Datacenters have been around since the 1940s, starting with the US military’s Electrical Numerical Integrator and Computer (ENIAC) in 1945 at the University of Pennsylvania. Google launched the first hyperscale data center in 2006, in the state of Oregon. It currently occupies 1.3 million square feet of space and employs a staff of ~200. Prior to that, in 1998 after the company was founded, it built its own racks and stored these in a leased facility in Santa Clara, CA.
Over the past 5 years, datacenter investments have exploded, propelled by the needs for ever-increasing data and compute capabilities to feed AI inference engines, and the growth of cloud computing and storage. Currently, the number of datacenters is ~12,000, spread across 179 countries, with 40% located in the United States. Most of these are owned or leased by corporations, and typically measure ~100,000 ft² in size. Hyperscale datacenters are significantly larger (10M ft², housing > 5000 servers), numbered ~1,100 in 2024 and belong to companies like Google, Amazon, Microsoft and Meta. An additional ~800 such datacenters are in the planning pipeline. The United States hosts ~50% of these, with other dominant global players like Alibaba, Tencent, and TikTok owner, ByteDance located in China and other regions. The capex investments are staggering – according to a recent Financial Times report, capital spending on datacenter construction by Google, Amazon, Microsoft and Meta from 2023 to the end of June 2026 hit $1.1Tn. It is expected to grow to ~$7 Tn by 2030, driven primarily by the growth of AI services. Hyper-scalers like Alphabet, Amazon, Meta and Microsoft have raised over $150 billion combined, by selling debt and equity to build datacenters and fuel the development of AI models and agents. Increasingly, AI infrastructure is being viewed as a new asset class by Wall Street and private equity players who view these systems and infrastructure as profitable investments, that can be securitized and sold to investors.
A modern datacenter consumes > 100 MW of power, can store and manage few hundred terabytes in a smaller facility to several exabytes (1 exabyte = 1 million terabytes) in a large one. A typical hyper-scale datacenter (think AWS, Microsoft, Google) has at least 5,000 servers managing this flow of data, and uses 100-500 acres (1 acre ~ 4,000 m² or 43,000 ft², and space scales with energy consumption). It also switches up to trillions of data packets across its network of servers, storage systems, and accelerated AI workloads. Reducing energy consumption, cooling needs, latency and space for every piece of the operation is critical – for cost, performance, sustainability and public acceptance.
Construction costs for a hyperscale datacenter (Figure 2) runs between ~$1-10B depending on the size (100 MW- 1 GW). This includes design, permitting, construction and installation of electrical, mechanical and cooling systems. The costs for the actual IT hardware – servers, networking cables, silicon chips, optical transmission and switching components, and GPUs (Graphical Processing Units) can be 2-3X this amount. Annual operating expenses for a 100 MW datacenter can run between $500M – $1B (including IT equipment amortization), with electrical energy accounting for ~10%. Reducing power consumption is critical – for reducing operating costs, reducing cooling and power needs, lowering pollution and gaining community acceptance.
Within a datacenter, managing the flow of information, storage and computing requires “Scale Up” and “Scale Out” architectures. Scale Up refers to dedicated nodes of multiple GPUs (Graphical Processing Units) packaged into multiple racks composed of 72 GPUs each. Each Scale Up node supports up to 15 such interconnected racks. Scale Out expands capacity by adding multiple such Scale Up nodes through TPU (Tensor Processing Units and additional servers) to support distributed computing environments. Scale Out enables the seamless integration of a large number of Scale Up nodes to meet demand in a modular fashion (Figure 3). The goal is to prevent bottlenecks, and provide elasticity and flexibility to handle dynamic workloads.
The amount of data that has to be transmitted is massive. High capacity fiber-optic links deliver data to and from datacenters today, a trend that was accelerated by the telecom revolution of the 1990s-2000s and progress in fiber-optic technology. Traditionally, the way this data was manipulated and processed inside a datacenter was through OEO (Optical to Electric to Optical) conversion, a highly lossy process consuming significant amounts of energy (lasers are still not very efficient, typically producing 20% optical energy relative to the electrical power it consumes). OEO needs significantly more lasers links, impacting cost and energy consumption. With modern day constraints on energy efficiency, the move towards purely optical data movement within Scale Up and Scale Out architectures is increasingly being deployed.
Within a datacenter the dynamic routing of data between Scale Up nodes and across Scale Out nodes is critical, in order to balance compute and storage loads. The traditional way of doing is through Electronic Packet Switching (EPS) , which is highly mature, but requires the dreaded OEO conversion, and its associated drawbacks. Additionally, an EPS for a high port count configuration consumes ~3000 W of energy, and constitutes ~5% of energy usage in a datacenter (in addition to the OEO losses discussed above). Optical Circuit Switching (OCS) reduces this drastically, consuming ~100W for an equivalent configuration. Elimination of OEO conversion can reduce energy consumption for switching to ~0.2%, a factor of 30X, which is significant. If only there were ways to control the angular direction of millions of optical beams dynamically – with low latency and loss.
Fortunately there are ways to do this ! A previous article discussed the history of voice and data switching, along with newer developments in OCS.
Figure 4 shows one way that the direction of optical beams can be controlled, using a prism to bend different wavelengths of light which bend in different directions. LiDAR (Light Detection and Ranging) companies like Baraja have used this effect to control how laser energy direction is controlled.
Other techniques to control light direction for scanning LiDAR are:
- Opto-mechanical, Oscillating Galvo or Magnetically Controlled Mirrors: use motor driven elements to reflect and receive light (Ouster, Aeva, Seyond, Robosense, Hesai, Innoviz).
- MEMs Mirrors: use mirrors embedded in Micro-Electrical Mechanical-Systems configurations in silicon wafer structures to bend light (Aeye, Hesai).
- Mechanically Rotating Assemblies: lasers and detectors mounted on a rotating element to transmit and receive optical energy (Velodyne and Argo, both bankrupt. Velodyne acquired by Ouster).
- Electronically Controlled Scanning of Optical Energy: uses electrically controlled laser array sources to illuminate and receive photons from specific regions of interest, with reflected photons captured by an array of Single Photon Avalanche Diode (SPAD) detectors (Opsys)
- Optical Phased Arrays in Silicon Photonics (SiPh): photons routed through waveguides within SiPh structures (Analog Photonics)
- Liquid Crystal Metamaterials: control light direction using properties of light sensitive metamaterials (Lumotive)
- Optical Switching in Silicon Photonics: control of optical energy direction through SiPh based waveguides and switching fabrics (Voyant Photonics).
Some of the techniques used for LiDAR beam control are highly relevant for optical switching in data centers. Items 6 and 7 above were discussed in a prior article.
(Disclosure – I am an advisor to Voyant Photonics, and have equity in Lumotive)
OCS – Market Growth
According to a recent report, the OCS market today is in the order of ~$600M today and expected to grow to $8B by 2030. This is a massive increase relative to prior reports which projected $2.5B by 2029. Currently, most of the OCS deployments are for Scale Out architectures. Scale Up generally did not use any type of switching. With the advent of next generation of GPUs from suppliers like Nvidia that will integrate OCS into its NVL576 architecture, OCS deployments for Scale Up will quickly rise in revenue, rivaling and ultimately exceeding Scale Out by the end of the forecast period.
The key advantages of the OCS (vs EPS) are energy savings (no OEO conversions), bandwidth and protocol agnostic operation (switching fabric remains the same as opto-electronic modules increase in data bandwidth), and a potential for semiconductor integration into opto-electronic modules (which makes it more deployable in Scale Up). A disadvantage is switching speed – traditional OCS systems typically reconfigure at the millisecond level, whereas EPS switching speeds are operate at the microsecond level. Given that all data center interconnects are expected to be optical in the next five years (which means no OEO or EDS), the imperative to deploy OCS is even more compelling.
Current OCS Manufacturers
Major current OCS manufacturers include Lumentum, Coherent and Huawei, primarily for Scale Out architectures.
Google pioneered the development of MEMs based optical switching through its groundbreaking Apollo program in 2018. Light enters the Project Apollo OCS via a bundle of fibers, and is reflected by an array of 136 functional MEMs mirrors patterned on silicon wafers. This yields the ability to switch ~19,000 (136×136) signals for a transmit-receive pair. Figure 5 is a schematic of the Google’s OCS optical architecture (reproduced from this comprehensive article about OCS fundamentals).
Although Google pioneered OCS and built it internally, it has a manufacturing partnership with Lumentum which has its own MEMs-based OCS products originally developed for telecom switching (Figure 6):
It currently has OCS configurations for 64×64 and 300×300 ports, ranging in price from $30,000 – $150,000. The company provided guidance of OCS sales by end 2026 of ~$400M, making it a leading supplier today. Apart from OCS, Lumentum is a dominant, full stack supplier of other optical technologies like lasers and CPO modules. Lumentum’s CEO Michael Hurston “recently warned that the next AI infrastructure bottleneck is indium phosphide (InP), a niche semiconductor material. The shortage, he said, will be worse than what the memory industry faces”. InP is the material used to make lasers that power CPOs and data transmission. OEO conversions that are necessary for EDS dramatically increase the need for more lasers. OCS architectures alleviate this issue, and also contribute to dramatically lowering energy and cooling requirements.
MEMs-based OCS pioneered by Google and Lumentum certainly was a revolution. However, key disadvantages include precision opto-mechanical alignment complexity, long term reliability of the MEMs structures, millisecond level switching latency, sensitivity to typical harsh temperature and vibration conditions in a data center, and switching link losses.
Coherent is a major optical and materials engineering company, supplying Google and Oracle with Liquid Crystal on Silicon (LCoS) based OCS solutions, with order volumes of ~3,000 units costing ~$100,000 each. It is is currently engaged with 10 OCS customers, including Google. LCoS-based OCS control voltage levels to a silicon backplane coated with liquid crystal to steer the light path. Relative to MEMs OCS with lifetimes of ~3 years, LCoS OCS can last for ~10 years, since there are no mechanical components involved. The main disadvantage is its relatively high switching latency, ~ tens of millisecond, which is not suitable for scenarios that require higher-speed dynamic reconfiguration.
Emerging Solutions for OCS
As discussed earlier, the following technologies that have been successfully deployed for LiDAR are emerging for the next generation of OCS solutions, enabling OCS suppliers skip years of trial and error and reach market sooner.
- Liquid Crystal Metamaterials (LCM): control light direction using properties of light-sensitive meta-materials (Lumotive). Lumotive’s 8XX-9XX nm ToF (Time of Flight) LiDAR has been commercialized in multiple industrial sensing applications. 13XX-15XX nm wavelength LCMs have recently been launched for FMCW (Frequency Modulated Continuous Wave) LiDAR and OCS has been recently launched.
- Optical Switching in Silicon Photonics: control light direction through SiPh based waveguides and switching fabrics (Voyant Photonics). The entire fabric has been proven in Voyant’s FMCW LiDAR products, and this experience can be leveraged for OCS applications.
- Optical Phased Arrays in SiPho: Analog Photonics has developed FMCW LiDAR based on this technology. Although OPAs could be potentially used to route photons through waveguides, there are no publicly announced efforts for using this, primarily because scaling OPAs in large formats is not practical today and waveguide losses are very high.
Other approaches for next-generation OCS include:
- Integrated MEMS: companies like nEye.ai are integrating MEMs mirrors within SiPh waveguides for an on-chip OCS. The company recently raised $80M in funding, adding to prior funding rounds totaling $72M. This approach promises high switching speeds and low power consumption.
- SiPho Integrated OCS: Companies like Salience Labs (based in the U.K.) and Ipronics ((based in Spain) are also pioneering optical switching through different mechanisms within a SiPho structure, delivering chip-scale, solid-state OCS implementations.
- New Material Platforms – Thin Film Lithium Niobate (TFLN) and Barium Titanate (BTO): multiple companies are engaged in leveraging the electro-optic properties of these materials for higher speed modulation and switching. Key advantages include high switch speeds and potential for chip-scale integration.
Microsoft, a leading provider of cloud services and AI technology, is investing heavily in datacenters. Its research division is actively looking at maximizing optical technologies within its datacenters, with Project Sirius focused on OCS technologies which promise nanosecond level switching speeds.
NTT, a leading communications and data services provider, headquartered in Japan, is also actively investing in OCS through its IOWN (Innovative Optical and Wireless Network) concept. It has partnered with Broadcom for high volume manufacturing of its photoelectric fusion switch.
Voyant Photonics, a SiPho company focused on LiDAR and datacenter OCS markets recently published a white paper on OCS technologies. The company has developed mature LiDAR products for robotic sensing (Figure 7), and is leveraging this experience for providing OCS solutions for the datacenter market.
Carbon LiDAR employs an on-chip beam-steering thermo-electric switch matrix arranged as a binary tree, within a silicon photonics chip SiPho fabric. Adapting this for datacenter OCS is relatively straightforward by extending the architecture into a non-blocking, reconfigurable switch network for OCS. Figure 8 shows how this can be integrated with high-speed CPO modules for Scale Up deployment, yielding a reconfigurable fabric for optimizing GPU and memory connections that is further enabled by the sub-microsecond reconfiguration speed,
The advantages of Voyant’s OCS solution include low optical link loss, microsecond level switching speeds (< 5 μsec), high reliability (no moving parts), high performance predictability (low sensitivity to shock, vibration and heat), low SWaP (Size, Weight and Power), semiconductor manufacturability scaling, and low cost. Apart from these, a big advantage is the ability to integrate seamlessly with CPO modules in Scale Up rack architectures.
Table 1 from Voyant’s recent white paper provides a performance comparison of OCS technology approaches:
Lumotive’s approach to OCS leverages its LCM (Light Control Metasurface) technology for solid-state LiDAR scanning and datacenter OCS implementation (Figure 9):
The technology has already been proven for multiple industrial LiDAR applications. For OCS, Lumotive advertises solid-state reliability, scalability to high port counts, low SWaP (a 256×256 2D LCM chip measures 4 x 3 x 2 cm³), low insertion loss (< 2 dB) and CMOS-scale manufacturability (Figure 10):
The company recently participated in authoring an Open Compute Project article on OCS technologies for hyperscale datacenters.
LiDAR certainly propelled the development of chip-scale semiconductor technologies for optical beam scanning over the past decade. It solved a major problem – elimination of opto-mechanical moving elements for scanning the Field of View (FoV). LiDAR integration for perception and localization in AoT® (Autonomy of Things) and physical AI applications like driverless vehicles, robotics and drones is slow as end users are still building their market entry plans. Physical AI also has other issues that take time to solve – pedestrian safety, worker safety, spoofing and cybersecurity. Optics in datacenters is here and now, the market is large, and growing exponentially. Its not surprising that solid-state LiDAR companies are increasingly pivoting to these markets. Apart from Lumotive and Voyant entering the OCS market, Aeva, a SiPh LiDAR company based in California is leveraging its LiDAR packaging and laser technology for Near Packaged Optics (NPO) and Co-CPO, which are deployed within Scale Up racks for datacenters.