Most of the compute in this year’s $450 billion AI infrastructure build never does useful work. RidgeScope diagnoses why in minutes, from telemetry alone: nothing to instrument
, /PRNewswire/ — ApexData Inc. today announced the general availability of RidgeScope, an AI-powered diagnostic platform that reads the telemetry of a GPU training run and determines what failed, why and what to do about it. It arrives as hyperscalers spend roughly $450 billion on AI infrastructure this year.
Behind that number sits what the industry calls the MFU gap – the shortfall between compute bought and compute used. Model FLOPs utilization, the share of a GPU’s theoretical compute that actually advances the model, typically sits between 30 and 40 percent; fleet studies measure averages closer to 20 percent. More than half of every dollar spent on GPU compute is lost to slow chips, starved data pipelines, communication stalls, hardware faults and runs that finish cleanly while learning nothing.
At about $3.26 per GPU-hour, a 128-GPU H100 cluster costs roughly $10,000 a day on demand, some $3.65 million a year. At 20 percent MFU it does half the training work the same hardware would deliver at the 40 percent a well-tuned run reaches: roughly $1.8 million of that annual bill buys nothing that better engineering could not recover. Closing the gap from 20 to 40 percent doubles the productive output of the same hardware without buying a single additional GPU.
“The industry is financing GPUs as if they were fully productive assets, yet most clusters deliver less than half of what the spec sheet promises,” said Evgeny Potapov, co-founder and CEO of ApexData. “This is not only a neocloud problem. It applies to every organization that trains models. You cannot close a gap you cannot see, and today almost nobody sees it.”
A lightweight agent on each server collects more than 12,000 signals per training run: GPU behavior, interconnect traffic, job logs and scheduler records. An AI engine weighs the evidence against more than 20 known failure patterns, ruling each in or out. Every case ends in a verdict: what the data shows, what it implies and what it cannot decide, each claim tied to the measurement behind it.
“GPU waste is quiet. One slow chip drags down 127 healthy ones, a job saves its progress so often that it stops making any, another sits on expensive hardware computing nothing, and every dashboard stays green,” said Andrey Shamakhov, co-founder and CTO of ApexData. “We name the failure, show the evidence and say what it costs. And when the data cannot answer something, the verdict says so instead of guessing.”
For GPU cloud operators, RidgeScope settles the question behind every support ticket: customer code or cluster hardware.
RidgeScope recognizes Hugging Face Transformers, Megatron-LM, PyTorch Lightning and Keras/TensorFlow.
For most teams, training data is the company’s most valuable intellectual property. RidgeScope reads only system telemetry and scheduler metadata – never datasets, source code or model weights – and can run entirely inside the customer’s own network: on-premise or air-gapped, with local LLM models, per-tenant isolation, SSO/SAML, and GDPR- and PCI DSS-grade encryption.
Generally available today, RidgeScope installs in minutes on Slurm and Kubernetes and returns a first verdict within 30 minutes. A public Failure Catalog and Evidence Index list every failure mode and every signal a verdict may cite. Demonstrations run live at https://ridgescope.ai.
About ApexData
ApexData Inc. (Sunnyvale, California) builds AI-powered observability for ML training and infrastructure, drawing on two decades of DevOps and distributed-systems experience.
Media Contact:
Andrei Surkov
+14083298998
[email protected]
SOURCE ApexData Inc
