← All speakers

Steven Latré

Conference affiliation: VP AI · imec · 2026

Steven Latré is vice president of AI at imec, where he connects AI applications with the design of the chips and systems that run them. Models change within months, while developing new silicon takes years. His work helps hardware designers understand what emerging applications need—and helps application developers compare hardware by the cost, speed and quality of the work it completes.

On this page

From network research to AI products

Steven completed his doctorate at Ghent University in 2011, researching automated management of multimedia networks and the quality experienced by their users. He led IDLab Antwerp at the University of Antwerp from 2013 to 2022, helping grow a small team into roughly 100 researchers working across communication networks and artificial intelligence.

Working with colleagues at Ghent University, his team became the only European finalist in a DARPA competition using machine learning for dynamic spectrum management: coordinating access to radio frequencies. Steven describes the result as a collaborative achievement across disciplines. His approach to research leadership emphasizes shared outcomes over individual performance measures, including the effort required to work with specialists who approach problems differently. He became an imec fellow in 2022 and took responsibility for bringing its AI research activities together.

In 2025, Steven joined Openchip as chief AI and software systems officer, stepping down as imec’s head of AI while remaining an adviser. He described the move as a chance to build AI products. Openchip’s ambition combined a European RISC-V AI accelerator with the software needed to use it; Steven helped build a team working on agentic systems and efficient model serving on new hardware.

When Openchip concentrated its efforts on high-performance computing and AI datacenter chips, its Belgian software and agentic-AI team stayed together. Steven announced his return to imec with all 17 team members, with a May 1 start. His remit as vice president of AI included bringing AI offerings to system and application customers making decisions about hardware, inference costs and unused performance.

Comparing hardware through real workloads

Steven helps lead AIStack, imec’s public benchmarking initiative. He leads AI strategy across silicon and software, alongside colleagues responsible for product, engineering and experimental work. AIStack measures applications end to end, giving model developers, infrastructure providers and chip designers a common basis for comparing systems.

Two strands explain the initiative’s approach:

  • Self-hosting economics depend on utilization. In AIStack’s comparison of APIs, rented GPUs and purchased hardware, the team ran 64 coding tasks across several model and hardware combinations. Running more jobs at once spreads hardware costs across more work, but eventually slows each task. Fitting a model in memory is therefore only a starting point: useful capacity also depends on completion time, solution quality, caching and how much of the day the machine is busy. The team found configurations where renting beat API prices and others where the model provider’s API remained cheaper.
  • Benchmark scores depend on what agents can access. AIStack’s investigation of coding-agent benchmarks began when the team could not explain its rankings. Agents were retrieving solution commits from repository history and upstream pull requests instead of independently solving the tasks. The researchers rebuilt task images to remove accessible answers and added explicit restrictions on retrieving solutions. Those interventions substantially changed the scores. Comparing hardware for coding agents requires checking that the measured work represents the task developers actually want completed.

1 conference talk

Loading more talks and highlights…

References