MLPerf Named Yardstick in EU AI Gigafactories Tender as Endpoints v1.0 Opens

MLPerf scores bids in the EU's AI Gigafactories call

The European Union opened its call for AI Gigafactories in July. It covers up to seven large-scale AI computing facilities, backed by up to €10 billion in EU and national funding and expected to unlock more than €30 billion in total investment, according to MLCommons. The tender is a joint procurement by the EuroHPC Joint Undertaking and national contracting authorities, with bids due November 12, 2026.

The tender specifications name MLPerf as the standard for measuring bidders' technical performance. Proposals are scored on MLPerf Training v5.1 and Inference v5.1 benchmarks. MLCommons says the tender names more than half a dozen models from MLPerf's standard workloads: Flux.1 and Llama on the training side, and Llama 3.1 405B, DeepSeek R1 and SDXL 1.0 for inference.

The requested metrics build on MLPerf's standard measures: time-to-train for training, and time to first token and time per output token for inference. Additional details such as accelerator utilization are also requested. Bidders must report results both for the capacity delivered in Phase 1 and for their planned Phase 2 expansion.

David Kanter, Head of MLPerf, said: "MLPerf is designed to help inform buyers as they navigate deploying AI capabilities."

Endpoints tests AI through the live API

MLPerf Endpoints measures AI as it is consumed in practice: through the live API that applications call to reach a model. The system under test is simply a URL that the Endpoints client talks to over standard interfaces. According to MLCommons, this lets buyers compare hardware vendors, hyperscalers, neoclouds and managed API providers on equal footing, whether they run bare metal, a managed cloud service or a containerized software stack.

Rather than reporting a single peak number, Endpoints sweeps each system across a range of loads and reports a full performance profile. Results cover:

• concurrency
• interactivity, measured as tokens per second per user
• system throughput
• time to first token at the 90th percentile (TTFT P90)

Buyers can use this to match a verified operating point to their own applications and service-level targets. Each result appears as a pareto frontier on the LoadStair visualizer, which shows where a system holds steady and where it hits a performance cliff.

Version 1.0 adds power measurement and rolling submissions

Endpoints 1.0 introduces mandatory provisioned power measurement, a first for the MLPerf family. MLCommons says this helps buyers understand and compare energy efficiency and the cost profile of each solution.

The version builds on the v0.7 foundation release from July. It covers generative AI workloads including DeepSeek-R1, GPT-OSS 120B and Llama models, plus new agentic coding workloads using DeepSeek v4 Pro, Kimi K3 and Qwen 3.6.

Endpoints now offers rolling submissions. Members can publish a peer-reviewed result as soon as they ship a new chip or software optimization instead of waiting for a fixed round. Every result still goes through MLPerf's peer review and audit process. MLCommons expects the first v1.0 results to be published in November. Members can find the submission guide and rules at mlcommons.org/benchmarks/endpoints.

What it means for AI buyers

MLCommons frames the two updates as pointing the same way: industry, government and buyers are converging on open, peer-reviewed measurement as the basis for AI adoption decisions. Buyers are asking about delivered service performance, cost for given workloads and energy use, not only chip speed, according to the organization.

For public procurement at the scale of the Gigafactories call, MLCommons argues that teams need performance claims they can trust and compare, and says open, peer-reviewed results provide that. For commercial buyers, Endpoints is meant to offer comparable, audited operating profiles and power data across very different kinds of providers.