The explosive growth in demand for AI compute has led Zettabyte to call for a new quality standard. The company, active in AI infrastructure, highlights how the market now faces a proliferation of hardware and software offerings that are difficult to compare.

Zettabyte and the call for a standard

Zettabyte's initiative does not come in a vacuum. Over the past two years, the race toward Large Language Models and fine-tuning has multiplied the need for computational resources. GPUs with ever-larger VRAM, networking architectures like NVLink, and optimized inference solutions are just some of the variables at play. Without shared metrics, the risk is choosing systems unfit for the workload, with direct consequences on TCO.

What AI compute quality means

When speaking of compute quality, raw power is not the only factor. Latency, throughput in tokens per second, energy efficiency, and the ability to handle different quantization levels become essential parameters. Zettabyte aims to define a standard that includes these aspects, enabling those evaluating on-premise, cloud, or hybrid deployments to make informed decisions. A common benchmarking framework would help avoid surprises in production, such as memory bottlenecks or unacceptable response times for real-time applications.

Impact on on-premise infrastructure

For organizations that choose to keep data in-house — for sovereignty reasons, GDPR compliance, or simply cost control — a quality standard is even more critical. In self-hosted environments, the choice of hardware and serving software cannot rely on marketing claims. The ability to compare uniform metrics would allow proper cluster sizing, reduce Total Cost of Ownership, and ensure predictable performance. AI-RADAR, in its dedicated analysis, examines precisely these trade-offs, highlighting how the lack of shared references often leads to overprovisioning or, worse, inadequate infrastructure.

The broader context

Zettabyte's proposal fits into a discussion already started by consortia like MLCommons with MLPerf. However, current benchmarks often fail to cover real-world deployment scenarios, especially for those operating on-premise with variable workloads. Standardizing compute quality could foster a more transparent ecosystem, where companies evaluate solutions with the same clarity used today for database performance. As demand continues to grow, Zettabyte's initiative signals a necessary market maturation, where quality can no longer be optional.