A field guide to DePIN unit economics: take rate, price per GPU-hour, utilization, and market cap vs revenue, with attributed, sourced DePIN economics.
A network can advertise hundreds of thousands of machines and still run a take rate that loses money on half its jobs. The market cap and the revenue tell two different stories.
Disclosure: we're building a venture in this category. Every figure below is attributed to a named, dated source, and we flag our own stake where it bears on the argument.
The compute market trains your attention on the wrong number. The big one gets the headline: the machine count, the token's market cap, the countries served. None of those tell you whether a single job clears a margin. For that you have to do something unglamorous. You have to read the invoices.
Let's walk the math that decides which story is real.
A compute marketplace makes money on the spread between what a buyer pays and what the supplier earns. That spread is the take rate, and the model you choose for it shapes everything downstream.
Two live examples, two philosophies. A major fiat marketplace runs an effective take of about 20%: its live GPU prices sit roughly 25% above what hosts earn, and it removed explicit hosting fees back in mid-2024 (a major GPU marketplace, May 2026). A large token-based network advertises a headline platform fee of just 2%, waived entirely if the renter pays directly in the network's token (a leading decentralized GPU network, 2025 year-in-review, Jan 2026).
Two percent looks unbeatable next to twenty. But the headline hides the real cost. On the token network, providers are paid in a cryptocurrency and bear its price risk, plus staking requirements and slashing penalties, so the economic burden carried by supply runs well above the visible 2% (per our competitive analysis of the token-based tier). Same word, "fee," very different bill. The clean planning band for a durable marketplace take sits somewhere around 15 to 25%, and the fiat end of that range is the more legible model.
The revenue per job depends on what the hardware rents for, and the spread is wide.
An RTX 4090 rents across these networks for roughly $0.18 to $0.53 per GPU-hour (the largest networks around $0.30, a major marketplace $0.37 to $0.53, aggregator medians near $0.40) (thundercompute, Jun 2026; a major GPU marketplace, May 2026; computeprices.com, 2025 to 2026). An A100 80GB runs about $0.53 to $1.45 (a major GPU marketplace low end to a leading decentralized GPU network high end). An H100 sits around $1.38 to $2.19 at the decentralized end, against $3.36 to $12.29 at the hyperscalers (thundercompute, Jun 2026; a leading decentralized GPU network, Apr 2026). That last comparison is the whole pitch in one line: decentralized and neocloud supply undercuts the big clouds by roughly 60 to 80%.
The discount is real. The question the discount can't answer by itself: is the machine busy enough to matter?
Utilization. What fraction of available GPU time is actually rented and paid for. It is the hinge of the entire model, and it is the figure most likely to be dressed up.
The spread is staggering. Idle-aggregation networks can show a very low real-paying share of registered supply, because most of what registered is unusable or simply off. One large network's own year-in-review put roughly 2,752 GPUs as verified against a registered base above 327,000, under 1% genuinely verified (a leading decentralized GPU network, 2025 year-in-review, Jan 2026). Curated, enterprise-supplied networks tell a different story: one reports roughly 38% utilization (House of Chimera, Q1 report, Apr 2025), another claims GPU utilization above 80% heading into 2026, though against a small base (BlockEden, Mar 12 2026). The gap from under-1%-verified to 80%-busy is not a measurement quirk. It tracks the difference between supply a buyer can trust and supply that just exists.
A machine count flatters. A take rate constrains. Utilization tells the truth. Read all three, in that order, and the real business stops hiding.
This is where the gap between story and income statement shows. The broader DePIN compute sector carries roughly $10B in circulating market cap against only about $72M in on-chain revenue for the year (BlockEden, "DePIN's Revenue Reckoning," Mar 12 2026). That's a revenue multiple that does not survive contact with a spreadsheet, and it's why consolidation is underway.
But underneath the froth, real businesses exist. The category's revenue leader booked $127.8M in revenue for the year, with 150-plus enterprise clients (BlockEden, Mar 12 2026). The lesson is narrower than "the model is broken." The headline number and the income statement live in different neighborhoods, and only one of them pays salaries.
We unpack the supply-quality root cause in the flagship, The sub-1% problem in decentralized compute, and the build-versus-rent side of the cost question in Buy, rent, or borrow the GPU.
So when a compute network asks for your money or your benefit of the doubt, make it show the work. Define the per-job margin before launch, not after. Choose supply you can trust, so utilization tracks the busy end of that range rather than the idle one. The invoice is the only slide that matters.
Nothing here is an offer to sell a security or investment advice.
Field notes on venture building, AI, and capital. No spam, unsubscribe anytime.
By subscribing you agree to receive AOS Insights e-mails. We use your address only for this newsletter - see our Privacy Policy.
