Provider record · updated 2026-09-10
Together AI (GPU Clusters) — GPU pricing & deliverability
12 tracked offerings · grade mix A:0 · B:2 · C:0 · D:6 · U:4. GPU Clusters page publishes 'high-speed InfiniBand' interconnect for all clusters, managed Kubernetes or Slurm-on-Kubernetes orchestration, on-demand with no commitment, and reserved terms up to 6 months with upfront payment. Capacity described as 25+ cities: USA (600MW), Europe (150MW+), Asia and Middle East. Also listed reserved-only with contact-sales pricing: HGX B300 (270 GB), GB200 NVL72 (186 GB, 512 to 1,000+ GPUs), GB300 NVL72 (288 GB) — not recorded as offerings since no public price. No A100, L40S, or MI300X on the clusters page. InfiniBand generation/bandwidth (e.g. NDR 3.2Tbps) not specified on fetched page.
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H100 | H100 (reserved) US, EU, Asia +1 more | $3.19 | reserved | D | publ | High-speed InfiniBand | ↗vendor |
| H100-SXM | Reserved H100, 31-90 days (31-90 days) | $3.45 | reserved · min: 31-90 days | U | unkn | not published | ↗vendor |
| H100-SXM | Reserved H100, 7-30 days (7-30 days) | $3.69 | reserved · min: 7-30 days | U | unkn | not published | ↗vendor |
| H200 | H200 (reserved) US, EU, Asia +1 more | $3.99 | reserved | D | publ | High-speed InfiniBand | ↗vendor |
| H100 | H100 US, EU, Asia +1 more | $3.99 | on-demand | D | publ | High-speed InfiniBand | ↗vendor |
| H200-SXM | Reserved H200, 31-90 days (31-90 days) | $4.15 | reserved · min: 31-90 days | U | unkn | not published | ↗vendor |
| H200-SXM | Reserved H200, 7-30 days (7-30 days) | $4.99 | reserved · min: 7-30 days | U | unkn | not published | ↗vendor |
| H200 | H200 US, EU, Asia +1 more | $5.99 | on-demand | D | publ | High-speed InfiniBand | ↗vendor |
| B200 | HGX B200 (reserved) US, EU, Asia +1 more | $6.79 | reserved | B | publ | High-speed InfiniBand | ↗vendor |
| B200 | Reserved B200, 31-90 days (31-90 days) | $7.79 | reserved · min: 31-90 days | D | publ | High-speed InfiniBand | ↗vendor |
| B200 | Reserved B200, 7-30 days (7-30 days) | $7.99 | reserved · min: 7-30 days | D | publ | High-speed InfiniBand | ↗vendor |
| B200 | HGX B200 US, EU, Asia +1 more | $8.19 | on-demand | B | publ | High-speed InfiniBand | ↗vendor |
Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Terms & trust (as published, 2026-09-14)
| Uptime SLA | GPU Clusters: none published — Terms of Service state 'No guarantees are made with respect to the Services' quality, stability, uptime, or reliability, unless otherwise agreed between the parties in an order form.' Inference only: Provisioned Throughput page publishes 'Together targets a 99% monthly success rate for eligible requests' (marketed as '99% uptime SLA'); official blog publishes contractual 99% (single-DC) / 99.9% (multi-DC) uptime SLAs for dedicated inference deployments — measured on: Inference only: eligible requests to the contracted model through the PTU endpoint within purchased capacity (monthly success rate); blog states uptime is 'measured at inference completion, not load balancer.' No SLA measurement basis published for GPU clusters (node/cluster) |
| SLA remedy | none stated — no credit/refund/remedy terms published on provider pages for either PTU or GPU clusters; ToS says fees are non-refundable; any SLA remedies live in unpublished order forms |
| Capacity guarantee | PTU: 'Reserved capacity means your traffic doesn't compete... Guaranteed tokens per minute' (send up to contracted normalized TPM; overage falls back to serverless best-effort with no availability guarantee). GPU Clusters: reserved capacity 'Locked in for your duration' (terms up to 6 months, up to 4,000+ GPUs reserved vs up to 256 GPUs on-demand); reserved credits charged upfront for the full duration once provisioned |
| Egress fees | Free — docs: 'Together AI does not charge for data transfer into or out of your cluster... You pay only for compute and storage'; Instant Clusters GA post: 'Data transfer: free egress and ingress' |
| Billing granularity | per-hour — GPU clusters priced and billed per GPU per hour (on-demand: pay only for time cluster is running; commitment options Hourly, 1-6 Days, 1 Week-3 Months; reserved deducted upfront for full duration); no per-second/per-minute billing published |
| Self-serve signup | partial |
| Security certifications | SOC 2 Type II and ISO 27001:2022 (badges on GPU Clusters page; SOC 2 Type 2 report announced on official blog, 'officially ISO 27001:2022 certified'); HIPAA compliance claimed (encryption in transit/at rest, audit logging, BAAs); GDPR referenced as supported; trust center at trust.together.ai |
Notes: Self-serve is yes for Instant Clusters (console/CLI/API/Terraform, 'Spin up a cluster in minutes', prepaid credit-card balance per ToS) but on-demand is capped at ~256 GPUs; larger/longer reserved capacity and dedicated inference require contacting sales — recorded as partial. Key nuance: Together's published SLA numbers (99% PTU success rate; 99%/99.9% single/multi-DC in blog) apply to INFERENCE offerings, not GPU cluster/dedicated compute rentals — for GPU clusters no uptime SLA is published and the ToS explicitly disclaims uptime guarantees absent an order form. PTU SLA excludes overage traffic and has a one-month minimum term. Support page offers tiered support (Silver/Gold) with response-time SLAs, not uptime SLAs. The '99.9% SLA enterprise tier with automatic service credits' claim circulating in third-party sources was not found on any together.ai page and is excluded. Instant Clusters storage $0.16/GiB-month.
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record (public status feed, last 90 days)
1 incident on Together AI (GPU Clusters)'s own status page in the trailing 90 days, collected 2026-09-15. Self-reported by the provider — a floor on true incident count, not a ceiling; it still beats the war-story thread you were going to read instead. High counts often reflect granular per-rack reporting, not worse infrastructure: read the incidents, not just the number.
| Date | Incident | Impact | Duration |
|---|---|---|---|
| 2026-07-03 | Scheduled network maintenance ↗ | — | — |
What Together AI (GPU Clusters) does not publish
- gpus_per_node
- intra_node NVLink (not explicitly stated for B200/H200/H100)
- InfiniBand generation/bandwidth
- form_factor
- exact reserved minimum term
- regions per SKU
Unpublished fields are recorded as unknown and grade conservatively. Sellers improve their grades by disclosing — that is the point.
Weekly context on every provider
The Delivered Compute Report tracks how these prices and grades move.
Subscribe free