There is no GPU on Comsky Cloud yet.
We could have written a page with a card grid, a “from” price and a waitlist form dressed up as a checkout. Instead: the console has three size controls — vCPU, memory and SSD — and the API takes those same three numbers. There is no accelerator field anywhere in the product. GPU instances are being brought online through 2026, and until one is racked this page is a statement of intent, not a rate card.
That is not a simplified diagram. That is the request body. A fourth term would have to be added to the price formula before a GPU could be billed at all, and it has not been added.
What we are evaluating, with the prices we do not have.
A shortlist, not a catalogue. Nothing below is racked, ordered or reserved, and no row of this table can be deployed. No rate is shown because a rate we have not committed to is worth less than no rate at all — and each one is published when that card goes live.
| Card under evaluation | VRAM | What it would suit | Per hour | Per month |
|---|---|---|---|---|
| NVIDIA L4 | 24 GB | Steady inference and video transcode at low power. The cheapest honest place to put a small model that has to answer all day. | On request | On request |
| NVIDIA L40S | 48 GB | Mid-size inference, fine-tuning a 7B–13B model, and rendering. The card most Indian ISV workloads actually need. | On request | On request |
| NVIDIA A100 80GB | 80 GB | Training runs that have outgrown a workstation, and batch inference on larger models. The oldest card on the list and the easiest to source. | On request | On request |
| NVIDIA H100 | 80 GB | Serious training and high-throughput serving. The card people ask us for by name, and the hardest of these to get in India at a price we could hold. | On request | On request |
| NVIDIA H200 | 141 GB | Large models that have to fit in one card’s memory rather than be split across four. Under evaluation, no further along than that. | On request | On request |
VRAM figures are each manufacturer’s own published specification, quoted so you can size a workload against them; they are not a Comsky measurement. Card names are NVIDIA Corporation trademarks and appear here only to say what we are assessing. Which of these five actually ships, in what order, and whether the list changes entirely, is not settled.
Three shapes of work, and which end of the list each one lands on.
Useful now, even with no prices: if you can place your workload in one of these three, you already know roughly what you will be asking us for — and we can tell you sooner whether we can take it.
Inference that has to answer all day
A model already trained, serving requests at a steady rate, where the cost that matters is the one running twenty-four hours a day. This lands at the small end — 24 GB to 48 GB of memory, one card, and a machine you never switch off. It is also the workload we are most likely to be able to serve first.
Fine-tuning and short training runs
Adapting an open-weights model on your own data over hours or days, rather than training from nothing over weeks. This wants 48 GB to 80 GB and tolerates being scheduled — which is exactly the kind of demand a single-zone provider can plan capacity against.
A model that must fit in one card
Large-context work where splitting across four cards costs you more in interconnect than the cards save you. This is the top of the list, the scarcest supply, and the request we are least able to promise. Ask early and we will be straight with you about the queue.
Rendering and video
Batch transcode, offline rendering and simulation — work that is bursty, tolerant of being queued, and priced by the hour rather than the month. A good fit for on-demand hourly metering, once there is something to meter.
Work that cannot leave the country
The reason most people ask us rather than a hyperscaler: the data is under an Indian obligation and the workload has to sit beside it. Comsky runs one zone, in India, and instances, volumes and snapshots do not leave it. That constraint does not change when a card is added.
Work with a budget, not a card on file
Comsky bills from a prepaid rupee wallet with no stored card. Whatever the GPU rate turns out to be, it will draw down the same wallet, carry the same 18 % GST inside the number and appear on the same tax invoice under SAC 998315 as everything else.
Four things have to be true. None of them is finished.
Written out so that “coming in 2026” means something you can check us against, rather than something we can quietly redefine in December.
Capacity, power and cooling in ZONE01
A rack of accelerators is not a rack of web servers. It needs the power draw provisioned, the cooling sized for it and the floor space contracted before a single card is ordered. This is the slowest of the four and it is a facilities problem, not a software one.
Passthrough or partitioning on the hypervisor
Comsky runs virtual machines. A card has to be either passed through whole to one instance or partitioned so several instances can share it, and whichever we choose has to survive live migration, a host reboot and a customer resizing the instance around it. Today it does neither, because there is no card.
A rate we can hold for a year
The whole argument of this product is that you can multiply three numbers and audit your own invoice. A GPU rate that moves every quarter because our supply cost moved would break that promise on the one page where it matters most. We would rather publish late than publish a number we have to withdraw.
An image and a driver story that is not your problem
A GPU instance that boots without the driver, the toolkit and a sane container runtime is a support ticket with an IP address. The template has to come up ready, and it has to keep coming up ready after the next driver release.
We are not going to pre-sell against these. There is no deposit to leave, no reservation to hold and no early-bird rate, because all three would be us taking money for a thing that does not exist.
A fourth term, or it is not a Comsky price.
Every other price on this platform is arithmetic you can do yourself. A GPU line that arrived as an opaque bundle would be the one place we stopped showing our working, so it will not arrive that way.
What we can already commit to
Whatever the rate is, it will be published in rupees, GST already inside it, on the same page as the other three rates. It will meter against the same prepaid wallet, raise the same tax invoice, and — if it is offered on demand — obey the same 120 % monthly cap as everything else. There will be no minimum term and no lock-in.
What we cannot commit to
The number itself, the date, or which of the five cards is first. Anyone quoting you a Comsky GPU rate today is quoting you something we have not agreed internally, and you should tell us about it. See what is actually priced →
Tell us the workload and we will say whether we can take it.
Not a waitlist — a straight answer, including “no” and including “not before next year”. Capacity in a single zone is planned against real demand, so a described workload is genuinely more useful to us than an email address.
Describe your workload → · If the answer is that we cannot take it, we will say which class of card it needs, so the conversation is still worth having.
Everything on this platform that is not a GPU.
CPU instances are live, priced by the resource and deployable now. If your work is inference-adjacent but not actually accelerator-bound — embedding search, OCR queues, batch ETL, a model small enough to serve on cores — it runs here today at a price you can multiply out yourself.
Live today
- vCPU, memory and SSD sized independently
- 1–64 vCPU · 1–512 GB RAM · 10–16,384 GB SSD
- Live resize without rebuilding
- Snapshots on demand, and 7-day rolling backups
- One isolated network per account, in ZONE01
Not on this page
- No GPU instance type
- No accelerator field in the console
- No GPU term in the price formula
- No GPU image or driver template
- No reservation, deposit or early-bird rate
Asked, and answered without hedging.
Can I deploy a GPU instance on Comsky today?
No. The deploy wizard sizes vCPU, memory and SSD, and POST /v1/instances accepts cores, ramGb and ssdGb. There is no accelerator anywhere in the product to select, price or bill.
What does “through 2026” actually mean?
It means we are working on it across that year and have not committed to a month. The four gates on this page — facility capacity, hypervisor passthrough, a rate we can hold, and a driver-ready image — are the real schedule. When the last one closes, this page becomes a rate card.
Can I reserve capacity or pay a deposit now?
No, and we will decline if you offer. Taking money against a product that does not exist is the thing this page is written to avoid. Describe the workload instead — that is what actually influences what we buy.
Which card will be first?
Not decided. The inference end of the list is the easiest to source and the easiest to price, so it is the likeliest — but that is a reasoned guess and we are labelling it as one rather than letting it read as a plan.
What will it cost?
Unknown, and published at launch. What we can say is the shape: rupees, 18 % GST already inside the number, SAC 998315 on the invoice, drawn from the same prepaid wallet, with no minimum term. If someone quotes you a Comsky GPU rate before that, it did not come from us.
Will it be in India?
Yes. Comsky runs one zone, ZONE01, in India, and instances, volumes and snapshots do not leave it. Adding accelerators does not add a region; if we ever open a second one, that will be its own announcement.
I need GPU capacity this quarter. What should I do?
Buy it somewhere that has it, and tell us what you needed. We would rather lose the order than hold you on a promise. If the blocker is specifically Indian data residency rather than raw capacity, write to us anyway — there may be a way to split the workload so only the part that must stay here stays here.
Does any Comsky product use GPUs today?
Not on this platform. Comsky Cloud sells CPU instances, storage and networking in ZONE01, and that is the whole of what is running.
Start with Comsky Cloud.
Grow into the ecosystem.
One account opens every Comsky product — and one invoice covers them, however many you run.