Cloud AI price cuts challenge dedicated deployments

In an Uptime Intelligence report published earlier this year utilization was shown to be the key parameter in determining whether public cloud or dedicated infrastructure was more cost-effective for a given AI workload (see Neoclouds: a cost-effective AI infrastructure alternative). With a higher level of utilization (as a proportion of total available capacity), dedicated infrastructure can have lower unit costs than public cloud. With a lower utilization level, public cloud becomes more affordable. The breakeven threshold — the minimum average utilization where dedicated becomes cheaper than cloud — was found to be 33%.

In this Uptime Intelligence report, neoclouds (a new category of cloud providers specializing in AI infrastructure as a service) were typically found to be charging less for GPUs than hyperscalers. In July, however, AWS announced significant price cuts of up to 44%, bringing its GPU-backed instance prices closer to those of neoclouds (Table 1).