Ask a room of executives what their mainframe costs and you will get the hardware lease, maybe the IBM bill. Ask what leaving it would cost and you will get silence, or a number someone heard at a conference. Neither answer survives contact with a CFO.
This page is the anatomy of both numbers — the fully loaded cost of staying, the fully loaded cost of leaving — and the arithmetic that connects them. Every industry figure below is flagged as an estimate; every worked example is labeled illustrative. When you want your own numbers instead of ours, the exit calculator models them in the open.
01 · The anatomy of the run cost
The run cost hides in four places. Only two of them appear on an invoice.
Hardware and facilities — the visible line
The machine itself is usually the smallest component: the lease or depreciation on the CEC, storage, tape and networking, plus the floor space, power and cooling around it. It is also the line with a built-in clock. IBM ships a new mainframe generation on a regular cadence — z16 arrived in 2022, z17 followed the cycle after — and each refresh is a forced decision point. Refreshing is usually cheaper than nursing aging hardware through a maintenance tail, but signing the refresh recommits the estate to the platform for another multi-year cycle. That is why the refresh window, not the outage or the audit, is when most exit conversations actually start.
Software licensing — where the money actually goes
On most estates, software is the largest single component of the run cost. Three mechanics explain why.
MLC. Core IBM software — z/OS, CICS, Db2, IMS, MQ — is billed as a Monthly License Charge: a recurring rental, not a purchase. You never finish paying for it, and the price is indexed to how hard the machine works.
Sub-capacity and the four-hour peak. Under sub-capacity licensing, usage is reported monthly (via SCRT) and the charge follows the rolling four-hour average peak — the busiest sustained window of the month, not the average. One heavy batch window, one quarter-end crunch, and the whole month is priced against it. Newer constructs such as Tailored Fit Pricing trade the peak mechanics for committed consumption, which smooths the bill without shrinking it.
IPLA and the ISV layer. One-time-charge (IPLA) products carry annual subscription and support, and the independent software vendors — the scheduling, monitoring, security and database tools stacked on the estate over decades — are commonly priced on capacity tiers. Grow the MIPS and several line items step up together.
Put together, published estimates for software cost alone commonly fall in the range of $2,000–$4,000 per installed MIPS per year on large estates, with fully loaded per-MIPS figures higher once everything else is counted. Treat any per-MIPS number as a placeholder for your own bill — but notice the shape: the cost scales with capacity, and capacity only ever drifts upward.
People — the retiring-SME premium
The payroll line understates the people cost in two ways. First, scarcity: the engineers who can read the estate are retiring faster than they are being replaced, and the market prices that scarcity into every contractor day-rate and every counter-offer. Second, concentration: on many estates, the working knowledge of a core system lives in the heads of a handful of people. That key-person risk never appears in the ledger — until a departure turns a routine change into a research project. The premium is not that mainframe salaries are extravagant; it is that every year the same knowledge costs more to keep and more to lose.
Opportunity cost — the invoice that never arrives
The fourth component has no line item at all. The ERP program that cannot cut over because the general ledger still closes on the mainframe. The S/4HANA migration with an unscoped mainframe dependency. The AI roadmap that stalls because the business logic it needs is written in code the models cannot parse and the remaining SMEs cannot fully explain. Staying does not just cost the run-rate — it holds every dependent program at the same red light. For estates feeding an ERP cutover, this is usually the largest number of the four, and the hardest to write down.
figure 01 · where the run-rate dollar goes
Illustrative composition of a fully loaded annual mainframe run cost on a mid-size estate. Your split will differ — model it in the calculator.
02 · Trimming MIPS is a treadmill
Because the bill is indexed to capacity, the first instinct is always MIPS cost reduction — and the levers are real. Tune the heaviest batch jobs. Move eligible work to zIIP engines, which sit outside the MLC calculation. Reshape schedules so the four-hour peak lands softer. Cap LPARs. Renegotiate the enterprise agreement. Archive cold data. Each of these works, and a disciplined program can take a visible slice out of the bill in the first year.
Then three things happen. Workload growth — a few percent a year on most estates — quietly compounds back everything you cut. The levers you pulled cannot be pulled twice: a tuned job is tuned, an offloaded workload is offloaded. And nothing about the exercise touched the other two costs — the people premium keeps climbing with every retirement, and the blocked programs stay blocked, because the estate they depend on is exactly as opaque as it was before the tuning began.
The compounding arithmetic is unforgiving. At 4% annual growth, capacity rises about 17% over four years — which means a hard-won 15% reduction is fully erased in under four years, and the estate then keeps growing from a higher base. A second optimization round starts from thinner pickings than the first, because the easy tuning is already done. Each cycle costs more engineering effort and returns less bill. That is the definition of a treadmill: real motion, no displacement.
MIPS reduction is worth doing — do it. But it is a tactic that buys time, not a strategy that ends the spend. You run faster to stand still, and the treadmill does not care how fit you are.
The strategic question is not “how do we shave 15% off the bill” — it is whether the estate should still exist in its current form at the next hardware refresh. That question has a doctrine of its own: the Exit Doctrine.
03 · What an exit costs
An honest exit budget has six phases. Vendors who quote you fewer are usually planning to discover the missing ones on your invoice later.
| phase | what it covers | what drives the cost | typical share |
|---|---|---|---|
| Discovery & rationalization | Inventory the estate, find the dead code, decide what moves, what retires, what stays. The Palm 360 stage. | Estate size, artifact hygiene, how much nobody remembers | 5–10% |
| Rules extraction | Lift the business rules out of the code with full source lineage. The Palm Ark stage. | Program count, logic density, dialect spread | 15–20% |
| Build / forward engineering | Redesign the rules for the target — cloud services, a modern stack, or an ERP fit-gap. The Palm Ray stage. | Target platform, integration count, data migration | 30–40% |
| Parallel run | Both systems live, outputs reconciled daily until the evidence is boring. | Length of the confidence window, reconciliation tooling | 15–20% |
| Cutover | The switch itself — rehearsed, reversible, scheduled twice before it happens once. | Rehearsal count, rollback design | ~5% |
| Decommission | Licenses terminated, hardware retired, archives retained for the regulator. | Contract exit terms, data retention obligations | 5–10% |
Before the table becomes a budget, one decision cuts it more than any negotiation: scope. The single largest cost lever in an exit is not moving code cheaply — it is not moving code at all. Rationalization work routinely finds that a double-digit share of programs in a decades-old estate is dead, duplicated, or feeding reports nobody has opened in years. Every program retired in discovery is extraction, build, and parallel-run spend that never happens. This is why the cheapest phase in the table exists, and why skipping it to “save time” is the most expensive decision on the sheet.
Two more observations from the table. First, the phases most often skipped — discovery and parallel run — are precisely the ones whose absence shows up later in the post-mortems. Cutting them does not remove the cost; it converts it into incident response, at a worse exchange rate.
Second, the build and parallel-run phases — half or more of the envelope — are hostage to one upstream variable: how accurately the rules were extracted. Rework is the hidden multiplier in every failed conversion. In our internal benchmark, deterministic extraction lands around 95% rule accuracy with 100% source lineage, where generic LLM approaches land near 30% — and every rule that arrives wrong is paid for three times: built wrong, caught in parallel run, built again. The benchmark page documents the methodology; the budget consequence is simple: extraction quality is the cheapest place in the whole program to spend well.
04 · The three numbers a CFO needs
Strip away the workshop decks and a mainframe exit is three numbers.
1. The run-rate. What a year of staying costs, fully loaded — software, hardware, facilities, people — growing along your MIPS curve. Not the IBM bill: the whole anatomy from section 01.
2. The exit envelope. The all-in program cost, discovery through decommission, parallel run included. A useful planning form is a multiple of one year’s run cost per year of program — it keeps the estimate honest as scope moves.
3. The break-even month. The month the cumulative cost of the exit path drops below the cumulative cost of staying. Everything before it is investment; everything after it is return. This is the number the board meeting is actually about.
worked example · illustrative only
A 2,500-MIPS estate: $3,000 per MIPS per year in software, $0.9M hardware and facilities, 24 FTEs at $130k loaded, 4% annual MIPS growth. A 30-month program at an envelope of 120% of one year’s run cost per program-year, landing at a post-exit run cost of 35% of current. Every line is an assumption — the calculator lets you change all of them.
Read the example coldly. Break-even at month 77 is nobody’s idea of a quick win — it is a six-and-a-half-year commitment, and the saving only starts compounding after cutover at month 30. That is the honest shape of a full exit at these assumptions. It is also the pessimistic frame, because the model prices none of the opportunity cost: no unblocked ERP cutover, no retired key-person risk, no AI program that finally has legible business logic to work with. Those arrive as upside the spreadsheet never claimed.
And the same arithmetic cuts the other way: shrink the envelope, shorten the program, or cut the post-exit percentage, and break-even walks left. This is exactly where extraction accuracy, scope discipline and target choice earn their keep — they are not quality abstractions, they are levers on one number.
It is worth knowing which lever moves the number most. In this model, break-even is most sensitive to program duration and envelope together — every extra program-year both adds spend and delays the savings — and to the post-exit percentage, which sets the slope of the payback line forever after. The run-rate inputs mostly scale both paths together. Practically: a program that finishes six months sooner beats one that costs 10% less, and a target that runs lean beats both. Test the sensitivity yourself; the model recomputes as you type.
05 · Getting your numbers
Ours are illustrations. Yours are three inputs away.
The Mainframe Exit Calculator runs this exact model in your browser — run-rate, envelope, break-even, and the seven-year curve — from assumptions you control. No email, no gate, formulas in the open. When the directional answer is interesting, a scoped discovery with Palm Key replaces the assumptions with an inventory.
06 · Questions, answered
How much does mainframe modernization cost?
There is no single number, and anyone who quotes one without seeing your estate is guessing. The cost scales with estate size, logic density, and how much of the estate actually needs to move — which is why discovery comes first. A useful planning frame is an exit envelope expressed as a multiple of one year’s run cost per year of program; the illustrative 2,500-MIPS estate above lands around $34.6M over 30 months. The honest answer is a modeled one, from your inputs.
What does a mainframe cost per MIPS per year?
Commonly cited estimates put software cost alone at roughly $2,000–$4,000 per installed MIPS per year on large estates, with fully loaded figures higher once hardware, facilities and people are included. The spread is wide because contracts, workloads and ISV stacks differ; use the range only as a starting assumption and replace it with your own bill.
Is MIPS cost reduction enough, or do I need to leave?
Do the reduction — tuning, zIIP offload, peak management and renegotiation buy real money in year one. Then notice what they cannot do: growth compounds the bill back, the levers only pull once, and neither the retiring-SME premium nor the blocked-program cost moves at all. Trimming is a tactic. Whether to leave is a separate, strategic question — the Exit Doctrine exists to decide it.
How long does a mainframe exit take to break even?
It is a function of four inputs: run-rate, envelope, program duration, and post-exit run cost. The illustrative model on this page breaks even around month 77; more aggressive assumptions land earlier, heavier ones later. The calculator computes the month from your inputs and shows the crossing on the curve.
What is MLC, and why does it dominate the bill?
MLC — Monthly License Charge — is IBM’s recurring charge for the core software stack: z/OS, CICS, Db2, IMS, MQ. Under sub-capacity licensing the charge follows the measured rolling four-hour average peak, reported monthly, so the bill is set by your busiest sustained window rather than your average load. It is a rental you never finish paying, indexed to a peak you can only partially control.
Does the z16-to-z17 refresh change the economics?
Each generation genuinely improves price-performance, and refreshing usually beats paying the maintenance tail on aging hardware. But the refresh also recommits you to the platform for another multi-year cycle — hardware, software agreements, and skills. That makes the refresh window the one natural moment when staying and leaving can be priced side by side, on the same page, before the signature closes the question for another cycle.