PsiQuantum, a quantum computing company whose first breakthrough is in AI data centers, says it has developed a device that routes light signals without converting them to electricity, allowing separate server racks to behave as one at speed-of-light latency with almost no heat and negligible energy use.
The AI boom depends on networking hundreds of thousands of AI processors together, but that networking is also the data center's weakest link. GPUs can crunch numbers at high speed yet are bottlenecked by the comparatively slow copper that carries the output from GPU to GPU, especially while training increasingly large AI models that require coordination between many processors.
Newer data center racks more often use optical cables to carry data between GPUs. But optical information must be converted into an electrical signal to be routed to its destination, converted back into light, and then converted once more when it arrives. In the largest data centers those conversions happen billions of times per second and require optical switches that generate heat, energy and bulk.
PsiQuantum says its new device can route the signal without ever converting photons into electrical signals, potentially creating speed-of-light connections between server racks that generate almost no heat and consume negligible energy.
Robert Soderbery, executive vice president of PsiQuantum, told Semafor that separate data center racks would effectively behave as one using the device, with data flowing between GPUs in separate racks at the same speed and latency as between GPUs interconnected inside a single rack. That matters because server racks are reaching the limit of compute density, size and energy efficiency.
PsiQuantum is not alone. The battle to convert data center networking to photonics has been raging since the ChatGPT moment kicked off the AI race. Earlier this year, photonic computing company Lightmatter helped launch an initiative to standardize optics in the data center space, and Nvidia has been spending billions on investments in optics companies such as Lumentum and Coherent.
The only major hyperscaler currently using purely optical circuit switching in an AI data center today is Google, which built a system that uses tiny mirrors known as MEMS to direct light signals. For now, the GPU still reigns supreme, but data centers are seeing a lot more of the light.
Analysis
Why This Matters
- Networking is the data center's weakest link in AI training, and the device removes repeated conversions between light and electricity that generate heat, energy and bulk.
- If separate server racks can behave as one, operators could push past current limits on compute density, size and energy efficiency.
- The claim puts PsiQuantum into an intensifying race to shift data center networking to photonics, alongside Google's MEMS approach, Lightmatter's standardization push and Nvidia's investments in optics companies.
Background
The AI boom rests on networking hundreds of thousands of processors together, and during training of increasingly large models the comparatively slow copper carrying data between GPUs becomes the constraint. Newer data center racks increasingly use optical cables, but light signals must be converted into electricity to be routed, converted back into light, and converted again at their destination, with those conversions happening billions of times per second in the largest facilities. PsiQuantum, a quantum computing company, says its new device performs the routing without converting photons to electrical signals. Google, described as the only major hyperscaler using purely optical circuit switching in an AI data center today, steers light signals with tiny mirrors known as MEMS.
What to Watch
- Whether the device moves from announcement to working deployment, where the claim that separate racks behave as one will be tested in real data centers.
- Progress on the industry standardization initiative Lightmatter helped launch earlier this year.
- What Nvidia does with its multibillion-dollar investments in optics companies such as Lumentum and Coherent.
Checked
12 claims in this story checked against 1 source: 11 supported.
Show each claim
- SupportedThe headline says PsiQuantum’s photonic device could ease the data center’s weakest link.Semafor: “But all that networking is also the data center’s weakest link. ... PsiQuantum says its new device can do that without ever having to convert the photons into electrical signals. It could be used to create speed-of-light connections between server racks in AI data centers”
- SupportedThe summary says the company claims separate racks would behave as one.Semafor: “Robert Soderbery, executive vice president of PsiQuantum, said in an interview with Semafor that separate data center racks will effectively behave as one, using PsiQuantum’s device.”
- SupportedThe summary says data would move between GPUs at the speed of light.Semafor: “It could be used to create speed-of-light connections between server racks in AI data centers”
- SupportedThe summary says the connections would generate almost no heat.Semafor: “It could be used to create speed-of-light connections between server racks in AI data centers that generate almost no heat and consume negligible energy.”
- SupportedThe opening describes PsiQuantum’s first breakthrough as being in AI data centers.Semafor: “Quantum computing company’s first breakthrough is in AI data centers.”
- SupportedThe opening says the device routes light signals without converting photons into electrical signals.Semafor: “PsiQuantum says its new device can do that without ever having to convert the photons into electrical signals.”
- Partly supportedThe opening describes the connection as having “speed-of-light latency.”Semafor: “It could be used to create speed-of-light connections between server racks in AI data centers”
- SupportedThe opening says the data center’s networking is a bottleneck for the AI boom, which depends on networking hundreds of thousands of processors.Semafor: “The AI boom wouldn’t have happened without the ability to network hundreds of thousands of AI processors together. But all that networking is also the data center’s weakest link.”
- SupportedThe story says comparatively slow copper bottlenecks GPUs, particularly when training increasingly large AI models that require coordination among many processors.Semafor: “While GPUs can crunch numbers at blazing speeds, they are bottlenecked by the comparatively slow copper that carries the output of that math from GPU to GPU. That’s especially true during the training process of increasingly large AI models that require the most coordination between large numbers of GPUs.”
- SupportedThe story says optical signals undergo repeated conversions, and that in the largest data centers this happens billions of times per second.Semafor: “But optical information must be converted into an electrical signal in order to be routed to its destination, then converted back into light, before it’s converted once again when it reaches its destination. All those conversions happening billions of times per second in the largest data centers”
- SupportedThe story says Google is the only major hyperscaler currently using purely optical circuit switching in an AI data center.Semafor: “The only major hyperscaler currently using purely optical circuit switching in an AI data center today is Google”
- SupportedThe story says Nvidia has been spending billions on investments in optics companies including Lumentum and Coherent.Semafor: “Nvidia has been spending billions on investments in optics computing companies like Lumentum and Coherent.”
