Pricing model
Each customer line is priced from its own cost, so the price follows the cost model. Human hours are never sold: setup, cycles and handoff are packages priced from standard effort, and automation shares its saving with the customer.
1. From cost to price
Read top to bottom. Every line is its cost divided by (1 − its margin). Margins differ by line type: packages 45%, managed compute 30% (the most contested part of the price), platform and Ari usage 60% (the differentiated part).
| Customer line | Cost, incl. shared | Price | Share of price |
|---|---|---|---|
| Setup package margin 45% on this line | $34,064 | $61,934 | 25% |
| Managed RL cycle packages margin 45% on this line | $77,298 | $140,542 | 57% |
| Managed compute margin 30% on this line | $15,057 | $21,510 | 9% |
| Platform and Ari usage margin 60% on this line | $6,463 | $16,157 | 7% |
| Handoff package margin 45% on this line | $3,907 | $7,104 | 3% |
| Total | $136,789 | $247,247 | 100% |
Shared overhead of $17,764 (storage, rework, insurance, support, legal) is allocated across the lines by direct cost. Blended margin 44.7% against a 45% target and 35% floor.
2. GPU type and capacity
A GPU type moves cost through its hourly rate and through how much work an hour buys. The table compares cost per unit of work, not sticker price. Existing capacity only counts for its own GPU type.
Example deal
| GPU | PayGo $/GPU-h | Throughput vs H100 | $ per unit of work | GPU cost | Share of COGS | Price |
|---|---|---|---|---|---|---|
| H100 | $5.49 | 1.0× | $5.49 | $8,400 | 6.1% | $247,247 |
| B200 | $8.50 | 2.2× | $3.86 | $12,982 | 8.8% | $264,613 |
| B300 | $11.00 | 3.0× | $3.67 | $12,320 | 8.3% | $266,095 |
Stress case: 8 RL cycles, 10x rollout attempts
| GPU | PayGo $/GPU-h | Throughput vs H100 | $ per unit of work | GPU cost | Share of COGS | Price |
|---|---|---|---|---|---|---|
| H100 | $5.49 | 1.0× | $5.49 | $113,288 | 25.8% | $692,557 |
| B200 | $8.50 | 2.2× | $3.86 | $64,835 | 18.1% | $577,390 |
| B300 | $11.00 | 3.0× | $3.67 | $83,825 | 21.3% | $628,452 |
Capacity verdicts other than FEASIBLE mean the planned weeks are not fully covered on that GPU type; the deal stays on HOLD until capacity is confirmed. Rates and throughput are illustrative placeholders.
3. Token economics for Ari
| Tokens per GPU-hour (measured) | Cost per million tokens | Price per million at 60% margin |
|---|---|---|
| 100,000 | $54.90 | $137.25 |
| 1,000,000 | $5.49 | $13.72 |
| 5,000,000 | $1.10 | $2.75 |
The current meter is $13.72 per million tokens, at the placeholder throughput of 1,000,000 tokens per GPU-hour. It is provisional until throughput is measured. For a $10 price at 60% margin, the GPU must serve about 1,372,500 tokens per GPU-hour; for $5, about 2,745,000.
Market reference: output prices in the Mercor case study run about $1 to $30 per million tokens.
4. Effort and automation
Standard effort sets each package. The effort tier, calibration against past engagements, and the share of hours Ari replaces all scale the labour inside the packages. Calibrate after each engagement by recording actual against standard hours.
| Human hours replaced by Ari | COGS | Price | Margin |
|---|---|---|---|
| 0% | $136,789 | $247,247 | 44.7% |
| 25% | $112,161 | $222,436 | 49.6% |
| 50% | $87,533 | $197,627 | 55.7% |
| 75% | $62,905 | $172,823 | 63.6% |
Shown at 50% of the saving passed to the customer. The customer price falls as Ari replaces hours, and Applied Compute keeps the rest of the saving.
5. Fixed-cost recovery
Gross margin per engagement is not enough on its own: the platform team and Ari development are fixed monthly costs. The figures below use placeholder costs and volume, and show the number of engagements a month that covers them.
| Contribution per engagement (price less COGS) | $110,458 |
| Fixed monthly cost (placeholder: platform team $60K, Ari development $40K) | $100,000 |
| Break-even engagements per month | 0.9 |
| Net per month at 2 engagements | $120,916 |
6. Stop points, overage and commitment checks
Early exit
| Stop point | Invoiced | Cost to that point | Margin |
|---|---|---|---|
| Stop after setup | $49,449 | $34,064 | 31% |
| Stop after cycle 1 | $107,140 | $67,003 | 37% |
| Stop after cycle 2 | $164,831 | $99,942 | 39% |
Overage against buying GPUs
A GPU-hour above the allowance is metered at $8, against a blended cost of $3 and a list price of $5 to buy one more. The meter covers the marginal price, so it never loses money.
Committed cycles
| Cycles | Commit discount | Quote | Blended margin |
|---|---|---|---|
| 3 | 0% | $247,247 | 44.7% |
| 4 | 5% | $289,650 | 42.6% |
| 6 | 10% | $374,868 | 40.0% |
The commit discount falls on the cycle packages only, so the blended margin falls as commitment rises. Each level shown stays above the minimum margin.
7. Defensibility
- Price band. $210,545 to $283,946 around $247,247 when throughput, GPU cost and effort each move 20% together. Quote the band until throughput is measured.
- Floor. No line can be priced below its cost. The blended margin must stay above the 35% floor, or the quote is held as HOLD - MARGIN.
- Competitor at 10% below. If a competitor prices at $222,522, margin falls to 38.5%. Match scope rather than rate: trading cycles or workflow scope keeps the economics intact.
- Holds. The quote is held while the handoff has no holdout evaluation cost (HOLD - COST GAP), while token prices are unmeasured (HOLD - THROUGHPUT), or while the model is unverified.