imec study: break-even GPU utilization for coding agents ranges 15%–89%
AI Engineer · youtube · 2026-10-31
Steven Latré, who leads AI at Belgian semiconductor R&D center imec, presented findings from thousands of benchmark runs quantifying when self-hosted GPUs beat APIs for long-running coding agents.
Key findings (open-weight models, three coding harnesses, SWE-Bench Pro):
- Hardware costs scale in steps, not a smooth curve
- Concurrency matters more than model accuracy for economics
- Break-even GPU utilization varies hugely: 89% on small systems vs 15% on 8x B200 machines
- The right harness can cut GPU time in half for the same result
The talk covers test setup, per-model accuracy, hardware cost modeling, API vs renting vs buying comparison, and harness/token-use analysis.
More from coding & agent
- ASO playbook lifts keywords 200+ spots, turned into a long-running SEO/GEO agent template — tinyfool · 2026-10-31
- Microsoft Research open-sources Agensh: 1,024 coding agents, no boss, 55% on Pandoc rebuild — JafarNajafov · 2026-10-31
- antirez: agent coding in Redis now hampered by constant security stops — antirez · 2026-10-31
- Google dumps in-house Pytype for Meta's Rust-based Pyrefly, which checks 1.85M lines/sec — bigaiguy · 2026-10-31
- Google open-sources EGM, the plugin its engineers use to stop AI shipping sloppy code — thisguyknowsai · 2026-10-31
- Dev calls out Anthropic for not shipping built-in memory wiki — because it sells tokens — arieljalali · 2026-10-31