Leadership1 publisher2 min readPublished Updated
Data-starved GPUs are NetApp's sales case for its Novus file system
NetApp is pitching Novus, a file system built to exceed 100Tbps, on its own figure that data-starved GPUs can run below 30% utilization. Buyers adding GPUs have reason to measure what their clusters already use, though every number in the pitch so far is NetApp's own.
The Board Room · Leadership desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- NetApp built Novus mainly for operators of AI factories, naming neoclouds and GPU-as-a-service providers as the intended buyers.
- NetApp's own figures say a cluster of 50,000 GPUs needs 100Tbps of cumulative bandwidth.
- Clusters that typically held around 10,000 GPUs are giving way to planned builds of 50,000 GPUs and upwards, ITPro reported.
- Deloitte projects AI factory adoption in financial services rising from 24% in 2025 to 63% by 2028.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- cost On NetApp's figure, the owner of a data-starved cluster is paying for more than 70% of its GPU capacity to sit idle.
- decision A request for more GPUs now comes with an earlier question: whether existing clusters are near NetApp's data-starved floor, where extra compute would hit the same storage limit.
- constraint Taking a cluster from 10,000 to 50,000 GPUs commits the owner to roughly 80Tbps more throughput on NetApp's ratio, so the GPU order also sets the storage budget that follows.
NetApp wants the AI budget question to be about how fast GPUs are fed [1]. Kurian, who gave the opening keynote at NetApp's Insight conference [3], stated the diagnosis plainly. "Most GPUs are not fed data at the rates that they need to be fed. They have a high degree of parallelism and concurrency, which makes traditional storage architectures not capable of addressing those requirements," he said [4].
Spread across 50,000 GPUs, NetApp's 100Tbps requirement [9] comes to 2Gbps per GPU on average [2]. Novus is built to exceed that same 100Tbps [6]. NetApp supplied both the requirement and the product sized to meet it [5].
The board-deck version is simple. A GPU running at the sub-30% utilization NetApp cites for data-starved clusters [5] is idle capital, and faster storage puts it back to work. That version is incomplete. The report does not say how often real clusters fall into that condition or how far Novus lifts them, and it does not include a price, a ship date or an outside benchmark.
A finance chief would object that the company diagnosing the bottleneck is also the one selling the fix [5]. I think the objection is fair, and cheap to settle. Whether a cluster is waiting on data is a question about hardware the buyer already runs. It can be answered before any storage order is placed.
Timing matters because clusters are growing fivefold [3], toward builds like the 50,000-GPU AI factory Nvidia and Samsung announced in October last year [8]. Suppose bandwidth needs scale in a straight line from NetApp's per-GPU ratio. Then a 10,000-GPU cluster needs about 20Tbps, a fifth of what the larger build requires [4].
Novus is aimed first at the providers who rent out GPU capacity [2], so enterprises that buy time from them run on whatever storage the provider chose. In my view, utilization figures now belong in provider due diligence. Deloitte projects AI factory adoption accelerating across industries, with the energy sector reaching 82% by 2028 [11].
What to watch
- An independent benchmark showing how far Novus lifts GPU utilization on a production cluster, measured against NetApp's sub-30% baseline.
- NetApp publishing price and availability for Novus, so buyers can weigh storage spend against the cost of more GPUs.
- Neocloud and GPU-as-a-service providers disclosing utilization figures to the enterprises that rent from them.