Operators and workload owners
Datacentre operators, and companies running steady batch or inference workloads who would like their real cost per completed job measured against something new.
India is building gigawatts of capacity in 40 °C ambient, on some of the world's most expensive industrial power, in cities that have run out of water. We're building compute designed for that reality — rather than imported from somewhere cooler and derated on arrival.
The problem
Put plainly: at a PUE of 1.7, forty paise of every rupee an Indian datacentre spends on electricity goes to removing heat rather than doing work. At the scale India is now building, that is not a detail of operations. It is the difference between a domestic compute industry and a rented one.
What we build
Our unit of design is the total cost of delivered compute at rack and facility scale — including every watt spent removing the heat the last watt made.
Designed around 40 °C and water scarcity from the first drawing, instead of adapted from equipment designed for a temperate climate and quietly derated for ours.
You buy compute against your own workload and pay for what it completes. You do not buy, rack, cool, patch or operate hardware — that is our problem, permanently.
Less exposure to dollar pricing, allocation queues, and roadmaps decided in another hemisphere. For a growing set of Indian buyers this has stopped being a preference.
Boundaries
Being specific about what we do not do is usually more informative than another paragraph about what we do.
How we work
Every number we publish will come from our own instrumented system under a real workload, not from a vendor datasheet or a spreadsheet nobody has tried to break.
A cost model that has only ever been used to sell something is a brochure. We would rather show the assumptions and let people attack them.
Benchmarks are easy to win and expensive to believe. The only figure that matters to a customer is what a completed unit of their work costs.
If the thesis is wrong we want to find out from someone qualified, in month three, and not in silicon two years later.
Who we would like to hear from
Datacentre operators, and companies running steady batch or inference workloads who would like their real cost per completed job measured against something new.
Anyone working on datacentre thermal engineering, systems, or open compute. We publish jointly, claim no exclusivity, and would rather collaborate than duplicate.
We are raising a first round. Materials, the cost model and its assumptions are available on request to people we are in conversation with.
This site is deliberately light on detail. Varuna is pre-incorporation, and the technical and commercial specifics stay private until that is finished. We share them directly and in full, under NDA, with people we are actually talking to.
If you are one of the three above, write and say which. We answer everything.
vb@varuna.systems →