Dipole Labs

[Dipole Labs @dipole_labs]

The most expensive computers ever built spend half their time waiting for data.
At Dipole Labs, we're building high-speed optical circuit switches that keep data as light. At gigawatt scale, they recover up to $1B of compute every year.
https://t.co/5ELNp4VfTe

[Dipole Labs @dipole_labs]

The most expensive computers ever built spend half their time waiting for data.
At Dipole Labs, we're building high-speed optical circuit switches that keep data as light. At gigawatt scale, they recover up to $1B of compute every year.
https://t.co/5ELNp4VfTe

[Seth Bannon @sethbannon]

“The same technology is how we’ll go after doing the computation itself with light. For now, though, back to bouncing lasers.”

Not only is this technology wildly cool, but it’s wildly valuable. It can save a 1 GW AI data center $1B every year.

How?

AI data centers are filled with extraordinarily expensive GPUs. But hey can sit 50% idle because the network can’t move data between GPUs fast enough.

Data travels between GPUs as light through fiber but whenever it reaches an electronic switch, it's converted from light into electricity, and then back into light.

Each conversion adds latency, consumes energy, and generates heat. Today’s fixed networks can’t reconfigure quickly as AI workloads change, so GPUs wait for data instead of actually computing.

@dipole_labs keeps the data as light.

They’re building software controlled, pixelated optical switches that bounce beams through open space, connecting any input fiber directly to any output fiber. No moving parts and can reconfigure hundreds of connections in less than a microsecond.

The result: ~25% more useful compute from the same GPUs. A 10,000-GPU cluster can produce the output of roughly 12,500 GPUs, without buying another 2,500.

At the scale of a 1 GW AI data center, that’s ~$1B of additional compute every year.

Dipole starts by making every GPU count. Over time, they'll make light the computer. Optical computing here we come. Congrats on coming out of stealth @d_thureja and @GabrielePasq!