Renting GPUs for Crypto: I Did the Math on Self-Hosting GLM-5.2

The same GPU node, running the same model, will quote you a token price that varies by a factor of 350. What decides it is not the rental rate.

I went through both sides of the crypto GPU rental market — renting your card out, and renting someone else's to host an open model — and carried the arithmetic all the way to dollars per million tokens. The workload is GLM-5.2: open weights, 753B parameters, released June 2026. Exactly the case decentralized GPU networks exist for.

All prices are as of September 2026.

Side one: renting your card out

A verified RTX 4090 on Vast.ai grosses about $6/day. That looks great — $180/month from a card that would otherwise gather dust.

Then the deductions start. The card draws 450W, which is 10.8 kWh/day around the clock:

Electricity rate	Per day	Gross minus power
$0.06/kWh	$0.65	$5.35/day
$0.08/kWh (Texas, Midwest)	$0.86	$5.14/day
$0.15/kWh	$1.62	$4.38/day
$0.30/kWh (Germany, UK)	$3.24	$2.76/day

Still respectable — but only because that table quietly assumes 100% utilization. It doesn't work that way.

Published reviews converge on a different figure: about $52/month net, after electricity, depreciation and platform fees, at typical load. Work that backwards and real utilization is around 40%, not 100. Two thirds of the time the card sits waiting for a tenant while continuing to age in your apartment.

The demand side is the actual risk here. In Q1 2026 Akash averaged 334 GPUs on the network with 84 actually working, and available capacity contracted by more than 57% quarter over quarter.

Payback: an RTX 4090 costs roughly $1,800. At $52/month net it pays for itself in 35 months — longer than the card survives in sellable condition under round-the-clock load.

So renting your GPU out is not passive income. It is converting depreciation into cash. The card loses value either way; the only question is whether it does so silently or hands you fifty dollars a month on the way down. Buying hardware specifically to rent it out is close to guaranteed loss. Renting out hardware you already own, on a 6–8 cent tariff, is reasonable.

Side two: renting someone else's to host GLM-5.2

GLM-5.2 is a 753B-parameter Mixture-of-Experts model, ~40B active per token, 78 layers, native 1M-token context. Open weights — that's the whole point of the exercise.

"Open weights" does not mean "fits anywhere":

Precision	VRAM needed	What that means
FP16/BF16	1,488–1,642 GB	two 8-GPU nodes minimum
FP8	~744 GB	8×H200 (1,128 GB) — production config
INT4 / Q4	372–411 GB	8×H100 80GB (640 GB) with KV headroom
2-bit dynamic	239–241 GB	quality tradeoff, experiments only

Renting a whole node:

Configuration	Per hour	Per month 24/7
8×H200, conventional cloud	$26.48	$19,330
8×H100 PCIe, conventional cloud	$23.12	$16,878
8×H100 on Akash ($2.09/card)	$16.72	$12,206
8×H100 at spot ($1.49/card)	$11.92	$8,702

The savings are real, and worth stating plainly: decentralized networks genuinely run 55–70% below the hyperscalers. The same H100 that Azure charges $6.98/hr for on-demand goes for $1.47–1.49 at io.net and Vast.ai. Akash lists H100 PCIe at $2.09, SXM5 at $2.69.

Case closed, seemingly. Except we haven't reached the token yet.

Token price: where it all flips

The hourly rate tells you nothing about the cost of work. Only one thing matters: how many tokens the node emits per hour.

For speed I use vLLM's benchmarks: a DeepSeek-class MoE (671B total, 37B active — nearly GLM-5.2's twin in shape) sustains 2.2k tokens/second per H200 in a production-like deployment. That's the best published result, after kernel work and Dual Batch Overlap; a year earlier it was around 1.5k.

Here's the caveat that headlines drop: that is aggregate throughput under high concurrency, not one user's response speed. vLLM publishes no per-user metric in that post.

Regime	Tokens/s	8×H200	8×H100 Akash
Node fully saturated	17,600	$0.42 / M	$0.26 / M
Single user, batch 1–2	~50	$147 / M	$93 / M

For comparison, Z.ai's official API for the same model: $1.40 per million input tokens, $0.26 cached input, $4.40 per million output.

So a saturated node delivers tokens 10× cheaper than the API — that's where the triumphant headlines come from, and they aren't lying. The same node under a single developer delivers tokens 15–33× more expensive than the API. The gap between two rows of one table is 350×.

Break-even: how many agents it takes

Working back from $4.40 per million output tokens:

Configuration	Must sustain	Roughly
8×H200 ($26.48/hr)	1,672 tok/s	~33 continuous streams
8×H100 Akash ($16.72/hr)	1,056 tok/s	~21 continuous streams
8×H100 spot ($11.92/hr)	753 tok/s	~15 continuous streams

And that is around the clock, no weekends. A node you shut down at night loses two thirds of its output.

Apply that to a solo developer. Even with heavy agentic work you land near 5M output tokens and a few hundred million input tokens a month. On the API with caching: about $82/month. Your own node on Akash for the same month: $12,206.

That's roughly 150×. No DePIN discount closes a gap like that, because the gap isn't about discounts.

The circle closes

• Rent out an RTX 4090: $52/month net at realistic load.
• Rent a 8×H100 node for GLM-5.2: $12,206/month at Akash pricing.
• To fund one rented node you would need to rent out 235 RTX 4090s — about $420,000 of hardware, in your apartment.

The market looks symmetric: supply on one side, demand on the other, a token in between. In practice the two sides are three orders of magnitude apart.

What crypto actually changes here

What it genuinely gives you. Access to hardware with no contract, no legal entity, no KYC. Per-minute billing with no quarterly prepayment. A supplier stake instead of a reputation system — those 200 IO per chip that burn on misbehavior. Geographic spread that three hyperscalers don't offer.

What it doesn't. No SLA. No guarantee the 8×H200 you need is free at the moment you need it — remember Akash's 84 working GPUs and −57% capacity in a quarter. And critically, no fast interconnect. That 2.2k tokens/s per card was measured on an InfiniBand cluster. Sharding an MoE model across heterogeneous machines over ordinary internet is not remotely the same regime — and multi-node serving is the one scenario where self-hosting beats the API at all.

The uncomfortable part: the DePIN discount does not come from the token. It comes from the hardware being someone else's, already depreciated, sitting in a cheap electrical socket. Vast.ai delivers exactly that with no blockchain at all — and in these tables its $1.49 spot H100 came in below Akash's $2.09. The token is a settlement and collateral mechanism, not a source of savings.

What to actually do

You're solo or a small team. Use the API. $82/month versus $12,000 is not a case that rewards optimization. Self-hosting is justified by exactly one argument here, and it isn't economic: data that must not leave your perimeter.

You run 20–35 agents around the clock. Now the math is worth doing seriously, and DePIN pricing genuinely moves break-even down — from 33 streams to 21. Verify two things before committing: actual availability of the configuration at that moment, and the interconnect between nodes. Without InfiniBand or equivalent, the economics you came for fall apart.

You have idle hardware and a cheap socket. Rent it out. Just model 40% utilization rather than 100%, and treat the result as recovered depreciation, not income.

If you need privacy specifically. A Mac Studio M3 Ultra with 512 GB of unified memory ($10–15k) runs GLM-5.2 at 6.5–9.5 tokens/second. That's indecently slow, but it is the only way to hold a model this size in one box, with no rentals and no one else's data center. Workable for background processing under NDA; not for interactive work.

The takeaway

Decentralized GPU networks are neither a bubble nor a revolution. They're a working, fairly narrow tool with an honest 55–70% discount to hyperscalers and equally honest problems: unstable capacity, no SLA, slow interconnect.

But the main conclusion of this arithmetic isn't about crypto at all. It's that inference cost is determined by utilization, not by hardware price. Until that lands, any "rent vs API" comparison returns a random answer — you can demonstrate a 10× saving and a 30× overspend from the same price list with equal honesty.

A cheap GPU doesn't make a token cheap. What makes a token cheap is a queue of work for that GPU — and you still have to get the queue from somewhere.

Originally published at klukyanov.ru.