Bloomberg AWS was in talks to sell Trainium for other companies' data centres, talks Amazon called early stage.
| Claim | Source | Status on 7 November 2026 |
|---|---|---|
| Amazon intends to make Trainium available outside AWS late in 2027 | Jassy call remarks, About Amazon, 30 October 2026 | Amazon statement |
| Custom silicon saved customers more than $2 billion in Q3 2026 | Same page | Amazon statement |
| Graviton and Trainium revenue commitments each up more than 20 times | Same page | Amazon statement |
| Prior-generation Trainium short-term capacity sells at a premium | Same page | Amazon statement |
| Selling Trainium racks to third parties is quite possible | Jassy shareholder letter, 9 April 2026 | Amazon statement, no date given then |
| Trainium3 nearly fully subscribed; Trainium4 about 18 months out | Same letter | Amazon statement as of April 2026 |
| AWS in early talks to sell Trainium for other companies' data centres | DeSantis to Bloomberg, reported by TechCrunch, 18 June 2026 | Named reporting |
| Form, buyers, price, support and Neuron terms off AWS | None | Not announced |
The wording matters. Amazon says available outside of AWS; press headlines say selling chips. The offer could be chips for other clouds, complete servers, racks for enterprise halls, or something else. Nor has Amazon named a generation. Eighteen months from the April letter lands near October 2027, but Amazon has not linked Trainium4 to the outside-AWS plan.
Supply is the other unknown. Every Trainium placed outside AWS is one not added to AWS's own fleet, and Trainium3 was already nearly fully subscribed in April. Our reading, not Amazon's, is that early outside-AWS volume may be limited and favour large buyers.
Amazon is not the only company moving custom silicon beyond its own cloud. Oracle and Google announced on 27 October 2026 that Google TPUs will be available on Oracle Cloud Infrastructure in calendar 2027. We compare that route with GPUs in our guide to Google TPUs on OCI from 2027.
Trainium is already on-premises: AWS AI Factories
AWS AI Factories, announced on 2 December 2025, are dedicated AWS AI infrastructure deployed in customers' own data centres. You provide the building and power; AWS deploys and runs the rest. The FAQ sets out the operating model:
- Location. AWS deploys AI Factories only in your data centres, whether you own the facility or lease colocation space.
- Hardware. Accelerators are Trainium (Trn2, Trn3) and NVIDIA (P6-B200, P6-B300, P6e-GB200 and P6e-GB300 UltraServers), and you can mix types in one AI Factory. AWS plans to add NVIDIA Rubin and Vera Rubin once generally available.
- Operations. Only AWS personnel may operate the infrastructure and services. AWS can add nationality or security-clearance requirements for local staff.
- Control plane. The AI Factory connects to a parent AWS Region you choose and appears beside that Region's Availability Zones in the same console and APIs. AWS installs a private Direct Connect point of presence on site.
- Services. EC2, EKS, ECS, AWS Batch, SageMaker AI and Amazon Bedrock (through the Bedrock distributed inference engine), plus EBS, FSx for Lustre and S3 Express One Zone. NVIDIA Run:ai and NVIDIA AI Enterprise can also run there.
- Timeline. About 3 to 6 months once the data centre is ready and handed over, depending on configuration and component availability.
- Sovereignty. Training and inference stay inside the AI Factory perimeter unless you integrate Region services. Keys can stay outside AWS through the KMS External Key Store, and the Nitro System is designed with no operator access to instance memory.
- Tenancy. Both single-customer and trusted-community multi-tenant set-ups are supported.
AWS says AI Factories are built for government workloads at every classification level up to Top Secret, and named HUMAIN in Saudi Arabia as partner for an AI Zone of up to 150,000 AI chips, including NVIDIA GB300 GPUs, in a HUMAIN data centre. It also says Trainium4 will support NVIDIA NVLink Fusion.
So the 2027 choice is not Trainium or nothing until late 2027. It is a choice between three routes, summarised from the sources above.
| Factor | AWS AI Factory (available now) | Trainium outside AWS (late 2027) | Self-owned GPU cluster |
|---|---|---|---|
| Who operates | AWS personnel only | Not announced | You or your integrator |
| Accelerators | Trn2, Trn3, NVIDIA B200, B300, GB200, GB300 | Generation not announced | Any NVIDIA or AMD part you can source |
| Software | Neuron SDK, EC2, EKS, SageMaker AI, Bedrock | Neuron terms off AWS not announced | CUDA or ROCm stack you manage |
| Control plane | Parent AWS Region console and APIs | Not announced | Your own |
| Pricing | Tailored after a joint assessment | Not announced | Capex plus power, cooling, staff |
| Lead time | About 3 to 6 months after site handover | Late 2027 at the earliest | Vendor delivery plus integration |
| Main risk | Dependence on AWS operations and the parent Region | Terms, allocation and support unknown | Integration, staffing, utilisation |
Trainium3 by the numbers
Trainium3, AWS's first 3nm AI chip, is the generation it is shipping now. The figures below are from AWS's Trainium and Trn3 pages, and every comparison with Trn2 is AWS's own.
| Specification | Trainium3 chip | Trn3 UltraServer |
|---|---|---|
| HBM3e capacity | 144 GB | Up to 20.7 TB |
| Memory bandwidth | 4.9 TB/s (1.7x Trainium2) | 706 TB/s aggregate |
| Compute | Up to 2x Trainium2 MXFP8 throughput | Up to 362 MXFP8 PFLOPs |
| Scale-up | NeuronLink-v4, 2 TB/s per chip | Up to 144 chips via NeuronSwitch-v1 |
| Scale-out | Not applicable | Up to 28.8 Tbps EFA per UltraServer |
| Data types | FP32, BF16, MXFP8, MXFP4 | Same |
| Versus Trn2 UltraServer (AWS claim) | Not applicable | 4.4x performance, 3.9x memory bandwidth, 4x performance per watt |
| Serving efficiency (AWS claim) | Not applicable | Over 5x output tokens per megawatt at similar per-user latency |
Peak PFLOPs are ceilings, not tokens per second on your model, and these pages do not publish the workloads behind each comparison. Use them to shortlist, then test. Our Trainium3 versus NVIDIA inference cost guide covers how to read these claims against GPU benchmarks.
Neuron SDK and model support
Software decides whether Trainium works for your models. Positions below are as of 7 November 2026.
Release cadence. Neuron 2.33.0, released on 19 October 2026, is the current release and covers Inf1, Inf2, Trn1, Trn2 and Trn3. Components include TorchNeuron, TorchTitan Neuron, JAX NeuronX, vLLM Neuron and vLLM Omni Neuron; TorchNeuron and vLLM Neuron are both labelled Beta.
PyTorch. TorchNeuron registers Trainium through PyTorch's standard extension point for new hardware, and runs both eager mode and torch.compile. It entered public beta with Neuron 2.33.0 and supports Trn2 and Trn3. Amazon says existing PyTorch code runs on Trainium after one configuration change and no rewrite, that TorchNeuron works with TorchTitan and FSDP, and that it tests TorchNeuron against hundreds of thousands of models on every build. Amazon set the Apache 2.0 open-source release for PyTorch Conference North America on 20 October 2026; Jassy confirmed on 29 October that it had happened, and the GitHub repository carries the Apache-2.0 licence.
Serving. vLLM Neuron launched in July 2026. Its documentation includes model recipes for Llama 3, GPT-OSS, Qwen3-32B, Qwen3-VL, Qwen3-Embedding and Qwen3-ASR, plus tutorials for Llama 3.3 70B, EAGLE3 speculative decoding, disaggregated inference and a GPT-OSS 120B deployment on Trn3. AWS's Trainium page says vLLM, Hugging Face Transformers and TorchTitan run natively, with Ray, EKS and AWS Batch for orchestration.
Kernels. The Neuron Kernel Interface (NKI) gives instruction-level access to the chip, and AWS publishes an open-source NKI Library of optimised kernels. Custom CUDA kernels will not run on Trainium; plan to rewrite them in NKI or replace them with framework operators.
Bedrock inside an AI Factory. AWS says it works with third-party and closed-weight model providers to validate which models run there. Ask for the exact Bedrock model list for your perimeter; do not assume parity with a public Region.
The open question for outside-AWS Trainium. Neuron today is built for EC2: the runtime is described as the driver and libraries for Inferentia and Trainium instances, and the machine images and containers are AWS-specific. Amazon has not said how the compiler, runtime, drivers and monitoring will be supported on hosts AWS does not operate, or what scale-out fabric replaces EFA. TorchNeuron helps at the framework layer, but the lower layers decide whether a non-AWS Trainium rack is supportable in your building.
Power, space and facilities
With an AI Factory you supply the building and the power, so facilities become your critical path. AWS has not published per-rack power, cooling method or floor loading for Trn3 on its Trainium, Trn3 or AI Factories pages. It does say Trn3 racks carry over twice the chip density of Trn2, so expect higher power density per rack. AWS plans long-term scaling with you so your site has power and space for expansion, and the 3 to 6 month clock starts only at handover.
Ask AWS for these in writing during the site readiness assessment:
- Power per rack and per hall for your planned mix of Trn3 and NVIDIA UltraServers.
- Cooling method (air, liquid to the rack, or both) and water or coolant supply specifications.
- Floor loading, rack dimensions and cable routing for NeuronSwitch and EFA fabrics.
- Power quality, redundancy and generator requirements that AWS treats as preconditions for its service SLAs.
- Space reserved for later generations, including Trainium4 and NVIDIA Vera Rubin.
AWS's claim of more than 5x output tokens per megawatt over Trn2 matters most on power-constrained sites, where megawatts, not rack count, cap capacity. If you are also planning NVIDIA racks, our GTC 2027 buyer guide covers NVIDIA's 800 VDC and liquid-cooling direction.
The contract model: what to negotiate
AWS says AI Factory pricing depends on location, scale, accelerators, services and your existing infrastructure, and is shared only after a joint initial assessment. EC2 Capacity Reservations are available, services carry their service-specific SLAs, and newer generations can be added. Negotiate these terms:
- Term and commitment. Length, minimum spend, and whether reserved capacity is tied to an instance type or fungible across Trainium and GPU.
- Refresh rights. Get the commercial terms for adding newer generations, and what happens to older Trn2 or Trn3 racks.
- Facility versus service SLAs. AWS's SLAs cover its services; your facility's power and cooling sit with you or your colocation provider. Map the gaps.
- Parent Region dependency. Ask what keeps running, and for how long, if connectivity to the parent Region is lost.
- Exit. Decommissioning, hardware removal, media destruction, data return and the notice period.
- Personnel controls. Nationality, clearance and on-site access rules for AWS staff, written into the contract rather than a side letter.
For outside-AWS Trainium the questions are more basic: purchase or lease, warranty and RMA, firmware patching, Neuron licensing and support, the network fabric, allocation priority, and whether AWS will operate hardware you own.
Worked planning example: sovereign inference live by Q4 2027
A regulated organisation wants a 70-billion-parameter open-weight model serving inside its own facility by 1 October 2027. The figures are AWS's; the scenario is illustrative.
Memory sizing. At BF16 (2 bytes per parameter) the weights need about 140 GB; at MXFP8 (1 byte) about 70 GB; at MXFP4 (half a byte) about 35 GB, plus a few per cent for the MX formats' shared scale factors. A Trainium3 chip has 144 GB of HBM3e. At BF16 the weights would leave about 4 GB per chip (144 minus 140) for KV cache, so plan at least two chips per replica. At MXFP8 one chip holds the weights with about 70 GB left for KV cache and activations, subject to accuracy testing at that precision. A full Trn3 UltraServer's 20.7 TB is far more than one replica needs, so size by concurrent users and context length, not by model size alone.
Timeline, AI Factory route. AWS quotes about 3 to 6 months after handover. Counting back from 1 October 2027, the facility must be handed over between 1 April 2027 (if deployment takes 6 months) and 1 July 2027 (if it takes 3). From 7 November 2026, 1 April 2027 is 145 days away, about 21 weeks. Assessment, contract and any power or cooling upgrades must finish before handover, so start the AWS conversation now, or in Q1 2027 at the latest.
Timeline, outside-AWS route. Amazon says late 2027. Even if units ship in Q4 2027, your team must then rack, network, validate and harden them. Our working assumption, not Amazon's, is production use in 2028 at the earliest.
| Milestone | AI Factory route | Outside-AWS route |
|---|---|---|
| Vendor terms published | On request after joint assessment | Not announced |
| Contract and site assessment | Q4 2026 to Q1 2027 | After Amazon publishes terms |
| Facility ready and handed over | 1 April to 1 July 2027 | Your schedule |
| Hardware live | By 1 October 2027 | Delivery late 2027 at the earliest |
| Production serving | Q4 2027 | 2028, our estimate |
India, EU and US notes
India. The DPDP Rules, notified in November 2025, phase in over 18 months, so most obligations apply from 13 May 2027 (see our DPDP Rules checklist). Sovereign inference designed now will run under those rules. We found no AWS announcement of an AI Factory in India as of 7 November 2026; ask your AWS account team directly. Compare the economics with domestic GPU options in our note on GPU cloud pricing in India.
EU and UK. The data plane stays inside your perimeter, but management runs through a parent AWS Region. Have your legal team assess that control-plane link against your sovereignty requirements; external key storage and personnel nationality controls are the levers AWS offers.
US public sector. AWS says AI Factories are built for workloads at all classification levels up to Top Secret. Accreditation for a specific site remains a separate process.
Decision table: which route fits
| Pick this | If |
|---|---|
| AWS AI Factory with Trn3 | You need dedicated in-country capacity in 2027 and accept AWS-only operations |
| AWS AI Factory mixing Trn3 and NVIDIA GB300 | Some models depend on CUDA kernels and others run cleanly on Neuron |
| Wait for outside-AWS Trainium | Your need is 2028 or later and you must own or operate the hardware yourself |
| Trainium in a public AWS Region | An existing Region meets your residency needs and you want to test Neuron without capex |
| Self-owned or colocated GPU cluster | You need full operator control, multi-vendor freedom, or a CUDA-only software estate |
Implementation checklist
- Inventory models, precisions, context lengths and custom CUDA kernels; flag anything TorchNeuron or vLLM Neuron, both Beta, cannot run.
- Benchmark representative models on Trn2 or Trn3 in a public AWS Region before committing to on-prem capacity.
- Request the Bedrock model list for an AI Factory under your chosen parent Region.
- Run a facility gap analysis against AWS's power, cooling and floor-loading requirements.
- Put the six contract terms above into the term sheet.
- Ask AWS in writing what the late-2027 outside-AWS offer will include, and revisit at re:Invent 2026 (30 November to 4 December, Las Vegas) and at each quarterly results call.
- Keep a GPU fallback in the plan until Trainium tests meet your latency and quality targets.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our data centre and cloud networking team helps buyers compare AI Factory, public-cloud and self-owned accelerator routes, from portability checks and capacity sizing to site readiness and hybrid links to AWS Regions. Talk to us at /contact-us/.
Last updated: 7 November 2026.
