One Third of This Museum Is a Data Center

The visit starts on an escalator, descending into a mirrored room where light moves like it's alive, shifting from cyber-graphic static to something closer to mycelium, then to saturated, moving forest. A neck-worn device releases a rotating scent. A wristband tracks heart rate and skin conductivity as the room responds. All around are people looking up and around and down. Some sit. Some lean against a wall, trying to find their bearings. It is kaleidoscopic and somewhat overwhelming. Nothing here is prerecorded in the way a film is prerecorded. It is being generated, live, by a model trained on more than a billion data points pulled from rainforests, museum archives, and sixteen live environmental sensors still feeding it new information right now.

"This is a laboratory of imagination," DATALAND’S creator Refik Anadol says. "We have walls, but no limits in the ideation." 

The premise is meant to run in both directions."For 5,000 years, humans have been emotionally moved by artworks, but the relationship has always flowed in one direction.”

While developing this project, Anadol, an artist who made his name by using AI models to generate his pieces, asked if it was possible for those artworks to feel us back.“We dreamed of a place where audience and artwork could merge, creating a feedback loop of collective emotion. Through collaborations with extraordinary innovators, and technologies woven directly into the architecture itself, that dream has now become reality."

That is DATALAND, Refik Anadol Studio's self-described "world's first Museum of AI Arts."

It opened on June 20, 2026, inside The Grand LA, the downtown Los Angeles development Frank Gehry designed shortly before his death in December 2025. The building draws obvious comparisons to teamLab's immersive spaces in Tokyo and Las Vegas's Sphere. The more useful comparison is a data center. By some accounts, roughly a third of DATALAND's 25,000 square feet is occupied by server rooms, cooling equipment, and networking infrastructure, rather than gallery space. 

Gensler designed the building. Arup engineered it -– structural, mechanical, electrical, plumbing, IT, acoustics, lighting, and security all under one roof. Neither firm was building a museum in the traditional sense. They were building a machine. DATALAND spreads that machine across five galleries. 

Visitors clip on a biometric wristband and a scent necklace at the entrance, then move into the main hall, where 84 Epson projectors and floor-to-ceiling MicroLED walls carry the inaugural exhibition, ‘Machine Dreams: Rainforest’. ‘The Latent Gallery’ lets visitors page through the underlying AI model's data on transparent touch panels, while the mirrored ‘Infinity Room’, evolved from Anadol's original 2014 UCLA installation, runs on generative AI that simulates real-world physics. An LED Cube adds motion, tilting the floor as it layers in rapid sound. Every room feeds visitor biometric data into the final space, the Sanctuary, which turns heart rate and skin conductivity into a live, evolving data sculpture.

DATALAND's compute stack has two distinct layers. An on-site edge computing facility in the building's basement is dedicated to live calculation. It runs on roughly 130 Nvidia Blackwell chips and around 150 computers total, valued at close to US$5 million. That's the hardware turning a generative model's output into a video signal that 84 synchronized Epson projectors and floor-to-ceiling Absen MicroLED walls can actually display, in real time. It skips the latency that a round trip to a cloud data center would introduce.

Separately, Anadol Studio's underlying AI system, the Large Nature Model, runs on Google Cloud infrastructure in Oregon. The studio says the facility runs on 87 percent carbon-free energy. That's the layer handling the heavier lift: training and updating a model built on more than a billion ecological data points and 500 million images from partners including the Smithsonian, Getty, and Cornell's Lab of Ornithology. It also draws a live feed from sixteen rainforest sensors.

Nate McWilliams, Refik Anadol Studio's director of technical system design and production, has an unusual vantage point on that split. He was Arup's audio-visual consultant on DATALAND, working the project from the engineering side, before the studio hired him away after the installation's success, according to a spokesperson for Refik Anadol Studio. Before Arup, McWilliams spent nine years as an audio-video engineer at Walt Disney Imagineering, integrating the show systems behind Disney's theme-park attractions.

Even with that inside view, he won't put a number on exactly how the workload splits between the two. "We don't share the split between on-prem and cloud, but we do have a substantial compute on-prem," he says. "Both on-prem and cloud systems have strengths and weaknesses, and I think we have done a decent job of making use of those. In the end, both are required to make a place like DATALAND work."

McWilliams says the hardest part of pulling DATALAND together was coordinating people, not the technology itself.Construction crews and media teams were competing for the same physical space on a short timeline, and their needs regularly collided. "If we needed to build a scaffold to do 'X,' it might mean the scaffold blocked the visibility for the media teams to do their work," he says. "It required daily communication and a heavy site presence for all team members."

The second half of that coordination problem was internal. DATALAND's deployment team drew from construction, software development, AV integration, systems design, and media design, and each discipline solved problems in its own idiom. AV observability alone, McWilliams points out, can run through a control system like AMX or Crestron, custom software, a node-based programming language, a show controller, or a logging analysis platform. His team had people who could build any of the five. 

"We had a lot of options for how we might tackle problems, with each team planning it in a way that was native to their own paradigms as an artist, developer, or engineer.”

Part of his job was making sure teams did not duplicate each other's work.That mix changed how McWilliams thinks about AV systems design. Proprietary control platforms like QSYS are a familiar presence on integration teams, while bespoke developers writing custom code for a single installation are not. "I thought it was a major step forward to write bespoke software for our own needs," he says.

 That decision shows up everywhere in how the building runs, from its rendering software to how its displays stay in sync.

DATALAND's real-time visuals run on VVVV and Unreal Engine, McWilliams says, not a proprietary black box built from scratch. The more consequential build is the layer wrapped around them. "We have a totally custom media pipeline for audio and video … Our media server system is largely designed and developed in-house, and is something that, I personally think as a system designer, I can appreciate as a work of art in and of itself."

That pipeline isn't custom for its own sake. McWilliams says the resolutions, color spaces, and bit depths DATALAND works in at points along that pipeline are pushing past what current software and industry standards are built to handle. That forces his team to solve problems most AV integrators never encounter on a single install.

The display and audio integration is where DATALAND reads less like art and more like a systems integration case study. Beyond the 84 Epson projectors generating more than 700 square meters of projected pixels, the museum runs Absen MicroLED walls floor to ceiling in multiple galleries and transparent OLED panels for interactive surfaces. The mix lets the same generative output move seamlessly from projected surface to solid display without visibly changing medium.

Audio runs on roughly 250 L-Acoustics loudspeakers across the building's galleries, tied together through L-ISA, L-Acoustics' spatial audio platform. Signal moves over Milan-AVB and AES67, the same networked-audio protocols increasingly standard in large-scale AV integration. The museum's Sanctuary space also carries the first permanent U.S. installation of L-Acoustics Ambiance, a system that actively shapes room reverberation in real time rather than relying on fixed acoustic treatment. Solotech, an L-Acoustics certified integration partner, handled the install.

Keeping all of that synchronized to a model generating new output continuously, rather than looping a fixed program, required McWilliams's team to build their own tools.It started with scale. DATALAND needed a rendering cluster larger than its manufacturer's supported configuration, so the studio built its own way to keep the machines aligned. Now, the moment a new machine is recognized on the network with a role assigned to it, a sequence kicks off automatically. The machine gets access to the shared data systems for real-time rendering, finds the frame it should be working on, and starts rendering. 

Because DATALAND runs on a video media fabric rather than fixed cabling, repatching a new machine into the video canvas takes seconds, not a site visit. Underneath that custom logic sits more familiar broadcast fundamentals: frame edge reference to keep video clocks aligned across machines, and positional reference so every system agrees on where it is in time. Audio stays locked in with regular re-triggering, tight enough that clocks never drift far enough to notice. McWilliams said the team has contingency tools ready to link audio rendering directly into the compute cluster if that light-touch approach ever stops being enough. None of it requires physical repatching, for audio, video, sync, or data, when machines change roles, or a spare gets swapped in.

DATALAND's sustainability pitch has centered on the cloud side of its infrastructure. A Google spokesperson has said the Large Nature Model uses roughly one smartphone charge's worth of energy to regenerate the artwork per visitor, running on Oregon's low-carbon grid, according to Hyperallergic.What's missing is a parallel figure for the building itself: the power and cooling load of an on-site facility running roughly 130 GPUs continuously. The building was never designed to be a data center. McWilliams closes that gap. 

"The power demand on-prem is not substantially different from any other world-class entertainment or multimedia immersive attraction with custom graphics."A GPU running one of our models versus rendering our complex graphics is not substantially different, and most of our GPUs aren't maxed out."

DATALAND's compute racks draw two or three 30-amp, 120-volt circuits each, below what McWilliams has seen in broadcast, data centers, and other immersive installations. He says rack density specs for particularly demanding deployments run above 20 kilowatts but nearly all of DATALAND's racks run below seven, and most operate under five. The racks themselves are unusually tall, 58 rack units against the more typical 44, spreading that load out rather than stacking it high.

Cooling is conventional. "Air-cooled via our HVAC system,” McWilliams says. “Our server room is well constructed but relatively typical. We have redundant CRAC cooling systems, hot aisle, cold aisle, etc. Nothing special here." No exotic liquid cooling, no engineering feat beyond what a well-built server room already does. By McWilliams's account, the mechanics underneath DATALAND's scale are ordinary.

DATALAND wasn't built with the redundancy required for true 100 percent uptime, McWilliams says, and the studio made that tradeoff on purpose. "Getting those extra few percent is so much more expensive, so we adopted the strategy of having cold spares or hot spares on hand.” 

Tech ops staff stay on-site to troubleshoot failures and swap devices into new roles on a modular system built for fast swaps.The building's data logging systems track dropped frames from every machine in the cluster. McWilliams's team has used that data more than once to catch a failing machine, pull it from the rendering cluster, and get it fixed before it caused a visible problem. The data set is not deep yet, he said, but the studio is already building toward predictive maintenance, catching a failure pattern before it becomes an outage.

A full power loss mid-show is the scenario McWilliams treats most seriously, and the one the building is least equipped to shrug off. "Running an experience like this on battery backup is not practical," he says. Some critical systems carry enough battery runtime for a graceful shutdown, but a full recovery still takes time to reboot. However, he says downtown Los Angeles' power service has been reliable throughout the project, and the studio has built automated control routines to make a restart as painless as possible. But if the power actually goes out mid-show, the fix is a busy night for the on-site tech team, not a seamless failover.

DATALAND won't stay unusual for long. It's just early. As more cultural institutions chase live, generative, GPU-hungry installations instead of fixed media, the questions Gensler and Arup had to answer stop being edge cases and start being a design brief. How much floor space compute actually needs. How you cool it. How you keep 84 projectors and 250 speakers synchronized to a model that never stops generating.

Shane Myrbeck, Arup's project director on the build, described the process as requiring "extraordinarily close coordination of technology, engineering, and architecture from the outset." The work, as he framed it, meant making a largely invisible system "technically harmonious, reliable, and buildable" around an artistic vision that kept evolving as the technology did.

Whether DATALAND's specific approach travels is an open question, even inside the studio. 

"While this isn't exactly site-specific, it is heavily customized for our particular video and audio canvases," McWilliams says. 

Adapting the studio's current exhibition, ‘Machine Dreams: Rainforest,’ to another building might be possible. But McWilliams suggests the more interesting question is how long a visitor has before DATALAND's own exhibition changes into something else entirely. The museum is built to keep evolving, not to stay fixed the way its walls suggest it should.

Walk back up that escalator, and the building looks ordinary again from the street. It isn't. A third of it is downstairs, running. It's kept alive by a team that decided the fastest way to build a museum like this was to write their own software first and ask the AV industry's usual playbook second.