Today @dipole_labs is launching out of stealth. They build high-speed optical circuit switches for AI data centers.
It's quite surprising, but GPUs in a datacenter spend about half their time idle, waiting for data. Between racks data already travels as light through optical fiber. But at every switch it gets converted to electricity and converted back. That burns energy and takes time.
Dipole's switch keeps the light as light. Software-programmable mirrors steer a light beam, with no moving parts and no energy conversion. These mirrors are so fast, the network can reconfigure over 1,000x faster than the optical switches running in data centers today.
The math on the other side is simple. If you get 25% more useful compute out of the same GPUs, a 10,000-GPU cluster behaves like a 12,500-GPU one. At gigawatt scale that's about a billion dollars a year of compute you already paid for and weren't using.
Deepankur and Gabriele met as postdocs at Harvard and have spent a decade in photonics between them. Deepankur did his PhD in quantum optics at ETH Zurich, published in Nature, and won the ETH Medal for his thesis. Gabriele did his PhD at EPFL in semiconductor physics, solved a decade-old problem in condensed matter physics in Nature Materials, and won the IBM prize from the Swiss and Austrian Physical Society. This is a company where the founders personally built the skills and tech to make this product successful.
They're not waiting on the data centers to prove it. There's a smaller market that needs exactly the same programmable mirrors right now: neutral-atom quantum computers, which steer lasers to hold individual atoms in place. Dipole has over $1M in LOIs from those customers.
If you're an AI cluster or network architect at a hyperscaler or neocloud, they want to hear where today's optical hardware falls short. Same for photonic foundries who can take a custom device from prototype to scale.