A Walk Through the ICML 2026 Sponsor Booths

Industry seems to have a growing influence on academia these days. At ICML 2026 in Seoul, the sponsor section took over an entire exhibition hall. Here are some impressions from walking around, with photos.

Attendees filling an aisle of the ICML 2026 sponsor hall — The sponsor hall — a crowd lining up for free hoodies...

Big Tech

Companies like OpenAI, Google, Apple, Meta, Amazon, and Microsoft had the largest footprints. They all ran things in a similar way:

Lightning talk sessions — researchers took turns presenting their work on the booth stage
Researchers on site, by team — you could walk up to a researcher in the area you care about and ask questions directly

Apparently Noam Brown also came by for a talk, but I couldn’t make it — no time. 😢

Recruiting felt like the biggest goal.

The ICML 2026 Google booth with a demo schedule covering the front of its pillar — The Google booth. The entire front of the pillar is the booth demo schedule. Microsoft's booth is visible behind it.

OpenAI booth schedule screen announcing an Agents Q&A session — OpenAI's booth schedule — sessions like Agents Q&A rotate through the day.

The ICML 2026 Meta booth — The Meta booth — Where ideas become reality.

A robotics lightning talk in progress on a booth stage — A lightning talk in progress on a booth stage.

Bonus: NAVER

NAVER also had a booth, run much like the big tech ones. Researchers from each area were around, so you could chat about whatever field you’re into.

MEET TEAM NAVER at ICML 2026 sign — NAVER's booth sign — Where AI research becomes reality.

Frontier Labs Outside the US

Frontier labs from outside the US were there too — Mistral, ByteDance, Alibaba, and Xiaomi. Mostly recruiting-focused as well.

The ICML 2026 ByteDance booth — The ByteDance booth, with demos of generative models like Seedance 2.5 Preview and Seedream 5.0 Lite.

Neoclouds, Infrastructure, and Inference

A range of neocloud companies had booths — Runpod, Together AI, Nebius, and more. Recruiting seemed to be the main goal here too.

Runpod booth — AI Infrastructure at Developer Speed — The Runpod booth. Research labs like OpenAI and Anthropic are listed as customers.

VESSL AI — a GPU cluster provider. A business where securing lots of good GPUs is what matters, positioned similarly to Lambda Labs.

VESSL AI booth — The AI Cloud for research teams — The VESSL AI booth — on-demand GPUs and dedicated clusters.

FriendliAI — inference optimization. The business model is leasing GPUs, serving models on them, and selling the API — a game of how efficiently (price, bandwidth, latency) you can serve.

FriendliAI booth — The Frontier Inference Cloud for Agents — The FriendliAI booth, advertising 2–5× faster output token speed and 50–90% lower inference cost.

A side note: this company has deep ties to vLLM, the open-source project that first comes to mind for inference optimization. FriendliAI was founded by professor Byung-Gon Chun of Seoul National University (currently on leave), building on his lab, the Software Platform Lab. Continuous batching (iteration-level scheduling) — now a standard technique in modern LLM serving engines including vLLM — was first proposed in the team’s Orca paper (OSDI 2022) (see the FriendliAI blog). Gyeong-In Yu, Orca’s first author and Chun’s former PhD student, is now the CTO — and vLLM co-creator Woosuk Kwon co-authored Nimble (NeurIPS 2020) in the same lab back when he was an undergrad at SNU.

Korean companies like HyperAccel (accelerators) and Nota AI (optimization) were there as well.

Data Collection Companies

Data companies stood out too — Scale AI, Voxel51, and Toloka.

Toloka booth — The human data layer for frontier AI — The Toloka booth, pitching RL gyms, coding (SWE-bench extensions), STEM reasoning, and Physical AI data.

Voxel51 booth — The multimodal data platform for Physical AI — The Voxel51 booth, with NVIDIA, LG, Hyundai, Microsoft, and Ford logos on the wall.

Handshake AI — a company that collects and sells data. They gather data from undergrads (math, languages, you name it) and sell it to frontier labs.

The Handshake AI booth — Advancing frontier AI with human expertise.

Other Booths That Left an Impression

General Intuition — a world-model lab spun out of the game-clip platform Medal. At the booth they were demoing MIRA, a multiplayer world model trained on Rocket League.

MIRA multiplayer world model demo at the General Intuition booth — The General Intuition booth, introducing MIRA, a multiplayer world model trained on Rocket League.

MIRA is a “playable multiplayer world model” built with Kyutai in collaboration with Epic Games. The blog post, technical report, and code are all public, and you can even play it in the browser. In short:

A 5B-parameter diffusion transformer plus a 600M video representation codec, running diffusion forcing in latent space
Trained on ~10,000 hours of Rocket League 2v2 matches collected via bot self-play — reproducing game dynamics like boosts and collisions purely through video prediction, with no physics engine or renderer
Conditions on the key inputs of up to four players at once, running in real time at 20 fps on a single B200 GPU
Despite training only on short clips, rollouts stay stable past the 5-minute measurement limit — in practice, for hours
The training/inference code is open source, along with the 1,000-match-hour Rocket Science dataset (~4,000 hours across four viewpoints) on Hugging Face

The funding story is a talking point too. They raised a $133.7M seed in October 2025 led by Khosla Ventures and General Catalyst (TechCrunch), and eight months later, in June 2026, a $320M Series A at a $2.3B valuation with participation from Jeff Bezos, Eric Schmidt, and others (TechCrunch). Medal’s action-labeled gameplay data (hundreds of millions of hours) is the core asset, but as I heard at the booth, they don’t sell the data — the plan is to sell the agent models themselves.

ElevenLabs — someone from the Scribe (STT) team was there and shared a lot.

METR — I was curious about them, but sadly no one was at the booth.

A row of smaller booths including METR, Huawei, and Nebius — The row of smaller booths further in — METR and Huawei are visible.

There were also interpretability and alignment booths like Goodfire and MATS.

Booth of interpretability research company Goodfire — The booth of Goodfire, an interpretability research company.

Booth of the AI alignment research program MATS — The booth of MATS, an AI alignment and security research program.

Quant and trading firms were out in force too — Citadel, Jane Street, Jump, and more. Chinese quant firm JoinQuant (聚宽) also had a large booth. They’re very much in the race for ML talent.

The large booth of Chinese quant firm JoinQuant — The booth of Chinese quant firm JoinQuant (聚宽) — one of the flashiest in the hall.

Jong Hyun Park