Words we use, in plain English

One line each. The same lines appear when you tap a dotted word anywhere on the site.

Ways to buy

Buy now: Order GPUs listed for sale at the price shown. We hold them for 12 hours while we confirm with the supplier, usually the same business day.Where you meet it →
RFQalso: Request a block, Request for quotes: Request for quotes: you describe the block you need, suppliers send prices, you pick. We aim for three firm quotes within 24 hours.Where you meet it →
Open marketalso: Secondary market: Spare GPUs that AI companies and other GPU owners list for weeks or a few months. Fewer servers, booked fast.Sellers list whole servers for the weeks they are free; buyers book them at the price shown or ask for quotes. For bulk or long terms, see Data-centre direct.Where you meet it →
Data-centre directalso: Direct contract, Primary market, Data centre direct: Bulk GPUs on a long contract, usually a year or more, signed straight with a partner data centre. We find the options and run the paperwork; you pay the data centre.Where you meet it →
Firm quote: A price the supplier is bound to until the date on it. You can accept it as it stands.
Indicative quote: A price to talk about, not an offer. Nothing can be accepted until the supplier makes it firm.
Hold: GPUs kept out of anyone else's order while yours is confirmed: 12 hours on Buy now, 5 business days after an LOI.
Notify mealso: Standing request: Leave your email and what you want. When matching GPUs are listed, we email you once per new listing. Unsubscribe from any email.
Reserved: A set number of GPUs for a set term, often months. You pay for the whole term whether you use it or not, and pay less per hour than on-demand.Where you meet it →
On-demand: Pay by the hour with no term and stop any time. It costs more per hour than reserved, and is not always free when you want it.Where you meet it →
Pay as you go: You pay for the GPU-hours you run, from a balance you top up. No term; stop any time.
GPU-houralso: $/GPU-hour, GPU·hr: One GPU running for one hour: the unit compute is priced in. A server of 8 GPUs for a day is 192 GPU-hours.
Price referencealso: Silicon Liquid price reference, Reference price: The middle of public on-demand prices for one GPU across providers, collected every day. It shows whether a price is fair.Where you meet it →
SL cleared printalso: Print: A price from a trade signed on Silicon Liquid, published without names. We show one only once 3 signed trades exist in 90 days.Where you meet it →

What you are buying

GPU: The chip AI models run on. Its memory (GB) decides how big a model fits.
Server: One machine with 8 GPUs in it, sometimes called a node.
Nodealso: Whole nodes: Another word for one server, usually with 8 GPUs. We sell whole nodes only, never a share of one.
Rack: A cabinet of servers wired as one unit; an NVL72 rack holds 72 GPUs.
Window: Whole servers on one fabric island, reserved for you for a set range of dates. Training capacity is sold this way.We never sell a share of a server or an idle percentage: a window is always whole servers for whole dates.
Pool: GPUs you take one at a time, each behind its own endpoint, for as long as you need. Inference is sold this way.
Fabric islandalso: Island, One fabric: A group of servers on the same fast network, so they can work on one job together. A training block sits on one island.
InfiniBandalso: IB: A very fast network between servers that lets them work as one. Needed to train across servers.
RoCE: Fast links between servers over Ethernet. Does the same job as InfiniBand; at large scale it is often a little slower.
Endpoint: A web address (HTTPS) your app sends requests to. The supplier runs the GPU behind it; you never manage a server.
Delivery class: How the GPUs reach you: an endpoint (M), a Slurm cluster (C), Kubernetes (B), virtual machines (D) or whole servers (A).
Class M: Ready-to-use endpoint. Your model or container runs behind its own web address (HTTPS) and key. No servers to manage, and no SSH.
Class C: Slurm cluster access. A reservation on the supplier's Slurm cluster: log in and submit jobs. Login nodes and storage are shared with the supplier.
Class B: Your own Kubernetes. A private Kubernetes cluster on servers kept for you: deploy with kubectl.
Class D: Virtual machines. Virtual machines from the supplier's cloud, with SSH.
Class A: Whole servers (bare metal). Whole servers, wiped and handed to you with root access. Setting up takes hours to days.
SSH: Logging in to a server's command line to run your own software on it.
Slurm: The job scheduler most training clusters use. You submit jobs; it runs them on the servers you reserved.
FP8: Storing each number of a model in 8 bits. Half the memory of the usual 16 bits, nearly the same answers.
Tier III: A data centre that can be maintained without switching your servers off.
Tier IV: A data centre built so that no single failure switches your servers off.
Data residency: A rule that your data must stay in a given country.

Checks and handover

Evidence levelalso: Checked: How much of what a supplier says we have checked ourselves, from Declared (the supplier says so) to Delivered (a rental here ran its term).Where you meet it →
Declared: The supplier says it has this capacity and has passed our company checks. We have not measured it ourselves.Where you meet it →
Profiled: We have the supplier's delivery profile on file: how it hands capacity over, and how long a changeover takes.Where you meet it →
Reporting: Our monitoring agent is receiving live readings from these GPUs.Where you meet it →
Steady: Seven days of live readings with at least 95% coverage, and nothing out of place.Where you meet it →
Certified: Passed our acceptance test on these machines in the last 30 days, with a signed certificate.Where you meet it →
Delivered: At least one rental here has run its term and been handed back with a signed certificate.Where you meet it →
Deverified: Live readings stopped or a spot check failed. Ask us why before you buy.Where you meet it →
Delivery Certificatealso: Certificate, Handover certificate: Our signed test result for the servers handed over. The contract, and your payments, start from it.Where you meet it →
Acceptance: You confirm the GPUs you were handed pass the tests on the certificate. Your term, and what you pay for, start from it.
Deemed acceptancealso: Accepted automatically: If every test passed and you neither accept nor object within 48 hours, the GPUs count as accepted and your term starts.
Return Certificate: Our signed record that the servers were handed back at the end of a term, wiped and in working order.
KYB: Know your business: the standard check of who a company is and who owns it.
Screening: Our check of a company against sanctions lists, and of where it is, before it trades. Bigger deals get deeper checks.Where you meet it →
SLA: Service level agreement: the uptime the data centre promises, and the credit you get if it misses.
Disputealso: Open a dispute: Telling us something went wrong with GPUs you bought. The supplier answers, and our desk decides on the tests and readings within 5 business days.Opening one inside the acceptance window stops automatic acceptance. Beyond our desk, the paper names an arbitrator.Where you meet it →
Replacementalso: Replaced: Equivalent GPUs we find when delivery fails: the same model, as many or more, for the rest of your dates. You accept or decline.We look at the same supplier's other capacity first. If the replacement costs more, the first supplier owes you the difference, up to what the original would have cost.Where you meet it →

Paperwork and money

LOI: Letter of intent: a short, non-binding note that you mean to buy. It holds the capacity for 5 business days while the contract is written.
NDA: Non-disclosure agreement: both sides keep what they learn private, and agree not to go around Silicon Liquid for 24 months.
Non-circumvention: A promise not to cut out the party that made the introduction, here for 24 months.
Buyer Credit Sheet: One page on your company for the data centre: who you are, how the contract is funded, who signs.
Data Centre Fact Sheet: One page on the data centre for you: power, cooling, network, certifications and track record.
Term sheet: A short, non-binding summary of the deal: price, length, start, upfront payment.
MSA: Master services agreement: the data centre's main contract, with the rules for the whole relationship.
Order form: The page that says exactly what you are buying: GPUs, servers, price, dates, payments.
Addendum: Silicon Liquid's short add-on to the data centre's contract: our certificate starts it, our monitor measures uptime, and our fee is written in.
Tranche: One part of a long contract delivered on its own date, such as 24 servers in March and 24 in May. Each has its own certificate.
Billing start: The day a tranche's certificate is accepted. The data centre invoices that tranche from this day; you pay them directly.
AUPalso: Acceptable use policy: Acceptable use policy: what you may and may not run on the servers.
DPAalso: Data processing addendum: Data processing addendum: how the data centre treats any personal data you put on the servers.
Take-or-pay: You pay for the whole contract even in months you do not use it. Most long contracts work this way.
Deposit: The part of the price paid before delivery, to the supplier. You get it back if the GPUs are not delivered.
Commissionalso: Venue fee, Our fee: What the supplier pays Silicon Liquid when a trade goes through. It is never added to the price you see.Where you meet it →
Credit notealso: SLA credit: A record of money a supplier owes you, such as an SLA credit or a refund. The supplier takes it off its next invoice or pays it back; we never hold it.Where you meet it →

A word missing, or a line that does not make sense? Tell us and we fix it.