The Los Angeles Convention Center opened its doors on Sunday morning for the 53rd SIGGRAPH, the first time the conference has returned to the city since 2023. By the time it closed on Thursday it had drawn more than 9,500 people from 69 countries to a show floor that ran the full range from single-room startups to the largest names in visual computing.
It also ran, for its final two days, directly opposite AMD’s Advancing AI event 380 miles up the coast in San Francisco. Anyone trying to cover both weeks of AI infrastructure news had to choose. That collision is worth noting, because the two events were closer in subject matter than the badge design suggested.
Sunday: the Games Summit
The conference opened with something it had never had before. The Games Summit is a full one-day track dedicated to game development, and Games Chair Emily Hsu framed it as an entry point for practitioners who have always been part of the SIGGRAPH community on the technical side without ever having a dedicated home track.
The gap was real. The “Advances in Real-Time Rendering in Games” course has run at SIGGRAPH for years, and a substantial share of the Technical Papers program has always been immediately applicable to game engines. What was missing was a room where the audience assumption was games first rather than film first. That room now exists, and the fact that it was scheduled on opening day rather than buried midweek says something about where the organizers think the growth is.
The shift toward the pipeline
Opening day set a tone that held through the week. Rendering and simulation advances remained central, but the center of gravity has been drifting for several years now toward production workflows, tool interoperability, and what emerging technology actually does to a working pipeline.
This is what a mature field looks like. The question in the Production Sessions was less often “can this be rendered” and more often “can this be rendered inside an existing facility’s toolchain without rebuilding it.” Los Angeles supplied a disproportionate share of the people asking, with contributors drawn from Lightstorm Entertainment, Paramount Animation, Sony Pictures Animation, Sony Pictures Imageworks, Netflix and its Eyeline Labs, Magnopus, Electronic Arts and Activision, alongside researchers from UCLA and USC. California’s creative industries account for 15.4 percent of the state’s total employment once indirect and induced jobs are counted, and the conference leaned into that.
Lanny Smoot took the stage to walk through a career of inventions built across 27 years at Walt Disney Imagineering R&D and two decades of research before that. It was the least AI-focused session of the week and, by several accounts, one of the most valuable.
Tuesday: NVIDIA takes the day
NVIDIA did not take a keynote slot so much as take a day. Physical AI Day was anchored by Ming-Yu Liu’s session on world models as the frontier of physical AI, unlocked by graphics — a framing that argues the rendering pipeline is not adjacent to AI research but the substrate underneath it.
The rest of the day followed the thesis. Aaron Luk carried OpenUSD into AI environments. Jonathan Stephens of Lightwheel and Steven Ren of Palatial worked through simulation-ready scene construction, the unglamorous problem of building environments that a robot policy can actually train inside. Miles Macklin introduced Newton, NVIDIA’s open-source physics engine, which underpinned the Physical AI Day program itself.
The sponsored keynote pulled the threads together, with Neil Ashton, Ming-Yu Liu and Edward Liu covering neural rendering, world models, and simulation methods built for AI by AI. Edward Liu runs applied deep learning research and the work behind the DLSS suite, which is the clearest existing proof of the underlying argument: the highest-volume neural rendering deployment in the world already ships inside a consumer graphics driver.
Tuesday afternoon: the Zeus bet
At 2:30, Bolt Graphics founder Darwesh Singh took the largest public stage the company has had.
His argument is that real-time path tracing and physics simulation are the actual goal of computer graphics, that rasterization-derived GPU architectures are the wrong shape for reaching it, and that Bolt’s Zeus design is built for that target from the ground up. The published performance claims are striking: ten times the rendering speed, six times the FP64 performance, and three hundred times the electromagnetic wave simulation speed of legacy GPUs.
Those are company numbers. No third party has verified them. Zeus completed a test chip tape-out as of April, and full delivery is targeted for the fourth quarter of 2027, which puts roughly eighteen months between the claims and any possibility of independent measurement. Standing on a SIGGRAPH stage to make an architectural argument against the incumbent, on a day the incumbent had booked the entire schedule, was either good positioning or very good nerve.
Wednesday: the Teapot Test
Tripo AI closed the sponsored keynote series on Wednesday afternoon with a session built around a deliberately chosen reference. The Utah teapot has been the standard object for testing whether a system can render since 1975; the pitch was that the field has moved from testing whether a system can show something to testing whether it can make something. Generative 3D, full-stack, running into world models.
Tripo is headquartered in Beijing. A Chinese generative 3D platform holding a sponsored keynote slot at the American computer graphics conference, in the current export-control environment, is a data point that will be read differently depending on who is doing the reading. The research on generative 3D and world models is worth taking seriously on its merits. The platform is worth evaluating carefully before it enters a production pipeline. Both things are true at once.
What the numbers say
More than 9,500 attendees from 69 countries is a strong showing, and well ahead of the roughly 8,800 who came to Denver in 2024. It is also meaningfully below the 12,400-plus that turned out for Vancouver in 2025. The return to Los Angeles did not reproduce that spike, which is worth watching rather than over-reading after a single year.
The official framing of the week was that AI ran through nearly every program at the conference, positioned as a creative partner rather than a replacement. That is the diplomatic version, and it was the honest one in most of the rooms. But AI was no longer a track at SIGGRAPH 2026. It was the load-bearing assumption underneath the papers, the production sessions, the hardware pitches, and the day NVIDIA bought outright.