Amid the AI boom, technology’s role in ensuring unimpeded access to data hinges on photonic-electronic integration.
SURESH VENKATESAN, POET TECHNOLOGIES
As the technology world settles into 2025, the race is on to achieve 1.6 Tbps and 3.2 Tbps — speeds that are critical for the future generations of high-performance computing solutions for the data center industry. Hyperscale data centers, which are increasingly driven by demands from artificial intelligence (AI) hardware and software developers, require far more bandwidth than early-iteration data centers to satiate the global desire for advancements in medical technology, wearables, robotics, automotive, financial services, and other sectors. Large language models, at the root of AI development and machine learning, can only get “smarter” by having more data and faster times.
Courtesy of iStock.com/ipopba.
During the past year, the optical interposer has established itself as the rising star in chip-scale architectures, enabling the next level of data speeds. The best optical interposers that are currently on the market do not require wire bonds or active alignments, which dramatically reduces costs and results in coveted power savings. Dependent on the efficient
performance of lasers, optical interposers are likely to power data centers in the future.
Though AI technology is still considered to be in its infancy, this era is advancing steadily. Current market projections forecast the global data center market to grow by >5× its current levels by 2032, due to the massive demand for AI. This projection points to an insatiable desire for more data. It also signals challenges for an industry that has struggled to get more performance out of traditional platform technologies.
A forecast takes shape
In a Forbes article published last fall, Kazuhiro Gomi, CEO of NTT Research, described how the ongoing demand for increased compute power has created bottlenecks that the components that are currently used cannot satisfy1. These component-level limits extend to current computer architectures. They also serve to underscore the tech sector’s overall
appetite for energy, Gomi said.
Simply, AI workloads are pushing the boundaries of what copper connections can manage. Optical fibers, on the other hand, can transmit more data per second and lose much less signal over the same distance.
This one-for-one comparison of components is a core driver behind the massive shift from AI data centers to fiber optics, including well-known examples, such as NVIDIA’s NVLink for chip-to-chip connections and InfiniBand, which is used to connect multiple racks.
One server rack contains 32 graphics processing units, all of which are connected to each other by 18 switch nodes. NVIDIA’s Blackwell chip — which it has introduced to enable companies to build and run real-time generative AI on trillion-parameter large language models — will hold 1296 optical transceivers
per Blackwell rack. This value corresponds to 1 million transceivers for every 800 racks, on average, with each to be powered by optical engines. For context, Amazon Web Services alone has at least 30 data centers around the world, each of which has around 2400 racks, or around
3 million transceivers.
This use of data centers is the key reason as to why the optical transceiver market is forecast to more than double during the next eight years. But at the same time, the market for 800-Gb optics is specifically projected to increase by >10× by 2029. This is the direction in which AI developers are urging the industry to move. Once 1.6-Tbps applications dominate the market, their growth path will be similar.
Meanwhile, 3.2 Tbps is a target that few companies can currently dream of reaching. Soon enough, however, it will be necessary to operate at that level, too.
On a forward path
Meeting such unprecedented need for speed depends on innovation — a reality that creates a conundrum on top of opportunity. An enormous volume of innovation flooding into the ecosystems of the world’s largest data center companies requires tremendous power to run and cool. For example, according to Goldman Sachs, a single ChatGPT query uses nearly 10× as much energy as a Google search. Meanwhile, the Boston Consulting Group estimates that data centers will account for 16% of the total U.S. power consumption by 2030, increasing from 2.5% in 2022.
Optical interposers, which rely on the efficient performance of integrated lasers, are positioned to power data centers in the AI era. Courtesy of POET Technologies.
An interposer. Flat, combined electronic and photonic capabilities power the modern data center. Courtesy of POET Technologies.
Photonic integrated circuits (PICs) offer a potential solution in this context. The PICs technology space has undergone considerable growth and diversification, and multiple market forecasts anticipate the PICs market to exceed $45 billion by the start of the next decade. According
to industry observer Akash Anand, growth in the PICs market between 2024 and 2032 will occur at a compound annual growth rate of 18.1%, owing to the components’ increased adoption in optical networks used for data transmission.
A burgeoning PICs space in support of existing data needs, however, only underscores the fact that the energy requirements of AI compute networking are more intensive than anything yet to be seen in the world. Existing AI programs must execute trillions of computations to achieve their optimal performance.
The price of delivering this kind of performance, as well as the environmental cost, are significant hindrances for traditional semiconductor devices and processes. As Juniper Networks said in a 2024 industry report, despite the efforts of semiconductor infrastructure vendors to design more efficient products, current AI model training demands are increasing power requirements2.
The optical interposer
An optical interposer populated with tightly integrated electronic and photonic components reduces power consumption, eliminates crosstalk from copper wire bonds, uses fewer components, and can be built with lower labor costs than optical transceiver modules that are produced any other way. As a critical enabling technology for future data centers, optical interposers fundamentally ensure lower levels of power consumption. This translates to the data center industry’s having a lesser effect on the environment, which is an increasingly important feature given how much data is being consumed by AI applications.
The reduction of these losses is key to photonics’ success in future data centers,
which heightens the significance of optimized components in controlling the flow and efficiency of lasers at chip scale — which is itself vital to producing a scalable photonic component for commercial applications.
One proven way to harness lasers is with optimally performing waveguides; a low-loss waveguide will usher each laser consistently through its lane, creating a steady, predictable flow of power that can be used to transfer data. And, unlike traditional silicon photonics, the interposer platform can be material agnostic, integrating with traditional silicon and indium phosphide platforms as well as alternatives, such as thin-film lithium niobate (TFLN), which is showing incredible aptitude to scale to higher speeds while reducing power consumption. Already, device developers are favoring TFLN modulators for their low levels of power consumption, paired with their high bandwidth capabilities and low insertion losses. The material also holds promise in aiding heterogeneous integration for packaging.
The optical interposer offers another critical benefit. Miniaturization through TFLN (or through any means) is essential to meet the needs of hyperscalers, and the optical interposer makes this achievable on commercial scales. The optical interposer introduces a novel way to address the problem areas of an industry that knows it must brave fresh paths in order to sustain Moore’s law.
As a result, companies including POET are pioneering the conceptualization and development of optical interposers, targeting applications that can scale to high-volume production and seamlessly integrate into existing infrastructure — without increasing capital expenditures or power consumption. The passive alignment of the optical interposer yields significant cost reduction compared with traditional active alignments as well as the ability to ramp to high volume with less capital equipment. At the component
level, optimized waveguides can be fabricated to ensure that their lasers deliver necessary levels of performance — namely, high efficiency and low error rates. Such an approach simplifies chip-level engineering and increases flexibility. With all components on POET’s optical interposer integrated onto a single chip, via an adaptation of existing CMOS manufacturing methods, assembly time and labor cost savings are additional benefits that the company has realized.
Of course, the critical parameter to gauge is the point of view of the end user. Whether it is POET’s optical interposer, or another commercially available solution, the goal remains the same: to design and deploy optical transceivers that are high-speed light-to-voltage and voltage-to-light converters or radio frequency (RF)-to-light and light-to-RF converters that can generate data speeds that are magnitudes higher than current generations and within budget constraints. How each optical interposer on the market gets to this detail depends on their respective designs and the Internet Protocol behind them. But without optical transceivers, graphic processing units will be unable to meet AI compute and/or AI workload market demands in a manner that is timely or cost-effective, and certainly not in a manner that delivers on both.
The case for photonics
As Gomi and other experts have said recently, the idea of optical computing predates by many years the current interest and attention that it is commanding from industry1. As the need for chip-level innovation intensifies, photonics has already been shown to drive innovation. Data centers are one of the first industries to see what photonics can achieve.
But other sectors will benefit. On the heels of AI, quantum compute, for example, will increase demand even further for optical computing. If AI is at its dawn during today’s digital technology transformation, then quantum computing is rising from the shadows, preparing to have a dramatic effect on all aspects of science. Eric Mounier, chief analyst of photonics and sensing at Yole Group, is among the experts who has shared projections that the quantum computing market will be worth hundreds of millions of dollars by 2034.
As for the immediate future, expect the photonics era to continue its trajectory in 2025 as the data center industry shifts its approach to fulfill what it sets out to: address the demands of a world that is highly dependent on connectivity and data.
Meet the author
Suresh Venkatesan, Ph.D., is chairman and CEO of POET Technologies. Previously, he was senior vice president of technology development at GlobalFoundries, responsible for the company’s technology R&D. He has more than 25 years of experience in semiconductor tech development; email: [email protected].
References
1. K. Gomi (September 2024). Optical computing: what it is, and why it matters. Forbes.
2. Juniper Networks (February 2024). 5 Key data center trends for 2024: the currents driving change in data center networking. Juniper Networks industry report, p. 3.