a supercomputer from Labs
Concordia Computer — Labs' planned multi-gigawatt supercomputer. The campus is the computer.
concordia campus · photoreal concept render
interactive concordia campus · real-time massing · drag to orbit
Concordia is a planned artificial-intelligence supercomputer: a unified computational infrastructure built to power frontier intelligence, advanced AI, physical intelligence, and Frontier Orchestration at massive scale — measured not in buildings, but in gigawatts, clusters, processors, networks, and intelligent compute.
Labs is designing the entire environment — power, network, cooling, silicon, and orchestration — as a single machine. Over time, Concordia becomes the substrate the Labs ecosystem runs on.
The industry has spent a decade describing a version of this idea — connected devices, instrumented buildings, sensors reporting telemetry back to a computer somewhere else. Compute brought around a facility. Concordia is not that, and the distinction is the entire point.
When Labs says the campus is the computer, it means the sentence literally. A wall becomes a place where computation happens. A table becomes a place where computation happens.
Chairs, walls, tables, floors and fixtures carrying processors, memory and state inside the same architecture that runs the data halls — participants in the machine, not endpoints reporting to it.
Structures, enclosures, racks and surfaces positioned, connected and reconfigured by the machines and robotics operating the site — not assembled once and left static for a decade.
Elements that energize themselves and negotiate their own draw, so a campus that computes is also a campus that can rearrange its own computation.
Labs Machines automates the facility, Labsbotics operates inside it, Factory5227 builds the hardware, and Concordia runs the compute — until the line between facility and machine stops being useful.
If a campus can be a computer, then what compute is becomes an open research frontier rather than a settled definition. Most computation today resolves to text, image, audio, video and control. Labs treats that list as incomplete.
Computation whose output is experienced rather than read — captured, represented, transmitted across a network and reconstructed on a device.
An initiative being explored within YuhmmyAI, asking whether taste and its surrounding sensory dimensions can be computed and carried over a network — so a person experiences a sensory element remotely rather than only reading a description of one.
Perception, manipulation, navigation and force — modalities that do not reduce cleanly to tokens, and that Physical Intelligence already pushes computation into.
Representations that do not yet exist, models trained on non-standard modalities, latency closer to interaction than to batch inference, and hardware willing to treat a sensory signal as a first-class object.
Concordia is being designed to host that exploration rather than foreclose it. An environment built only for today's modalities would be obsolete against the frontier it is meant to serve.
Concordia takes physical form as branded, purpose-built facilities — power-dense, liquid-cooled, and secured. The visible face of the campus that is the computer.
concordia · facility · photoreal concept render
interactive concordia facility · real-time massing · drag to orbit
A simulated interior — rank after rank of accelerated compute, cooled and orchestrated as one. Move through it.
concordia data hall · photoreal concept render
interactive · concordia data hall · fly-through
Concordia is engineered as one machine — from the campus to the rack. A Concordia server: dense accelerated compute, sleeved and liquid-cooled, built to be racked by the thousand.
nx server · photoreal concept render
interactive · concordia server · turntable
Concordia is designed around its own silicon — the Nx family. Two coupled compute classes, one memory fabric, one compiler — built for the movement, memory, and computation of intelligence, not for any single model.
nx accelerator · photoreal concept render
concordia · nx accelerator · design concept
The throughput class — training, prefill, and large-batch inference. Tensor-dense and HBM-rich, with collectives and expert routing native to the die.
The latency class — decode, agent loops, and interactive reasoning. SRAM-forward, with hardware-native state and speculative execution.
A hybrid HBM/SRAM memory that treats live inference state — KV pages, prefixes, agent context — as a first-class object.
Expert routing, KV transfer, and collectives as native operations. Deterministic islands inside; a dynamic scheduler outside.
Nx silicon racked at gigawatt scale — models, agents, memory, and tools orchestrated as one workload.
Concordia's first-generation Nx silicon — designed on Concordia's own workloads, and re-designed by them.
Phase-specialized silicon — NxP for throughput, NxU for latency — coupled over one memory fabric. The developer writes a model; the runtime places the work.
Concordia is the environment where each Nx generation is designed. Real workloads become the blueprint for the next chip.
Concordia is built to close the recursive self-improvement loop. The intelligence running on Nx silicon is pointed back at the silicon: profiling its own kernels, proposing architectures, and verifying them in simulation — so each generation shortens the distance to the next.
Models explore the architecture space and generate candidate Nx designs.
Concordia trains the next frontier model on the silicon the last loop produced.
Every kernel, cluster, and watt is instrumented — the machine profiles itself.
Findings return to the design step as constraints, and the loop closes tighter.
Most of the industry waits on an outside vendor cycle between each turn of that loop. Concordia runs every step of it inside one facility, so improvement compounds instead of queuing.
Each Frontier Lab pursues a different frontier, and each Frontier Intelligence a different expression of intelligence within it. Left alone, that produces 19 separate computational programs and 19 incompatible ways of describing the same primitives. Concordia is the layer that prevents it.
Models, agents, tools, memory and state addressable across shared fabric — rather than across bespoke pipelines built one pair at a time.
The same orchestration semantics, state formats, runtime contracts and accounting for compute, so a perception system built for one frontier is addressable by an agent built for another without translation.
A supercompany needs what a collection of companies does not: not merely shared services, but a shared machine.
Concordia's most distinctive workload is neither training nor inference. It is orchestration — the continuous coordination of many models, agents, tools, memories and states toward outcomes, at scale. Labs treats it as a discipline to be industrialized rather than a feature to be shipped.
Orchestration is becoming the difficult part of applied intelligence. Not obtaining a model — conducting many of them together, reliably, economically and accountably.
Concordia will be built, supplied, and scaled with the world's leading infrastructure partners — bringing frontier silicon, power, networking, cooling, and construction into one facility, and adding AI infrastructure as the frontier advances.
Concordia is a planned system. Partnerships and infrastructure are being explored and will be announced as they are confirmed.
Concordia Computer — Labs' planned multi-gigawatt supercomputer — runs as a standalone supercomputer lab, designed, built, and integrated under Labsintuition's Physical Intelligence architecture.
The building isn’t housing the computer. The campus isn’t housing the computer. Together, they are the computer.
A unified computational infrastructure designed to power frontier intelligence, advanced AI, physical intelligence, and Frontier Orchestration at massive scale — measured in gigawatts, clusters, processors, networks, and intelligent compute. Over time, Concordia will also serve as a development and deployment environment for Labs’ proprietary compute architecture.
A planned build-out — breaking ground in 2030, reaching full campus scale around 2032.
Site works begin. First Concordia Campus foundations, multi-GW power intake, and network trenching.
Initial Concordia Clusters and a Frontier Orchestration Cluster energize; the network fabric comes online.
The campus reaches planned multi-gigawatt scale; C2I unifies compute, network, and power as one system.
Concordia Compute opens capacity externally; NxUs and NxPs move toward production for orchestration intelligence.
Concordia Panels are an envisioned line of solar-powered intelligent panels — generation, storage and load management designed as computational devices, rather than electrical hardware with software attached afterward.
At multi-gigawatt scale, energy decisions and compute decisions are the same decision. Scheduling a workload is scheduling power. Placing a cluster is placing load. So the point where energy enters the system has to be addressable by the same layer that places a model or an agent.
Concordia is planned with Labs Impact and the Labsintuition Foundation from the first line of the site plan — with the goal of sustainably and cleanly protecting the environment it operates in, and of putting frontier hardware to work on that goal across every Labs ecosystem.
The first Concordia campus is in siting — a city partner is still being decided. Environmental and community commitments are being written into that selection, not added after it.
Labs holds a specific view of where this goes: that open-source and open-weights artificial intelligence will define the most significant era of the field so far, and that the arrival of AGI is more likely to happen in the open than behind a single closed door.
If frontier weights become broadly available, the model stops being the durable advantage. What stays scarce is everything underneath it — compute at scale, energy, orchestration, memory, state, the fabric between them, and the capability to run the whole arrangement reliably for years.
From silicon to state to fabric to facility to power — so the system is valuable enough that others would rather use it than reproduce it.
An exploration of an exchange layer where compute capacity — measured, scheduled and accounted for through C2I — could be contributed, drawn, allocated and exchanged, rather than only rented from one provider on fixed terms.
Capacity is never consumed uniformly. Workloads surge and idle; some hold reserved capacity they don't use while others need capacity they don't hold.
Labs expects the AGI wave to be constrained by compute long before it is constrained by ideas — and participation in frontier compute should reach past those able to build campuses.
Concordia Panels and CompEx are at the research and exploration stage. No exchange has launched, no capacity is currently offered externally, and nothing described here is available for purchase.
Generative AI changed how much compute the world consumes and how it consumes it — increasingly continuous, distributed and incremental. At the same time robotics, autonomous systems, intelligent hardware, vehicles, machines and connected environments are expanding the population of things that require compute at all, creating a layer of physical demand that did not previously exist.
The Internet of Things connected the physical world. The Computeconomy makes that world computationally and economically active — devices, machines and facilities becoming consumers and contributors of compute.
Demand pushes compute across devices, facilities, regions, edge systems and terrestrial data centers — and in time toward orbital infrastructure.
As capacity is built along that gradient, location stops being a deployment detail. Proximity, connectivity, energy, availability, capability and demand begin to decide where a workload runs and what capacity is worth.
Labs takes the position that computational capacity can be individually owned, coordinated, exchanged and monetized wherever it exists — becoming as fundamental to the global economy as capital.
Alongside Subscriptions, Advertising and Transactions, Exchanges emerge as a distinct commercial model — continuously pricing, coordinating and exchanging distributed computational capacity.
Concordia Panels describe generation at the edge of the system. Supply describes what feeds it — and supply arrives in two forms.
Compute already present in devices in circulation — phones, laptops, vehicles, appliances, wearables, and eventually almost anything with processing capacity. It requires no acquisition, and for most of its life it sits idle.
Additional physical compute someone acquires and introduces into that environment when they want more of it — hardware bought to contribute, not only to consume.
A solar panel cannot enter the energy economy until it is built and installed. Compute is not bound that way: capacity can be contributed incrementally from hardware that already exists, without waiting for purpose-built infrastructure to be financed and constructed.
The Computeconomy, Native and Added Compute are positions Labs is researching, not products. No exchange has launched, no capacity is currently offered externally, and nothing described here is available for purchase.
Concordia runs as a standalone supercomputer lab — but it is designed, built, and operated with the combined capabilities of Labsintuition, the operating lab and systems integrator for Physical Intelligence.
Integrated by Labsintuition — the operating lab & systems integrator for Physical Intelligence.