Every machine we own is rented. We need more of them.

One platform, five ways to earn from it.

The cloud rental is how customers find us. The same DGX Spark then carries them into owning hardware and paying for our help, and the GPUwerk Platform lets us add capacity without buying every box ourselves. The same software already runs other GPU classes, so the fleet can grow beyond the Spark.

Status as of 1 October 2026. Only the cloud rental has revenue today. The other lines are built or offered and are what the round helps us sell.

We're not the only Spark host. Here's where we stand.

A handful of small providers rent the DGX Spark or its GB10 siblings. The large hourly marketplaces mostly don't: on 18 September Vast.ai listed no GB10 offers and RunPod and Lambda had no GB10 product. Meanwhile H200 and B200 lead times through hyperscaler channels ran 36 to 52 weeks (SemiAnalysis, March 2026), and small teams are at the back of that line.

Small-provider details checked on their public sites on 15 September 2026; large-cloud details on 18 September 2026.

Against the large clouds, the Spark is the cheapest way to get a lot of memory for inference, long context or mixture-of-experts models:

Dollars per gigabyte of memory per hour, from public pricing pages captured 18 Sep 2026. This compares memory capacity, not speed: the Spark's 273 GB/s LPDDR5x is not HBM, so our customers run inference and large-context work, not large training jobs. Vast.ai's indicative GB10 price (~$0.30/h) reflects past home-hosted supply; it listed no offers on the day we checked.

A $4,470 node, network included, that pays back in about 14 months.

Since launch, customers occupied 76% to 91% of the hours each Spark was online. The model prices every occupied hour at $0.79, our rate per node-hour.

Scenarios start at $0.79 per occupied hour and assume the rate falls 15% every year as newer hardware arrives, to about $0.41 in year five, while running costs stay flat. Hardware is $4,470 per node including network. Revenue and contribution per node-month above are for year one. ROI = (contribution over the period − hardware) ÷ hardware. Five years matches how large operators depreciate servers: Meta, Microsoft and Google use five to six years. The three-year figure is the conservative case. Company costs of about $1.8k/month (automated accounting, go-to-market, control-plane hosting; no founder salary from the round) are kept separate from node contribution. Colocation is priced from Coolhousing's public rates in Prague: a 42U rack from $499/month with 2 kW, extra power at $9.24 per 100 W up to 6 kW, and a 10 Gbps uplink, about $853/month on a 12-month term, with 24/7 access and KVM over IP. Power is metered on top at $0.26/kWh. Hardware is $4,387 per Spark excluding VAT, our best current price, plus one ~$1,000 MikroTik cluster switch per twelve nodes. Resale value at the end of the period is not counted, so the return is understated.

The hardware is getting more expensive

The same memory shortage that is raising GB10 prices is raising rental prices and lead times across the GPU market. For us that means three things. Nodes bought now cost less than nodes bought next year. The machines we already own would cost more to replace than we paid for them, which supports their resale value. And waiting to raise makes the same fleet more expensive. It is not only GB10: a used RTX 3090, a card from 2020, now sells for about $1,000. The model above uses our best current price, $4,387. Prices exclude VAT.

Why utilisation should rise as the fleet grows

• Sign-ups convert when there is a machine to use. New accounts that found no free Spark were put in a queue, so there was little reason to add credit: 3 of 44 have. With free capacity at sign-up, they go straight to a running instance.
• Snapshot and offload frees idle nodes. 13% of billed hours so far were customers holding a stopped Spark so they wouldn't lose their workspace. Customers will be able to snapshot an instance, release the machine, and restore later on any free Spark. That time becomes sellable, and customers stop paying to keep a machine idle. The workspace archive and 6-hourly backups this builds on already run in production.
• Reservations fill whole nodes. A reserved node is paid for every hour of the month. A bigger fleet lets us take reservations without starving on-demand customers.
• More nodes mean more configurations. Every rack has one four-node cluster with 512 GB of unified memory for models too large for one or two machines; the other nodes are sold alone or as linked pairs. Only the cluster nodes need 200 Gb/s networking, which keeps the network cost per node low.

Twelve-month plan, base case

Partner nodes: GPUwerk keeps 30% of $0.79/h at 70% utilisation, about $121 per partner node per month. On-premise sales, projects and managed services are upside and are not in these figures. With no founder salary paid from the round, company costs are small and the fleet covers them from about 6 own nodes; profit then grows with every node added.

Demand isn't the bottleneck. Capacity and operations are.

Money is released against utilisation, not hope.

If utilisation doesn't hold at 10 nodes, we stop buying and say why. At the base case, 47 nodes bring in about $19k a month; see the business plan above.

• Competition on price. AxForge already undercuts us, and nothing structural stops RunPod or Lambda from adding a GB10 product. Our edge is an EU base, operating experience on this exact machine, measured benchmarks, and a book of reserved contracts. Scale is what makes that edge last.
• NVIDIA reprices or retires GB10. We already run multi-vendor hardware, Lenovo and ASUS GB10 systems side by side, and the platform supports other GPU classes, so we aren't tied to one OEM or one product. All of it is still NVIDIA silicon.
• Rental prices will fall. GPU rental rates drop as newer hardware arrives, so the model cuts our rate 15% every year; at 30% a year, base-case payback moves from 14 to 15 months and the 5-year return halves. Front-loaded payback is what protects us.
• Hardware keeps getting more expensive. If prices rise faster than rental rates, payback on new nodes stretches. Buying in tranches limits how much we commit at any one price.
• Supply at volume is unconfirmed. We have bought single units from several vendors, not volume. Quotes for 40+ units are the first task after a term sheet.
• New lines are unproven. The Platform, on-premise, project and professional-GPU offers have no revenue yet, and outside the Spark we compete directly with RunPod and Vast.ai on price. Partner hosting also brings tax and liability work that we are handling country by country.
• It's early. Three and a half weeks of data, three machines, concentrated usage. We'll report utilisation, MRR and queue length to investors every month.

Talk to us.

Want to discuss the plan or follow our progress? Get in touch. We share detailed data, including the fleet model, in one-to-one conversations.