Startup Spotlight: Dipole Labs builds optical switches to keep AI cluster traffic in light

Why it matters

AI infrastructure spending increasingly depends on extracting more work from installed accelerators. Dipole is betting that workload-aware optical switching can improve utilization without adding another row of GPUs.

Deepankur Thureja and Gabriele Pasquale are building Dipole Labs around an expensive piece of dead time inside AI clusters: processors waiting for the network to deliver data.

The Boston-based startup, founded in 2026 with a European base in Zurich, is developing optical circuit switches that route traffic between racks without converting the signal from light into electricity and back again at each switching point. Dipole pairs that hardware with a control layer designed to learn GPU communication patterns and reconfigure the optical connections around training and inference workloads.

That makes Dipole a co-design bet spanning photonic devices, networking hardware and cluster software. It is also a technically credible attempt to commercialize the founders' academic careers. Thureja and Pasquale met as postdoctoral researchers at Harvard after building quantum and nanophotonic devices at ETH Zurich and EPFL, respectively.

Dipole joined Y Combinator's Summer 2026 batch, which held Demo Day on September 10th. TechCrunch subsequently included Dipole among the batch's nine most frequently cited companies in conversations with early-stage investors. Cisco Investments also lists Dipole in its portfolio, although the financing amount and structure have not been published.

From quantum devices to cluster plumbing

Dipole's founders did not arrive from the usual data-center networking pipeline. Both spent their careers manipulating light and electronic states at the scale of individual materials and devices.

Thureja completed a physics doctorate at ETH Zurich, where his research focused on electrically controlling excitons, the bound electron-hole pairs that carry optical energy in semiconductors. A 2022 Nature paper led by Thureja demonstrated electrically tunable confinement of neutral excitons in two-dimensional semiconductors, reaching confinement lengths below 10 nanometers.

ETH Zurich awarded Thureja its 2024 ETH Medal for his thesis, "Electrically tunable quantum confinement of neutral excitons." He later worked at Harvard on quantum optics and photonic devices before co-founding Dipole.

Pasquale earned his doctorate in applied physics and materials science and engineering at EPFL. Harvard's Low-Dimensional Quantum Materials Laboratory welcomed him as a postdoctoral researcher on July 9th, 2024, after doctoral work on the electronic and optical properties of indium selenide.

His research produced a method for detecting spin polarization through chirality-induced tunneling currents. The resulting Nature Materials paper, published in January 2025, reported electrical detection of light's chirality in a two-dimensional material and experimental evidence for a previously predicted spin-polarized state. EPFL later recognized the work with its 2025 IBM Prize.

Those papers do not prove that Dipole can ship a reliable data-center switch. They establish that its founders have spent years building and measuring difficult optical devices, where nanoscale fabrication errors can determine whether an experiment works at all. That background matters in a market where packaging, yield, insertion loss, thermal behavior and control electronics can erase a photonic device's theoretical advantage before it reaches a rack.

The network Dipole is trying to reshape

Data already moves between many data-center components through optical fibers. Conventional network architectures still route much of that traffic through electrical packet switches, requiring conversions between optical and electrical signals. Those conversions consume power, add components and generate heat.

An optical circuit switch creates a direct light path between selected endpoints. The trade-off is that the path must be established and changed as communication patterns move across the cluster. A switch that reconfigures too slowly can struggle with short, irregular transfers. A fast device without sufficiently capable control software can spend its advantage connecting the wrong machines.

In its YC launch materials, Dipole says its switches are targeting sub-microsecond reconfiguration at large port counts. Dipole's planned control layer would observe GPU communication patterns and coordinate each optical configuration with the training or inference job running above it.

Dipole says this architecture could give a 10,000-GPU cluster the effective performance of one containing 12,500 GPUs, representing up to 25% additional useful compute from the same hardware. Dipole also models as much as $1 billion in extra annual compute capacity for a one-gigawatt data center.

Those figures are projections supplied by Dipole, rather than results from a published deployment. The model does not specify the workload mix, baseline network, current GPU utilization, optical port count or operating assumptions behind the improvement. A hyperscaler evaluating the product will want results across collective communication patterns, model sizes and failure scenarios, followed by total-cost comparisons that include the switch, fibers, control systems and integration work.

Dipole is opening design partnerships with hyperscalers and neocloud operators, alongside collaborations with photonic foundries intended to move its custom devices from prototype toward volume production. That is an invitation to help define the product, rather than evidence of a completed commercial rollout. It also reveals the founders' immediate priority: secure access to real cluster workloads before locking in an architecture that could take years and substantial capital to manufacture.

Optical switching already has production proof

Dipole is entering a category that has moved beyond laboratory demonstrations. Google has used optical circuit switching in production data-center networks for years.

A 2022 Google research paper described how optical switches and software-defined networking helped its Jupiter architecture deliver five times greater speed and capacity while reducing capital costs by 30% and power use by 41%. Google's results came from an architecture developed for its own infrastructure, so they should not be treated as a benchmark for Dipole. They do confirm that dynamic optical topology can operate at hyperscale.

The market has also begun organizing around shared interfaces. The Open Compute Project launched an optical circuit switching subproject in July 2025, with initial participants including Google, Microsoft, Nvidia, Lumentum, nEye, iPronics and Oriole Networks. The project is working on open technologies and management interfaces for optical switching in AI infrastructure.

That standardization creates an opening and a constraint for Dipole. Common interfaces could reduce the work required to place a new switch inside an existing network. They can also make the physical device easier to substitute if several vendors meet the same control and reliability requirements.

Dipole's proposed defense is tighter coordination between the switch and the workload. Its control layer is supposed to predict or learn GPU traffic and synchronize the optical topology with the job. The value will depend on how far that integration goes. Google already combines optical switching with centralized traffic and topology engineering, while other suppliers are pairing photonics with their own scheduling software. Software-directed switching is becoming an expected capability across the category.

A field with funded competitors and published specifications

Dipole will have to distinguish itself from optical networking companies that have had longer to build products, raise capital and disclose technical specifications.

nEye announced an $80 million Series C on April 14th, 2026, bringing its stated total funding to $152 million. It is developing an optical-circuit-switch-on-a-chip intended to move from laboratory development into semiconductor fabrication.

Salience Labs publishes specifications for a 32-port all-optical switch module, including claimed latency of 10 nanoseconds, reconfiguration below 300 microseconds, insertion loss below 2 decibels and power consumption below one watt per port. Oriole Networks is pursuing a broader full-stack photonic network with interface cards, photonic switches, passive routing hardware and software integrations for collective communication libraries.

These vendor specifications are not directly comparable. Latency through an established optical path differs from the time needed to reconfigure that path, and system performance depends on port density, optical loss, traffic scheduling, fault recovery and workload behavior. Still, the detail offered by established competitors sets the disclosure bar Dipole will eventually face.

Dipole has announced a sub-microsecond reconfiguration target, but it has not published port count, bandwidth, insertion loss, switching mechanism, power draw or independently measured cluster results. Its product materials refer to optical switching and compute modules, leaving the longer-term computing architecture less defined than the networking product.

The hardware roadmap will require Dipole to make decisions that software startups can postpone. The founders must choose a fabrication process, qualify foundry partners, package photonic devices, control coupling losses, design electronics and demonstrate repeatable operation across temperature and vibration ranges. Customers will expect predictable failure behavior and an upgrade path that does not strand existing network investments.

Cisco is watching the same bottleneck

Cisco Investments' presence is strategically useful because Cisco sells the switching, routing and optical systems that Dipole could complement or challenge. Cisco's portfolio page identifies Thureja and Pasquale as Dipole's leadership and categorizes the investment under silicon and optics.

The backing gives Dipole a relationship with an organization that understands data-center procurement and network qualification. It does not establish a product integration, customer deployment or distribution agreement. The investment is best read as evidence that a major networking incumbent wants exposure to Dipole's approach while the architecture remains unsettled.

For Dipole, the incentive is straightforward. A young photonics vendor needs access to customers, foundries, packaging expertise and network operators far earlier than a conventional software company needs a channel partner. Cisco can help with those connections. Cisco also gains an early view into a technology that could shift value away from electrical switching layers if optical circuit switching captures a larger portion of AI-cluster traffic.

The bet is co-design, not a standalone box

Dipole's strongest version is a network built as a coordinated system: custom photonic hardware that can reconfigure quickly, software that understands the communication structure of AI workloads and integrations that let cluster operators adopt the technology without rebuilding every layer above it.

That approach fits the founders. Thureja and Pasquale come from research environments where materials, devices, measurement systems and control software are developed together. Dipole is carrying the same method into data-center infrastructure, where the system is vastly larger and the commercial constraints are less forgiving.

Design partnerships are the correct next step because optical networking requirements are shaped by the workloads and physical layouts of a small number of large buyers. They also create concentration risk: a young hardware supplier can spend years adapting a design to one operator's architecture without reaching a repeatable product.

Dipole's founders have chosen a bottleneck with real demand and formidable incumbents. Google has proved optical circuit switching can work in production. The Open Compute Project is turning it into a shared industry program. Better-funded photonics companies are publishing products and raising the capital required to manufacture them.

Dipole now has to show that its founders' device-level expertise can produce a switch that reconfigures quickly, integrates with AI workloads and survives ordinary data-center operations. The academic credentials explain why Thureja and Pasquale are credible people to attempt it. Design partners, measured specifications and manufacturing yield will determine whether the light stays on.