Rent vs Own AI Hardware: a 36-Month Cost Comparison

Most teams first meet AI compute as an hourly price. You spin up a GPU instance, run the experiment, shut it down and pay for the hours. That model is hard to beat for a one-off job. It gets harder to defend once the "experiment" becomes a model your team queries every working day, an agent pipeline that runs all night, or a fine-tuning loop you go back to every week.

This article puts numbers on that change. We compare owning a DGX Spark (bought outright, racked and managed by us) with renting cloud GPUs by the hour over 24 and 36 months. We show every assumption, so you can swap in your own numbers, and we're clear about when renting is still the right answer.

The two machines are not the same, and that matters

A fair comparison has to start here. A DGX Spark is not an H100. They're built for different jobs.

• DGX Spark has 128 GB of unified LPDDR5x memory with 273 GB/s of bandwidth, a GB10 Grace Blackwell Superchip, and up to 1 PFLOP of FP4 compute with sparsity (NVIDIA DGX Spark User Guide). Its strength is capacity: NVIDIA rates one unit for models up to 200B parameters and fine-tuning up to 70B (NVIDIA DGX Spark product page).
• A datacenter GPU such as an H100 or A100 has far more memory bandwidth and raw compute, but less memory per GPU: 80 GB on the H100 and A100 instances we price below. A cloud H100 will generate tokens from a model that fits in 80 GB much faster than a Spark will.

So there are two honest framings:

1. "I need a box that can hold a large model and serve it to a few people." Here the Spark competes with a cloud instance that has enough memory. It doesn't need to match its speed.
2. "I need maximum throughput on a model that fits in 80 GB." Here a cloud H100 does more work per hour, and the comparison per hour flatters the Spark. If this is you, weigh throughput, not just hours.

Below, we compare against the cheapest single-GPU instances on two public price lists with at least 80 GB of GPU memory, plus Lambda's GH200, which has 96 GB. None of them holds as much as a Spark's 128 GB. None of them is as slow, either.

The cost of owning

Owning means a one-time hardware price plus a monthly management and rack fee:

• DGX Spark hardware: $6,500 per unit (one-time)
• Management and rack: $99 per unit per month, billed from the day the machine is racked

The fee covers power within an agreed envelope, cooling, a shared 1 Gbps port with 2 TB outbound a month, private remote access with full root, monitoring of temperature, disks and reachability, reboots on request, one approved OS reimage a month and 15 minutes of remote hands a month, with a four-business-hour response target. See what management includes.

Over the two horizons:

Prices exclude taxes and import duties. The machine is still yours at the end. We come back to what that's worth below.

The cost of renting

We took on-demand prices from two public price lists on 31 Oct 2026:

Lambda's prices are listed as "plus applicable sales tax/VAT/GST". Cloud prices change often, so check the live pages before you decide.

Two utilisation levels

• Working hours: 8 hours a day on 22 working days = 176 hours a month. This is a team that uses the model during the day and shuts the instance down at night (and actually remembers to).
• Always on: 24 hours × 365 days ÷ 12 = 730 hours a month. This covers a shared endpoint, an overnight agent pipeline, or a model loaded and waiting for requests.

What renting costs at each level

Each cell is hours × hourly rate, multiplied by 36 for the three-year columns. These figures cover compute only. Persistent storage, snapshots and data transfer are priced separately by each provider.

Break-even: how many hours a month?

The cleanest way to compare is to ask how many rented hours a month cost the same as owning.

Over 36 months, owning costs $10,064, or $279.56 a month. Divide that by each hourly rate:

• GH200 at $2.29: $279.56 ÷ $2.29 ≈ 122 hours a month (about 5.5 hours per working day)
• H100 PCIe at $3.29: $279.56 ÷ $3.29 ≈ 85 hours a month
• A100 Secure at $1.79: $279.56 ÷ $1.79 ≈ 156 hours a month
• A100 Community at $1.39: $279.56 ÷ $1.39 ≈ 201 hours a month

Over 24 months, owning costs $369.83 a month, so the break-even rises: about 162 hours a month against the GH200, 112 against the H100 PCIe, 207 against the Secure A100 and 266 against the Community A100.

Payback period

Another way to read it: if renting costs more per month than the $99 fee, the monthly difference eventually pays back the $6,500 up front.

• Always on (730 h) against the GH200: $1,671.70 − $99 = $1,572.70 saved a month, and $6,500 ÷ $1,572.70 ≈ 4.1 months to pay back.
• Always on against the Community A100: $1,014.70 − $99 = $915.70 a month, which gives ≈ 7.1 months.
• Working hours against the GH200: $403.04 − $99 = $304.04 a month, which gives ≈ 21.4 months.
• Working hours against the Community A100: $244.64 − $99 = $145.64 a month, which gives ≈ 44.6 months. That's longer than our 36-month horizon, so renting wins in this case.

That last line is the important one. If your use really is part-time, and the cheapest 80 GB instance is good enough, renting a community-cloud GPU for working hours costs less over three years than owning a Spark.

What the arithmetic leaves out

What the cloud gives you that ownership doesn't

• Elasticity. You can rent eight GPUs for one afternoon. A Spark is one Spark.
• No upfront capital. $6,500 today is a real cost for a small team, and that money could earn something elsewhere.
• Speed on models that fit. As noted above, an H100 is much faster than a Spark on a model that fits in 80 GB.
• No commitment to one architecture. When a new GPU generation ships, a renter switches by changing an instance type.

What ownership gives you that the arithmetic doesn't show

• The memory. 128 GB on one machine means models that need two rented GPUs, or a quantization compromise, fit on one Spark. Two to four linked units reach the ~405B to ~700B class at FP4. See what you can run on 1, 2, 3 or 4 units.
• Availability. A machine you own is never "out of capacity in this region".
• A flat bill. No meter runs while a model sits loaded, so you don't have to remember to shut instances down, and an agent that loops overnight doesn't produce a surprise invoice.
• Data stays on your machine. It's exclusive bare metal that is never shared, with full root access.
• A residual asset. After 36 months you still own the hardware.

Resale value

Compute hardware loses value as newer generations arrive, but it doesn't fall to zero. A machine that still runs current software usually has buyers. How much it's worth depends on the market at the time, the machine's age and its condition, and nobody can promise a figure three years out.

Every dollar of resale value lowers the true cost of owning. If you sell the unit for an amount R at month 36, your net cost is $10,064 − R. Each $1,000 you recover lowers the effective monthly cost by $1,000 ÷ 36 ≈ $27.78.

A NiceRack buyback programme is coming later. When it opens, offers will be based on what the same machine in the same condition is selling for at the time, its age, and any repairs or missing parts. There will be no guaranteed buyback price. More on that in buying hardware you can sell back.

The hybrid that often wins

Many teams end up somewhere in between. They own enough capacity for the steady base load and rent for bursts. A Spark serves your everyday model behind a private endpoint, and a hosted provider takes overflow or the occasional frontier model you can't run locally.

The routing is the fiddly part. A local OpenAI-compatible gateway such as LocalRouter can send requests to your own hardware first and fall back to a hosted provider when the local box is busy or the model isn't there. Your application keeps calling one endpoint either way.

A quick decision guide

Renting is likely better if:

• you use a GPU for fewer than about 85 to 200 hours a month, depending on which instance you'd otherwise rent (see the break-even list above)
• your workloads are bursty, short-lived or experimental
• you need peak throughput on models that fit in 80 GB
• you can't or don't want to spend capital up front

Owning is likely better if:

• a model is loaded most of the day, or all of it
• you need more than 80 GB of memory per machine for the models you care about
• you want a predictable monthly bill and data on hardware nobody else uses
• you expect to use it for two years or more

Next steps

• Compare configurations on the hardware page, or go straight to the single DGX Spark.
• See every price, including management, on pricing.
• Read how managed hosting works, or check the FAQ.
• Tell us what you want to run on the contact page. We'll confirm availability and send an invoice, and nothing is charged online.