The Enterprise AI Compute Gap: Spending Outpaces Visibility

VentureBeat Pulse Researchの最新レポートは、企業がAIインフラ、特に専門クラウドに急速に投資している一方で、コストとGPU利用率に関する基本的な可視性を欠いている実態を明らかにした。

The central finding, dubbed the 'compute gap,' illustrates a disconnect: while investment in AI infrastructure accelerates, most organizations cannot effectively track its economics. Only about one in five enterprises (21%) currently run AI in production at scale, yet spending intentions significantly outpace this maturity.

Despite relying heavily on familiar hyperscalers like Google Cloud, Microsoft Azure, and AWS, and major model APIs from OpenAI and Anthropic for current AI deployments, enterprises are signaling a significant shift. The largest planned area for evaluation over the next year is AI-specialized clouds (45%), a segment almost none of these enterprises currently use.

This indicates a coming re-platforming, with a clear majority (64%) planning to switch or add an infrastructure provider within twelve months. However, immediate switching intentions (within the next quarter) primarily involve reshuffling among existing incumbent providers.

Purchasing decisions for AI infrastructure prioritize integration with existing stacks (41%) and total cost of ownership (35%), rather than headline cost per million tokens (8%). This preference for TCO is notable, given the widespread inability to rigorously measure it.

The GPUs enterprises already own run at half utilization or less for the overwhelming majority, and fewer than half can rigorously track what their compute costs or returns.

Indeed, the report highlights substantial inefficiency: 83% of enterprises report GPU utilization of 50% or less, with nearly half running at 25% or below. This means significant existing capacity sits idle while companies plan further investment. Furthermore, fewer than half (44%) rigorously track the cost and return of their AI compute, indicating a major gap between stated priorities and actual measurement capabilities.

The next major constraint in large-scale inference — the shift from GPU compute to memory bandwidth — is also largely unaddressed, with roughly one in five enterprises either unaware or unprepared. This suggests that the current 'compute gap' of cost visibility may soon be compounded by new technical bottlenecks, arriving before the first problem is solved.