Insights ·

Affordable compute is becoming a condition for using AI. How to get it without losing control

GPU prices diverge, hyperscalers spend over $700 billion and DePIN networks offer idle capacity that is hard to trust. How the Smart Compute Router in TrustTwin OS makes distributed capacity usable for valuable data.

WEREALeggi in italianoLire en françaisAuf Deutsch lesenLeer en español

For most of the last decade, the question for a company adopting new software was which tool to buy. With AI it has become a different question: where the computation will run, and at what price. A model that reads contracts, a pipeline that turns thousands of photographs into a 3D twin, an agent that works through a backlog overnight: all of them consume GPU time, memory and energy, and that consumption does not stop after the pilot.

Access to compute at a sustainable cost is turning into a precondition for using AI at all. The organisations that have it can experiment, iterate and put workloads into production. The ones that do not stay stuck at the demo.

What the numbers say in 2026

The market is sending mixed signals, and both sides matter.

  • Older GPUs are getting cheaper, the newest are not. The AIMultiple Cloud GPU Price Index, which tracks 75 providers, puts the September 2026 on-demand median at $3.25 per hour for an H100 and $4.40 for an H200, while a B200 sits at $6.52. The median for the newest generation of chips has more than doubled since October 2024, from $2.12 to $4.72 per hour.
  • The same chip can cost six times more depending on who sells it. In the same index, an H200 hour ranges from about $2 to almost $14. Where you buy matters as much as what you buy.
  • The giants are spending at a scale no one else can match. Analyst tallies of company guidance put 2026 capital expenditure by Amazon, Alphabet, Microsoft and Meta above $700 billion combined. Most of the newest capacity is committed to them and their largest customers.
  • Power is the next constraint. The International Energy Agency expects electricity use by data centres to more than double by 2030. New capacity now waits on grid connections, not only on chips.
  • The price of a token keeps falling, the bill does not. The Ramp AI Index reports that the average price US businesses pay per million tokens fell 41% between March and September 2026. Usage grows faster than prices fall, especially once agents start running multi-step tasks on their own.

The effect shows up in adoption. According to Eurostat, 55% of large EU enterprises used AI in 2025, against 17% of small ones. The gap is not about interest. It is about who can afford the infrastructure, the people to run it and the risk of moving data.

Three options, none of them a good fit

An organisation that holds large, valuable datasets (scans, archives, drawings, models, research data) usually faces three choices.

  • Rent from a hyperscaler. Capacity is there, but the pricing is built for large accounts, and the data has to move to someone else's infrastructure before any work can start.
  • Buy the hardware. GPUs sit idle most of the month between peaks, and someone in house has to run, update and secure them.
  • Send the data out to a service provider. Control is gone. You no longer decide where the files go, and you pay again for every job.

Public programmes help at the margin. The EuroHPC AI Factories offer free access modes for SMEs and startups, and the call for AI Gigafactories closes on 12 November 2026. They are valuable for research and model training, but they are allocated through applications and are not designed to run a company's daily production workloads.

DePIN: plenty of capacity, not enough trust

Decentralised physical infrastructure networks (DePIN) take a different route. They aggregate idle GPUs from data centres, companies and individuals and rent them out, often well below hyperscaler prices. The idea is sound: a lot of computing power in the world sits unused for most of the day.

The problems are just as concrete, and they are documented by independent research:

  • Declared capacity is not real capacity. A node can advertise a GPU it does not have, or one that is shared, throttled or incompatible with the job.
  • Availability varies. Nodes come and go. A long rendering or training job needs checkpoints and somewhere to continue when a node disappears.
  • Location is unknown. For many datasets, running "somewhere on the network" is simply not allowed.
  • Outputs are hard to verify. Proving that a computation was carried out correctly on an untrusted machine is still an open research problem.

Cheap capacity that you cannot trust is not usable for data that matters. The question is not whether distributed capacity is cheaper. It is whether it can be governed.

Have a heavy workload and nowhere sensible to run it? Tell us about the data and the constraints. Talk to the TrustTwin OS team.

How the Smart Compute Router and DePIN nodes fit together in TrustTwin OS

TrustTwin OS treats compute as one execution continuum. Work can run on our own managed nodes, on nodes installed on the customer's hardware, or on idle capacity that customers and partners make available to the network, including DePIN capacity. What changes is not where the GPUs are. It is that every job crosses the same governance before, during and after execution.

The Smart Compute Router takes each workload through six stages:

  1. Classify. What kind of work it is and how sensitive the data is.
  2. Constrain. Residency and security policy applied as hard limits. An EU-only profile excludes non-European providers entirely. If a residency rule cannot be guaranteed, the workload does not run.
  3. Preflight. GPU, VRAM, CUDA, network and a minimal execution verified on the candidate node before the job is committed. Declared capacity is checked, not trusted.
  4. Route. Cost, time and risk optimised across local, cloud, edge and distributed capacity, among the nodes that passed the first three stages.
  5. Recover. Retries, checkpoints and fallback to another node when a provider fails.
  6. Validate. Output completeness and quality checked, with a record of where each step ran and which decisions were made.

This is what makes DePIN capacity usable for serious work. Preflight answers the problem of capacity that exists only on paper. Recover answers nodes that drop out. Constrain answers location. Validate answers the question of whether the result can be trusted. The router does not make distributed compute trustworthy by assumption; it only uses it where the checks pass.

Digital twin of a statue of Saint George, shown half as rendered surface and half as wireframe mesh
Heavy 3D processing in the AerariumChain pipeline: scans are processed unattended on private and partner nodes, and twins come back with a record of where they ran.

Why it lowers cost over time

The economics follow from the architecture. Heavy processing runs unattended on our own and partner nodes, at a fraction of hyperscaler cost. Customers can make their own computing capacity available to the network when they do not need it, so a workstation that would sit idle at night contributes capacity instead of depreciating. As more nodes join under the same policies, unit cost falls for everyone on the network.

We tested this where data are hardest. In AerariumChain, scans of irreplaceable objects are processed on private and partner nodes under strict rules about where the data may go. SweetHive brings the same OS to small and mid-size businesses, with local nodes that keep files and models on the company's machines.

Five questions to ask any compute provider

  1. Can you guarantee where my data is processed, and what happens if you cannot?
  2. Do you verify a machine's real capacity before my job starts?
  3. What happens to a long job when a node fails halfway?
  4. How do you check that the output is complete and correct?
  5. Can I see, afterwards, where each step ran?

If the answers are vague, the low price is not the whole price.

Sources

Talk to us about your workloads All insights