Set your monthly token consumption per person. Compare 16 or 64 concurrent requests per node, the number of B300 nodes your team requires, and the cost against Kimi, GPT, and Claude APIs.
A planning model, not a performance or savings guarantee. Different models can require different tokens and retries to complete the same task.
Installed allowance / node$650,000
Annual operations / node$150,000
Selected net cost / node / month$23,333
Start with your demand.
Monthly demand is people × tokens per person. For each concurrency setting (16 or 64), we automatically round up the number of nodes needed to cover that demand. Tokens include input and output, split 8:1 to match the benchmark workload. The starting value of 5B per person is an editable scenario, not typical usage.
1 person × 5B per person = 5B tokens/month.
The busy-time ceiling reserves room for idle periods and demand variation; it is not GPU utilization or a latency guarantee. Capacity scales by this fraction of benchmark-rate operation. Actual required busy time can be lower after rounding up to whole nodes. The defaults of 70% busy time and 90% API-cache reads are assumptions, not customer averages.
Adjust acquisition and operations
Annual operations default to $150,000 per node, independent of acquisition value. The allowance includes hosting, power, cooling, support, and management, and applies in both business-hours and 24/7 scenarios. This is a planning assumption, not a quote.
Off-hours sell-back sensitivity
No confirmed Shadeform B300 host payout is available in this model. The $0 rate means unconfirmed, not free capacity. Enter a current marketplace reference and a separately negotiated payout percentage; if you have a net host rate, enter it with 100% payout. The 80% payout, 50% occupancy, $0.50 extra cost, $0.05 software fee, 44-hour buffer, and 20% annual rate decline are editable assumptions, not Shadeform terms. The software fee is modeled across all 720 hours per month for each installed GPU, including unsold hours; set it to zero if already included in your net rate or operations allowance.
Business hours reserve 176 of a normalized 720-hour month for your team. Only the remaining 544 hours can be offered, less the reclaim buffer. No daytime spare capacity is sold. Booking duration, timely reclamation, secure tenant isolation and wiping, colo/network permissions, and (for leased hardware) lessor permission must be confirmed. Additional cost covers only expenses beyond the annual operations allowance. Resale income is uncertain and shown before income taxes.
Daily VM spot resale and reclamation were described in an exploratory provider discussion, with a software fee plus revenue sharing. No signed pricing or service-level commitment is established here. The serving benchmarks below do not validate performance under that VM platform; passthrough, isolation, model reload, and actual capacity need testing before deployment.
Shadeform provider requirements: Tier 3/4 data centers and provider onboarding. B300 availability. Provider acceptance, actual payout, and reclaim terms remain subject to agreement. Checked September 13, 2026.
Lease and depreciation-benefit sharing
70–80% means a negotiated share of the lessor’s depreciation tax benefit, not a 70–80% tax saving. At 21% tax and 100% eligible basis, it equals 14.7–16.8% of acquisition value before financing charges. The five-year planning horizon is not a tax depreciation schedule. The scenario assumes an eligible taxable lessor can use the deduction immediately; eligibility and tax capacity require professional review.
This is an amortizing rent proxy with zero residual value, no down payment, and no automatic transfer of ownership, not a lease quote. Gross rent uses the financing rate before benefit sharing. The shared benefit is spread evenly as a rent reduction; do not subtract it again from a quote already reflecting tax benefits. Deposits, lender fees, insurance changes, early termination, and end-of-term purchase costs are excluded. Actual lease pricing must cover the lessor’s costs and return.
Purchase is shown without the buyer’s own depreciation benefit; lease is shown after the assumed commercial rent discount. This is not a complete after-tax comparison. No lessee depreciation or rental-payment tax deduction is credited. An owner with usable tax deductions may benefit directly instead. State taxes, timing/NPV, income taxes on resale, and recapture are not modeled.
Check the numeric inputs and their allowed ranges. Percentages must be 0–100 (tax and financing rates 0–50), resale rates 0–100, and the buffer 0–544 hours. Results retain the last valid settings.
Same token volume. Different economics.
Monthly costs in USD. Savings compare each API bill with the full levelized cost of all required nodes.
Swipe the table horizontally to compare all three APIs.
API costs price only the selected demand, not spare capacity. Selected funding and net resale cover every required node. Savings are not latency-matched.
Concurrent requestsPer node
Capacity / nodeTotal tokens / month
Nodes requiredAuto-sized to your team
Purchased fleetNet levelized / month
Kimi K3Official API
GPT-6 AstraStandard API
Claude Opus 5Standard API
Concurrent requests are not employees. Independent replicas are assumed to divide traffic evenly; node counts are volume-based minimums, not peak-load or redundancy sizing. API prices are token charges, not ChatGPT or Claude subscriptions. Latency parity is unverified separately for Kimi, GPT, and Claude.
Full-horizon cost and cash outlay
Concurrent requests
Purchased fleet
Kimi K3
GPT-6 Astra
Claude Opus 5
Annual cash flow per node
USD per node. Positive net outlay is cash paid; negative is a modeled net inflow. No income-tax deductions credited to the buyer or lessee.
Year
Capital / rent + operations
Resale payout
Extra resale costs
Net cash outlay
Performance, with the evidence attached.
SGLang unified serving, TP8, MXFP4, no speculative decoding. One node with eight B300 GPUs. The model uses 8,192 input and 1,024 generated tokens per request, including reasoning.
Both scenarios use reported per-node benchmark throughput. More nodes increase monthly capacity; per-stream latency remains the scenario assumption.
Concurrent requests
First-token delay
Speed per stream
Node output throughputOutput tokens only
Modeled completion time
Concurrency 16: low-latency TP8, P50 first token 3.539 s, P50 time per output token 19.47 ms; 785 total tokens/s/GPU. Concurrency 64: balanced TP8/DCP8, 11.635 s, 40.19 ms; 1,395 total tokens/s/GPU. Aggregate throughput includes input and output. Multiply by eight GPUs and divide by nine to get node output tokens/s for this 8:1 workload.
P50 inputs do not establish mean throughput or P95 latency. For 16/64, capacity uses measured aggregate throughput, not the median-latency formula. All monthly capacities are extrapolations.
The 16-request benchmark misses a strict 2-second first-token target. First token may be reasoning, not the visible answer. Long coding contexts, uneven batches, and queues need separate tests. No speedup from cache reuse or speculative decoding is credited.
Prices in USD per million tokens: first-party, standard processing, global routing, short text context. Latency and speed: independent measurements of those providers’ APIs, not official guarantees. Snapshots checked September 13, 2026.
Latency snapshot: max reasoning effort, 10,000 input tokens. Lower effort can respond sooner. The benchmark workload differs from the 8,192-input-token SGLang test above; these are not latency-matched comparisons.
Latency and speed are Artificial Analysis median measurements for the reasoning settings shown. Cache writes replace base input charges; they are not added on top.
Model / latency setting
Input
Cache read
Cache write / miss
Output
First-chunk delaySeconds · P50
Output speedTokens / second
First chunk is not necessarily the first answer token: it may contain reasoning, or arrive after silent thinking. It is not full-response time. Values come from each source’s “Key Comparison Metrics & API Features” table, not its summary panels. Artificial Analysis reports trailing 72-hour medians; performance varies with prompt, effort, traffic, and routing. No P95 or matched-workload guarantee is established. These reference values do not change the cost or capacity calculations.
GPT-5.6 Sol has promotional pricing available at least through November 21, 2026. Holding that rate for the entire ownership horizon is a sensitivity, not a price commitment.
Zero residual value. Purchase defaults to acquisition plus annual operations over five years, with no buyer tax shield. Optional lease-benefit sharing and off-hours resale are separate, assumption-bound scenarios. Extending the purchase horizon does not reduce the upfront payment.
Except for the explicit lease-financing proxy, funding costs are excluded. Taxes, paid API tools, overflow, failover, migrations, equipment replacement, and operating overruns are excluded. One node is a single point of failure. No hardware replacement is budgeted over either horizon.
Same tokens, not equal outcomes
The Kimi comparison prices the same model family. GPT and Claude are token-cost alternatives, not validated substitutes. Tokenizers, reasoning, output length, retries, and task acceptance can differ. Validate cost per accepted task before making a purchase decision.
Usage, performance, and API prices are held constant throughout the horizon. Cache reuse changes API billing only; self-hosted throughput remains fixed. Neither cache-hit rates nor sustained request occupancy are guaranteed.
View the formulas
Monthly demand
People × selected monthly tokens per person (billions) × 1,000,000,000. Input = demand × 8/9; output = demand ÷ 9.
No cache: base rate. With cache: read share × read rate + remaining share × write/miss rate.
Monthly fleet cost
Required nodes × [(purchase or modeled lease payments + horizon operations − net resale over the horizon) ÷ (12 × years)].
Selected fleet saving
API cost − fleet cost. Percentage = saving ÷ API cost.
Resale offset
Business hours only: (720 − 176 − reclaim buffer) × sold fraction × 8 GPUs × [reference rate × payout fraction − extra cost/GPU-hour]. Annual rate decline applies to the reference rate, not the extra cost. Subtract the software fee × 720 hours × 8 GPUs each month. Sum each year’s net receipts, then levelize over the horizon.
Lease proxy
Gross monthly rent = acquisition × monthly financing rate ÷ [1 − (1 + monthly financing rate)−months]; at 0% use acquisition ÷ months. Shared benefit = acquisition × eligible basis fraction × lessor tax rate × pass-through fraction. Net rent = gross rent − shared benefit ÷ months. No second lessee tax deduction is applied.