GPU model
AWS Trainium2 (Trn2): specs, VRAM and rental prices
AWS Trainium2 (Trn2): 96 GB of chip memory (HBM3), no public rental price listed. Specs from AWS, prices as of Oct 29, 2026.
Trainium2 is AWS’s second-generation machine-learning chip. AWS made Trn2 instances generally available on Dec 3, 2024, through EC2 Capacity Blocks for ML. A trn2.48xlarge instance holds 16 chips. The AWS Price List shows no on-demand price for it, so no rental price is listed here.
Prices were last checked Oct 29, 2026. A listing not re-checked for 7 days is greyed and left out of every lowest-price figure; after 14 days it is removed from the page.
AWS Trainium2 (Trn2) specifications
| Specification | Trainium2 (Trn2) | Source |
|---|---|---|
| Architecture | AWS Trainium2 | aws.amazon.com |
| GPU memory | 96 GB | aws.amazon.com |
| Memory type | HBM3 | aws.amazon.com |
| Memory bandwidth | 2,875 GB/s (derived: 46 TB/s ÷ 16 chips) | aws.amazon.com |
| FP16 / BF16 peak | Not published | aws.amazon.com |
| FP8 peak | 1,300 TFLOPS (derived: 20.8 PFLOPS ÷ 16 chips) | aws.amazon.com |
| Chip-to-chip interconnect | Not published | aws.amazon.com |
| Typical board power | Not published | aws.amazon.com |
AWS prints memory per chip for the 1-chip size (trn2.3xlarge) and compute and bandwidth for 16-chip instances. Figures marked derived divide a 16-chip total by 16. AWS publishes no chip power rating. Not published on the pages cited here: FP16 / BF16 throughput; Chip-to-chip bandwidth; Chip power.
Sources: AWS Trn2 instances page · Trn2 general availability announcement (Dec 3, 2024) · AWS Trainium page. Every figure above is copied from these AWS pages; nothing is estimated.
Trainium2 rental prices by provider
No provider currently lists a public per-GPU-hour price for the Trainium2 that the Census has been able to confirm, so there is no rental price to show. The specifications above are still AWS’s own.
Compare the Trainium2 with other GPUs
- Trainium2 vs H100: specs side by side with the lowest listed rental price for each.
- Trainium1 vs Trainium2: specs side by side with the lowest listed rental price for each.
- Trainium2 vs Trainium3: specs side by side with the lowest listed rental price for each.
Price history
The history chart appears once there are two daily snapshots of an on-demand price. Tracking began Oct 15, 2026; there is no on-demand snapshot for this GPU yet.
Embed this chart
<iframe src="https://datacentercensus.com/embed/gpu-prices/trainium2" title="AWS Trainium2 (Trn2) rental prices" width="100%" height="500" style="border:0;max-width:640px" loading="lazy"></iframe>
<p style="font:13px/1.4 system-ui,sans-serif;margin:6px 0 0">Source: <a href="https://datacentercensus.com/gpu-prices/trainium2">Data Center Census</a></p>
<script>addEventListener("message",function(e){if(e.origin==="https://datacentercensus.com"&&e.data&&e.data.dcc){document.querySelectorAll('iframe[src^="https://datacentercensus.com/embed/"]').forEach(function(f){if(f.contentWindow===e.source)f.style.height=e.data.h+"px"})}})</script>Free with the credit line. Preview the embed.
How these prices are collected
- What counts
- Only prices a provider publishes on a public pricing page or in its public price list, opened by the Census on the day shown. A marketplace is listed under the same rule as every other provider: it must publish per-GPU-hour prices publicly, with no login, and its rows are shown with the type Marketplace. Quotes and “contact sales” rows are not listed, and neither are purchase prices.
- $ per GPU-hour
- The listed price per GPU-hour. Where a provider lists a price per instance, the Census divides it by the GPUs in that instance and shows the original figure under it. Nothing else is converted or estimated. Google TPUs and AWS Trainium and Inferentia chips are priced per chip-hour on the same terms, with “chip” in place of “GPU”.
- Pricing types
- On-demand, spot (including tiers a provider calls preemptible or low-priority), reserved with the term the provider states, and marketplace. Reserved rows without a stated term are not listed. A marketplace row is the lowest per-GPU-hour price its public page shows for that GPU, which comes from third-party hosts and changes often. Marketplace rows are a separate type: they appear in the table, filters, downloads and feeds, but never in the “lowest on-demand” or “lowest spot” figures.
- Variants
- SXM, PCIe, NVL and similar versions are kept separate when the provider names them. A provider that does not name the variant is shown without one.
- Freshness
- Every listing carries the date it was last confirmed, and the daily run re-opens each one. If a listing goes 7 days without a re-check it stays on the page greyed, labelled with the date it was last confirmed, and drops out of the “lowest” figures and the calculator’s default results. At 14 days without a re-check it is taken off the site; the daily history still holds it.
- History
- Each daily run saves a snapshot, so the model pages can chart the lowest listed on-demand price over time and the change log can show what moved. History began Oct 15, 2026.
- Differences between providers
- Listed prices are not like for like: regions, minimum instance sizes, networking, storage, commitments and availability differ and are not priced here. Check the provider before relying on a figure.
Providers checked but not listed
- Vast.ai: Its pricing page loads live "from" prices from a public JSON file whose terms allow display and citation with attribution but require a data license for bulk collection, redistribution or use in any index, benchmark or derivative product. The Census republishes prices as open data, so it does not list Vast.ai without that license.
- Voltage Park: Its pricing pages say "Contact for pricing" for every GPU; one homepage line says "Starting at $1.99/hr" without a GPU model or unit.
Trainium2: common questions
How much VRAM does the Trainium2 have?
AWS lists 96 GB of GPU memory for the AWS Trainium2 (Trn2), which is HBM3, with memory bandwidth of 2,875 GB/s (derived: 46 TB/s ÷ 16 chips). The specification table on this page shows each figure with the page it came from.
How much does it cost to rent a Trainium2 per hour?
No provider currently lists a public per-chip-hour price for the Trainium2 that the Census has been able to confirm, so no rental price is shown.
Can I rent a Trainium2 by the hour?
Not at a public price the Census can confirm. No tracked provider lists a public per-chip-hour price for the AWS Trainium2 (Trn2) right now. The Census adds a price as soon as a provider publishes one on a public page or price list.
Which architecture is the Trainium2 built on, and when was it announced?
The AWS Trainium2 (Trn2) uses the AWS Trainium2 architecture and was announced or launched in 2024, according to the AWS page cited in the specifications.
Related: All GPU prices · GPU comparisons · Rental cost calculator · H100 · H200 · B200 · B300 · GB200 NVL72 · GB300 NVL72 · A100 · L40S · RTX PRO 6000 · L4 · T4 · MI300X · MI325X · MI355X · Trainium1 · Trainium3 · Inferentia2 · TPU v5e · TPU v5p · TPU v6e · Ironwood · MI450 · Vera Rubin · Rubin CPX · Rubin Ultra
Cite this page
Last updated Oct 29, 2026. Figures may change as records are updated — include the date you retrieved them.
| APA | Data Center Census. (2026). AWS Trainium2 (Trn2) specs and rental prices. Retrieved from https://datacentercensus.com/gpu-prices/trainium2 |
|---|---|
| MLA | “AWS Trainium2 (Trn2) specs and rental prices.” Data Center Census, Oct 29, 2026, https://datacentercensus.com/gpu-prices/trainium2. |
| Link | https://datacentercensus.com/gpu-prices/trainium2 |
Republishing a figure or table? Attribute it to Data Center Census with a link to this page. The underlying data is downloadable as JSON and CSV; the press kit has standing figures and contact details.