News

Google to skip MLPerf benchmark submissions

Sunday, February 15, 2026 at 09:12 PM

Google plans to refrain from submitting performance benchmarks to MLPerf for its machine learning hardware.

Context

Google is reportedly skipping upcoming MLPerf benchmark submissions in early 2026, signaling a strategic retreat from general-purpose industry standards. This move marks a shift away from direct per-chip competition with NVIDIA, which remains aggressive in the MLPerf arena with its Blackwell Ultra (GB300) platform. Instead, Google is prioritizing a closed-ecosystem narrative, emphasizing the vertical integration of its Trillium (TPU v6) hardware with internal models like Gemini 3 Deep Think. This pivot matters as Google refocuses on system-level Total Cost of Ownership (TCO) and cluster-scale efficiency. Recent internal metrics show Trillium delivering 4.7x peak compute gains and 67% better energy efficiency over prior generations. By bypassing MLPerf v6.0, Google is betting that investors and cloud customers will prioritize the 84.6% reasoning accuracy achieved by its integrated stack over third-party standardized metrics. This "efficiency-first" strategy aligns with a broader 2026 market trend focusing on ROI and sustainable scaling for 100,000-chip TPU clusters.

Sources (1)

Related Companies

Google
Google
GOOGL
US