CCIR tier rates are productivity-adjusted reference rates. There is no clean throughput number to settle the question — usable compute per dollar depends on workload, framework, and scale efficiency — so CCIR does what commodity markets do when output is hard to measure: grade the inputs. Crude oil prices light sweet, medium, and heavy sour as distinct feedstock grades because refinery yield is hard to standardize but feedstock is observable. Compute is the same: fabric, host topology, and operational bundle are observable; effective throughput is not. MLPerf submissions on identical silicon show realized throughput varies materially with system configuration — fabric, host topology, and operational scaffolding — which is what the tier construct grades.
Top-grade fabric, dedicated single-tenant bare-metal, full enterprise bundle — managed operations, tiered support, observability, egress allowance, regulated-workload compliance. The rate buys capability that translates directly into useful compute per dollar at production scale.
Dedicated single-tenant bare-metal with disclosed fabric, basic support, no managed-ops wrapper. Where most production training spend lands. The middle grade — the rate is lower, but each $/GPU-hr buys less of the bundle that turns silicon into output.
Co-located or marketplace clearing on variable Ethernet fabric. Self-serve, no support tier, interruptibility common. The lowest rate band — and the lowest effective throughput per nominal $/GPU-hr once interrupt risk is priced in.
| Attribute | T1IF | T2IF | T3IF |
|---|---|---|---|
| Tenancy | Dedicated bare-metal, single-tenant | Dedicated bare-metal, single-tenant | Shared-pool or marketplace |
| Fabric | IB-NDR or eq. (3.2 Tbps RDMA, 8×400 Gbps) | IB-HDR / RoCE-v2 / EFA / mixed disclosed | Ethernet / variable / undisclosed |
| Bundle depth | Managed ops, tiered support, observability, egress allowance, compliance | Basic support, no managed ops | Self-serve, no support tier |
| Typical operators | AWS · Azure · OCI · GCP · CoreWeave · Nebius | Lambda OD · RunPod Secure · Crusoe · Hyperstack | RunPod Community · Vast.ai · Together MoE |
| Use case | Production training, regulated workloads | Dedicated training, less ceremony | Cost-sensitive batch, dev/test, spot |
Each tier is a capability bundle. The matrix describes what the rate at each tier actually packages — fabric, host, bundle depth, typical operators, use case.
Median rate per chip across each Intelligence Factory tier. n is the number of independent operators contributing to that median. An em-dash means the cell has no Published row in today's snapshot.
| Silicon | T1IF Tier 1 IF | T2IF Tier 2 IF | T3IF Tier 3 IF | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Median $/hr | Trend | n | Median $/hr | Trend | n | Median $/hr | Trend | n | |
| GB200 | $13.52 | 3 | — | — | — | — | |||
| B200 | $8.60 | 3 | $6.59 | 5 | $3.77 | 4 | |||
| H200 | $7.11 | 4 | $4.79 | 5 | $3.44 | 7 | |||
| H100 | $6.52 | 4 | $3.59 | 7 | $2.69 | 9 | |||
| A100 | — | — | $1.99 | 3 | $1.07 | 6 | |||
The same chip can trade across all three tiers. An H100 SXM in a Tier 1 hyperscale deployment, a Tier 2 dedicated cluster, and a Tier 3 marketplace listing all carry identical silicon at fundamentally different rate bands — because the deployment is the product. A buyer can nominally pay less at T3IF and receive less usable compute per dollar after interrupt risk, support gaps, and operational overhead are priced in.
Tier assignment is per-SKU and grade-based — driven by the capability bundle each SKU exposes, not by which operator hosts it. An interruptible SKU from any operator routes to T3IF; a dedicated-host SKU with disclosed IB-class fabric and full enterprise bundle routes to T1IF; everything in between routes to T2IF. The same operator can therefore appear in two or three tiers across its SKU catalog, depending on what each SKU actually packages.