Live signals Morning Signal 07:00 GMT
Become a member

Ayar Labs: optical I/O chiplets, funding, partners and rivals

Ayar Labs makes TeraPHY optical chiplets and SuperNova lasers to link AI chips with light. Here is what they do, who funds the firm and how rivals compare.

A computer generated image of a spiral of lines

Ayar Labs is a San Jose chip company that makes optical input/output (I/O) parts, the TeraPHY optical chiplet and the SuperNova light source, which let AI processors send data over fibre instead of copper. As of 3 November 2026 it has raised $650 million in 2026 alone, a secondary share sale has valued it at more than $5 billion, and it says its products will be ready for volume production by the end of 2027.

What Ayar Labs makes

The company sells two components that work as a pair. Neither is a finished product on its own: Ayar supplies the optical parts and its customers build them into their own chip packages. As chief executive Mark Wade put it to Reuters in September 2026, “We build the optical chip, and our optical chip gets fed into our customer’s product.”

TeraPHY optical chiplet

TeraPHY is a small silicon photonics die, or chiplet, that sits in the same package as a GPU, accelerator or switch chip. On one side it talks electrically to the main processor; on the other it converts that traffic to light and sends it over optical fibre. In March 2025 Ayar announced a version with a UCIe interface, the industry standard for connecting chiplets from different vendors, and described it as the first optical UCIe chiplet. The company quoted 8 Tbps of bandwidth for that part and said it can link chips over distances from millimetres to kilometres.

SuperNova light source

Silicon is poor at generating light, so TeraPHY needs an external laser. SuperNova provides it: a multi-wavelength light source that feeds the chiplet over fibre. The 2025 version supplies 16 wavelengths, and each wavelength carries its own data stream down the same fibre. Keeping the laser outside the processor package means a failed laser can be replaced without scrapping an expensive chip. At OFC 2025 the manufacturer Jabil showed an external laser array built from 64 SuperNova sources, which the companies said could support up to 1 petabit per second of bidirectional bandwidth.

Why co-packaged optics matter for AI data centres

Large AI models are trained across thousands of GPUs that have to exchange data constantly. Today most of those links inside a rack use copper, and links between racks use pluggable optical transceivers that sit at the front of a switch, some distance from the chip. Both approaches spend a lot of power pushing electrical signals across circuit boards, and copper’s reach shrinks as speeds rise.

Co-packaged optics (CPO) move the electrical-to-optical conversion into the chip package itself. Vendors claim this cuts power per bit, raises bandwidth density and lets processors spread across a larger physical area while still behaving like one system. Nvidia, for example, claimed 3.5 times better power efficiency and four times fewer lasers for its CPO switches compared with traditional methods. These are manufacturer figures rather than independent measurements.

Ayar’s focus is “scale-up” networking: the dense links between accelerators in the same rack or pod, where copper is still the default. That is the main point of difference from switch-based CPO, which targets the network between racks.

Funding and investors

Ayar Labs was founded in 2015. Its funding has grown quickly since late 2024.

  • Series D, December 2024: Ayar raised $155 million in a round led by Advent Global Opportunities and Light Street Capital. AMD Ventures, Intel Capital, Nvidia, 3M Ventures and Autopilot joined as new investors, alongside existing backers including Applied Ventures, GlobalFoundries, Lockheed Martin Ventures and Playground Global. The company said this took total funding to $370 million and its valuation above $1 billion.
  • Series E, March 2026: On 3 March 2026 Ayar closed a $500 million Series E led by Neuberger Berman at a $3.75 billion valuation, taking total funding to $870 million. Participants included the Qatar Investment Authority, ARK Investment and strategic investors AMD, Alchip Technologies, MediaTek and Nvidia.
  • Series E extension, September 2026: Reuters reported on 10 September 2026 that Ayar had added $150 million to the Series E, bringing its 2026 total to $650 million, with the Taiwanese server maker Wiwynn joining as an investor. A separate $225 million secondary purchase of shares from early employees and investors, led by Antero Peak Group at Artisan Partners, valued the company at more than $5 billion.

Ayar says it is spending the money on high-volume production and test capacity. Reuters reported that the funding also paid for a new chip design centre in Bengaluru, India.

Partners and customers

Ayar does not publish a customer list. Its partner relationships, as disclosed by the company, include:

  • Nvidia: an investor since 2024. In June 2026 Ayar joined the NVLink Fusion ecosystem, the programme that lets companies plug custom CPUs and accelerators into Nvidia rack-scale systems. Ayar said it will make its products electrically and optically compatible with Nvidia’s optical and SerDes technology.
  • Foundries and packaging: the UCIe chiplet announcement named GlobalFoundries (its GF Fotonix platform), TSMC (its COUPE silicon photonics technology) and ASE as ecosystem partners.
  • Chip designers and system makers: AMD, Alchip and MediaTek are strategic investors; Alphawave Semi and d-Matrix were named as partners in 2025; Jabil built SuperNova into a laser array; and Wiwynn joined as an investor in 2026.

Leadership

According to the company’s leadership page, as of November 2026:

  • Mark Wade, co-founder and chief executive
  • Vladimir Stojanovic, co-founder and chief technology officer
  • Chen Sun, co-founder and chief scientist
  • Lisa Cummins Dulchinos, chief financial officer
  • Sejal Patel Daswani, chief people and operations officer
  • Scott Clark, vice president of manufacturing and operations, and Taiwan country manager

The board includes former Intel chief executive Pat Gelsinger, Jordan Katz of Advent Global Opportunities, Craig Barratt, Will Graves, Ganesh Moorthy and Geoff Tate.

How Ayar Labs compares with other optical interconnect efforts

Several large chip companies and well-funded start-ups are working on the same problem from different angles. Nvidia and Broadcom put optics inside network switches. Ayar, Lightmatter and Celestial AI (now part of Marvell) aim to bring optics right up to the compute chip.

Company Approach Headline figures (company claims) Status as of November 2026
Ayar Labs TeraPHY optical chiplet plus external SuperNova laser, sold to chip makers for their own packages 8 Tbps per UCIe chiplet; 16 wavelengths per SuperNova (2025 parts) Independent; aims to be qualified for volume production by end of 2027
Nvidia Co-packaged optics in its own Spectrum-X (Ethernet) and Quantum-X (InfiniBand) switches Up to 400 Tbps per Spectrum-X Photonics switch; 144 ports of 800 Gb/s on Quantum-X Photonics Announced March 2025; Quantum-X due late 2025 and Spectrum-X in 2026, per Nvidia at launch
Broadcom Co-packaged optics on its Tomahawk Ethernet switch chips Tomahawk 6 Davisson: 102.4 Tbps, 16 optical engines of 6.4 Tbps, 200 Gbps per lane Third-generation CPO switch, sampling to early-access customers from October 2025
Lightmatter Passage M1000 optical interposer and Passage L200 co-packaged optics chiplets M1000: about 114 Tbps per tile; L200: 32 Tbps; L200X: 64 Tbps M1000 slated for summer 2025; L200 series planned for 2026
Celestial AI (Marvell) Photonic Fabric optical links for scale-up connectivity between compute and memory No bandwidth figures in Marvell’s completion announcement Acquired by Marvell, completed 2 February 2026; first revenue expected in second half of Marvell’s fiscal 2028

Sources for the table: Nvidia, Broadcom, The Register on Lightmatter and Marvell. Bandwidth figures are not directly comparable because they describe different products: a switch chip’s total capacity is not the same thing as one chiplet’s I/O.

Key differences

  • Switches versus processors: Nvidia’s and Broadcom’s CPO products are complete switches. Ayar sells a building block that goes next to someone else’s GPU or accelerator, so its success depends on chip designers adopting it.
  • Independence: Celestial AI is now inside Marvell, which said the deal cost it $1 billion in cash plus newly issued shares. Ayar and Lightmatter remain independent, and Ayar counts Nvidia, AMD and Intel’s venture arm among its backers.
  • Standards: Ayar has bet on UCIe and on compatibility with Nvidia’s NVLink Fusion, which could let one optical chiplet serve several customers’ designs.
  • Timing: Reuters reported that the industry, Ayar included, is aiming for optical scale-up products to reach the market as early as 2028. Marvell’s own forecast for Celestial AI revenue points to the same window.

What to watch

Ayar’s next milestone is moving from demonstrations to qualified volume parts by the end of 2027, the deadline Wade gave Reuters for customers planning 2028 and 2029 systems. The company has not named those customers. Until it does, or until a shipping AI system uses TeraPHY, the funding rounds and partner announcements show strong investor and industry interest rather than commercial scale.

Sources