The Problem: The Memory Wall Is Breaking the Economics of AI Compute

AI’s promise is unprecedented, but its economics are broken. Copper-based interconnects cannot keep up with modern models, limiting data movement between and within compute and memory. The result is high latency, inefficient hardware utilization, and unsustainable inference costs—creating a hard ceiling on AI margins.

The Solution: Grow-Style Optimization for the Interconnect

Today’s data-center compute and networking were not designed for inference. Slow, power-hungry copper interconnects, along with costly signal processing and memory overhead, drive high latency and high cost.

High-performance inference requires purpose-built hardware and a vertically optimized software stack. Piris Labs builds this optimized stack by combining proprietary optical interconnects with software designed to maximize hardware utilization. We vertically optimize the interconnect layer—just as the compute layer was vertically optimized by Groq (acquired by NVIDIA for $20 B).

Piris Labs combines proprietary optical interconnects with a purpose-built software stack designed to maximize hardware utilization, delivering 5x lower latency, 10x lower power per bit, and 2x lower cost per token

Our Launch Video
https://youtu.be/PW1PbMYw2c0

The Team

We are a team of specialists at the intersection of hardware and AI:

• Ali Khalatpour: MIT-trained engineer and Harvard–Stanford–trained optical scientist; developed the first room-temperature terahertz semiconductor laser; led the development of GUSTO’s optical engine for NASA
• Keyvan Rezaei Moghadam: AI scientist and engineering leader with experience building 0-to-1 AI infrastructure at Meta and Twitter.

Our Ask

AI Products: If you want to scale your inference workloads while slashing costs, reach out to us at contact@pirislabs.io.

Chip Makers & ODM Partners: We are also looking for a couple of additional partners who get the chance of trying our optical solution first hand. Shoot us an email at founders@pirislabs.io.