tl;dr: Data centres are becoming factories whose product is the token. Producing a token efficiently involves distinct stages with very different hardware demands, yet today most stages run on the same general purpose silicon. OLIX builds its systems around a novel photonic interconnect that enables each inference stage to run on chips designed for it; enabling a step change in inference cost & performance for the most demanding frontier workloads without hard coding model architectures into silicon.
Tokens are now a unit economics problem
AI has shifted from research and experimentation into deployment. The models work, and workloads are compounding in size, volume and complexity: agents that run for hours, contexts that span entire codebases.
All of that output is billed in one unit, the token. Capabilities determine what a token is worth and what the world will pay for it. Deployment adds a second question: what does a token cost to manufacture, in time, and energy, at quadrillion scale volume? The data centre is now a factory, and it must be assessed the way factories are: on the unit economics of what it produces.
Producing a token involves stages with different requirements, yet they all run on the same general purpose chip
Producing a token means running the same sequence of recurring operations across a set of model layers. Early models varied only within layers: the attention and FFN stages had significantly different requirements but layer to layer the model stayed the same.
This is no longer the case. In the search for increased performance per unit inference cost/compute/time, model layers have become increasingly heterogeneous. A model today can contain a mixture of global and sparse attention, dense FFNs followed by MoE layers and different model dimensions across the depth of the model.
In any other industry, mass production work like this is organised as a production line: each stage is assigned a specialised machine to maximise efficiency. A car plant does not press doors and paint them with the same machine. AI data centres are organised the opposite way, ignoring two centuries of accumulated production engineering wisdom.
Instead, models run on a rack containing multiple copies of the same silicon; each one expected to execute every stage efficiently. While increases in scale up domain size over time have unlocked greater token production efficiency through superior scheduling and parallelism, the token factory still relies on one design for everything.
Why it's built this way: moving data between chips is expensive
The single machine design isn't a mistake. It's the rational response to a constraint.
Moving data between chips has historically been expensive in both energy and latency, whereas moving from one stage to the next within a chip costs orders of magnitude less on both counts. Pipelining work between specialised chips would therefore be impractically slow & inefficient, particularly for small models and models that did not have large inference volume, supporting the entire work remaining on one chip. A chip that has to execute every stage of a model has to be general purpose, like a prototyping tool, not a machine optimised for mass production.
The first specialised systems have arrived… with significant tradeoffs
Recently, specialisation has started to emerge in the token factory through coarse disaggregation. Prefill decode disaggregation has become standard at the frontier to extract scheduling efficiencies but typically reuses the same general purpose chip for both stages. Repurposing general purpose SRAM chips as decode accelerators can significantly improve latency; but because that silicon was never designed for the workload, it does not produce an overall efficiency gain.
The next stage of proposed disaggregation splits decode itself, between attention and FFN (AFD). This forces two journeys through the top of rack switch network for every model layer, imposing significant system wide latency penalties and load balance challenges on hardware that was never designed to be deployed this way. Without a fast and efficient production line, the token factory cannot extract the full benefits of specialisation.
What OLIX is building
OLIX is building the token factory the way every mature factory is built: specialised machines, a line that moves work between them cheaply, and a manager that balances the flow. This will make today's frontier AI more abundant and affordable and, more importantly, enable the viability of even more powerful models in the future.
The line. Our novel Slow & Wide photonic interconnect moves data directly from die to die at ultra low latency and low energy cost. It is made possible by rack-scale codesign of every part of the link including the SerDes and laser's atomic structure. The OLIX interconnect allows over 100,000 chips to be connected together in a multi-rack scale-up domain to run frontier models at an efficiency and speed that are not feasible on traditional systems.
The machines. With cheap movement, each stage can get its own dedicated chip for maximum efficiency. OLIX is building a suite of accelerators, one for each stage, starting with DX-1. DX-1 has been designed to deliver the best possible decode performance across every metric that matters: with simultaneously the highest interactivity, highest throughput per watt and the lowest token cost. OLIX's chips are designed with a flexible compute fabric to provide support for any model and do not hard code any particular model network architecture into silicon. Scale matters more than single token performance, and DX-1 is designed to reach GigaWatts of compute in spite of the supply chain shortages in frontier logic, memory & advanced packaging.
The line manager. OLIX's systems are fully deterministic at multi-rack scale to make coordinating thousands of chips in unison tractable. The compiler, OLIX's superhuman line manager, is aware of the time required for every operation and ensures the line and the machines are aligned. Following production manufacturing best practices, buffering work-in-progress is treated as waste and minimised to ensure tokens are produced just in time. OLIX's compiler is already today fully simulating complete models end to end with cycle accurate instruction allocation in under an hour of wall clock time to produce bit accurate simulated tokens at over 10,000 tokens per second per user.
High Speed unlocks High Throughput on SRAM
SRAM stores roughly 100x less data per unit area than HBM, so the typical objection to an SRAM-based accelerator is it has insufficient memory capacity to store the KV cache needed to serve a large number of users simultaneously, and therefore an SRAM accelerator cannot have high throughput.
However, the number of users resident in the system is not fixed, but depends on the target throughput multiplied by the time each user occupies it. The faster each user is served, the fewer need to be resident at the same time to sustain the same aggregate token throughput rate.
An extremely fast SRAM system can serve users rapidly one after the other, so that only a handful need to be resident at any one moment and high throughput is achieved without large batches.
OLIX draws this speed from three compounding sources: an interconnect fast enough that splitting a single layer across many chips remains productive, rather than being consumed by data movement; on chip latency reduction techniques borrowed from high-frequency trading; end-to-end determinism that leaves no queueing jitter to pad against. In cycle-accurate simulation, this is sufficient to close the density gap, delivering high throughput and fast individual responses at the same time.
The production line has changed many industries before
Ford's early cars were assembled like houses: the chassis stood still and the workers walked to it, carrying parts and tools. In 1913 his engineers made the work move instead, sliding it past workers who each performed one step. This simple change turned work that took hours into work that took minutes and enabled the automobile revolution.
We are building the token factory infrastructure for frontier AI. OLIX.