Piris Labs Is Betting the Future of AI Runs on Light

The most valuable chips in the world have a surprisingly mundane problem: they spend too much time waiting.

Modern AI accelerators can perform staggering amounts of computation, but they can only work on the data they can reach. As models have grown and thousands of chips have been connected together, moving that data has become increasingly difficult. Information has to travel constantly between processors and memory, between chips, and across racks of machines. Every trip consumes time and power.

The industry calls this the “memory wall.” Increasingly, adding more compute alone can’t solve it.

Piris Labs is attacking the problem with light.

The company is building optical networking hardware that connects accelerators, memory, servers, and racks, paired with a software stack designed specifically around distributed AI workloads. The goal is to keep expensive compute fed with data while reducing the power and latency required to move it.

It sounds like an esoteric infrastructure problem until you follow the economics. Inference happens every time someone talks to a model, generates an image, runs an AI agent, or uses a product with AI embedded inside it. As those interactions multiply, the cost of moving data becomes part of the cost of running AI.

Piris believes optical interconnects can change that equation.

When the Chips Are Faster Than the Wires

The semiconductor industry has become extraordinarily good at computation. The connections surrounding those chips haven’t advanced at the same rate.

That imbalance matters more with each generation of AI. Today’s largest models are spread across enormous pools of memory and accelerators, so performance increasingly depends on how quickly those components can communicate. A faster processor only helps if the data it needs arrives fast enough to keep it busy.

Copper eventually runs into physics. Moving more information through electrical connections consumes power and generates heat, two resources that are already becoming scarce inside modern data centers.

Light offers a different set of properties. Fiber optics transformed long-distance communication because photons can move enormous amounts of information quickly and efficiently. Piris wants to bring those advantages much closer to the compute itself.

The company is developing optical interconnects alongside a purpose-built software stack, designing the two together around the demands of distributed inference. Piris is targeting 5x lower latency, 10x lower power per bit, and 2x lower cost per task.

Those numbers matter because inference is ultimately an economics problem. The companies building AI products don’t just need better models. They need to be able to afford to run them millions or billions of times.

If Piris can meaningfully change those economics, the interconnect stops looking like a component and starts looking like critical AI infrastructure.

Two Sides of the Same Problem

Building that system requires expertise that rarely lives under one roof.

Ali Khalatpour has spent more than a decade working at the edge of optical science. He earned his Ph.D. at MIT, where he led development of NASA’s terahertz laser program and demonstrated the world’s first high-temperature semiconductor terahertz laser. He later led light-engine development at Stanford, worked on low-loss photonics at Harvard, and served as Head of Metasurface Compute at Imagia.

Keyvan Moghadam came to the same problem from the AI side. He has spent more than a decade building and scaling AI systems, including work on AI leadership teams at Meta and X, where he built AI infrastructure and engineering teams. At Piris, he leads the software stack that makes the optical hardware usable for cloud and bare-metal operators.

Put them together and the company makes sense. Ali understands what light can do at the level of physics and hardware. Keyvan understands where modern AI systems are constrained and how a new architecture has to interact with the software running above it.

Neither side is enough on its own. A breakthrough optical device that doesn’t fit the way AI systems operate is unlikely to matter. Software can’t eliminate the physical limits of the hardware underneath it.

Piris is being built across that boundary from the beginning.

Why Light, Why Now

Photonics has held enormous promise in computing for decades. AI is creating the urgency to put more of it to work.

The scale of today’s systems is exposing limitations that were easier to tolerate when workloads were smaller. Clusters keep growing, accelerators keep getting faster, and power has become one of the defining constraints on new data-center capacity. At the same time, enormous amounts of capital are being spent on AI infrastructure, creating a rare moment when fundamental assumptions about how computing systems are designed are back on the table.

The memory wall sits directly in the middle of those trends. More accelerators create more communication. Larger models require more data movement. Faster chips put more pressure on the connections feeding them. And every watt spent moving information is a watt that can’t be spent computing.

Inference makes the opportunity even larger because it turns AI infrastructure into a continuous operating cost. Training a frontier model may require an enormous burst of computation, but inference happens for the entire life of the model. Every prompt, agent action, generated image, and AI-powered feature creates another workload that has to be served.

As AI becomes embedded throughout software, reducing the cost of that workload becomes one of the industry’s most important infrastructure problems.

Why We Invested

We are drawn to infrastructure companies built around constraints that become more important as the market grows. The memory wall fits that pattern unusually well.

Every major trend in AI puts more pressure on data movement. Models are getting larger. Inference workloads are growing. Clusters are adding more accelerators. Power and cooling are becoming harder constraints. Piris is going directly at the point where those trends collide.

The team made the bet particularly compelling. Ali has spent his career pushing the boundaries of what can be done with light. Keyvan has spent his building the AI systems that increasingly need those breakthroughs. It’s a rare pairing of deep optical physics and real-world AI infrastructure experience, aimed directly at a problem that requires both.

There is a long road from a new architecture to deployment throughout the world’s AI infrastructure. Piris will have to solve difficult problems across physics, manufacturing, software, integration, and adoption.

But the first wave of the AI infrastructure buildout was defined by an insatiable demand for compute. As those systems grow, the connections between the chips are becoming just as important as the chips themselves.

For decades, light transformed how information moves around the world.

Piris is betting it can do the same thing inside the machines running AI.

Visit Piris Labs at https://www.pirislabs.io.