Simulated for an anonymized client. Your workspace runs the same steps on your own demand, prices, and limits.
1 The company
An HPC startup in energy, between its Series A and Series B. Its researchers run physics simulations on rented cloud servers. The money raised lasts 36 months, and the hiring plan grows the team from 30 to 120 researchers in that time.
How should it pay for cloud compute so it costs least, even if hiring slows, with nothing owed after month 36?
2 Where the cost comes from
Researchers30 → 120The hiring plan
× Simulations19Per researcher, per month
× Node-hours8Per simulation: one server for one hour
× Price$5.65 → $7.58Per node-hour. Steps up in month 6; cause not yet known.
= Compute$26k → $138kA month, month 1 → month 36
Hiring drives the cost: four times the researchers means four times the compute. Storage and the data platform come on top, and $648k of cloud credits pay the first months.
AWS regionPrices, results, and the decision below follow it.
3 The simulation
The simulator replays this demand hour by hour for 36 months, busy weekdays and quiet nights, and pays for it 59 ways: the five options below, and 54 variants of D with more or less commitment and Spot. It then runs every plan again with demand 15% below the forecast from month 12, as if hiring slowed: the low end of the plan's own forecast range. The on-demand price is the client's own, from step 2, in both regions. The discounts are AWS's published rates for us-east-1.
Discount
Off on-demand
1-year plan, paid monthly
−28.3%
1-year plan, prepaid
−33.1%
3-year plan, paid monthly
−51.6%
Spot, can be interrupted
−70%
30-day commitments Unquoted
−29%
4 The options
Option
Cost, 36 months
If demand falls 15%
Largest payment
Owed after funding
A. On demandNo commitments: credits first, then on-demand prices.
$3.69Mthe baseline
$3.31M
$197kmonth 35
$0k
B. Prepaid 1-year plansOne-year Compute Savings Plans prepaid in months 11 and 23 for 55% of average compute, plus 30-day commitments for another 30%. The plan that was delivered. Rests on an unquoted 30-day rate.
$3.14M−$547k vs A
$2.83M
$525kmonth 23
$0k
C. One 3-year planOne three-year Compute Savings Plan from month 13 for 55% of average compute, paid monthly. The first draft.
$3.21M−$476k vs A
$2.90M
$177kmonth 35
$224k
D. Monthly-paid 1-year plansOne-year Compute Savings Plans in months 11 and 23 for 55% of average compute, paid monthly, with 50% of burst work on Spot.
$2.89M−$800k vs A
$2.61M
$174kmonth 35
$0k
E. 30-day commitments30-day commitments for 85% of average compute, resized every month, with 50% of burst work on Spot. Rests on an unquoted 30-day rate.
$2.67M−$1.03M vs A
$2.42M
$147kmonth 35
$0k
D at 65% coverage RecommendedLike D, with the one-year plans covering 65% of average compute instead of 55%.
$2.86M−$828k vs A
$2.59M
$174kmonth 35
$0k
Cost is the sum of every payment over the 36 months, so option A matches the plan on the home page.
B and C rebuild the home page's delivered plan and first draft on published discounts, with C sized to the same 55% coverage as the others, so their figures differ from the home page's.
Monthly cash out · each row on the same scale (largest bar $525k) · months 1–5 paid by credits
5 The decision
Drop plans that would owe money after month 36: C.
Drop plans that depend on a price not yet quoted: B and E.
Of the 56 plans left, pick the one that stays closest to the cheapest, whether demand follows the forecast or falls 15%.
The result: D at 65% coverage, at most $3k above the cheapest either way. D, at 55%, is up to $32k above.
Recommended · us-east-1
D at 65% coverage
Decision required
Approve one-year Compute Savings Plans in months 11 and 23, each covering 65% of average compute, paid monthly. Run 50% of burst work on Spot with on-demand fallback.
Approver
CFO, with the Head of Engineering
Status
Pending approval
Pending quote E. 30-day commitments would cost $197k less if its 30-day rate is confirmed.
3 The simulation
The simulator replays this demand hour by hour for 36 months, busy weekdays and quiet nights, and pays for it 59 ways: the five options below, and 54 variants of D with more or less commitment and Spot. It then runs every plan again with demand 15% below the forecast from month 12, as if hiring slowed: the low end of the plan's own forecast range. The on-demand price is the client's own, from step 2, in both regions. The discounts are AWS's published rates for eu-central-1.
Discount
Off on-demand
1-year plan, paid monthly
−23.9%
1-year plan, prepaid
−29%
3-year plan, paid monthly
−48.1%
Spot, can be interrupted
−65%
30-day commitments Unquoted
−29%
4 The options
Option
Cost, 36 months
If demand falls 15%
Largest payment
Owed after funding
A. On demandNo commitments: credits first, then on-demand prices.
$3.69Mthe baseline
$3.31M
$197kmonth 35
$0k
B. Prepaid 1-year plansOne-year Compute Savings Plans prepaid in months 11 and 23 for 55% of average compute, plus 30-day commitments for another 30%. The plan that was delivered. Rests on an unquoted 30-day rate.
$3.19M−$500k vs A
$2.87M
$552kmonth 23
$0k
C. One 3-year planOne three-year Compute Savings Plan from month 13 for 55% of average compute, paid monthly. The first draft.
$3.25M−$443k vs A
$2.93M
$179kmonth 35
$240k
D. Monthly-paid 1-year plansOne-year Compute Savings Plans in months 11 and 23 for 55% of average compute, paid monthly, with 50% of burst work on Spot.
$2.97M−$716k vs A
$2.68M
$175kmonth 35
$0k
E. 30-day commitments30-day commitments for 85% of average compute, resized every month, with 50% of burst work on Spot. Rests on an unquoted 30-day rate.
$2.70M−$986k vs A
$2.46M
$149kmonth 35
$0k
D at 65% coverage RecommendedLike D, with the one-year plans covering 65% of average compute instead of 55%.
$2.95M−$737k vs A
$2.67M
$175kmonth 35
$0k
Cost is the sum of every payment over the 36 months, so option A matches the plan on the home page.
B and C rebuild the home page's delivered plan and first draft on published discounts, with C sized to the same 55% coverage as the others, so their figures differ from the home page's.
Drop plans that would owe money after month 36: C.
Drop plans that depend on a price not yet quoted: B and E.
Of the 56 plans left, pick the one that stays closest to the cheapest, whether demand follows the forecast or falls 15%.
The result: D at 65% coverage, the cheapest in both cases. D, at 55%, is up to $20k above.
Recommended · eu-central-1
D at 65% coverage
Decision required
Approve one-year Compute Savings Plans in months 11 and 23, each covering 65% of average compute, paid monthly. Run 50% of burst work on Spot with on-demand fallback.
Approver
CFO, with the Head of Engineering
Status
Pending approval
Pending quote E. 30-day commitments would cost $249k less if its 30-day rate is confirmed.
What this rests on
Demand, prices, other charges, and credits come from the delivered plan shown on the home page: researchers × 19 jobs × 8 node-hours a month, at the client’s implied price per node-hour.
The implied price steps from $5.65 to $7.58 per node-hour in month 6; the cause is not yet explained by the client’s spend model.
The client’s price is used in both regions; the region changes only Savings Plan discounts, the Spot price share, and the interruption rate.
Burst profile (synthetic): weekday hours 09:00–19:00 at twice the night rate, weekends at 60%. Replaced by the client’s hourly usage when available.
Coverage is a share of the purchase month’s average hourly compute: 55% for the one-year plans (as delivered), 30% for option B’s 30-day layer, 85% for option E.
30-day guaranteed commitments: 29% net discount (Archera headline, not region-specific) on an assumed 3-year No Upfront Compute Savings Plan; resized monthly. Needs a quote.
Spot: half of burst usage, restart 3 minutes, checkpoint every 15 minutes, independent interruptions at the research run’s hazard; correlated reclaims are not modeled yet.
Costs are plain cash totals: no discount rate, so a dollar paid in month 36 counts the same as one paid in month 1.
What if hiring slows: demand 15% below the forecast from month 12, the month after the first purchase. 15% is the low end of the delivered plan's own forecast range (its 10th percentile is 15% below the central forecast). Commitments bought after the drop are sized from actual usage.
Read option E with care: it leads only because of the unquoted 29% net discount, which the research flagged as probably blended across regions.
Spot looks better the more of the burst it takes because capacity shortfalls and correlated reclaims are not modeled yet; the grid’s best points sit at the edge of the Spot range tested.
Prices from the research run 2026.10.14.AWS.Arm.HPC.pricing.inputs. Your workspace runs the same comparison on your own demand.