GPU Crack Spreads
A B200 serving DeepSeek V4.1-Flash makes $23.37 of tokens an hour at the prices hosts are actually paid, and costs $7.87 an hour to rent: a crack of +$15.50. Like a refiner’s crack spread, what the machine makes less what it runs on, for seven GPUs and three open models, with yields from SemiAnalysis InferenceX.
Crack per GPU-hour on DeepSeek V4.1-Flash, the most-used open model on OpenRouter: what the GPU’s benchmark output sells for at the prices hosts are paid, less its on-demand rent; “owning” sets the same tokens against SemiAnalysis’s cost of owning the GPU. Indicative, on token prices from Nov 20, 11:14 UTC. † InferenceX measured this GPU only at faster speeds than the target, so its yield, and crack, are if anything understated. ‡ Provisional rent: GB300 isn’t in the GPU index yet.
How the crack works
A refinery’s crack spread is what the products from a barrel of crude sell for, less the barrel. This tracker does the same for GPUs. The output of each GPU on each model comes from SemiAnalysis InferenceX, which benchmarks open models on every major GPU, continuously. Its agentic workload replays real coding-agent sessions, in which nearly all of each request’s input is already cached from earlier turns. We take the best configuration InferenceX has run for each GPU at the speed OpenRouter’s hosts actually deliver, 65 output tokens a second per user for V4.1-Flash. Each GPU’s yield is its input and output tokens a second and the share of input served from cache.
Those tokens are priced at what OpenRouter’s hosts are paid: each host’s prices, weighted by the requests it serves. The hour is priced at the GPU’s rent from our GPU Compute Price Index, on demand and, for commitments, on each provider’s own discount for the term. Token value less rent is the crack. A positive crack says renting the GPU and selling its tokens pays at the benchmark’s load. The break-even load says how busy the GPU must be to cover the rent. Power, networking and staff aren’t in it.
The board
Crack per GPU-hour against on-demand rent, with the load at which the GPU covers its rent. Where InferenceX has no result, the output a GPU would need to cover its rent, in output tokens a second, next to what a comparable GPU makes. Spot and commitment rents are in the curve below.
| GPU | Rent | V4.1-Flash | GLM-5.3 | Kimi K3 |
|---|---|---|---|---|
| H100H100 SXM | $3.82spot $2.49 | −$2.44†short of rent at full load | needs 40 tok/sno InferenceX result | needs 17 tok/sno InferenceX result |
| H200H200 SXM | $5.56spot $3.51 | −$2.88short of rent at full load | needs 58 tok/sB200 makes 298 | needs 25 tok/sB200 makes 73 |
| B200B200 HGX | $7.87spot $5.15 | +$15.50†covers rent at 34% load | +$19.88†covers rent at 28% load | +$8.10covers rent at 49% load |
| B300B300 HGX | $9.41spot $6.35 | +$69.82†covers rent at 12% load | +$15.51covers rent at 38% load | +$5.28covers rent at 64% load |
| GB200GB200 NVL72 | $16.00spot $6.22 (1 provider) · 3 providers | +$71.07†covers rent at 18% load | +$36.08†covers rent at 31% load | +$1.14covers rent at 93% load |
| GB300GB300 NVL72 | $16.35no spot price · provisional, 2 providers | +$175.51†‡covers rent at 9% load | +$71.84†‡covers rent at 19% load | −$14.21†‡short of rent at full load |
| MI300XAMD MI300X | $4.39spot $1.39 (1 provider) | −$3.20short of rent at full load | needs 46 tok/sno InferenceX result | needs 20 tok/sno InferenceX result |
Per GPU-hour, US dollars. Rent: the GPU index, global, all providers, on demand, with spot below; GB300 isn’t in the index yet, so its rent is provisional (Oracle Cloud $18.00 on-demand, Verda $14.71 1-month commitment, the median of the two). † Measured only at faster speeds than the target: understated. Download every daily fixing: gpu-crack-daily.csv.
What an hour makes, and what it costs
Token value per GPU-hour at the prices hosts are paid, against on-demand rent and SemiAnalysis’s cost of owning the GPU, for every GPU and model with an InferenceX result. Past the rent line, renting pays; past the ownership line, owning does.
The crack across the rent curve
DeepSeek V4.1-Flash: the crack per GPU-hour against each way of renting the GPU, against the cost of owning it, and the most a host could pay for it at 50% load. Commitment rents are on-demand rent times each provider’s own discount for the term, the median across providers; * where too few providers quote the term for this GPU, the discount across all GPUs.
| GPU | Spot | On-demand | 1 month | 1 year | 3 years | Owning | Pays at 50% load |
|---|---|---|---|---|---|---|---|
| H100†$1.38 of tokens | −$1.11on $2.49 | −$2.44on $3.82 | −$1.84on $3.22 | −$1.26on $2.64 | −$0.30on $1.68 | +$0.21on $1.17 | $0.69short of a 1-year lease |
| H200$2.68 of tokens | −$0.83on $3.51 | −$2.88on $5.56 | −$2.04on $4.72 | −$1.15on $3.83 | +$0.11on $2.57 | +$1.46on $1.22 | $1.34short of a 1-year lease |
| B200†$23.37 of tokens | +$18.23on $5.15 | +$15.50on $7.87 | +$15.51on $7.87 | +$17.75on $5.62* | +$19.87on $3.50* | +$21.64on $1.73 | $11.69covers a 1-year lease |
| B300†$79.22 of tokens | +$72.88on $6.35 | +$69.82on $9.41 | +$70.01on $9.22* | +$72.51on $6.72* | +$75.04on $4.18* | +$76.96on $2.26 | $39.61covers a 1-year lease |
| GB200†$87.07 of tokens | +$80.85on $6.22 | +$71.07on $16.00 | +$71.39on $15.68* | +$75.65on $11.42* | +$79.96on $7.11* | +$85.21on $1.86 | $43.54covers a 1-year lease |
| GB300†‡$191.87 of tokens | — | +$175.51on $16.35 | +$175.84on $16.03* | +$180.19on $11.68* | +$184.60on $7.27* | +$189.56on $2.31 | $95.93covers a 1-year lease |
| MI300X$1.20 of tokens | −$0.19on $1.39 | −$3.20on $4.39 | −$3.11on $4.31* | −$1.94on $3.14* | −$0.76on $1.95* | +$0.25on $0.95 | $0.60short of a 1-year lease |
Cost of owning: SemiAnalysis’s total cost of ownership per GPU-hour, as InferenceX publishes it. The last column is token value at 50% load, the rent a host could pay before running at a loss.
What the cracks say
Hosts are paid less than they post. Weighted by the requests each serves, V4.1-Flash sells for 10% more than the median host’s posted prices; the busiest host, DeepInfra, carries 15% of its requests.
On V4.1-Flash, B200, B300, GB200 and GB300 cover their rent; H100, H200 and MI300X don’t, even at full load at the speed hosts serve: an H100 makes $1.38 of tokens an hour against $3.82 of rent. Owning changes that: at SemiAnalysis’s cost of ownership the H100, H200 and MI300X earn +$0.21, +$1.46 and +$0.25 an hour. On those GPUs the lease takes the margin.
If renters compete the on-demand crack away, the break-even load is the utilization the market is pricing in: 34% for a B200, 12% for a B300 and 18% for a GB200 on V4.1-Flash.
Daily crack fixings
Crack per GPU-hour against on-demand rent at 16:00 UTC each day, from the first fixing on October 28. Each series follows a class of model, not one model, and moves to a successor by rule (marked). Each panel has its own scale.
What moved the V4.1-Flash crack since the previous fixing
| GPU | Crack | Token prices | Yield | Rent | Over 7 days |
|---|---|---|---|---|---|
| H100since Nov 18 | +$0.16 | +$0.06 | $0.00 | +$0.09 | +$0.14 |
| H200since Nov 18 | −$0.03 | +$0.13 | $0.00 | −$0.16 | +$0.28 |
| B200since Nov 18 | +$1.02 | +$1.11 | $0.00 | −$0.10 | +$2.09 |
| B300since Nov 18 | +$3.60 | +$4.11 | $0.00 | −$0.51 | +$53.01 |
| GB200since Nov 18 | +$4.33 | +$4.33 | $0.00 | $0.00 | +$71.92 |
| GB300since Nov 18 | +$5.94 | +$5.94 | $0.00 | $0.00 | +$11.81 |
| MI300Xsince Nov 18 | +$0.04 | +$0.04 | $0.00 | $0.00 | +$0.08 |
Yields over time
Output tokens a second per GPU from InferenceX’s best configuration at each model’s target speed, at every InferenceX run, log scale. The same GPU makes more as engines improve; the crack picks that up on the day InferenceX publishes it.
Token prices
Per million tokens. Realized: each host’s prices weighted by the requests it served in the 24 hours to the print, which is what the crack uses. Posted: the median host’s list, each host counted once at its cheapest standard endpoint. Speed: the p90 speed of hosts’ requests (90% reach it or beat it), weighted the same way, beside the target the yields are taken at.
| Model | Hosts | Input | Cached input | Output | Posted (median host) | First party | Speed |
|---|---|---|---|---|---|---|---|
| V4.1-Flash4.09M requests in 30 min | 43 | $0.139 | $0.00564 | $0.648 | $0.15 / $0.0036 / $0.65110% of posted, for this traffic | $0.15 / $0.003 / $0.60DeepSeek off-peak | 72 tok/starget 65 |
| GLM-5.3230K requests in 30 min | 36 | $0.636 | $0.144 | $2.80 | $1.05 / $0.155 / $4.1781% of posted, for this traffic | $1.40 / $0.26 / $4.40Z.ai list | 56 tok/starget 45 |
| Kimi K333K requests in 30 min | 20 | $1.77 | $0.252 | $14.27 | $2.92 / $0.30 / $15.0079% of posted, for this traffic | $3.00 / $0.30 / $15.00Moonshot list | 32 tok/starget 25 |
| GLM-5.3-Flashprices only: no InferenceX result | 35 | $0.111 | $0.0275 | $0.382 | $0.12 / $0.03 / $0.5090% of posted, for this traffic | $0.15 / $0.03 / $0.50Z.ai list | 53 tok/s |
Every V4.1-Flash host (43) · 1 left out as outliers
| Host | Share of requests | Input | Cached input | Output | Discount | Speed |
|---|---|---|---|---|---|---|
| DeepInfrafp8 | 15% | $0.098 | $0.00294 | $0.294 | 51% | 67 |
| InferenceNetfp8 | 12% | $0.08 | $0.008 | $0.666 | — | 88 |
| Wafermxfp4 | 12% | $0.036 | $0.01 | $0.50 | — | 78 |
| Togetherunknown | 7% | $0.30 | $0.006 | $1.20 | — | 114 |
| CoreWeavefp8 | 5% | $0.13 | $0.002 | $0.65 | — | 78 |
| Pearl Researchfp8 | 4% | $0.05 | $0.003 | $0.30 | — | 48 |
| Nebiusfp4 | 4% | $0.18 | $0.0036 | $0.72 | 40% | 87 |
| Coherefp8 | 4% | $0.115 | $0.0025 | $0.40 | — | 64 |
| Morphfp8 | 4% | $0.059 | $0.005 | $0.52 | — | 70 |
| Relacefp4 | 4% | $0.12 | $0.003 | $0.50 | — | 72 |
| Fireworksunknown | 3% | $0.15 | $0.003 | $0.60 | — | 60 |
| Alibabaunknown | 3% | $0.18 | $0.0036 | $0.72 | — | 64 |
| PrimeIntellectunknown | 3% | $0.0437 | $0.0024 | $0.70 | — | 66 |
| DigitalOceanunknown | 2% | $0.18 | $0.0036 | $0.72 | — | 78 |
| Baidufp8 | 2% | $0.30 | $0.006 | $1.20 | — | 28 |
| OpenInferencefp4 | 2% | $0.06 | $0.002 | $0.80 | — | 78 |
| Ionstreamunknown | 2% | $0.21 | $0.0105 | $0.62 | — | 84 |
| Riverunknown | 2% | $0.12 | $0.0024 | $0.48 | — | 60 |
| StepFunfp8 | 2% | $0.105 | $0.0021 | $0.42 | 65% | 51 |
| Particle AIunknown | 1% | $0.06 | $0.006 | $0.50 | — | 63 |
| Novitafp8 | 1% | $0.15 | $0.003 | $0.60 | 50% | 81 |
| Parasailfp8 | 1% | $0.30 | $0.006 | $1.20 | — | 91 |
| Rekaunknown | 1% | $0.12 | $0.008 | $0.64 | 20% | 60 |
| BaseTenunknown | 1% | $0.30 | $0.007 | $1.20 | — | 79 |
| Crusoefp8 | 1% | $0.29 | $0.007 | $1.20 | — | 71 |
| ModelRunfp4 | 1% | $0.17 | $0.006 | $0.67 | — | 76 |
| Cloudflareunknown | 1% | $0.10 | $0.003 | $0.60 | — | 36 |
| AtlasCloudfp8 | 0% | $0.135 | $0.0135 | $0.54 | — | 40 |
| Modalunknown | 0% | $0.25 | $0.005 | $1.00 | — | 86 |
| Mistralfp8 | 0% | $0.30 | $0.007 | $1.20 | — | 52 |
| Ambientfp4 | 0% | $0.045 | $0.0025 | $0.25 | — | 22 |
| StreamLakefp8 | 0% | $0.147 | $0.00294 | $0.588 | 51% | 44 |
| GMICloudfp8 | 0% | $0.18 | $0.0036 | $0.72 | 40% | 89 |
| DekaLLMunknown | 0% | $0.10 | $0.003 | $0.60 | — | 41 |
| Venicefp8 | 0% | $0.30 | $0.0075 | $1.20 | — | 70 |
| Googleunknown | 0% | $0.30 | $0.03 | $1.20 | — | 42 |
| Makorafp8 | 0% | $0.26 | $0.0058 | $1.17 | — | 74 |
| SiliconFlowfp8 | 0% | $0.15 | $0.003 | $0.60 | — | 65 |
| Tenstorrentfp8 | 0% | $0.10 | $0.003 | $0.40 | — | 30 |
| Inceptronfp4 | 0% | $0.10 | $0.006 | $0.60 | — | 36 |
| Ant Digitalfp8 | 0% | $0.12 | $0.0024 | $0.48 | 60% | 50 |
| PrimaLabsunknown | 0% | $0.24 | $0.0048 | $0.96 | — | 44 |
| Near AIunknown | 0% | $0.30 | $0.006 | $1.20 | — | 17 |
Left out, priced more than 4 times above or below the median host for this traffic: Mancer 2.
Every GLM-5.3 host (36)
| Host | Share of requests | Input | Cached input | Output | Discount | Speed |
|---|---|---|---|---|---|---|
| Waferfp4 | 27% | $0.0699 | $0.0689 | $1.40 | — | 56 |
| Decartfp4 | 9% | $0.56 | $0.10 | $1.81 | — | 66 |
| InferenceNetfp4 | 8% | $0.084 | $0.079 | $3.60 | — | 49 |
| Togetherunknown | 7% | $1.40 | $0.26 | $4.40 | — | 69 |
| Tenstorrentfp8 | 6% | $0.60 | $0.10 | $2.50 | — | 49 |
| Z.AIfp8 | 5% | $1.40 | $0.26 | $4.40 | — | 30 |
| CoreWeavenvfp4 | 5% | $0.75 | $0.19 | $3.00 | — | 63 |
| DigitalOceanunknown | 4% | $0.91 | $0.169 | $2.86 | — | 60 |
| Coherenvfp4 | 3% | $0.50 | $0.10 | $3.50 | — | 24 |
| Sail Researchfp8 | 3% | $0.75 | $0.12 | $3.00 | — | 41 |
| Fireworksunknown | 2% | $0.70 | $0.13 | $2.20 | — | 56 |
| Friendliunknown | 2% | $1.26 | $0.234 | $3.96 | 10% | 66 |
| Novitafp8 | 2% | $0.63 | $0.117 | $1.98 | 55% | 33 |
| Cloudflareunknown | 2% | $0.70 | $0.13 | $2.20 | — | 12 |
| BaseTenfp4 | 2% | $1.40 | $0.14 | $4.40 | — | 47 |
| Nebiusfp4 | 1% | $1.40 | $0.26 | $4.40 | — | 58 |
| DeepInfrafp4 | 1% | $0.563 | $0.125 | $2.50 | 38% | 35 |
| Inceptronfp4 | 1% | $1.05 | $0.125 | $3.40 | — | 43 |
| Mistralnvfp4 | 1% | $1.54 | $0.154 | $4.84 | — | 21 |
| Relacefp4 | 1% | $0.0703 | $0.0703 | $4.38 | — | 44 |
| Makorafp4 | 1% | $1.50 | $0.26 | $4.40 | — | 96 |
| Morphfp8 | 1% | $1.04 | $0.212 | $4.88 | — | 46 |
| ModelRunfp4 | 1% | $2.10 | $0.40 | $6.60 | — | 76 |
| Parasailfp4 | 1% | $1.40 | $0.26 | $4.40 | — | 79 |
| Modalunknown | 1% | $1.40 | $0.26 | $4.40 | — | 57 |
| Crusoefp4 | 0% | $1.40 | $0.26 | $4.40 | — | 45 |
| Amazon Bedrockunknown | 0% | $1.68 | $0.312 | $5.28 | — | 52 |
| GMICloudfp8 | 0% | $0.70 | $0.13 | $2.20 | 50% | 20 |
| AkashMLfp4 | 0% | $0.19 | $0.19 | $4.40 | — | 35 |
| Featherlessunknown | 0% | $0.42 | $0.078 | $1.32 | 70% | 28 |
| Veniceunknown | 0% | $1.40 | $0.26 | $4.40 | 20% | 52 |
| Doublewordnvfp4 | 0% | $0.25 | $0.13 | $1.00 | — | 35 |
| AtlasCloudfp8 | 0% | $1.40 | $0.26 | $4.40 | — | 19 |
| SambaNovaunknown | 0% | $1.40 | $0.07 | $4.40 | 50% | 30 |
| Alibabaunknown | 0% | $1.19 | $0.238 | $3.74 | — | 34 |
| Baidufp8 | 0% | $1.40 | $0.26 | $4.40 | — | — |
Every Kimi K3 host (20)
| Host | Share of requests | Input | Cached input | Output | Discount | Speed |
|---|---|---|---|---|---|---|
| Wafermxfp4 | 18% | $0.276 | $0.13 | $9.60 | — | 32 |
| InferenceNetfp4 | 16% | $0.245 | $0.214 | $15.00 | — | 48 |
| Moonshot AIunknown | 14% | $3.00 | $0.30 | $15.00 | — | 20 |
| Nebiusfp4 | 9% | $3.00 | $0.30 | $15.00 | — | 41 |
| Togetherunknown | 8% | $3.00 | $0.30 | $15.00 | — | 28 |
| Fireworksunknown | 7% | $3.00 | $0.30 | $15.00 | — | 26 |
| Relacefp4 | 6% | $0.18 | $0.18 | $11.01 | — | 42 |
| DeepInframxfp4 | 5% | $2.85 | $0.285 | $14.25 | — | 46 |
| Riverunknown | 4% | $3.00 | $0.30 | $15.00 | — | 74 |
| Sail Researchfp4 | 3% | $0.345 | $0.172 | $16.58 | — | 39 |
| Parasailfp4 | 3% | $3.00 | $0.30 | $15.00 | — | 27 |
| KTunknown | 2% | $0.45 | $0.30 | $15.00 | — | 26 |
| Makorafp4 | 2% | $1.00 | $0.50 | $18.00 | — | 36 |
| Io Netfp4 | 1% | $1.90 | $0.19 | $14.00 | — | 45 |
| Amazon Bedrockunknown | 1% | $3.30 | $0.33 | $16.50 | — | 42 |
| BaseTenfp8 | 1% | $3.00 | $0.30 | $15.00 | — | 32 |
| Modalmxfp4 | 1% | $3.00 | $0.30 | $15.00 | — | 46 |
| Alibabaunknown | 1% | $3.00 | $0.30 | $15.00 | — | 26 |
| DigitalOceanunknown | 0% | $2.55 | $0.255 | $12.95 | — | 30 |
| Chutesmxfp4 | 0% | $1.85 | $0.185 | $9.50 | — | 18 |
Every GLM-5.3-Flash host (35)
| Host | Share of requests | Input | Cached input | Output | Discount | Speed |
|---|---|---|---|---|---|---|
| CoreWeavenvfp4 | 24% | $0.12 | $0.02 | $0.38 | — | 53 |
| Coherenvfp4 | 10% | $0.105 | $0.02 | $0.34 | — | 52 |
| Wafernvfp4 | 9% | $0.042 | $0.041 | $0.184 | — | 60 |
| Togetherunknown | 7% | $0.15 | $0.03 | $0.50 | — | 92 |
| DigitalOceanunknown | 7% | $0.113 | $0.03 | $0.375 | — | 61 |
| Fireworksunknown | 5% | $0.15 | $0.03 | $0.50 | — | 56 |
| Relacefp4 | 4% | $0.0399 | $0.03 | $0.0784 | — | 32 |
| Particle AInvfp4 | 4% | $0.13 | $0.04 | $0.50 | — | 70 |
| Z.AIfp8 | 4% | $0.15 | $0.03 | $0.50 | — | 24 |
| InferenceNetfp4 | 3% | $0.108 | $0.035 | $0.35 | — | 57 |
| OpenInferencefp4 | 3% | $0.05 | $0.025 | $0.45 | — | 82 |
| Novitafp8 | 3% | $0.084 | $0.0168 | $0.28 | 44% | 10 |
| DekaLLMunknown | 2% | $0.10 | $0.03 | $0.50 | — | 16 |
| BaseTenfp8 | 2% | $0.15 | $0.03 | $0.50 | — | 57 |
| Riverunknown | 2% | $0.12 | $0.024 | $0.40 | — | 54 |
| Friendliunknown | 2% | $0.15 | $0.03 | $0.50 | — | 60 |
| Modalfp4 | 2% | $0.15 | $0.03 | $0.50 | — | 67 |
| Near AIfp8 | 1% | $0.08 | $0.012 | $0.50 | — | 9.0 |
| DeepInfrafp4 | 1% | $0.101 | $0.0201 | $0.335 | 33% | 46 |
| Makorafp4 | 1% | $0.14 | $0.03 | $0.24 | — | 59 |
| StreamLakefp8 | 1% | $0.0885 | $0.0177 | $0.295 | 41% | 16 |
| Rekaunknown | 1% | $0.12 | $0.03 | $0.30 | — | 43 |
| Mistralunknown | 1% | $0.15 | $0.03 | $0.50 | — | 60 |
| GMICloudfp8 | 1% | $0.075 | $0.015 | $0.25 | 50% | 21 |
| Inceptronfp8 | 1% | $0.16 | $0.04 | $0.50 | — | 47 |
| Crusoefp4 | 0% | $0.15 | $0.03 | $0.50 | — | 62 |
| Parasailfp8 | 0% | $0.15 | $0.03 | $0.50 | — | 10 |
| Morphfp8 | 0% | $0.103 | $0.052 | $0.298 | — | 24 |
| Tenstorrentfp8 | 0% | $0.075 | $0.015 | $0.25 | — | 23 |
| KTunknown | 0% | $0.15 | $0.03 | $0.50 | — | 37 |
| Sail Researchfp4 | 0% | $0.175 | $0.0225 | $0.625 | — | 44 |
| Veniceunknown | 0% | $0.15 | $0.03 | $0.50 | — | 51 |
| Darkbloomfp4 | 0% | $0.065 | $0.012 | $0.18 | — | 2.0 |
| Nebiusfp4 | 0% | $0.15 | $0.03 | $0.50 | — | 53 |
| Baidufp8 | 0% | $0.15 | $0.03 | $0.50 | — | 24 |
Yields
From SemiAnalysis InferenceX’s agentic benchmark: the best official configuration per GPU at the model’s target speed per user (p90), from its latest run. Total includes cached input; the crack prices cached and uncached input separately.
| Model | GPU | Output tok/s | Total tok/s | Cached | Configuration | Run |
|---|---|---|---|---|---|---|
| V4.1-Flashat 65 tok/s per user | H100 | 130† | 20K | 93% | vllm, fp4† measured only at 68 tok/s per user and faster | Nov 20 |
| V4.1-Flashat 65 tok/s per user | H200 | 328 | 52K | 97% | vllm, fp4 | Nov 20 |
| V4.1-Flashat 65 tok/s per user | B200 | 2,800† | 388K | 95% | vllm, fp4† measured only at 107 tok/s per user and faster | Nov 20 |
| V4.1-Flashat 65 tok/s per user | B300 | 13,908† | 828K | 98% | dynamo-vllm, fp4† measured only at 84 tok/s per user and faster | Nov 20 |
| V4.1-Flashat 65 tok/s per user | GB200 | 13,354† | 817K | 96% | dynamo-vllm, fp4† measured only at 126 tok/s per user and faster | Nov 20 |
| V4.1-Flashat 65 tok/s per user | GB300 | 15,785† | 1.17M | 91% | dynamo-vllm, fp4† measured only at 142 tok/s per user and faster | Nov 20 |
| V4.1-Flashat 65 tok/s per user | MI300X | 145 | 21K | 96% | vllm, fp4 | Nov 20 |
| GLM-5.3at 45 tok/s per user | B200 | 298† | 20K | 94% | dynamo-sglang, fp4† measured only at 86 tok/s per user and faster | Nov 18 |
| GLM-5.3at 45 tok/s per user | B300 | 240 | 37K | 95% | dynamo-sglang, fp4 | Nov 18 |
| GLM-5.3at 45 tok/s per user | GB200 | 768† | 42K | 93% | dynamo-sglang, fp4† measured only at 92 tok/s per user and faster | Nov 18 |
| GLM-5.3at 45 tok/s per user | GB300 | 1,145† | 64K | 88% | dynamo-sglang, fp4† measured only at 89 tok/s per user and faster | Nov 18 |
| Kimi K3at 25 tok/s per user | B200 | 73 | 10K | 95% | dynamo-vllm, fp4 | Nov 4 |
| Kimi K3at 25 tok/s per user | B300 | 63 | 8,333 | 91% | vllm, fp4 | Nov 4 |
| Kimi K3at 25 tok/s per user | GB200 | 76 | 11K | 94% | dynamo-vllm, fp4 | Nov 4 |
| Kimi K3at 25 tok/s per user | GB300 | 10† | 1,564 | 97% | dynamo-vllm, fp4† measured only at 433 tok/s per user and faster | Nov 4 |
GLM-5.3: InferenceX benchmarks GLM-5.2 and GLM-5.3 as one model: they share an architecture.
Method
The crack. For each GPU and model: token value per GPU-hour = 3,600 × (output tokens a second × the output price + input tokens a second × (uncached share × the input price + cached share × the cached-input price)) ÷ 1,000,000, with prices per million tokens; crack = token value − rent. The break-even load is rent ÷ token value: the share of the benchmark’s output the GPU must sustain to cover its rent. The owner’s crack uses the cost of owning instead of rent. “Pays at 50% load” is token value × 0.5.
Yields. For each model and GPU, the highest official throughput per GPU among the configurations InferenceX runs (engine, precision, parallelism) at the model’s target speed per user, from InferenceX’s runs in the 14 days to the day, so a run that covers only some configurations doesn’t knock a GPU’s yield down. A configuration measured only at speeds below the target is left out, since its throughput there would overstate what the GPU does at the target. One measured only at speeds above the target counts at its slowest measured point, which understates it; those yields are flagged. InferenceX’s agentic benchmark (AgentX): replayed agentic coding sessions, multi-turn and long-context, with most of each request’s input already in the cache from earlier turns. Each result gives input and output tokens a second per GPU and the share of input served from cache, and those are what the crack prices. Each model’s target is the median p90 speed of its standard OpenRouter hosts, weighted by their requests, rounded to the nearest 5 tok/s. It is set at the first collection of each calendar month and holds for that month. p90: the speed 90% of requests reach or beat, the measure InferenceX uses for agentic work. Yields change only when InferenceX publishes a new run, and on that day the crack picks it up; the daily series shows how much of each move came from yields.
Token prices. From OpenRouter’s list of every host serving each model. Prices are what each host charges, net of any discount it advertises. The realized price of each kind of token is the average over the model’s standard endpoints weighted by the requests each served in the 24 hours to the fixing (OpenRouter reports requests in the last 30 minutes at every collection). Without request data it falls back to OpenRouter’s default routing, which weights hosts by the inverse square of their price. The median host’s posted prices are kept beside it. Left out: free endpoints, faster or cheaper service tiers (priority, flex), hosts OpenRouter marks as degraded or down, and endpoints whose price for the model’s traffic mix is more than four times above or below the median host’s. A model needs at least 5 priced hosts for a print.
Rents. The inthesingularity.ai GPU Compute Price Index, global, all providers: spot and on-demand indices, and commitments of a month, a year and three years at on-demand times each provider’s own term discount, the median across providers (pooled across GPUs where fewer than three providers quote a GPU’s term). GB300 isn’t indexed yet, so its rent is provisional: each provider’s on-demand price, or its shortest commitment where it posts none, from the index database, and the median across providers. Cost of owning: SemiAnalysis’s total cost of ownership per GPU-hour, as published in InferenceX.
Models and rolls. Each series follows a class of model: efficient (the most-used open-weight model on OpenRouter: DeepSeek V4.1-Flash), large (a large open model, several hundred billion parameters: GLM-5.3) and frontier-size (the most-used open model of a trillion parameters or more: Kimi K3). Each series follows a class rather than a model, like a continuous futures contract. A class moves to a successor once the successor has carried more paid tokens on OpenRouter than the current model for seven straight days and InferenceX has agentic results for it on at least four of the GPUs here. The move takes effect at the next fixing, is logged below and is marked on the charts. Also collected, without a crack: GLM-5.3-Flash. The second most-used open model on OpenRouter. InferenceX doesn’t benchmark it, so its prices are collected but it has no crack.
Fixing. One a day at 16:00 UTC, with the GPU index’s fixing, from the token prices collected at or before then (within 36 hours), InferenceX’s latest run on or before the day, and the index’s rents. A final fixing stands as published. Between fixings, the figures above show the same calculation on the latest prices.
What it leaves out. Power and networking where they’re billed separately, staff and idle time. InferenceX measures tuned configurations at full load, and real fleets run below it, which is why the break-even load matters.
Updates. Prices are collected through the day with the GPU index, and InferenceX is checked at every collection. Download every daily fixing and the latest print: gpu-crack-daily.csv. The raw collections (every host’s prices, traffic and speed, every InferenceX result) are in the site’s database.
Changes. Oct 27, 2026: First version: yields from NVIDIA’s Dynamo recipes and one vLLM run, a standard 64K-input trace, median host prices, on-demand rent. Oct 28, 2026: Before the first fixing, the method changed. Yields now come from SemiAnalysis InferenceX by rule (one benchmark for every GPU, at the speed hosts deliver), so H100 and B300 get cracks. Token prices are weighted by the requests each host serves, with the median host’s posted prices kept beside them. GLM-5.3 replaces GLM-5.3-Flash, which InferenceX doesn’t benchmark. MI300X added. New measures: cracks against commitment rents, the owner’s crack, the highest rent a host can pay at half load, and each series follows a model class that rolls to successors.
Yields: SemiAnalysis InferenceX, which describes its data as open source and freely available; credited wherever used. Cost of ownership: SemiAnalysis, as published in InferenceX. Token prices, traffic and speeds: OpenRouter’s listings of each host. Rents: the inthesingularity.ai GPU Compute Price Index. Cracks are inthesingularity.ai calculations. Not investment advice. See also the Who makes money serving open models? analysis.