Last updated: September 9, 2026
Kernpool is a service operated by Light & Honor Technologies Pte. Ltd., a company registered in Singapore under UEN 202245535R. In these Terms, "Kernpool", "we", and "us" mean that company; "you" means the person or organisation using the Service.
By creating an account, connecting a cloud account or cluster, or submitting a workload, you agree to these Terms. If you are agreeing on behalf of a company, you confirm you have authority to bind it. If you do not agree, do not use the Service.
Kernpool is a scheduler and control plane. You connect the cloud accounts, Kubernetes clusters, Slurm queues, or machines you already have, and Kernpool decides where each workload runs, provisions it there, and tracks what it costs.
Your compute stays yours. Workloads launch inside your own accounts and clusters, on machines you rent directly from your provider.
We do not resell compute, do not hold your cloud credentials beyond what scheduling requires, and take no share of what you pay your provider. The management fee in section 6 is the only thing we charge for the Service. We do not charge for closing an account, and where a refund is due under section 10 it goes back in full — there is no exit fee and no handling fee.
We may refuse registration, or suspend an account, where we reasonably believe these Terms or applicable law have been breached.
Because everything runs in your own environment, the parts of security and compliance that sit inside that environment remain yours.
Kernpool charges a graduated management fee on the GPU capacity it schedules for you.
The pricing page is part of these Terms. The band rates, the worked examples, the reference source, and the fixed rates it carries all form part of this agreement. Where the pricing page and this section disagree, this section wins. Section 13 governs how either of them changes.
How the fee is calculated. Kernpool knows the accelerator type, the region, and how long each machine ran. It applies the band rate to a reference cost for that machine. How that reference cost is arrived at depends on whose machine it is.
Capacity you rent from a public cloud. The reference cost is the public spot rate: the spot price published for that accelerator type in that region by the reference source named on the pricing page, as it stood at the moment the machine was provisioned. That figure is then held for the life of that machine, so a price move part-way through a run does not change what you are charged for it.
Spot is the basis even when you run on-demand or reserved capacity. That normally means the figure we charge against is at or below what you actually paid, since spot is usually the cheapest rate on the open market. The exception, stated plainly rather than left to be discovered: a large committed discount or provider credit can put your real cost below spot, and the fee does not follow it down.
Hardware you own, and Slurm or Kubernetes clusters you already run. These have no rental cost and no cloud region, so there is no spot price to look up for them. The pricing page carries a separate fixed rate per accelerator per hour for hardware you own, and that rate is what the band is applied to. You are paying for scheduling, not for the machine.
Anything not on either list. If your accelerator appears on neither the spot reference source nor the fixed-rate tables, we will agree a rate with you in writing before scheduling it, and we will not charge a fee for it until we have. Where the pricing page maps an unlisted accelerator onto a listed one, it uses the closest published model by memory and interconnect, and it sets out the mapping.
We do not read your cloud provider's billing or cost APIs. Your invoices, spending commitments, and negotiated discounts stay between you and your provider.
Committed or reserved fleets at scale are usually better served by a separate agreement. Talk to us before you commit.
How the bands are applied. Bands are assessed on the GPUs running at each moment of billing, not over a month. In practice we take a snapshot every minute and bill each snapshot separately. The GPUs live in that snapshot are ranked by reference cost from highest to lowest and filled into the bands in that order, so your effective rate moves with the size of the fleet: grow it and more of the total lands in the cheaper bands; shrink it and the rate rises again. Nothing is carried over or reset — there is no monthly settlement.
You keep all rights in the code, data, and model weights you bring. We claim no ownership and no licence to use them for our own purposes. In particular, we do not use your code, your data, or your model weights to train our own models, and we never will without a separate written agreement.
You grant us the narrow licence needed to operate the Service: to submit your workload to the machine it was scheduled onto, and to store the job metadata and logs shown in your dashboard.
Your workloads run on storage inside your own account, so what happens to a disk after a job ends is governed by your provider's terms and your own retention settings, not ours.
We may suspend an account immediately, without notice, where we reasonably believe conduct in this section is occurring and there is risk of harm to other users or to the network.
The Service is provided without an uptime commitment unless you have a separate written agreement with us. Kernpool schedules onto capacity your provider makes available; quotas, regional shortages, and spot reclamation are theirs, not ours.
We may perform maintenance on the control plane and will give notice where practical. If the control plane is unavailable, workloads already running in your account continue — scheduling of new work pauses. Design long-running jobs to checkpoint.
Suspension. Where we suspend an account rather than terminate it, scheduling of new work stops and the dashboard stays available. Machines already running in your own account are not touched — we do not shut them down, and your provider keeps billing you for them, so check whether you want them stopped. Autostop policies you have already set continue to run. The balance is not affected by a suspension. We will tell you why, and what would lift it.
Closing your account. You may close your account whenever you like. Any remaining balance ends with the account and is not paid out, so if you are finished with us, spend it down first. There is no hurry, and no need to close at all: an idle account costs you nothing and the balance keeps indefinitely.
Accounts that go quiet. We do not close accounts for inactivity. After 24 months without activity we send one email to remind you the balance is still there. Ignoring it changes nothing. We do clear out data that is no longer doing anything, which the privacy policy sets out; the balance is not affected.
If we end it and you did nothing wrong. If we terminate your account for reasons of our own — we no longer want to serve your use case, or we leave your market — we return the unused balance in full to the card you paid with, with no deduction. If we discontinue the Service altogether we give at least 90 days' notice and refund every account in full.
If we end it because you broke these Terms. We may suspend or terminate your account for breach of these Terms, or where the law requires it. Except where the breach falls under section 8, or is otherwise serious or urgent, we will tell you what the problem is and give you a reasonable opportunity to put it right before terminating. Breach of the acceptable use rules in section 8 ends access immediately. Either way the balance ends with the account, on the same footing as if you had closed it yourself — nothing is deducted from you that you would have kept by leaving voluntarily. Losing the Service is the consequence of breaking the rules; there is no financial penalty on top of it.
Fraud. Where we terminate for fraud, for chargebacks we can show to have been fraudulent, or for circumvention of metering or billing carried out to avoid paying us, access ends immediately and without notice, and the balance ends with the account. This case requires intent to defraud and evidence of it. An ordinary dispute about metering is not fraud and falls under the paragraph above.
Nothing accrues against you. We do not invoice in arrears, so there is no running debt: the balance drops to zero, scheduling stops, and you owe us nothing. The one exception is a payment that is taken back after the balance has already been spent — a chargeback, or a payment your bank recalls. In that case the account carries a deficit equal to the amount reversed. We will tell you, scheduling stays off until it is cleared, and you can clear it by topping up or by resolving the reversal with your bank. We will not set the deficit off against anything else of yours.
What survives. On termination your right to use the Service ends immediately. Any provision that is meant to operate after the account has gone continues to do so — in particular the ones defining the balance and the fee, setting out what happens to the balance, limiting liability, and governing law and disputes.
The Service is provided "as is". To the extent the law allows, we exclude all implied warranties, including fitness for a particular purpose and uninterrupted availability.
We are not liable for indirect or consequential loss, for lost profits, or for lost or corrupted data. Our liability for all claims taken together is capped at the management fees charged to your account in the twelve months before the first event giving rise to a claim — counted as charged, whether or not any of them were later refunded to you — not the balance you topped up, and not what your cloud provider billed you.
Nothing here excludes liability that cannot be excluded by law, including for fraud or for death or personal injury caused by negligence.
You will indemnify us against claims arising from your workloads, from content you brought to the Service, from the cloud accounts you connected, or from your breach of these Terms — including reasonable legal costs.
We carry the equivalent risk in one direction: if someone claims that the Service itself, as we supply it, infringes their intellectual property, that is ours to defend and ours to pay for. We will not ask you to indemnify us for the thing we built, and the cap in section 11 does not apply to this obligation — a cap measured in management fees would make the promise worth nothing against a patent claim.
That promise works on the usual terms. Tell us about a claim once you know of it — if you are late, we are only released to the extent the delay actually damaged the defence. Let us run the defence and any settlement, do not admit anything or settle without asking us first, and give us the cooperation and information we reasonably need, at our cost. If we decline to take the defence on, or do not take it up within a reasonable time, you may run it yourself and we will pay your reasonable costs and any award.
If a claim looks like it will stick, we may at our own cost get the right for you to keep using the Service, change it so it no longer infringes without making it materially worse, or, if neither is reasonably available, end your account and refund your unused balance in full.
None of this covers a claim that arises because you modified the Service, combined it with something we did not supply, or used it outside these Terms, where the claim would not have arisen otherwise.
We may change these Terms, and we may change the fee rates and reference rates on the pricing page. Either way, for material changes we give at least 30 days' notice by email and update the date at the top of the page concerned. Continuing to use the Service after a change takes effect means you accept it.
A rate increase does not reach money you have already paid in. Balance on the account when the change takes effect is metered at the old band rates until it is spent; the new rates apply to what you add afterwards, and where both are present the older money is spent first. A rate cut applies to everything at once, including balance you topped up before it — there is no sense in making you wait for a lower price.
What that does and does not fix. What is held is the band rate — our percentage. The figure it is applied to is the reference cost, and for rented public-cloud capacity that is the public spot rate, which moves with the market and is not ours to hold. If spot prices for your accelerator rise, the same run costs you more in fee, and that is not a change under this section: no notice, no right to refuse it. We mention it because "your rate is protected" would otherwise promise more than it delivers. The fixed rates on the pricing page, including the ones for hardware you own, are ours, and a change to those does fall under this section.
If you do not accept a change, tell us before it takes effect and we will return your unused balance in full to the card you paid with. Your money is already with us, so a change you did not agree to should not leave you holding a balance that can only be spent on terms you rejected. The payment and refund policy sets out every route by which money comes back, and the deadline for each.
These Terms are governed by the laws of Singapore, without regard to conflict of law rules. Disputes are subject to the exclusive jurisdiction of the courts of Singapore.
Before filing, please write to us — most disputes are resolved faster that way.
Light & Honor Technologies Pte. Ltd.
UEN 202245535R
[email protected]