The purchasing opportunity: Novation’s open model
The purchasing opportunity
This is our model, opened up. Every input is yours, the arithmetic is printed below it, and it runs entirely in your browser — nothing is sent anywhere and there is no form.
It prints the purchasing opportunity against your incumbent — baseline 1 minus baseline 3 in our method — not our incremental value over the best alternative you could reach alone, which is agreed with you per engagement and computed there. If it returns no opportunity on your numbers, that is the answer and you should not call us; if it returns a positive number below our 12% screen, the saving is real on the model's terms and we would still tell you not to add a supplier for it. We would rather you find that out here, in two minutes, than four weeks into an engagement. On its default assumptions the frontier is narrow, and it is a curve rather than a cutoff. The price gap and the share actually moved trade against each other: 1.25× needs about 77% of the work that can move to actually migrate to pass the 12% screen, 1.35× about 62%, 1.45× about 53%. Fleet size trades against both, and two floors matter. On the defaults the model is net-positive above roughly $7M of annual spend, where the $150k subscription floor and the fixed costs are repaid; it passes the 12% screen above roughly $24M, falling to about $11M at a 1.6× price gap and 70% capture. Two cautions on those figures: they hold the other inputs at default, and they hold cross-supplier data at 250 TB/month while shrinking the fleet, which this page's own note tells you to scale. Scale it and the floors fall further. Read the grid rather than the three numbers.
Your fleet
Interruptible capacity — both sides
—
Fabric
Frictions
The bridge
| Work that can move | — |
| Held back to retire commitment | — |
| Actually captured | — |
| Gross saving vs your blended rate | — |
| Interruptible discount captured | — |
| Gross benefit, before any deduction | — |
| Rerun waste | — |
| Checkpoint write overhead | — |
| Fabric drag | — |
| Data logistics | — |
| Vendor onboarding & management | — |
| Your cost of adopting Novation | — |
| Net before our fee | — |
| Novation's subscription | — |
| Net modeled saving | — |
| as % of work that can move | — |
| as % of total fleet spend | — |
Adjust the inputs on the left.
The cheaper thing to try first
Before adding a supplier, most buyers can move the same movable work onto their incumbent's interruptible tier. No second contract, no egress between clouds, no vendor review, no fee to us — but no supplier price gap either, and the same rerun and checkpoint costs. Run on your inputs:
| More interruptible capacity at the incumbent | — |
| Novation, net modeled | — |
| Share of our number that alternative gets | — |
—
Buy or hire?
The other way to get this function is to hire a senior compute-procurement or FinOps lead — roughly $350–500k fully loaded, on the $300–355k base the frontier labs post for the role — who cannot re-quote every supplier weekly but does bring judgment and standing you cannot buy.
| Novation, all-in per year | — |
| Plus your adoption cost, year one | — |
| Against a hire at $350–500k | — |
—
The frontier — net % of the work that can move
Price gap across the top, share actually moved down the side, at your current fleet size and all other inputs as you have set them. Green passes the 12% screening rule; the gold cell is where you are.
Note what this shows: read it as a curve, not a cutoff. At 40% moved nothing in this band passes; at 80% moved everything does. The price gap alone never decides it, so we print the whole grid rather than a single make-or-break number.
The arithmetic, in full
Three conventions the model follows, each of them against us. Deductions are priced at the supplier's rate rather than yours; charging them at your blended rate would overstate every friction by exactly the price gap. Interruptible-instrument frictions are priced at the discounted rate that work actually runs at. And the fee is the flat subscription our proposed tiers charge the fleet you entered — $360k on $60 million — shown as its own line after the net before fees, so gross benefit, execution cost, implementation cost, the fee, and the net can never be read as one number. The subscription is the only Novation fee in the model. Onboarding ($40–120k, one-off) and any fixed-fee purchase decisions are priced per engagement on the pricing page and are not deducted here; the engineer-days above are your side of adoption, not our onboarding fee. Add them if you want an all-in first year. One thing the model does not do: it does not credit only the increment of interruptible discount over spot you already run. Because supplier_spend is obtained by dividing captured spend by a blended price gap, existing spot is already reflected in the baseline, and crediting only the increment would understate the spot line by about $554k and net by about $499k on the default fleet.
Break-even: interruptible capacity
What this omits, and it matters. Each preemption is treated as an independent fixed charge, and the requeue delay is not modeled at all: in a constrained market the gang may not re-form for hours, which can dominate everything above. Treat p* as an upper bound on tolerance, not a comfort. The gang-size field does not change the arithmetic: it prints, beside the input, the preemption rate a gang of that size would see on the published measurements, so you can see how far below it the rate you entered sits. Raise the rate yourself if the warning applies to you; the next panel shows what that does to the net.
Failure cost: what is deducted, and what the scale gap is worth
The model deducts both the checkpoint write and the restart, and the restart is the one people expect to be missing. rerun carries the restart at 20 minutes plus half a checkpoint interval of lost work — 35 minutes per preemption, which is longer than the 7–12 minutes per failure the published measurements report even where checkpoints write in under three seconds. The write is deducted separately and is the larger line of the two: on the default fleet, $219k of checkpoint stall against $20k of rerun.
The real gap is scale, and the warning beside the gang-size field shows it. The default of 1.5 preemptions per 72-hour job implies a mean time to failure of 48 hours — which on Meta's RSC measurements corresponds to a gang of roughly 170 GPUs. Our target workloads are larger than that, and failure rate rises roughly in proportion to job size. So the model is optimistic by a knowable margin, and here it is:
| Gang | MTTF | Preemptions / job | Net on the default fleet |
|---|---|---|---|
| 256 | 31.6 h | 2.3 | $2.57M −0% |
| 1,024 | 7.9 h | 9.1 | $2.49M −4% |
| 4,096 | 2.0 h | 36.5 | $2.16M −16% |
| 16,384 | 0.5 h | 145.8 | $1.12M −57% |
Meta RSC reliability study, HPCA 2025 (11 months, 4M jobs, >150M A100 GPU-hours): MTTF 47.7 days at 8 GPUs against 7.9 hours at 1,024, projected 1.8 hours at 16,384. We scale inversely with job size between those anchors. The bottom row is the pathological case where rerun saturates at 100% of interruptible spend — at that scale nobody should be running the work that can move on interruptible capacity at all, which is what the row is telling you. Two things we still do not model: the requeue delay when a gang cannot re-form, and hardware failures on firm capacity, since rerun and checkpoint are applied only to interruptible spend. The second is deliberate — you already checkpoint on the capacity you have, so only the incremental checkpointing belongs in an incremental comparison — but it is a judgement, not a fact, and you should know we made it.
Break-even: slower fabric
At x=0.20 and y=0.35: f* ≈ 71%. Below that exposed-comm fraction the cheaper fabric wins. Caveat we will not bury: exposed-comm fraction is not invariant under fabric degradation — congestion breaks the overlap that hid the communication, and synchronous training is governed by the slowest rank, not the mean. Treat f* as an optimistic bound and measure your own.
What this model still does not do
It does not price capacity availability, which in a constrained market is often worth more than rate. For many fleets the binding question is whether a supplier has 512 contiguous cards on Tuesday, not whether they are 30% cheaper, and a model whose output is a price delta cannot answer that. It is the largest known omission.
It does not price your best alternative without us. The incumbent’s interruptible tier is approximated above, from your blended rate rather than your incumbent’s firm rate, and on the defaults it captures about a third of the opportunity at zero egress, zero onboarding, and zero fee. A negotiated second supplier plus a router — the second baseline in our method — is not modeled here at all, because it depends on your suppliers and your team. It is agreed with you before an engagement and computed on it, and our fee is judged against that number, not this page’s. It also has no gang-size or requeue term; it treats the commitment holdback as a simple carve-out rather than optimizing retirement against shortfall risk; it ignores tail risk from a supplier failing mid-campaign; and it is a first-year view, so onboarding and integration fall away in year two and a marginal year one can still be a good three-year decision.
It is a screening tool, not a proposal. Its job is to tell you quickly whether the conversation is worth having.