KraneShares AI ETF AGIX has taken a position in private photonics company Ayar Labs (0.64% weight in AGIX as of 7/15/2026¹). Additionally, KraneShares launched Photonics ETF LUMA on 7/15/2026.

---

AI has a new bottleneck: getting chips to talk to one another.

Consider what has happened to artificial intelligence (AI) over the past two years.

AI models have grown larger, the computing clusters used to train and run them have expanded, and the data centers housing those clusters have grown with them.

Increasingly, the challenge facing AI infrastructure is not simply how fast an individual chip can calculate; it is how quickly thousands of chips can connect and work together as one system.

Today, many of those connections still rely on copper.

NVIDIA's NVLink 6, one of the industry's most advanced copper-based interconnect systems, does not work efficiently beyond one meter. As AI systems expand from dozens of accelerators (chips specifically designed to process AI workloads), to hundreds and potentially more than a thousand², distances between these chips increase to tens or hundreds of meters. At this distance, maintaining extremely high performance with copper becomes increasingly difficult and costly.³

Copper-based interconnects, on this trajectory, become a bottleneck; we believe only one technology has the potential to solve it: photonics.

Instead of moving data with electrical signals, photonic interconnects use light. The goal is straightforward: transfer more information, using less energy. Transitioning to this technology requires a replumbing of the AI value chain, in which the medium of communication shifts from electrons to photons.

The industry is spending billions because they need to.

Goldman Sachs, in its May 2026 Global Tech note, put a number on what the move to photonics looks like at the market level.

Goldman projects a $154 billion total addressable market for AI optical networking, of which $91 billion sits in co-packaged optics (optical connectivity directly alongside the chip rather than in a separate pluggable module).^4,5 Goldman also estimates that the total networking cost per computing unit could increase approximately 29 times between the current (NVIDIA's GB300 NVL72) and the projected next-generation architectures (Rubin Ultra NVL576).⁵

The AI infrastructure industry's investments and acquisitions also demonstrate the importance of photonics in scaling AI capabilities:

• American semiconductor giant Marvell paid up to $5.5 billion for Celestial AI, a company still without proven revenue, to acquire its advanced Photonic Fabric technology outright.⁶
• NVIDIA committed $4 billion between optical component leaders Lumentum and Coherent and locked in laser supply through 2028.^7,8
• Ciena, a global leader in networking systems, bought Nubis Communications (which specializes in high-performance, ultra-compact, low-power optical and electrical interconnects designed specifically to handle surging AI-related data traffic) for $270 million.⁹
• According to a press release from Tower Semiconductor, it signed contracts totaling $1.3 billion for 2027 with its largest customers and received $290 million in customer prepayments.¹⁰

In under twelve months, at least twelve billion dollars of corporate action have been directed at a supply chain that, from the outside, still looks like a research topic.

We believe two important trends are emerging.

Historically, optical components were often viewed as one part of the networking supply chain.

However, as AI clusters grow, the connections between chips, racks, and data centers increasingly influence the performance of the entire system. The faster these chips become, the more costly it is to leave those chips waiting for data. The roadmap of every major graphics processing unit (GPU) and switch supplier has been rewritten around that fact.

NVIDIA's upcoming platforms: Spectrum-X Photonics in H2 2026, Vera Rubin Ultra in 2027, and Feynman in 2028 are less product launches and more so pit stops on a forced migration. We believe the question is no longer whether the migration happens; it is who supplies the pieces.

Second, and this is the point that can easily be missed, the migration coincides with a quiet re-fragmentation of the AI accelerator market. While NVIDIA remains a dominant force in AI computing, the largest cloud companies are developing their own specialized AI chips: Google's TPU, Amazon's Trainiums, Microsoft's Maia, and Meta's MTIA

These chips are designed for different workloads and may use different systems to communicate with the rest of the data center. Put simply, there is a growing number of powerful machines that do not necessarily speak the same language. NVIDIA's NVLink is designed around NVIDIA's ecosystem, while other semiconductor and networking companies are developing their own architectures. For companies building custom AI chips, a more flexible optical connection is critical.

Why Ayar Labs

Ayar Labs is a Silicon Valley-based semiconductor company pioneering optical interconnects and co-packaged optics (CPO) to overcome the power, bandwidth, and latency bottlenecks of traditional copper interconnects in AI and high-performance computing.

Ayar's in-package optical input/output (I/O) device (interconnect) uses almost 85% less energy than traditional pluggable optical modules⁴, which convert electrical signals into light and vice versa. Co-packaged optics integrate optical connections directly beside the chip, enabling roughly 10× higher bandwidth along the package edge and up to 100× higher bandwidth within the same physical footprint.⁴

In simple terms, Ayar Lab's technology enables more data to move through a smaller space while using less power.

Ayar's core product, the TeraPHY optical I/O chiplet, is also designed to support multiple communication protocols and work within advanced semiconductor packaging systems. Ayar is trying to build an optical module compatible with an increasingly diverse AI chip market.

Furthermore, the company's SuperNova light source keeps the laser outside of the high-power processor package.¹¹ This separates a temperature-sensitive component from one of the hottest parts of an AI system and is designed to improve the thermal characteristics of the optical architecture.

NVIDIA, Advanced Micro Devices (AMD), and Intel Capital have all invested in Ayar Labs.

The company also has potential manufacturing pathways involving the New York-based contract semiconductor manufacturer GlobalFoundries and the broader Taiwan Semiconductor (TSMC) ecosystem, through companies such as Alchip and Global Unichip. As the industry consolidates into two or three vertically integrated production and design chains, Ayar is the rare independent platform we believe no cloud company can overlook.

Conclusion

Photonics is moving closer to the center of the AI infrastructure story. We believe Ayar Labs is positioned at one of its most important intersections. As the challenge of moving data becomes increasingly central to AI performance, we believe photonics represents a compelling and growing part of the opportunity AI ETF AGIX, the KraneShares Public-Private AI & Technology ETF, is designed to capture.

Additionally, KraneShares recently launched Photonics ETF LUMA, the KraneShares Photonic and Optical ETF, an ETF that seeks to provide exposure to both public and private companies worldwide that develop optical interconnects, transceivers, fiber-optic cables, and other light-based infrastructure. Portfolio companies in LUMA are actively developing technology to help solve the bandwidth bottleneck posed by copper interconnects.

---

Holdings are subject to change.

For AGIX standard performance, top 10 holdings, risks, and other fund information, please click here.

For LUMA standard performance, top 10 holdings, risks, and other fund information, please click here.

Citations:

1. Data from Bloomberg as of 7/15/2026.
2. Data from "NVIDIA NVLink and NVLink Switch," NVIDIA website, retrieved 7/15/2026.
3. Data from "NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer," NVIDIA Technical Blog, as of 3/16/2026.
4. Data from "How Optical I/O Is Enabling the Future of Generative AI: A Q&A with Vladimir Stojanovic," Ayar Labs, as of 11/8/2023.
5. Data from "Optical Networking: The Next Mega Trend in AI Infrastructure," Goldman Sachs Research, as of 5/12/2026.
6. Data from "Marvell to Acquire Celestial AI, Accelerating Scale-Up Connectivity for Next-Generation Data Centers," Marvell Investor Relations, as of 12/2/2025.
7. Data from "NVIDIA Announces Strategic Partnership With Lumentum to Develop State-of-the-Art Optics Technology," NVIDIA Newsroom, as of 3/2/2026.
8. Data from "NVIDIA and Coherent Announce Strategic Partnership to Develop Optics Technology to Scale Next-Generation Data Center Architecture," NVIDIA Newsroom, as of 3/2/2026.
9. Data from "Ciena to Acquire Nubis Communications to Expand Its Inside the Data Center Strategy and Further Address Growing AI Workloads," Ciena, as of 9/22/2025.
10. Data from "Tower Semiconductor Signs Customer Contracts for $1.3 Billion Silicon Photonics Revenue for 2027," Tower Semiconductor Investor Relations, as of 5/13/2026.
11. Data from "Optical I/O Solutions Optimized for AI Workloads," Ayar Labs website, retrieved 7/15/2026.