say “swap-meet”

The marketplace for idle silicon.

Rent out the CPU, RAM and GPUs you're not using. Reclaim them the moment you need them back. Buyers get cheap compute; sellers earn on hardware that would otherwise depreciate to zero.

Request access → See how it works
Hardware-attested isolation Reclaim in seconds Metered & escrowed
Built on AMD SEV-SNPIntel TDXNVIDIA H100 CC Attested with SPIFFEMigrated with CRIU
The problem

Two failures of one market, on opposite sides.

Supply side

Capacity that evaporates

Fleets are sized for peak. Nights, weekends, between training runs, most of it idles — and idle silicon depreciates whether or not it runs.

60–70%typical off-peak idle on large fleets · illustrative
Demand side

Compute that's scarce

Startups, researchers and AI teams queue for GPUs and pay hyperscaler premiums — while usable capacity sits dark a network away.

2–4×on-demand premium vs. the marginal cost of idle hardware · illustrative

The blocker was never wiring the two together — it's trust (a stranger's code on your fleet) and reclaim (getting your machine back instantly). swpmt is built around exactly those two.

How it works

Submit a job. It runs on someone's idle hardware. It survives reclaim.

Submit

Spec resources, price ceiling, preemption tier & attestation policy — CLI, SDK or console.

Attest & match

The scheduler binds an attested stall; your data key releases only inside the verified enclave.

Run & checkpoint

Work runs in a TEE; state is checkpointed on your interval, so loss is always bounded.

Reclaim & migrate

When the owner wants it back, the job drains, hands off to another stall, and resumes — automatically.

Sellers & buyers

One floor, two ways to win.

▦ vendors

Sell idle capacity

Install the node agent, set your hours, floor price and how much warning you'll give. Earn while attested workloads run in isolation.

  • Zero risk to production — tenants never touch your network or other tenants.
  • Reclaim on demand — one click drains and hands off within your pledged window.
  • Real revenue on hardware that idled at 2am.
Set up a stall →
◈ shoppers

Buy cheap compute

Submit a workload with your requirements and max price. Pay only for metered usage, drawn from escrow.

  • Your IP stays yours — the host can't read your weights or data.
  • Interruption is safe — you resume automatically, losing at most one checkpoint interval.
  • Prices well below on-demand, because idle hardware is nearly free at the margin.
Shop the floor →
Why sellers say yes

Confidential computing makes strangers-on-your-fleet safe — both directions.

◇ enclaves

Encrypted in use

Buyer memory is encrypted by the CPU/GPU (SEV-SNP, TDX, H100 CC). A curious host can't read weights or data.

✓ attestation

Proven, not promised

Data keys release only to an enclave whose measurement matches policy — which also kills the “claims H100, delivers T4” fraud.

⇥ egress lockdown

No phoning home

Default-deny egress on a managed overlay; your production subnets are never routable from a tenant.

attestation flow
measure platform→ verify vs. endorsed refs + revocation→ issue short-lived identity→ release data key into enclave
Pricing

Interruption is priced, not prevented.

Sell the same hardware at three interruption guarantees. Pick the loosest tier your workload tolerates and pay the lowest price.

cheapest

Spot

from $1.50 / H100-hr

Hard preempt. Reclaimed with a short deadline.

  • 30–120s warning
  • Best for checkpoint-friendly work
  • Serverless, batch, frequent-checkpoint ML
Start on Spot
popular balanced

Standard

from $2.10 / H100-hr

Soft preempt with a guaranteed drain window to checkpoint & migrate.

  • ~5 min guaranteed drain
  • Warm live-migration when possible
  • Containers, VMs, medium-checkpoint ML
Choose Standard
premium

Reserved

from $3.00 / H100-hr

No preemption within the lease. Provider stakes collateral and is slashed if broken.

  • No interruption
  • Stake-backed guarantee
  • Latency-sensitive & non-interruptible
Reserve capacity

Price to the buyer, per H100-class GPU-hour

illustrative — shape, not a quote
On-demand
$4.10
1-yr reserved
$2.60
GPU-cloud spot
$2.20
swpmt spot
$1.50
Early access

Bring what you're not using. Take what you need.

We're onboarding a first cohort of vendors and shoppers. Tell us your fleet or your workload and we'll get you on the floor.

Or email hello@swpmt.com.

Herman, the swpmt hermit-crab mascot
meet Herman

A hermit crab swaps shells.
swpmt swaps hosts.

Herman's our mascot for a reason. When a seller wants their machine back, your workload just scuttles into a new shell — checkpointed, migrated, and running again — without ever losing its home.