← All articles

Your GPU is depreciating whether you use it or not

The case for owning large-model hardware is real: private data stays in the building, there are no API terms of service, and you control the stack. The financial case is harder, and it depends almost entirely on one number: how many hours a day the box does useful work.

This post works through the payback math from Kai's Every Way to Get 512GB of VRAM for Local AI (at Full Context) (October 17, 2026), then adds a simple depreciation view. Anything labeled estimate is our own arithmetic with stated assumptions. This is not tax or accounting advice.

The video's payback math

The video uses four NVIDIA DGX Sparks ($6,950 each, $27,800 total, before the switch) running DeepSeek V4.1 Flash, and compares them with DeepSeek's API at $0.15 per million input tokens and $0.60 per million output tokens (off-peak).

  • One user, flat out all year: 87.7 tok/s × 31.5 million seconds ≈ 2.77 billion output tokens. At $0.60/M that is about $1,659 of API spend, so the hardware takes about 17 years to pay back.
  • Sixteen busy sessions, all year: 342.7 tok/s ≈ 10.8 billion tokens ≈ $6,485 of API spend, so about 4.3 years to pay back.
  • Electricity: about $1,122 per year at average rates, per the video.

The video's conclusion: "the hardware earns its keep when it's private and busy at the same time."

Including electricity (estimate)

The headline payback figures appear to compare hardware cost against gross API value. If you subtract the video's own electricity figure:

  • One user: $1,659 − $1,122 ≈ $537/yr net, so about 52 years to pay back.
  • Sixteen sessions: $6,485 − $1,122 ≈ $5,363/yr net, so about 5.2 years.

These figures also count output tokens only and leave out the switch. Input tokens and cache reads would add API value; networking would add cost. The direction does not change: a single user never pays it back, and a busy shared box gets into the range of a hardware lifetime.

A simple depreciation framing (estimate)

Payback compares against an API. Depreciation asks a different question: what does an hour of this box cost me, used or not?

Assumptions (all estimates):

  • 3-year useful life, straight-line, zero resale value. Real resale value may be higher or lower: the video shows used DDR5 RDIMMs selling for 5× what the whole $6,000 EPYC build cost in January 2025, and DGX Spark prices rising 74% since launch. Hardware markets move both ways.
  • Electricity at the video's $1,122/yr.
  • 8,760 hours per year.

Four DGX Sparks:

  • Depreciation: $27,800 ÷ 3 ≈ $9,267/yr ≈ $1.06/hour
  • Power: $1,122 ÷ 8,760 ≈ $0.13/hour at full load
  • Total ≈ $1.19 per hour of availability

That hourly cost is mostly fixed. Now spread it over tokens:

UtilizationOutput tokens/hourCost per 1M tokens (estimate)
1 user, 24/7≈ 316K≈ $3.76
16 users, 24/7≈ 1.23M≈ $0.96
16 users, 25% of the day≈ 0.31M averaged≈ $3.55 (depreciation does not shrink when idle; power counted at 25%)

Two honest takeaways:

  1. Utilization dominates. Going from one user to sixteen cuts the cost per token by about 4×. Going from 24/7 to a quarter of the day gives most of it back.
  2. On pure token cost, a cheap hosted API can still win. Even fully used, our estimate of about $0.96/M is above DeepSeek's off-peak $0.60/M output price. Owning makes sense for privacy, control, latency, models or fine-tunes the API doesn't offer, or sustained high utilization. It is not a guaranteed saving.

The same lens on a pricier build (estimate)

Four RTX PRO 6000 Max-Q cards are $64,000 at the video's $16,000 each, before the workstation. Over three years that is ≈ $2.44/hour in card depreciation alone. RunPod rents the same four cards at $9.96/hour. The video notes that $64,000 buys about 6,400 rental hours, roughly 268 days around the clock. If you use the cards less than about a quarter of the time over three years, renting is cheaper. If you use them more, owning can be. Sharing is how an owner gets from the first case to the second.

Idle hours are the expensive ones

The point is simple: depreciation is charged per calendar hour, not per used hour. A box that runs your agents eight hours a day and sits idle for sixteen is paying full depreciation on the idle two-thirds.

Ways to fill those hours:

  • Lease out a share of the day. Your nights, weekends, or a fixed block of hours go to someone else as time-based leases.
  • Serve a model to many renters. Keep a well-sized model loaded and let other users' requests batch alongside yours. The 4.3-year figure is a 16-session figure for a reason.
  • Do both. Keep your private hours private and share the rest.

There are real costs: wear, power, the attention needed to keep a shared box healthy, and making sure your own data stays separate from renters' work. Those belong in your own numbers.

Where Hypermesh fits

Hypermesh connects hosts who have enrolled, checked hardware with renters who need time or tokens on it. Today that means posted prices per hardware class and model, time-based leases and per-token access. Renters pay in USDC (or buy USDC through Stripe's onramp), and hosts are paid through Stripe Connect. Demand-based pricing and day-ahead booking are listed on our pricing page as coming soon.

If your hardware has idle hours, or your work would fit in someone else's, see how it works at hyperme.sh.


Source: Kai, "Every Way to Get 512GB of VRAM for Local AI (at Full Context)", October 17, 2026, which cites DeepSeek API pricing and RunPod pricing. Depreciation, net-of-power payback and cost-per-token figures are our estimates from the stated assumptions. Prices change; check them on the day you buy.