EN
 |  Sign in · Register

How CMB works — explained in 5 minutes

What is CMB?

CMB — ComputeMoreBusiness is a marketplace for time-limited GPU compute. Providers (data centres, companies with spare capacity) auction their GPU hours, buyers (AI teams, businesses) bid for them — or publish their demand as a tender and let providers compete on price. Auctions and requests are priced per GPU-week or GPU-month, slots per GPU-hour; billing is always by delivered hours (1 week = 168 h, 1 month = 730 h): price per hour × number of GPUs × delivered hours.

Role: buyer

You need compute. You can bid on lots and create your own tenders.

Role: provider

You have spare GPU capacity. You can list lots (auction them) and submit offers on tenders.

EU-sovereign — what does it mean?

On CMB all compute offered is EU-sovereign. This is not an optional label but a requirement for every provider (Terms, clause 5f):

Providers can also state ISO certifications (e.g. ISO/IEC 27001 information security, ISO/IEC 27017/27018 cloud, ISO 50001 energy). Buyers can require certifications in requests and filter by them in the marketplace. Certificates must be proven within five business days on request; false statements are treated as improper provisioning (penalty).

1 · Create an account

1

Click Register in the top right corner.

2

Choose your role: buy compute or provide compute — the role determines which features you see.

3

Enter your company details (incl. VAT ID and registered address — required for legally valid contracts), email and password. You are signed in immediately.

2 · The open auction: bidding on lots

A lot is a concrete capacity package, e.g. "8× H100, Saturday 00:00 to Monday 00:00".

1

Open a lot under Compute Offers ("View"). You see the details, time remaining and the current highest bid.

2

Enter your bid — it must be at least highest bid + minimum increment (the valid minimum is pre-filled). A confirmation popup summarises price, total value and the binding nature of your bid.

3

When the clock runs out, the highest bid wins. Anti-sniping: bids in the final 2 minutes automatically extend the auction by 3 minutes.

⚡

Buy now: if the provider has set a buy-now price, you can take the lot immediately at that price without bidding — the auction ends at once with the award to you, then buyer protection payment, access and start check follow as usual. The option disappears as soon as a bid reaches the buy-now price.

4

Under My dashboard you can always see your position: leading, outbid, or awarded.

For providers: create lots via Create offer. Note: the auction must end at least 5 days before the delivery window starts — leaving the 72-hour payment deadline plus buffer for transfer and provisioning.

3 · The tender: providers underbid each other

The reverse route — ideal for larger or longer contracts: you define the requirements, providers compete for the job.

1

As a buyer: Create request — define requirements (GPU type, quantity, region, EU sovereignty), delivery window, maximum price and offer deadline.

2

Providers submit offers. Each new offer must be at least one decrement below the current best. Providers only ever see the best price, never their competitors.

3

After the deadline the selection phase begins: you choose from the best 3 offers — not necessarily the cheapest, but the most economical.

4

Your award automatically creates the contract including the delivery process (see next section).

4 · Delivery & proof: from auction to compute

Every award creates a delivery with five stages — visible to provider and winner under "My dashboard" → "Delivery":

1

Payment (buyer protection): The buyer pays the full order value up front into the buyer-protection account at our payment provider Stripe (SEPA transfer, SEPA direct debit or card) — within 72 hours of the award, at the latest before delivery starts. No payment, no activation. If payment is not received, the order is cancelled automatically: a 10 % cancellation fee is due to the provider. The money stays with CMB until release and is paid out to the provider only after verified delivery (buyer protection).

2

Access provisioning: The provider deposits the access credentials on the platform — the timestamp is recorded. All common access types can be offered:

Possible access types: SSH (VM or bare metal — full access, the B2B standard) · Containers (Docker/Kubernetes — buyer ships their image, isolated & secure) · Jupyter/web interface (URL with token, ideal for data science) · Inference API (API endpoint with key only, no machine access) · Remote desktop (RDP/VNC for workstation scenarios) · VPN/private interconnect (private tunnel, enterprise/EU-sovereign). Provider and buyer agree on the access type in the lot or request.
🔒 How access credentials are protected: They are stored encrypted (AES-256), appear in no e-mail and are visible only to buyer and provider after signing in — only after clicking “Show”, and every retrieval is logged. The provider can update the access during delivery (new password/token); 48 hours after the end of delivery or on cancellation the data is deleted automatically. Buyers can deposit their public SSH key in advance — then no password needs to be transmitted at all. Passwords, tokens or private keys in the chat are blocked.
3

Verification (real hardware): The provider starts the telemetry agent on the GPU server. Its first report contains, for every GPU sold, the model, memory and a unique GPU ID as well as – if the GPU is idle – a short performance measurement. The platform compares this with the lot: if count, model, minimum memory and minimum performance match, verification passes and the reported GPUs are stored as the order's baseline. If not, there is a grace period of 2 hours — then termination with a 10 % penalty (clause 5e).

4

Hourly telemetry: The agent sends a keep-alive (all GPUs with ID, utilisation, temperature, memory errors and uptime) every 10 minutes and a spec check every hour. Each hour of the delivery window is rated: delivered if there is no gap over 20 minutes, all GPUs sold are reported and no new uncorrectable memory errors occur — otherwise not delivered. If the provider has entered a check address, the platform additionally checks, at most every 15 minutes, whether the access is reachable. The hour list is visible live to both sides on the delivery page.

Important for providers — the telemetry agent: It runs on the GPU server for the entire delivery window. The download button and the ready-made start command including the order token are on the delivery page.
  • Linux (recommended, mandatory for multiple GPUs): cmb-agent.py — needs only Python 3, no installation. Reports all GPUs (NVIDIA via nvidia-smi, AMD via rocm-smi). With --systemd-installieren it keeps running even after a reboot; with --gpus 0,1,2,3 it reports only the GPUs sold from a larger server. If PyTorch is available, it measures performance – but only while the GPU is idle, so the buyer's jobs are not disturbed.
  • Windows: cmb-agent.ps1 (PowerShell, one GPU, performance measurement via OpenCL).
  • Reachability (recommended): Enter the check address of the access on the delivery page (e.g. gpu01.example.eu:22) and allow connections from the CMB server. No data is sent and no credentials are used.
The agent is also your proof of payment: hours without keep-alives count as not delivered and are not paid. If no metrics at all arrive within 24 hours of the delivery start, the order is automatically reversed: advance payment refunded to the buyer plus a 10 % non-delivery penalty (clause 5b). If different GPUs are reported during the delivery window than at verification, or performance drops below 90 % of the baseline, the buyer may terminate immediately — also with a penalty (clause 5e).
5

Final report & settlement: Delivered hours × award price, hardware proof, event list. Undelivered hours are refunded from the buyer-protection account; the provider receives a self-billed credit note for the delivered hours.

4a · For providers: how to release your GPUs

The technical release happens on your infrastructure — CMB sets the frame (lot = GPU type, quantity, window, access type) and documents the provisioning. Four proven approaches:

1

Whole machine (bare metal): the machine belongs to the buyer for the window. Release = create a user account (Windows user + remote desktop, or Linux user with SSH key), deposit the credentials here, deactivate the account after delivery. The simplest route for single cards and workstations.

2

Virtual machine with GPU passthrough: create a VM on the host (Proxmox, KVM, vSphere) and pass the sold GPUs through exclusively. The buyer receives the VM credentials; after expiry the VM is deleted — cleanly isolated, nothing remains.

3

Containers (Docker/Kubernetes): start a container with exactly N assigned GPUs (--gpus); the buyer gets SSH into the container or ships their own image. Fastest provisioning, good isolation — the standard of the large providers.

4

GPU partitioning (datacenter cards): H100/A100 can be split via NVIDIA MIG into up to 7 independent instances for different customers. Enterprise option for data centres.

The wrapper is always the same: create account/VM/container → deposit credentials on the delivery page → run the telemetry agent on the machine (your proof of delivery per hour) → revoke access after delivery ends. Verification and the hourly spec check document contractual provisioning (clauses 5e/6 of the terms).

4b · Extension option — extending a running order

If the buyer needs the capacity for longer, the order can be extended without a new auction: the buyer requests a new delivery end on the delivery page, the provider confirms or declines, and the buyer pays the additional amount into the buyer-protection account within 48 hours — at the original winning-bid price. Access, telemetry and hourly rating continue seamlessly; the agent does not need to be restarted. There is no automatic renewal: without mutual confirmation and payment the order simply ends as scheduled — no cancellation fee, no penalty.

4c · Large orders — payment in sections

If the order value exceeds 25.000 €, the full amount is not paid up front. The delivery window is split automatically into sections of at most one week (up to 60). Each section is paid into the buyer-protection account before it starts — the first within the usual 72 hours of the award, each following one at least 48 hours before the section begins — and settled from telemetry after it ends, with its own invoice and payout — paid out to the provider 48 hours after settlement unless either side objects to that section. You get an e-mail reminder 72 hours before each payment deadline — or you set up a SEPA direct debit mandate once and each section is collected automatically 9 days before it starts.

Monthly rental: From 30 days of duration, offers and requests can have a price per GPU-month (1 month = 730 hours; billing is still based on delivered hours). In the marketplace, monthly offers are shown in a separate grid. Both types — weekly and monthly price — are auctions; optionally the provider sets a buy-now price. The buyer always takes the entire lot (all GPUs for the whole delivery window), so the provider sets the minimum commitment via the delivery window. With automatic renewal the order continues month by month until the buyer or provider cancels on the delivery page — observing the notice period the provider sets in the offer (7, 14, 21 or 28 days before the end of the current period; default 7 days).

Rental until cancelled: auctions with weekly or monthly price have no end date. The provider sets the start, a minimum term and a notice period. You prepay the minimum term within 3 days after the purchase; afterwards the rental renews weekly (weekly price) or monthly (monthly price) and each period is paid by its start — with 1 click or by SEPA direct debit if you like. The “Cancel” button on the delivery page becomes active once the notice period before the end of the current period begins (at the earliest before the end of the minimum term; for weekly prices at most 3 days). The cancellation takes effect at the end of that period; the already created next period is dropped.

Slots: in the third grid providers offer free GPU time as a window from–to, with a fixed price per GPU-hour and minimum/maximum length. You choose start and end yourself and book instantly — no bidding. Start at the earliest 72 hours after booking; each booking becomes its own order with buyer protection, the rest of the window stays bookable for others.

Both sides benefit: the buyer never risks more than one section, and the provider is paid continuously instead of only at the end. If a payment fails to arrive, the order ends with the last paid section; the 10 % cancellation fee then applies to that section alone — everything already delivered remains unaffected.

5 · The hourly principle — the fairness rules

First: you only pay for delivered hours. Each hour of the delivery window is either delivered or not — decided by the signed telemetry (keep-alives without a gap over 20 minutes, hardware as verified, performance ≥ 90 % of baseline). Undelivered hours are never charged and are refunded from the buyer-protection account. There are no more percentage SLA credits — they are redundant when outages are not paid for in the first place. How much the buyer uses the GPU is irrelevant: provisioning is paid, not utilisation.

Example: 1,000 GPU hours ordered at €2.00, 30 of them without keep-alives. Charged: 970 hours = €1,940.00; €60.00 is refunded to the buyer from the buyer-protection account.

Second: right to terminate on outages. If 2 consecutive hours or 3 hours in total are not delivered, the buyer may terminate the order on the delivery page: delivered hours are settled, the rest refunded, no penalty.

Third: wrong hardware is no delivery. If the agent reports different GPUs during the delivery window than at verification (different ID or different model) or less than 90 % of baseline performance, provisioning is improper: the buyer may terminate, gets a refund plus a 10 % penalty owed by the provider (clause 5e) — mirroring the buyer's cancellation fee. At verification there is a grace period of 2 hours. If, on the other hand, a GPU sold fails or new uncorrectable memory errors occur, only the affected hour counts as not delivered — without a penalty.

Fourth: below 50 % delivered hours the service counts as not rendered — full reversal instead of pro rata settlement: advance payment back to the buyer plus a 10 % penalty (clause 5c).

And if the measurements miss the problem? That is what the objection is for: either side may object within 48 hours of completion — for example if access was unreachable or the buyer rebooted the machine themselves (every heartbeat carries the uptime; reboots show up as events). The payout is then held until CMB decides: release, reduction or reversal.

For large amounts CMB deliberately does not decide. If the amount in dispute exceeds 50.000 €, the platform merely withholds the payment — acting as a payment agent, not an arbitrator. Funds are released once both sides declare it jointly via the platform or a settlement or judgment is presented. The signed measurements remain stored and are provided as evidence.

When is the invoice issued? If all hours were delivered, invoicing and payout happen immediately. If hours were not delivered, invoices are issued after the objection period. Should an objection lead to a reduction after invoicing, it is corrected by a credit note.

Frequently asked questions

What does "EU sovereign" mean?
The capacity is hosted by a provider within EU jurisdiction without US CLOUD Act exposure — important for privacy-critical workloads and the EU AI Act.

Can I bid on my own lot?
No — nor on your own tender. This prevents price manipulation.

What happens if nobody bids?
The auction ends without an award, the tender as "no result". Simply relist, possibly with a different starting or maximum price.

Is this real money?
Yes. Awards create binding contracts with CMB — CMB sells to the buyer in its own name and buys from the provider. The buyer pays the order value before delivery into the CMB buyer-protection account at Stripe; after delivery and the objection period, delivered hours are paid out to the provider and undelivered hours are refunded automatically.

Ready? Register now — or sign in if you already have an account.