> ## Content Index
> Fetch the complete content index at: https://nexi.fund/llms.txt
> Use this file to discover other available public pages before exploring further.

# iPronics Raises $125M to Make AI’s Network Rewire Itself
- URL: https://nexi.fund/ipronics-optical-switching-2026/
- Published: 2026-09-11T10:45:48.000Z
- Updated: 2026-09-11T10:45:48.000Z
- Description: iPronics raised $125M in Series B funding co-led by Maverick Silicon and Light Street Capital, with NVIDIA joining, to scale a rack-mounted optical switch that reroutes AI traffic as light instead of electricity.
- Author: Nexi.fund Labs
- Tags: AI & Infrastructure, #mode-1, #hook-statistic, #track-E, #brand-heavy

$125 million for a networking switch. That is the size of the Series B iPronics closed on September 2, co-led by Maverick Silicon and Light Street Capital, with NVIDIA joining the round. The object of the bet is a rack-mounted box that routes data between GPUs as light instead of electricity.

🎯

AI clusters are scaling faster than the electrical switches that connect them, pushing optical circuit switching out of hyperscaler labs and into a merchant market.  
  
NVIDIA's participation in iPronics' round is the clearest signal yet that the interconnect layer, not the accelerator, is the binding constraint on AI capacity.  
  
Several photonics suppliers sell that layer now. The differentiator is programmability and cost per port, not raw switching speed. 

## The bottleneck moved from the chip to the wire

For three years the AI build-out has been told as a silicon story. More accelerators, faster accelerators, more of them per rack. The constraint that now binds sits one layer below. A cluster of tens of thousands of chips performs only as fast as the fabric connecting them, and that fabric has historically run on copper and electrical packet switching.

Copper carries a penalty that compounds with scale. It burns power, produces heat, and forces data through electrical conversion at every hop. The startup frames its goal in plain terms: networking systems able to link hundreds of thousands of computing chips inside one data center.

As we wrote in September, inference has already overtaken training as the dominant AI workload. [That earlier analysis](https://nexi.fund/inference-economy-compute-2026/) predicted the crossover would redirect capital. This round is one of the first allocations that fits the forecast. Inference is far more sensitive to network latency than training, because a serving system answers queries in sequence and stalls the moment data arrives late. The wire, not the chip, sets the floor on how fast a model responds and how many GPUs a cluster keeps busy.

The financial signal is the round itself.

$125M Series B · Sept 2026 

#### iPronics Series B round

Co-led by Maverick Silicon and Light Street Capital, with NVIDIA joining. Total funding now $177M. · *Optics.org, 2026*

The $125 million brings iPronics' total funding to $177 million. That gap is telling. The company had raised roughly $52 million before this round, so a single financing more than doubled its lifetime capital. Investors rarely commit that hard to a niche component supplier unless the niche has stopped being niche.

## Why NVIDIA wrote the check

iPronics began in 2019 as a spin-off from the iTEAM Photonic Research Laboratories at the Universitat Politècnica de València. Christian Dupont is chief executive. Dr. Daniel Pérez-López, a co-founder, is chief technology officer and the public face of the technical argument.

> With this investment, we can significantly accelerate our commercialization, delivering the scale our customers need to drive broad adoption and ensuring data centers can fully utilize their GPUs without compromising performance.— Christian Dupont, CEO, iPronics

Read that quote again and the product logic surfaces. The promise is utilization. The GPUs a customer already owns stop sitting idle waiting for data. One rack of accelerators can cost more than most companies earn in a year, and utilization decides whether a cluster is an asset or a cost center.

NVIDIA's participation turns a supplier story into a strategic one. The accelerator vendor has spent two years pushing into the layers around the chip, and the network is the layer it cannot easily own alone. A company that controls part of the interconnect becomes relevant to whoever sells the compute behind it.

The board changes reinforce the same reading. Manish Muthal, senior managing director at Maverick Silicon, joins the board. Geoffrey Tate, an advisor to Light Street Capital and the founding chief executive of Rambus, joins him. Young Sohn of Catalight Capital becomes an advisor. He was formerly corporate president and chief strategy officer at Samsung and chief executive of Inphi, and he sits on the boards of Arm and Cadence Design Systems.

#### What optical circuit switching actually does

An optical circuit switch (OCS) creates physical paths of light between endpoints instead of processing every packet in electronics. iPronics ONE is a rack-mounted optical switch built on silicon photonics, with integrated control, telemetry, and APIs. It reconfigures connectivity in real time, so a cluster matches its network to the workload in front of it. The payoff is higher GPU utilization and less energy per bit moved. The company's pitch is simple: the network stops being a fixed tax on every job.

## Google proved the concept. The question is who sells it.

Optical circuit switching gained prominence through Google, which built the technology into the networking architecture of its AI supercomputers. That proof point is real, and it is why every hyperscaler keeps a team on the problem. It is also a warning for anyone selling into the market. The largest customer has already shown it can build this on its own.

A cohort of photonics suppliers is selling into the gap that leaves. Molex introduced a high-radix optical circuit switch platform in March. Marvell and Lumentum demonstrated optical circuit switching across their respective hardware at OFC 2026\. Coherent brought its AI datacenter photonics portfolio to ECOC in September. Oriole Networks is integrating its photonic networking platform with AMD Instinct accelerators inside a UK research lab. On the component side, Quintessent raised $40 million for quantum-dot comb lasers, and Sivers Semiconductors committed $30 million to expand indium phosphide manufacturing in Glasgow.

| Parameter            | Optical circuit switching   | Electrical packet switching  |
| -------------------- | --------------------------- | ---------------------------- |
| **Switching medium** | ✔ Light path                | ✗ Electrical signal          |
| **Reconfiguration**  | ✔ Programmable in real time | ✗ Fixed topology per job     |
| **Power per bit**    | ✔ Lower at cluster scale    | ✗ Rises with conversion hops |
| **Scaling ceiling**  | ✔ Tens of thousands of GPUs | ✗ Limited by copper reach    |

Vendor specifications and OFC 2026 conference demonstrations, 2026

The pattern is a layer filling in around the GPU. Lasers, switches, co-packaged optics, and the control software that ties them together now come from a growing field of specialists. The winners will be decided by integration and cost per port, not by one dramatic breakthrough.

## The economics that pulled $177 million in

Financing a physical-layer company is not the same as financing software. It requires fabs and packaging partners, plus a qualification cycle that runs eighteen months or longer before a hyperscaler routes production traffic through a new component. Capital is the entry ticket, and iPronics now has enough of it to survive the wait.

Power per bit has moved from an engineering footnote to a procurement line item. Operators facing grid constraints and interconnection queues cannot simply add capacity at the edge of every cluster. They have to make the data path cheaper. An optical switch that raises utilization on hardware already bought is one of the few levers that pays back without waiting for new power.

That is the case Maverick Silicon and Light Street made when they co-led the round. Muthal framed optical circuit switching as critical because AI workloads have pushed traditional network infrastructure to its power and bandwidth limits. Shef Osborn of Light Street called iPronics the most scalable and cost-effective approach to solid-state optical switching at the interconnect layer.

The company is also moving toward its customers. It opened a Santa Clara office and is hiring for product strategy, customer deployments, and AI infrastructure partnerships. For a European spin-out selling to American hyperscalers, proximity is part of the product.

## What to watch

📊

**Key signals to track**  
  
Hyperscaler disclosures of optical switching in production clusters, in capex notes and technical talks.  
  
Whether NVIDIA folds a photonic switching roadmap into its own networking platforms.  
  
Cost per port on merchant optical switches as volume rises through 2027.  
  
Follow-on rounds across the component cohort: lasers, indium phosphide, and co-packaged optics. 

### Does optical switching reach the mainstream AI cluster by 2028?

🔮

**By the end of 2028, at least one top-five cloud provider will run an all-optical switching layer in a production training cluster, and merchant optical switch vendors will ship at multi-thousand-unit annual volumes. Horizon: 2028**  
  
Probability: 65% — the power math forces the transition and the technology is already proven at hyperscale, but hardware qualification cycles run long and in-house programs could absorb the demand. 

#### ✅ Arguments for

Copper is running out of headroom on power and reach exactly as cluster sizes keep climbing.  
  
Google's production use of optical circuit switching removes the technical doubt.  
  
NVIDIA's backing gives iPronics a route into the accounts that buy the most compute.  
  
**Confirmation criteria:** a named hyperscaler discloses an all-optical switching layer in a live training cluster, and merchant vendors report port shipments in the thousands. 

#### ❌ Arguments against

The biggest buyers may keep building optical switching in-house, leaving a thin merchant market.  
  
Reliability at hundreds of thousands of ports is unproven over multi-year service life.  
  
Co-packaged optics could absorb the switching function before a standalone market matures.  
  
**Disconfirmation criteria:** repeated qualification slips, or hyperscalers consolidating on integrated optics programs and skipping merchant switches. 

### Development scenarios

#### 🟢 Optimistic scenario (35%)

Two or more hyperscalers qualify merchant optical switching for production, and cost per port falls fast enough to pull in second-tier operators.  
  
**Implications:** iPronics and its peers scale into a durable component market, and utilization gains become a standard line in every AI capex calculation. 

#### 🟡 Base-case scenario (50%)

Optical switching spreads through the largest training clusters while the rest of the market stays on electrical networking for several more years.  
  
**Implications:** A real but concentrated market. The winners are decided by which vendors land the handful of hyperscale accounts. 

#### 🔴 Pessimistic scenario (15%)

Hyperscalers keep optical switching captive, merchant demand stays small, and component suppliers consolidate or get acquired before reaching scale.  
  
**Implications:** The technology wins inside a few data centers while the standalone market never justifies the capital now flowing into it. 

[ iPronics raises $125M to scale optical networking for AI data centers Trade-press account of the Series B, including the investor list, the iPronics ONE product positioning, and the company's commercialization targets. Optics.org ](https://optics.org/news/ipronics-raises-125-million-to-scale-programmable-optical-networking-for-ai-data-centers?ref=nexi.fund) 

The cleanest primary write-up of the round and what the capital is meant to buy.

[ iPronics Raises $125 Million to Scale Programmable Optical Networking for AI Data Centers The company's own announcement details the iPronics Optical Networking Engine and the board and advisor additions that came with the round. iPronics Newsroom ](https://ipronics.com/ipronics-raises-125-million-to-scale-programmable-optical-networking-for-ai-data-centers?ref=nexi.fund) 

Primary source for the technical claim: what the hardware does, not just what it raised.

[ iPronics Raises $125M Series B to Scale Optical Networking for AI Data Centers Analyst coverage that frames optical networking as becoming as critical as compute hardware for AI performance. Unite.AI ](https://www.unite.ai/ipronics-raises-125m-series-b-to-scale-optical-networking-for-ai-data-centers?ref=nexi.fund) 

Confirms the strategic read: programmable optical switching is maturing into core AI infrastructure.