Cost Optimization and Technology Financial Management Questions

Direct answer

Provider X is the pick for a three-year non-production workload. On-demand it totals $529.6k at 10% yearly utilisation growth, against $546.2k for Y and $579.2k for Z, and X stays cheapest at every growth rate from 0% to 30% when all three cost lines grow together. Reserved compute (a one-year commitment priced 30% below on-demand, where on-demand means paying the full list rate by the hour with no commitment) cuts X to $410.4k, but only pay for it if the compute actually runs most of the year, which non-production rarely does.

Assumptions (five, stated so a reviewer can change them)

1. Prices are flat for three years; no provider price changes.
2. Compute, storage and egress (data leaving the provider's network) all scale with a utilisation index (a multiplier on year-1 usage) of 1.00, 1.10, 1.21 for years 1 to 3 (10% growth a year).
3. The reserved case applies the 30% discount to all compute and is renewed each year sized to that year's forecast, which is slightly optimistic if forecasts miss.
4. No discounting. It cannot change the ranking here: X is the cheapest provider in every single year (year 1 is $160k, $165k and $175k), so any positive discount rate keeps X first.
5. Workloads perform the same on all three, and the cost of moving in or out is ignored.

The comparison and the sensitivity

providers = {            # year-1 list cost in $k per year
    "X": {"compute": 120, "storage": 30, "egress": 10},
    "Y": {"compute": 95,  "storage": 45, "egress": 25},
    "Z": {"compute": 110, "storage": 25, "egress": 40},
}

def tco(p, growth, reserved):
    total = 0.0
    for year in range(3):
        scale = (1 + growth) ** year          # utilisation index: 1.00, 1.10, 1.21
        compute = p["compute"] * scale * (0.70 if reserved else 1.0)
        total += compute + p["storage"] * scale + p["egress"] * scale
    return total

for reserved in (False, True):
    label = "1-yr reserved compute" if reserved else "on-demand compute"
    print(label)
    for g in (0.0, 0.10, 0.20, 0.30):
        row = {n: round(tco(p, g, reserved), 1) for n, p in providers.items()}
        best = min(row, key=row.get)
        print(f"  growth {g:>4.0%}: " + "  ".join(f"{n}={v:>6.1f}" for n, v in row.items()) + f"  -> best {best}")

# Stress test 1: only compute usage grows; storage and egress stay flat.
def tco_compute_only(p, growth, reserved):
    total = 0.0
    for year in range(3):
        compute = p["compute"] * (1 + growth) ** year * (0.70 if reserved else 1.0)
        total += compute + p["storage"] + p["egress"]
    return total

print("compute-only growth, on-demand")
for g in (0.10, 0.20, 0.30):
    row = {n: round(tco_compute_only(p, g, False), 1) for n, p in providers.items()}
    print(f"  growth {g:.0%}: " + "  ".join(f"{n}={v:.1f}" for n, v in row.items()) + f"  -> best {min(row, key=row.get)}")

# Smallest growth rate (0.1% steps) at which Y beats X on compute-only growth.
for label, reserved in (("on-demand", False), ("reserved", True)):
    flip = next((g / 1000 for g in range(0, 1000)
                 if tco_compute_only(providers["Y"], g / 1000, reserved)
                 < tco_compute_only(providers["X"], g / 1000, reserved)), None)
    print(f"Y beats X on compute-only growth ({label}) from growth of about {flip:.1%}" if flip else f"no flip ({label})")

# Stress test 2: egress volume multiplier at 10% growth, on-demand.
print("egress multiplier at 10% growth, on-demand")
for k in (1, 2, 4):
    row = {}
    for n, p in providers.items():
        q = dict(p, egress=p["egress"] * k)
        row[n] = round(tco(q, 0.10, False), 1)
    print(f"  egress x{k}: " + "  ".join(f"{n}={v:.1f}" for n, v in row.items()))

# Reserved pays off only if the instance runs more than 70% of hours.
print("reserved vs on-demand, X compute, year-1 $k at utilisation of hours")
for u in (0.30, 0.50, 0.70, 1.00):
    on_demand = 120 * u
    reserved = 120 * 0.70
    print(f"  {u:.0%} of hours: on-demand {on_demand:.1f}  reserved {reserved:.1f}")

on-demand compute
  growth   0%: X= 480.0  Y= 495.0  Z= 525.0  -> best X
  growth  10%: X= 529.6  Y= 546.2  Z= 579.2  -> best X
  growth  20%: X= 582.4  Y= 600.6  Z= 637.0  -> best X
  growth  30%: X= 638.4  Y= 658.4  Z= 698.2  -> best X
1-yr reserved compute
  growth   0%: X= 372.0  Y= 409.5  Z= 426.0  -> best X
  growth  10%: X= 410.4  Y= 451.8  Z= 470.0  -> best X
  growth  20%: X= 451.4  Y= 496.9  Z= 516.9  -> best X
  growth  30%: X= 494.8  Y= 544.6  Z= 566.6  -> best X
compute-only growth, on-demand
  growth 10%: X=517.2  Y=524.5  Z=559.1  -> best X
  growth 20%: X=556.8  Y=555.8  Z=595.4  -> best Y
  growth 30%: X=598.8  Y=589.0  Z=633.9  -> best Y
Y beats X on compute-only growth (on-demand) from growth of about 18.9%
Y beats X on compute-only growth (reserved) from growth of about 59.6%
egress multiplier at 10% growth, on-demand
  egress x1: X=529.6  Y=546.2  Z=579.2
  egress x2: X=562.7  Y=628.9  Z=711.7
  egress x4: X=628.9  Y=794.4  Z=976.5
reserved vs on-demand, X compute, year-1 $k at utilisation of hours
  30% of hours: on-demand 36.0  reserved 84.0
  50% of hours: on-demand 60.0  reserved 84.0
  70% of hours: on-demand 84.0  reserved 84.0
  100% of hours: on-demand 120.0  reserved 84.0

What the numbers say

• Pick X. It wins at all four growth rates tested (0%, 10%, 20% and 30%) when everything scales together. Its compute line is the highest of the three ($120k). Against Y, X pays $25k more for compute but saves $15k on storage and $15k on egress, so it is $5k a year cheaper at year-1 levels, and because every line scales with utilisation, the gap scales with it (X leads Y by $16.6k over the three years at 10% growth).
• Egress is the swing cost (the line whose uncertainty can change the ranking). Z's headline compute is mid-priced but its egress of $40k is four times X's. If egress volume doubles, Z goes from $579.2k to $711.7k while X goes to $562.7k; at four times, Z reaches $976.5k. A low hourly price paired with a heavy per-gigabyte egress rate loses to a higher instance price with generous in-region transfer when most traffic leaves the provider's network. To model that, split egress into in-region volumes (traffic that stays inside the provider's own region, often cheap or free) and out-of-region volumes with separate rates. As an illustration, if half of every provider's egress stays in-region at a quarter of the price, year-1 egress becomes X $6.25k, Y $15.6k and Z $25.0k (for example Z: $40k x (0.5 + 0.5 x 0.25) = $25k), and year-1 totals become X $156.25k, Y $155.6k and Z $160k, so X's $5k lead over Y disappears. That is why the split is worth asking for before signing.
• The choice does flip, but only on a narrow case. If just compute usage grows and storage and egress stay flat, Y overtakes X from about 18.9% yearly growth on-demand (at 20%, Y is $555.8k against X at $556.8k). With reserved compute the flip moves out to about 59.6%, a growth rate this workload is unlikely to see, because the 30% discount shrinks the compute gap that Y's low rate exploits.
• Reservations. A reservation costs 70% of the always-on on-demand price whether you use it or not, while on-demand costs only the hours you run. So the two cost the same when the instance runs 70% of the hours, and the reservation pays off only above that. Taking the $120k as the always-on on-demand cost, the reservation is $120k x 0.70 = $84.0k; at 30% of hours on-demand costs $36.0k, at 70% it is $84.0k, and at 100% it is $120k against $84.0k reserved. An office-hours workload (10 hours x 5 days = 50 of 168 hours, about 30%) sits on that first row: with a scheduled shutdown (stopping the instances outside working hours) on on-demand it costs about $36k, against $84k reserved. Reserve only an always-on baseline.

Trade-offs and pitfalls

• A three-way comparison on year-1 list prices alone says X wins by only $5k over Y, which is within the noise of a forecast; the lead only becomes convincing once the growth and egress cases are run. The decisive stress test is compute-only growth, the only growth case that flips the ranking, so that is the one to watch; the in-region egress split described above can also erase X's year-1 lead over Y, which is why it is the second thing to ask each provider about.
• Do not treat "utilisation grows 10%" as applying to every line without saying so; assumption 2 is what decides the ranking.
• Non-production often has a cheaper lever than provider choice: auto-shutdown schedules and smaller instance types.
• What would change my call: egress being mostly in-region, a quoted rate card (the provider's published price list) that differs from the figures given, or compute growing alone beyond about 19% a year.