The race to replace copper inside AI data centers

Overview

Every time engineers double the data rate on a copper wire, electrical noise doubles too, cutting the usable cable length in half. That physics problem is now strangling the AI industry. As graphics processing units (GPUs) push toward 224 gigabits per second per lane, passive copper cables inside data centers can reach less than one meter before the signal degrades. Ayar Labs, a startup born from research at the Massachusetts Institute of Technology (MIT) and the University of California, Berkeley, just closed $500 million in Series E funding at a $3.75 billion valuation to mass-produce chips that replace those copper links with light.

In This Story

3 Videos

3 People

4 Orgs

Optical startups become essential AI suppliers, Ayar Labs heads toward IPO

3 Context

Fiber optics replaces copper in telecommunications (1980s-2000s)

Key Indicators

$500M

Series E raised

Largest single funding round for a co-packaged optics startup, bringing Ayar Labs' total to $870 million

$3.75B

Ayar Labs valuation

Up from just over $1 billion at the Series D in December 2024, roughly tripling in 15 months

4-20x

Throughput-per-watt gain

Claimed improvement of optical over copper interconnects, depending on configuration and distance

<1 meter

Copper reach at 224 Gbps/lane

Maximum distance for passive copper cables at the data rates next-generation AI clusters require

~30%

Interconnect share of data center energy

Proportion of total data center power consumed by moving data between chips, a figure optical links aim to cut in half

Interactive

Exploring all sides of a story is often best achieved with Play.

Ever wondered what historical figures would say about today's headlines?

Debate Arena

Two rounds, two personas, one winner. You set the crossfire.

People Involved

Mark Wade

Chief Executive Officer and Co-Founder, Ayar Labs (Leading Ayar Labs through production scale-up)

Vladimir Stojanovic

Chief Technology Officer and Co-Founder, Ayar Labs (Leading technology development and Alchip partnership)

Gabe Cahill

Managing Director, Neuberger Berman (Taking board observer seat at Ayar Labs)

Organizations Involved

Ayar Labs

Semiconductor Startup

Status: Scaling production after $500M Series E

Develops optical interconnect chiplets that replace copper wiring inside AI servers with fiber-based connections, claiming 4 to 20 times more throughput per watt.

Nvidia Corporation

Semiconductor Company

Status: Strategic investor in Ayar Labs; developing own co-packaged optics platforms

The dominant supplier of AI accelerator chips, NVIDIA is both a customer for optical interconnects and a developer of its own co-packaged optics networking platforms.

Lightmatter

Semiconductor Startup

Status: Leading competitor; $4.4B valuation as of October 2024

Develops photonic computing platforms including 3D photonic interposers, positioned as Ayar Labs' closest independent competitor with $850 million in total funding.

Marvell Technology (acquirer of Celestial AI)

Semiconductor Company

Status: Completed acquisition of Celestial AI; building optical interconnect portfolio

Acquired photonic fabric startup Celestial AI for $3.25 billion to $5.5 billion in December 2025, gaining a 16-terabit-per-second optical chiplet platform.

Timeline

Ayar Labs closes $500M Series E at $3.75B valuation
Funding

Led by Neuberger Berman with investment from NVIDIA, AMD, MediaTek, Qatar Investment Authority, ARK Invest, Insight Partners, and Sequoia. Total funding reaches $870 million. Funds will scale production and expand the Hsinchu, Taiwan office.
Yesterday
Marvell acquires Celestial AI for up to $5.5 billion
Acquisition

Marvell Technology bought photonic fabric startup Celestial AI, whose 16-terabit-per-second chiplet technology removes a key independent competitor from the optical interconnect market.
December 2nd, 2025
Ayar Labs and Alchip demonstrate 100 Tbps optical I/O on TSMC COUPE
Technology

The companies showed a fully integrated co-packaged optical I/O subsystem on TSMC's COUPE platform, supporting more than 256 optical scale-up ports per accelerator.
August 1st, 2025
Ayar Labs launches first UCIe optical chiplet
Technology

The second-generation TeraPHY became the industry's first optical chiplet using the Universal Chiplet Interconnect Express standard, delivering 8 terabits per second and enabling interoperability across chip vendors.
April 1st, 2025
NVIDIA unveils silicon photonics networking platforms at GTC
Technology

NVIDIA announced Quantum-X InfiniBand and Spectrum-X Ethernet platforms using TSMC's COUPE co-packaged optics technology, targeting 1.6-terabit-per-second post-copper networking for AI clusters.
March 18th, 2025
Ayar Labs raises $155M Series D, crosses $1B valuation
Funding

Led by Advent Global Opportunities and Light Street Capital, with investment from AMD, Intel Capital, and NVIDIA. Total funding reached $370 million.
December 1st, 2024
Lightmatter raises $400M Series D at $4.4B valuation
Funding

Led by T. Rowe Price, the round brought Lightmatter's total funding to roughly $850 million, making it the highest-valued independent photonics startup at the time.
October 16th, 2024
First terabit optical link demonstrated
Technology

Ayar Labs demonstrated a fully functional TeraPHY chiplet running error-free at 1.024 terabits per second with less than 5 picojoules per bit of energy, in partnership with MACOM.
June 8th, 2021
Ayar Labs closes $35M Series B
Funding

The round funded development of the TeraPHY optical engine on GlobalFoundries' silicon photonics manufacturing process.
November 1st, 2020
MIT/Berkeley team publishes light-enabled microprocessor in Nature
Research

A DARPA-funded collaboration between MIT, UC Berkeley, and the University of Colorado Boulder demonstrated a 3-by-6 millimeter chip combining 70 million transistors with 850 photonic components. The result became the foundation for Ayar Labs.
December 1st, 2015
Ayar Labs founded
Corporate

Mark Wade, Alex Wright-Gladstein, Vladimir Stojanovic, and Chen Sun incorporated Ayar Labs after their MIT photonic chip won a clean energy competition. Prize money helped launch the company.
December 1st, 2015
Intel demonstrates first fast silicon optical modulator
Research

Intel researcher Mario Paniccia published results in the journal Nature showing a silicon modulator converting laser light into digital pulses at one gigahertz, proving silicon photonics was feasible.
February 1st, 2004

Scenarios

Optical startups become essential AI suppliers, Ayar Labs heads toward IPO

Discussed by: ARK Invest, Sequoia Capital, SemiAnalysis, and venture investors who view optical interconnects as a structural shift comparable to the rise of GPUs for AI

If AI cluster sizes continue doubling and copper's distance limits make optical the only viable scale-up interconnect, Ayar Labs and Lightmatter could become critical suppliers embedded in the GPU ecosystem. Ayar Labs' projection of 100 million annual unit shipments by 2028 would support a multi-billion-dollar revenue run rate, making an IPO the likely next step. This scenario depends on volume manufacturing yields being high enough to meet hyperscaler procurement standards and on Ayar Labs maintaining its technology lead as chip giants invest in competing approaches.

NVIDIA, Broadcom, and Marvell develop in-house optical, squeezing startups

Discussed by: LightCounting, EE Times, and semiconductor analysts who note that NVIDIA's COUPE-based platforms and Marvell's Celestial AI acquisition give incumbents viable in-house alternatives

NVIDIA is building co-packaged optics directly into its Quantum-X and Spectrum-X networking platforms using TSMC's manufacturing process. Marvell now owns Celestial AI's photonic fabric technology. If these incumbents move faster than expected, independent startups like Ayar Labs could find their addressable market shrinking to niche applications. The risk is highest for the scale-out networking segment, where incumbents already have strong customer relationships.

Copper innovations and alternative approaches delay optical adoption

Discussed by: Intel, Point2, and engineers who argue that linear-drive pluggable optics (LPO), active copper cables, and radio-based interconnects can extend copper's runway by several years

The semiconductor industry has repeatedly extended copper beyond its predicted limits. Active copper cables, advanced signal processing, and radio-based startups like Point2 offer intermediate solutions that could slow the transition to co-packaged optics. If AI cluster growth decelerates or if copper workarounds prove cheaper at scale, the optical interconnect market could take longer to materialize than current projections suggest, stranding companies that scaled production prematurely.

Ayar Labs acquired by a chip giant before reaching independent scale

Discussed by: CNBC, The Next Platform, and M&A analysts who cite Marvell's $5.5 billion Celestial AI deal as establishing a valuation benchmark for optical interconnect acquisitions

Marvell's acquisition of Celestial AI set a clear precedent: established semiconductor companies will pay billions to secure optical interconnect capability. NVIDIA, AMD, Intel, or Broadcom could acquire Ayar Labs to lock in supply and prevent competitors from doing the same. With Ayar Labs now valued at $3.75 billion and backed by both NVIDIA and AMD, any acquisition would involve complex strategic dynamics between rival investors.

Historical Context

Fiber optics replaces copper in telecommunications (1980s-2000s)

1980-2000

What Happened

Telephone companies and internet backbone operators spent two decades replacing copper trunk lines with fiber-optic cables. Corning Glass Works (now Corning Incorporated) developed low-loss optical fiber in 1970, but it took until the late 1990s dot-com boom for massive capital investment to lay fiber across continents. Companies like Global Crossing and WorldCom spent billions on fiber networks.

Outcome

Short Term

The transition created enormous wealth for fiber manufacturers and installers, but also fueled overbuilding that contributed to the dot-com bust when demand didn't immediately match supply.

Long Term

Fiber became the unquestioned backbone of global communications. The excess capacity laid in the 1990s eventually proved essential as internet traffic grew exponentially. Copper persists only for short last-mile connections.

Why It's Relevant Today

The current copper-to-optical transition inside data centers mirrors the telecom shift but at chip-level distances. The same physics—light carries more data, farther, with less energy than electricity in copper—drives both transitions. The overbuilding risk also rhymes: optical startups raising billions before volume revenue materializes.

Intel's silicon photonics program (2004-2024)

2004-2024

What Happened

In 2004, Intel researcher Mario Paniccia demonstrated that silicon could modulate light at gigahertz speeds, publishing the results in Nature. Intel invested heavily in silicon photonics for over two decades, building transceivers for data center networking. In 2006, Intel Senior Vice President Pat Gelsinger predicted optics would become "the mainstream of every chip that we build."

Outcome

Short Term

Intel proved the technology worked in the lab and shipped pluggable optical transceivers, but the products remained a niche business within Intel's portfolio for nearly 20 years.

Long Term

Intel eventually spun off its silicon photonics assets. The technology proved viable but struggled commercially until AI training created data movement demands that copper could not serve. The lesson: demonstrating feasibility is different from achieving mass adoption, and adoption requires an external forcing function.

Why It's Relevant Today

Ayar Labs' core technology descends from the same field Intel pioneered. The difference is timing: AI infrastructure spending now provides the demand signal that Intel lacked for two decades. Whether Ayar Labs captures the moment or repeats Intel's slow commercialization arc depends on whether production can scale as fast as orders arrive.

Qualcomm's rise as a wireless infrastructure supplier (1990s)

1990-2000

What Happened

Qualcomm developed Code Division Multiple Access (CDMA) technology for wireless communications and licensed it to the telecommunications industry. Major carriers initially resisted switching from existing standards. Qualcomm's founders, Irwin Jacobs and Andrew Viterbi, raised successive rounds of venture funding and went public in 1991 at a $486 million valuation, then fought a multi-year standards battle to establish CDMA as the basis for 3G mobile networks.

Outcome

Short Term

Qualcomm faced skepticism from European carriers favoring the competing GSM standard and was forced to divest its phone manufacturing business to focus on chips and licensing.

Long Term

CDMA became the foundation of 3G and later 4G/5G wireless standards. Qualcomm grew into a $180 billion company by owning the foundational technology that every smartphone requires.

Why It's Relevant Today

Ayar Labs occupies a similar structural position: a company with a potentially foundational technology trying to become an essential supplier to an industry undergoing a major infrastructure transition. The parallel highlights both the upside if optical I/O becomes standard and the risk that competing approaches or vertically integrated incumbents could limit the opportunity.