True pricing, on demand: the floor that speculative datacenters skip
AI needs silicon, power, and people who keep the boxes healthy. Those costs do not disappear when a rack sits half empty, and they do not get cheaper when someone announces a campus that is still waiting on a grid interconnection. A sustainable price for inference has to start from the full economic floor of the hardware that actually runs, not from a promise of future capacity.
This post pairs with our depreciation framing: idle hours are expensive because depreciation is charged per calendar hour. Here we take the next step—true pricing and an on-demand model that fills hours on hardware that already exists, instead of treating speculative datacenter buildout as the default path. Figures labeled estimate are our own arithmetic with stated assumptions. Cited numbers belong to their sources; they change, so check them on the day you decide.
Two ways to add capacity
There are two broad ways the world adds AI compute:
- Build ahead of demand. Raise capital, buy GPUs, secure land and power, pour concrete, and hope utilization catches up. That is the hyperscale and speculative-campus path.
- Fill hours that already exist. Put work on boxes that are already bought for private use, labs, or edge deployments, and that sit idle part of the day. That is the on-demand marketplace path.
Both can be honest businesses. Only the second one prices from hardware that is already depreciating somewhere.
What “full economic cost” means
A floor that hosts and renters can live with for years has to cover more than the sticker price of a card.
Rough components of a true hourly floor (all estimates; your numbers will differ):
| Cost | What it covers | Behavior |
|---|---|---|
| Depreciation | Capex ÷ useful life | Fixed per calendar hour, used or not |
| Electricity | Power at the wall, including PSU and cooling losses | Mostly load-dependent; still nonzero at idle |
| Network and space | Bandwidth, uplink, rack/desk real estate | Mostly fixed for a given setup |
| Host attention | Updates, monitoring, failure recovery | Soft cost; still real |
| Platform and payments | Marketplace take, card/crypto rails, tax where owed | Variable with revenue; Hypermesh take is 12% on hourly / Your Model and 18% on per-token / Full Model; host bank payouts are weekly with a $5 minimum |
In our depreciation post, four DGX Sparks at the video’s $27,800, over three years with the video’s electricity figure, landed near $1.19 per hour of availability before network, attention, or platform. That is not a Hypermesh catalog price. It is a floor sketch: anything you charge below that, sustained, is a gift from the host’s capital or from someone else’s subsidy.
True pricing means the posted rate for a lease or a token sits at or above that floor after the host’s own utilization assumptions. Underpricing to win a race against speculative capacity just transfers the loss to whoever owns the box—or to the next funding round.
What speculative buildout actually pays for
Announced AI campuses are not the same thing as powered racks. Oxford’s Smith School of Enterprise and the Environment finds that only about 35–50% of planned AI data-centre capacity is translating into delivery on schedule, with permitting, grid queues, equipment, and labour as the binding constraints—not capital (Smith School summary; full paper). Markets often price the announcement curve; the delivery curve is slower and lumpier.
Even where silicon is installed, utilization is often soft. A VentureBeat report on production AI users found that 81% of respondents running GPUs in production reported average utilization of 50% or less (VentureBeat). Underused GPUs still drew the power, cooling, and capital that went into them.
Energy demand is rising either way. The IEA’s update on energy and AI projects data-centre electricity roughly doubling from 2025 to 2030, with AI-focused facilities growing faster still, while bottlenecks make the most aggressive near-term scenarios less likely (IEA executive summary). Building more empty or half-empty capacity does not make that curve kinder; it front-loads the power and materials bill.
None of that is an argument against large facilities for training or for workloads that need them. It is an argument against treating “announce a campus” as the only way to serve intermittent, local, or privacy-sensitive inference—and against prices that only work if someone else eats the idle.
How on-demand helps
An on-demand marketplace does not invent silicon. It moves work onto hours that would otherwise sit empty:
- A host keeps private hours private and lists the rest.
- A renter buys a block of time or pays per token on a model that is already loaded.
- Price discovery can track demand (busy evenings cost more; quiet nights cost less) without pretending the floor is zero.
That is the sustainability claim we are willing to make, and no larger one: better utilization of hardware that is already manufactured and installed is cheaper in capital, materials, and often energy per useful token than building new capacity for the same intermittent load. It does not erase the need for new datacenters. It reduces how much spare capacity the industry has to pour for spiky, edge, and private work.
It also lines up with the economics in the sibling posts. Tokens per second per user shows why batching raises aggregate throughput. Clustering and sharing splits time-based leases from shared-model, per-token access. Both only pay off if someone fills the idle slice.
Demand-based pricing without lying about the floor
A healthy on-demand market has two numbers in view at once:
- Floor: full economic cost at the host’s expected utilization. Below this, the host is subsidizing the renter.
- Clearing price: what renters will pay for that hour or that token given alternatives (APIs, other hosts, waiting).
When clearing price is above the floor, both sides can stay. When it is below for long stretches, hosts leave or cut corners—or speculative operators raise more capital and push the race further. Demand-based and day-ahead booking are the tools that let quiet hours clear near the floor and busy hours clear above it, instead of one flat rate that is wrong twice a day.
Hypermesh lists demand-based pricing and day-ahead booking as coming soon on our pricing page. Time-based leases and per-token access on enrolled, checked hardware are what we offer today, with posted prices per hardware class and model. The platform take is 12% on hourly / Your Model leases and 18% on per-token / Full Model access; hosts receive bank payouts weekly with a $5 minimum.
What this means if you rent or host
If you rent, ask what is included in the rate: power, network, and whether the host can keep the box healthy for the lease you buy. A price that looks far below any honest floor is a signal, not a bargain.
If you host, start from your own floor—capex life, power at your tariff, uplink, and your time—then decide how many hours a day you can truly fill. Share the idle slice; do not pretend depreciation stops when you walk away from the desk.
If you are comparing a marketplace hour with a new rack in a speculative campus, count the same things on both sides: silicon, power, cooling, utilization, and who eats the empty hours. The cheaper story on paper is often the one that left something off the invoice.
Try it
Hypermesh connects hosts who have enrolled, checked hardware with renters who need time or tokens on it. Renters pay in USDC (or buy USDC through Stripe’s onramp), and hosts are paid through Stripe Connect. For the depreciation math that sits under this post, see Your GPU is depreciating whether you use it or not. For how sharing actually moves tokens, see Tokens per second per user. Or go straight to hyperme.sh.
Sources: Oxford Smith School, “35–50% of planned AI data centres are behind schedule” and Impact on Power and Commodity Markets; VentureBeat on GPU utilization (report); IEA, “Key Questions on Energy and AI”. Hourly floor figures marked as estimates follow the assumptions in our depreciation article. This is not tax, accounting, or investment advice.