The archive · Hardware & Devices · Strategic decision · 2020–2022
TensorDock launches GPU cloud from $0.29/hr after a 7-month beta; Show HN 147
A GPU cloud born from a Boston crypto-mining rig sells data-center GPUs from $0.29/hr with resizable VMs; its 2022-07-28 Show HN drew 147 points.
TensorDock
What the business is
TensorDock rents GPU and CPU virtual machines for machine learning — from $0.29/hour for an NVIDIA RTX 4000 up through V100s, A100s and ten GPU types — on its own hardware plus resold capacity from large data-center partners, with a REST API, community CLI and, later, a marketplace where hosts list their own compute.
How it started
TensorDock started as the founders' own crypto-mining operation just outside Boston. When wholesale consumers approached in 2020 during pandemic surges, they added business internet and power backups and found themselves running a rudimentary office data center, with reseller sites selling on their fully-owned hardware. Boston electricity at $0.23/kWh pushed them toward cheaper tier-3/4 data centers, and the founders set out to build a platform that combined the functionality of large clouds with pricing somewhere between a dedicated server and the big clouds.
What happened
After about seven months in beta, Core Cloud launched on 2022-07-28 with GPU VMs deployable in 45 seconds: CPU-only servers from $0.027/hour, RTX 4000s from $0.29/hour and Tesla V100s from $0.52/hour, across three US locations (New York, Chicago, Las Vegas). Differentiators included storage-only billing when a VM is stopped ($0.073/GB/month), editing a VM after creation — for example downsizing an A6000 to an A5000 without recreating it — 3x-replicated NVMe network storage, 10 Gbps networking, ready ML images, a REST API and a community CLI. In the thread the founders said they owned hardware in Boston and at Equinix Singapore and resold capacity from four large 1,000+ GPU operators, and they planned a marketplace launching later in August where hosts running OpenStack would set prices and customers would deploy directly.
How it ended up
TensorDock is still operating: the homepage fetched on 2026-09-05 sells GPU cloud globally — H100s from $2.25/hour, up to 30,000 GPUs through close partners across 100+ locations, instant VMs in beta and a host marketplace where independent suppliers set prices — evidence the marketplace bet materialized.
Background
TensorDock began as a crypto-mining operation just outside Boston. When wholesale customers approached the founders in 2020 during pandemic surges, the rig grew business internet and power backups and became a rudimentary office data center whose fully-owned hardware reseller sites were already renting out.
The bet was that machine-learning teams wanted big-cloud functionality at dedicated-server prices. After seven months in beta, the founders launched Core Cloud on 2022-07-28 with GPU VMs deployable in 45 seconds — RTX 4000s from $0.29/hour, V100s from $0.52/hour, ten GPU types, three US locations — plus features the big clouds make painful: storage-only billing while a VM is stopped, and resizing an instance (say A6000 down to A5000) without recreating it.
The Show HN drew 147 points and 52 comments on 2022-07-28. Commenters benchmarked TensorDock against Lambda Labs, Coreweave and Paperspace across five GPU types, asked pointed questions about security on resold hardware, and several said they were signing up that weekend; the founders answered that they owned hardware in Boston and Singapore, resold capacity from four 1,000+ GPU operators, and planned a marketplace where hosts would set their own prices.
The marketplace bet materialized: the homepage fetched on 2026-09-05 describes TensorDock as a marketplace of independent hosts who compete on price, offering up to 30,000 GPUs through close partners across 100+ locations, H100s from $2.25/hour, consumer GPUs from $0.12/hour, instant VMs and a CPU cloud.
What has to be true
- The origin was real infrastructure, not a pitch: owned mining hardware had already become a paying data center, giving unit economics to price between dedicated servers and hyperscalers.
- The wedge was developer flexibility: resizable VMs, networked storage, storage-only billing and 45-second deploys addressed the pain of migrating data when right-sizing GPU workloads.
- Dated traction: the 2022-07-28 Show HN drew 147 points and 52 comments, with commenters comparing prices card-by-card and several committing to try the service.
- The two-sided marketplace removed the capex ceiling: partners supplied 1,000+ GPU fleets while TensorDock supplied software, an 'Amazon of compute' play without buying every machine.
What can be applied
Turning your own underused hardware into a service gives a small player pricing power against hyperscalers, and adding a host marketplace then scales supply without your own capex.
Aftermath
As of 2026-09-05, TensorDock is still operating: its homepage sells 'affordable GPU servers for everything AI' with H100s from $2.25/hour, consumer GPUs from $0.12/hour, 45 GPU models, instant VMs in beta and a CPU cloud, run as a marketplace of independent hosts across 100+ locations in 20+ countries with up to 30,000 GPUs through partners, naming airgpu, ELBO and Creavite as customers. The 2022 Show HN documents the launch state: 45-second deploys, ten GPU SKUs, three US locations, storage-only billing and a marketplace planned for late August 2022.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card