Hilbeon Request a demo

Cost & deployment

The Cluster You Already Own

By Hilbeon · 4 July 2026 · 7 min read

TCO On-prem vs cloud R&D economics

The best cluster for most quantum chemistry in 2026 is the workstation already under your desk. Not because clusters are bad — because the jobs got smaller and the desktop got faster. A 16-core box now converges a fullerene at B3LYP in about three and a half minutes and a cholesterol-sized local-correlation (DLPNO-MP2) job in four. Once the compute you need fits on hardware you own, the whole cost conversation flips from “how many core-hours do we buy” to “why are we renting at all.”

Let's put real numbers on it — both the throughput and the total cost of ownership.

What a 16-core desktop already does

These are measured Hilbeon wall-clock times on a single 16-core node, each written up in its own post so you can check the setup:

Job Size Wall time, one desktop Write-up
C₆₀ fullerene, B3LYP/3-21G 60 atoms 215 s C60 in 215 seconds
Cholesterol DLPNO-MP2 (74 atoms) local-correlation energy 247 s · 99.7% recovered DLPNO on the desktop
Orforglipron, RHF 113 atoms · 1006 basis functions converged, bit-identical ×2 machines Inside orforglipron

A fullerene, a cholesterol-sized local-correlation job, and a 113-atom oral drug — none of them needed a queue, an allocation request, or a data-transfer plan. They needed a desktop and a few minutes.

The rental meter

Now price the same 16 cores in the cloud. A fair match to a 16-core desktop is an AWS c7i.8xlarge — 32 vCPUs (16 physical cores with hyper-threading) and 64 GiB of memory — at roughly $1.428 per hour on-demand. So a single C60 run costs about $0.085 of compute. Cheap! That is exactly the number a cloud calculator shows you, and exactly why the per-job framing is a trap.

You don't run one job. You run a research program. Keep one 16-core node genuinely busy just four hours a working day — a conservative load for anyone doing method development or a screening campaign — and the meter reads:

Line item Own the workstation Rent the equivalent (cloud)
Hardware ~$2,500 once $0 capex
Compute, 1 year @ 4 h/day (880 h) ~$100 electricity $1,257 (880 × $1.428)
Compute, 3 years ~$300 electricity $3,770
Pulling results back (egress) $0 — already local $0.09 / GB, one way
Queue wait none — interactive hours, shared allocation
3-year total ~$2,800 ~$3,770+

At a modest four hours a day the workstation breaks even in its third year and comes out roughly $1,000 cheaper over three. Double the load to eight hours — normal for an active screening or method-dev user — and the cloud bill is about $2,513 a year: the desktop pays for itself in around thirteen months, then keeps running for the price of electricity.

Assumptions: cloud compute at the c7i.8xlarge on-demand rate; electricity ~$100/year (≈250 W under load, 4 h/day, ~$0.15/kWh — rate-dependent). The throughput jobs above are single-point energies (see the linked posts for geometry and level-of-theory details); they measure wall-clock, not full optimizations.

The costs that never make the calculator

The table above is generous to the cloud, because it only counts core-hours. Three real taxes sit outside it:

  • Egress is a one-way street. Wavefunction dumps, orbital and density cubes, NCI grids, MD trajectories — a live orbital cube is ~2 MB, an NCI grid ~16 MB, a screening campaign tens of gigabytes. At $0.09/GB every result you want to actually look at is metered on the way out. On the box under your desk it is already on your disk.
  • Queue latency dwarfs run time. A 215-second job that waits three hours for a shared allocation has a 98% idle wall-clock. Per-second billing doesn't fix that — only zero-queue, interactive hardware does. Iteration speed, not core price, is what actually gates a project.
  • Someone runs the cloud. Machine images, storage lifecycle, spot-interruption handling, IAM — that is salaried DevOps time the core-hour price silently omits. A workstation you turn on.

What this means for whoever signs the invoices: for the steady diet of small-to-mid molecules that most medicinal-chemistry QM actually is, a local workstation plus a tool that uses it well is cheaper and faster to iterate than renting — and the data never leaves your building, which your IP and compliance teams will also appreciate.

Where the cloud still wins — honestly

This is not an anti-cloud argument. The cloud is the right tool for two cases, and we'll say so plainly: rare massive bursts — thousands of cores for a single afternoon, where owning that capacity would be absurd — and genuine zero-capex situations where a one-time hardware spend is the blocker. If your workload is spiky and enormous, rent it. If it's a steady stream of drug-sized molecules, the economics point the other way, and they point hard.

Hilbeon is built for that second world: a package simple enough that the desktop under your desk becomes the whole facility. The throughput above is real, measured, and reproducible — and it runs on-premises, on hardware you already have, with your data staying exactly where it is.

Turn the box under your desk into the facility

Start a 30-day guided pilot — every method, every core, your GPU — and see what your own workstation does with your compounds.

Request a demo See the C60 benchmark

References & sources