About DataStelio
DataStelio is an early-stage company with one product, InfraPulse, and one question behind it: when a GPU fleet reports 100% busy, how much did it actually deliver?
We started with the same assumption most people hold — that "GPU utilization" is a single, well-understood number — and found instead that it is three numbers hiding under one word. Allocation is what the scheduler shows. Busy is what the device reports. Delivered is what the silicon completed. The gap between the second and the third is where a large share of AI infrastructure budgets goes, and almost no fleet plots it, because the fine-grained telemetry that exposes it sits unread in DCGM and amd-smi.
What we believe
Measure before you recommend. There is no shortage of tools that will move a workload or drain a node. There is a real shortage of tools that can say, with a signal timeline and a dollar figure, what a fleet delivered and why it fell short. InfraPulse is read-only by design until an operator has confirmed the measurement is real.
Scarcity makes yield matter more, not less. In a sold-out market nobody feels slack. But new racks arrive in quarters and a yield point lands this month. Semiconductor fabs never stopped tracking yield because demand was high; they tracked it harder.
Published sources only. Every number in our writing links to where it came from. When we cannot find a public source — as with DCGM throttle-reason data, which appears in no public dataset we could locate — we say so, because that absence is part of the point.
Who is building it
DataStelio was founded by Shobhit Paliwal, a software engineering leader with more than fifteen years in high-scale distributed systems, eventing platforms, observability, and bare-metal infrastructure. The path to InfraPulse ran from observability of software systems to curiosity about the hardware underneath the current wave of AI, and then to a simple discovery: the utilization number everyone quotes is not measuring what they think it is.
Where we are
Design-partner phase. We are working with a small number of GPU operators — NVIDIA and AMD — trading written root-cause analyses for real fleet telemetry. If you run GPUs and want a second pair of eyes on numbers nobody usually reads, the offer is here. If you think the three-numbers picture is wrong, we would especially like to hear that.
Send 24 hours of DCGM or amd-smi output from any fleet you run and we will send back a written analysis of where your effective compute is going. No agent to install, no access to anything, no NDA needed, nothing published without your say-so.