pricing · the managed fleet
$400 a month per managed 8-GPU cell.
That is the price, and it is one number before any estimator. A single-cell account is $149; a fleet is banded by cell count at the same rate per cell. The GPU-hour is not in it — you pay the marketplace, at the marketplace's rate, on your own account.
What the fee buys.
Not a GPU, and not a discount on one. The fee is for the thing actually scarce: a node when there isn't one, behind an endpoint that doesn't move. Everything else in this list is the machinery that makes that sentence true.
The rotation
Add before remove, drain to zero, and the move scheduled against measured session gaps rather than a price tick — a median tool gap of 0.3 s, p90 13.6 s. The cell under your endpoint changes; the endpoint does not.
The continuity and park mechanism
A cell that has to go is parked rather than destroyed, so the weights and the cache survive the move instead of being re-fetched on the paid clock. It is the half of a rotation a buyer pays for twice when it is missing.
The ledger
Fifty-one rotations in the record, forty-two clean, six failures named beside the wins. A record that renders only wins is marketing, and this buyer prices marketing at zero.
The contract-end roster
Every live contract with its end date in one list, so a rotation is planned rather than discovered. The cheapest layer is not spot: committed blocks run 49% below on-demand at one month and 65% at twelve, and both read out of stock the day they were checked.
Availability is GPU Finder's 30-day duration-weighted provider telemetry, September 2026, H100 and B200 at one GPU. Rotation counts are the estate's own ledger. The committed-block discounts are Latitude.sh's published rates, read on 22 October 2026 — and both were out of stock on the day they were read.
Why a flat fee, and not a share of the savings.
The obvious shape for this business is contingency pricing: take a fraction of what the rotation saves. It reads well and it is bimodal. A buyer either finds a cell and saves sixteen to twenty-one percent, or finds nothing and saves nothing. The estate's own line paid $515 one month and $0 the next, across twelve price or cost events in about forty-one days with two genuinely clean snipes. A bill that swings between those two values cannot be budgeted, and it cannot fund a recurring service.
So the contingency was demoted from the product to an opt-in second line, and one worked sum says why. An 8-GPU cell at the $3.29 H100 provider median is $19,214 of GPU spend a month — 8 GPUs at 730 hours. A five-percent delta on that is $961 of savings, and half of it is $480.
It is honest in the other direction too. The fee does not grow when your fleet gets more expensive, so there is no reason to keep you on a pricier cell — and no delta to argue about at the end of the month.
A page that hides what is excluded is the thing it criticises, so they are printed below as a section rather than a footnote.
What is not included.
Four lines that belong to somebody else. None of them is billed here, and none is netted against the fee.
- The GPU-hour
- Billed by the marketplace, on your account, at the marketplace’s rate. The fee never touches it — so a cheaper cell you find is entirely yours, and the fee stays flat either way.
- Bandwidth and egress
- The two providers’ line, metered by whichever cells the traffic lands between. A rotation crosses two of them and neither is on this invoice.
- Storage
- Everything past the cell’s own disk — checkpoints, weight caches, the dataset. Whoever holds the volume bills it.
- Your account discount
- Whatever rate your marketplace account already carries stays yours, on your own invoice. It is never netted against the fee, and never counted as ours.
Three bands by fleet size.
One rate per managed cell. A band is a count of cells, not a fraction of your spend, so the bill does not move when the market does.
One cell
$149a month — the single-cell floor
One managed 8-GPU cell: the availability guarantee, the rotation, the ledger, the contract-end roster and the continuity/park mechanism.
$149 ÷ 8 GPUs — $18.63 a GPU-month, 0.8% of the $2,402 a GPU costs at the $3.29 H100 median.
the entry bandA small fleet
$400a month per managed 8-GPU cell
The standard band, and the number the rest of this page defends: one rate per cell, whatever the market is doing that morning.
$400 ÷ 8 GPUs — 1.6% of a $3,103 GPU-month at the $4.25 cross-class median; 2.1% of $2,402 at the $3.29 H100 median.
the default bandA large fleet
by fleetquoted against cell count, at the standard rate
Owned or on-prem capacity as the committed block beside the rented peak, with one endpoint and one console over both.
4 cells × $400 — $1,600 a month — still $50 a GPU-month. A band is a count of cells, not a share of spend.
talk to usA GPU-month here is 730 hours, and a GPU-month's cost is the hourly rate times 730. The fee itself is flat, per cell, per month: not a share of your savings, and not tied to your spend.
What it is measured against.
The number is not invented. Three published fee bases translate to a band of $30 to $80 per GPU-month, and $50 sits inside it. A fourth is where the floor comes from.
- Cast AI
- $5 a managed CPU-month — a per-managed-CPU rate. Carried onto the same unit: $320 a node-month on a 64-vCPU / 8-GPU node, so $40 a GPU-month.
- CloudHealth
- about 3% of tracked spend — a share of tracked cloud spend. Carried onto the same unit: $2,577 a GPU-month at the $3.53 weekly median, so $77.
- managedgpus
- $1,000 a month per 4 servers — a flat fee on hardware somebody else owns. Carried onto the same unit: $250 a server-month, so $31 a GPU-month at eight GPUs to a server.
- CloudFix
- from $149 a month — the category’s entry point — where the floor comes from. Carried onto the same unit: tiered on spend, billed through the marketplace. The same floor this page prints for one cell.
Each row is the vendor's own published price, carried onto the same unit. The translation assumes eight GPUs to a server and 730 hours to a month; the printed price is the vendor's, the carried figure is the record's arithmetic on it.
$400 per 8-GPU cell is $50 per GPU-month — inside the $30-to-$80 band, and printed as one number rather than a range to negotiate in. The band is the record's, not a sales range: it is what three independent published fee bases converge on once each is carried to the same unit.
Six questions a buyer asks.
Short answers. The long ones are on the rest of this page, and the record behind it is named at the foot.
What am I actually paying for?
Availability, and the record of it: a node when there isn’t one, the rotation that keeps your endpoint still, the ledger and the contract-end roster. Not the GPU-hour — that line belongs to the marketplace.
Why flat, and not a share of the savings?
A share is bimodal: the estate’s own line paid $515 one month and $0 the next, across twelve price or cost events in about 41 days with two clean snipes. $400 is a number a buyer can budget; a percentage of a delta is a number neither side can forecast.
What does the invoice not cover?
The GPU-hour, bandwidth and egress, storage past the cell’s own disk, and any account discount you already hold. The first three are the marketplace’s lines; the last is yours, and nothing is netted against the fee.
What happens when a cell has to move?
Add before remove, drain to zero, and land the move inside a measured session gap rather than on a price tick — median 0.3 s, p90 13.6 s. Forty-two of the fifty-one rotations in the record were clean, and the six failures are named.
Is there an uptime SLA?
Bounded windows only. An open-ended SLA on spot-backed capacity is a promise nobody can keep, so the guarantee is a node for a stated window and a rotation when the window ends.
Why is one cell $149 and a fleet $400?
Because the floor is the category’s own entry point — CloudFix opens at the same $149 — and one cell is one endpoint to watch. A fleet is priced per cell at $400, so the bill is a count of cells rather than a fraction of your spend.
Where the numbers come from.
Every figure on this page is a capture from one of five places. Where a number is a translation — a GPU-hour carried to a month, a per-server fee carried to a GPU — the arithmetic is printed beside it rather than folded into it.
- the position
- The ruling itself: the $400 fee, the $149 floor, three bands by fleet size, the two anchor translations, the availability table, the 51-rotation record and the $515 bimodal line.
- the seat arithmetic
- The retest seat behind that ruling: the anchor translations in full, the $19,214 sanity check that turns a $0-to-$480 contingency into a $400 line, and the $149 floor’s source.
- the ledgers
- Dated daily price ledgers for the $3.29 H100 provider median, the $1.80 on-demand floor and the $0.94 spot floor; GPU Economy’s published weekly series for the $4.25 and $3.53 medians.
- the vendor pages
- Cast AI, CloudHealth, managedgpus and CloudFix, each read at its own published price list rather than through a summary of it.
- availability
- GPU Finder’s 30-day duration-weighted provider telemetry, September 2026, H100 and B200 at one GPU a month.
one number, no table
$400 a month, and the market is somebody else's morning.
The kit is free. The managed fleet is one flat number per cell. Pick the one that fits how much of the watching you want to keep.