Neocloud

A hyperscaler sells hundreds of services: storage, databases, email, analytics and, among them, GPU machines. A neocloud mostly sells one thing: access to large numbers of recent accelerators, often whole clusters wired together with fast interconnects for training. CoreWeave, which calls itself “the essential cloud for AI,” is the best-known example; many smaller providers rent capacity by the GPU-hour.

The appeal is availability and price. When chips were scarce, neoclouds often had capacity sooner than the big clouds and offered simpler, cheaper terms for raw GPU time. The trade-off is what you do not get: usually fewer managed services around the hardware, and a younger provider whose finances and support you have to judge for yourself. NVIDIA’s own annual report names “Neocloud builders” among the customers buying its chips indirectly, which shows how established the category has become.

For a team choosing between them, the practical questions are the same as for any capacity deal: which chips and how many, how they are networked, what uptime and support are promised, where the data sits for data sovereignty rules, and how long the commitment runs. Measure your own workload’s throughput and cost on the actual cluster before signing.

The word is industry shorthand, not a formal category, and its boundary with the big clouds is fuzzy. Neoclouds have not been discussed by name on Chain of Thought so far.