AMD Instinct MI355X Kimi K3 Can Generate Up to $32B of Revenue per GigaWatt per Year

Summary

One utility gigawatt of MI355X capacity running Kimi K3 2.8T on vLLM generates up to $32.6 billion of token revenue per year at the 45 tok/s/user operating point InferenceX uses for agentic workloads. The figure comes from measured AgentX trace replays on an 8-GPU MI355X node (FP4 weights, MTP speculative decoding, TP8), priced at Kimi K3's OpenRouter rates ($3.00 input, $0.30 cached input, $15.00 output per million tokens) and sold at 60% utilization. Two cost profiles bracket the operator's outcome. Owning the fleet at hyperscaler volume costs $1.50 per GPU-hour all-in and leaves $16.5 billion of profit after a 30% model license fee, a 50.7% margin. Renting the same fleet at $3.00 per GPU-hour, the August 2026 price for a 3-year MI355X contract, leaves $10.3 billion, a 31.5% margin.

Key findings

1. 1

   $32.6B of revenue per GW-year. 478,469 MI355X GPUs fit in one all-in utility gigawatt at 2.09 kW per GPU (chip plus its share of node, network, and cooling). Each GPU produces 6,115 tok/s at the P90 45 tok/s/user point, which prices at $12.97 per GPU-hour gross and $7.78 after the 60% utilization haircut.

2. 2

   Compute is the smaller cost. At $1.50/GPU/hr the fleet costs $6.3B per year, less than the $9.8B paid to the model lab as a 30% license fee. Doubling the compute cost to $3.00/GPU/hr adds $6.3B of expense and cuts profit from $16.5B to $10.3B; the operator still keeps 31.5% of revenue.

3. 3

   The 45 tok/s/user point is the middle of the measured curve, not its peak. The MI355X vLLM frontier spans P90 interactivity from 10 to 116 tok/s/user. Loosening the target to 30 tok/s/user raises revenue to $42.8B per GW-year; tightening it to 60 tok/s/user drops revenue to $17.3B and pushes the rental profile to break-even.

4. 4

   Utilization break-even is low. The owned fleet covers compute at 16.5% utilization and the rented fleet at 33.0%. Every 10 points of utilization adds $5.4B of revenue and $3.8B of profit per GW-year under either cost profile.

5. 5

   Prefix caching carries the workload. The agentic traces are 99.2% input tokens and the vLLM server hits its GPU prefix cache 93.7% of the time at this point, so the blended sale price is $0.59 per million tokens even though output tokens list at $15.

Method

1. 1

   GPU-hours per GW-year = 1,000,000 kW / 2.09 kW per GPU x 8,760 h = 4.19 billion GPU-hours.

2. 2

   Revenue = $/GPU/hr gross x GPU-hours x utilization.

3. 3

   Compute expense = cost tier $/GPU/hr x GPU-hours, paid whether or not the fleet is busy.

4. 4

   Model license fee = revenue x 30%.

5. 5

   Profit = revenue - compute expense - license fee.

6. 6

   $/GPU/hr gross = tok/s/GPU x 3,600 x blended $/M tok, where the blended price weights input share (99.2%), cache hit rate (93.7%, cached input billed at $0.30) and output share (0.8%) at the interpolated point.

7. 7

   Interpolation: upper-left Pareto frontier on (P90 interactivity, tok/s/GPU), monotone cubic Hermite spline (Steffen 1990), no extrapolation.

Assumptions

Sources

• InferenceX profit estimator per gigawatthttps://inferencex.semianalysis.com/profit-estimator-per-gigawatt
• InferenceX AgentX dashboardhttps://inferencex.semianalysis.com/agentx
• OpenRouter Kimi K3 pricinghttps://openrouter.ai/moonshotai/kimi-k3
• SemiAnalysis AI Cloud TCO Modelhttps://semianalysis.com/ai-cloud-tco-model/