Retym Is Betting That AI's Next Bottleneck Is a Chip Most People Never See

The most expensive seat in an AI data center may be the one occupied by a chip nobody can point to. We like to picture AI as a mountain of GPUs. But those processors have to exchange information, often across racks, buildings and campuses. Between the electrical conversation of a server and the pulses of light in a fiber sits an optical transceiver. Inside some of those modules sits a digital signal processor, cleaning, encoding and recovering the message. Retym wants to make that processor.

Its wager is specific enough to be interesting. The company develops programmable coherent DSP technology for high-speed optical links used within and between cloud and AI data centers. In coherent optics, information is carried in properties of light that demand substantial signal processing at both ends. That engineering can buy reach and capacity. It also consumes power and space, two things a crowded module has very little to spare.

The distance that changed the business

When Gandhi and El-Bahar started Retym in 2021, coherent optics was already familiar in long-distance networks. El-Bahar later wrote that it was beginning to move into links shorter than 100 kilometers, including data center interconnects. They saw a second feature of that market: the powerful DSP chips were controlled by a small group of large semiconductor companies. Retym's answer was to build an independent supplier around analog design, conversion between electrical and digital signals, and algorithms for equalization and error correction.

Then AI changed the size of the audience. Generative AI did not invent the need to connect data centers, but larger clusters and faster network upgrades made that need harder to ignore. Gandhi told SiliconANGLE that network speed steps which once took roughly four years were arriving about every two. His forecast is a founder's view, not an industry stopwatch; it does explain why a company begun before the AI frenzy could emerge from stealth with an AI infrastructure pitch.

The DSP is the tiny translator with a demanding job description. This diagram shows the intended position, not a reported customer deployment.

The customer is the company building the module

Retym's immediate customer is not someone training a model. Its stated plan is to work with makers of high-speed optical transceivers and other ecosystem partners that integrate a DSP into a finished module. A cloud operator or network builder may ultimately specify the link, but the module maker must fit speed, power, heat, reach, cost and interoperability inside a physical product. A promising chip is only one component of that negotiation.

That is why the word neutral matters in Retym's materials. Some established suppliers sell multiple parts of the optical stack, including modules or systems. Retym argues that an independent DSP supplier can give module makers more room to choose other components and differentiate their own products. The commercial logic is straightforward: if the chip at the center of the module is available to more designers, more designers can attempt the module. Whether that produces better prices or more deployments will be decided by customers, not by a pitch deck.

“We're excited to collaborate with customers and ecosystem partners to integrate our DSPs into high-speed transceiver designs.”Sachin Gandhi, March 2025 announcement

There are alternatives. Marvell and Ciena are established coherent DSP players; Coherent makes optical products around its own DSP technology. On some shorter links, direct-detect optics may be the more sensible answer. Retym's difference is less a claim that coherent optics always wins than an attempt to be a focused, independent source of the processor when coherent optics does win. The same distinction makes its challenge plain: it has to prove its chip, its power envelope and its supply reliability against companies that already ship at scale.

The capital is real. Public unit pricing, revenue and deployed chip counts are not available.

The people behind the invisible chip

Gandhi has been through the awkward passage from chip idea to market before. Retym's biography says he co-founded Xpliant, later acquired by Cavium, and dryv.io, later acquired by Synaptics. He also worked at Marvell and Cisco. El-Bahar, Retym's CTO, previously led engineering teams at Samsung, Intel and Huawei and was CTO of Huawei's Israel research center. He teaches advanced analog design at the Technion. Their backgrounds do not guarantee a working product, but they explain why this company talks about both circuits and the businesses that buy them.

The team spans Cupertino, Austin, Israel and Armenia. On its site, Retym describes a mix of analog, DSP, VLSI and optical communications specialists; its engineering vice president, Yuval Bechori, calls for an agile, collaborative culture. That international footprint is useful because building a coherent DSP is several jobs pretending to be one: analog front-end design, algorithms, digital implementation, firmware, verification and packaging all have to agree. A beautiful algorithm that cannot meet the power budget is an academic result. A fast circuit that cannot be validated is a manufacturing problem.

The money buys a test, not an outcome

Retym emerged from stealth at the Optical Fiber Communication conference in March 2025. It said it had raised more than $180 million across rounds, with a new $75 million Series D led by Spark Capital and joined by Kleiner Perkins, Mayfield and Fidelity Investments. Spark general partner James Kuklinski joined the board. The company said the money would support scaling to production and a multi-generation roadmap.

That is a large amount of capital for an idea that is still mostly invisible to the public, but chip programs are expensive long before a customer can unpack a finished module. Engineers, design tools, prototypes, tests and production preparation arrive before volume sales. Retym has not published a named product family, a measured performance table, a unit price, a named paying customer or a shipment count. It is more accurate to say that the company has financed the attempt than that the market has settled the verdict.

The useful lesson is in the sequence. Retym did not begin with the broad claim that AI needs more infrastructure. Its founders identified a particular component, a particular buyer and a supplier concentration problem in 2021. When the market's urgency changed, the original technical thesis had a new setting. A company looking for a similar opening can copy the method: find the part of a growing system where the buyer has too few choices, then ask whether an independent supplier can meet the incumbent's performance and delivery bar. The method fails when independence alone is the product, or when the incumbent's integrated package saves more power, money or risk than the new option can.

The name is pronounced “Re-Time,” a neat pun for a business that handles signals measured in unforgiving fractions of a second. Its next chapter will be less poetic. Module makers will want chips that fit, run cool, interoperate and arrive when promised. AI can make the market larger. It cannot waive those tests.