Home / About
Kaldera is a GPU compute marketplace. The company is being incorporated and the product is in Phase 0. This page states what we want to build, which principles we are bound by, and what we are not as of today.
This is a draft marketing site; the figures on this page are examples.
Why
Kaldera came out of three concrete problems, not out of an idea. All three exist today; none of them alone justifies a new company, but together they do.
Rigs left over from mining, render studios, university clusters. The card works but cannot find a job; and when it does, payment, invoicing and trust are all handled by hand.
On cheap marketplaces a broken node is billed by the hour too, and surprise line items show up at the end of the period. Lowering the price is easy; keeping the invoice honest is hard.
KVKK (Turkey's personal data protection law), GDPR and the AI Act are on the same table. "Which country does it run in, which law are you subject to" has become part of the purchase decision; we want to answer that with architecture from the start.
Three principles
Each of the three sentences on the home page has a data structure and a delivery slice behind it. Below you get both the sentence and where it stands today.
A second that fails the health check enters no invoice line at all — not as a discount, but as if it had never been measured. Metering and the health gate are one of Phase 0's delivery slices; they are not running in production yet.
The cap is a brake, not a warning: when it fills, jobs are paused safely, the checkpoint is kept and a notification goes out. Raise the limit and it continues where it left off. The cap is not a guillotine.
Measuring the seconds you use instead of rounding to minutes or hours. A live counter and the breakdown on one screen; every line item is visible, including the storage charge on a stopped pod, and the commission rate is written down too.
Approach
We chose to get the marketplace experience right before building our own GPU network. The user sees one account, one API and one invoice; whose card spins underneath changes over time.
One contract and one console on top of providers that have a public provisioning API. Kaldera sells one endpoint, one invoice and one support line; the providers behind the supply are disclosed as a sub-processor list rather than offer by offer.
Our own benchmark bot scores the supply continuously: uptime, benchmark consistency, job acceptance rate, network. The score determines placement order — an input inside the code, not a marketing line.
When the host agent goes live, community machines will not sit in a separate list: they compete in the same marketplace under the same scorecard system. No privileged listing.
Region pinning is enforced server-side; a UI filter does not count as a security boundary. Identity data stays in our own database.
Roadmap
The phases below follow the order in our technical plan. Durations are relative to the start of the plan; we give no calendar dates, because giving them would mean inventing them.
One account, one API, one invoice and one console on top of existing provider APIs. Price and scorecard comparison across providers, a live cost counter, a hard budget cap. Done criterion: a user adds a card and can spin up GPUs from different providers within a few minutes, and sees a single invoice at the end of the period.
A single-binary host agent goes live; idle cards join the pool. Metering and billing run entirely on our own infrastructure. Done criterion: community machines are listed in the marketplace and the guarantee holds on that supply too.
An OpenAI-compatible endpoint that scales to zero: bring a model, get a URL. The model cache sits on the host's disk, with image pre-pulling for a low cold-start target. Done criterion: a customer pastes a model name and gets a production endpoint.
An EU legal entity, region pinning, the KVKK/GDPR/AI Act package, starting the SOC 2 Type I process, and an SLA-backed "Verified" tier made of audited data centres. Done criterion: the first enterprise contract with a sovereign region clause.
Honesty
This is the most important section on the page. The easiest lie a marketing site can tell is describing something future in the present tense. The following do not exist today.
The product is in Phase 0: the control plane, the agent and the console are skeletons. Console links point at a local development address. There is no public sign-up flow.
If you do not see a reference, a logo wall, a case study or a user quote on this site, the reason is simple: there aren't any yet. When there are, we will write them with their name and their permission.
We make no percentage uptime commitment, no credit table and no support response time promise. The SLA-backed Verified tier is a Phase 3 subject; until that day there is no commitment either.
The serverless endpoint is Phase 2; multi-node clusters, the model catalog and bare metal come after it. They sit in the menu with a "soon" badge, and they are not working endpoints today.
The host agent arrives in Phase 1. In Phase 0 the supply comes from the public APIs of third-party providers. We are not hiding that — it is written on this page; but on individual offers we show what is measurable instead of a provider name: country, reliability score, network and price.
SOC 2 Type I has not been obtained; the process stands as a target to be started in Phase 3. We claim no audit report, no security seal and no award.
The trade name, address, registry and tax details will be published on the contact page once they are settled. Until then the company fields in the legal texts are placeholders.
We do not invent names, titles or photos; there is no investor, press or advisor list either. When the team is settled it will appear here with real names.
Next
You don't pay for seconds that don't work, a hard budget cap, per-second billing. The day we fail to keep those three, this page is worth nothing.
The console currently runs at a local development address; there is no public version live yet.