After Samuel Insull built the modern U.S. electrical infrastructure system in the 1890s, the constraint stopped being how much coal you could burn. It became how far the grid could carry electricity and how many stations it could stitch together. By the 1920s, the real center of gravity in American capital spending had already shifted from boilers to high-voltage transmission lines, substations, and interstate interconnection.

Artificial intelligence (AI) infrastructure in the second quarter of 2026 sits at a similar turning point. After a lofty run in memory prices, the supply chain may have digested demand expectations and adjusted its expansion plans and roadmaps. We think the shortage is not over, and it may persist for years, but the marginal tightness is easing. Meanwhile, the architectural constraint that determines whether clusters (connected groups of powerful servers and specialized chips) can keep getting bigger may have migrated somewhere else: to the interconnect.

Why Memory Became The AI Bottleneck & How The Price Surge Happened

Start with HBM, the high-bandwidth memory that sits pressed against various types of AI accelerator chips, like graphics processing units (GPUs), as its working memory. As GPUs have scaled, the HBM riding on each one has swelled in both capacity and value. This is evident by looking at Nvidia's GPU lineup: there are 80GB of HBM on the H100, 192GB on the B200, 288GB on Blackwell Ultra, a 3.6x increase in barely two generations.¹ Prices climbed alongside volume. Samsung and SK Hynix have raised 2026 contract prices for HBM3E (the fifth generation of the advanced HBM standard) by nearly 20%. Now, the market expects sixth-generation 12-high HBM4 stacks to settle at $500–600 or more per unit, another 55–70% above HBM3E.² The arithmetic compounds: the total HBM market is projected to grow from $35 billion in 2025 to $54.6 billion in 2026, an increase of more than 50% in a single year, with $100 billion (According to Micron CEO) in sight by 2028.³

Dynamic random access memory (DRAM) tells a parallel story. AI servers need HBM at the GPU level, but they also need enormous quantities of ordinary server-grade DDR5 (Double Data Rate 5) on the central processing unit (CPU) side, for data preprocessing, model loading, and scheduling, and multi-GPU coordination. The agent boom may have made this worse. Agentic applications hold long-running context, accumulate multi-step tool-call records, and maintain deep memory caches, all of which live in CPU-side DDR5. SK Hynix has disclosed that AI inference demand is driving high-density server DDR5 demand up 50% quarter over quarter.⁴

Here is where the trouble compounds. HBM expansion crowds out commodity DRAM directly. Each gigabyte of HBM consumes roughly four times the wafer capacity of standard DRAM, and with HBM yields still running at 50–60%, every AI-grade chip effectively eats the capacity of three to four sticks of ordinary PC memory.⁵ On top of that, SK Hynix, Samsung, and Micron have redirected cleanroom capacity that once produced consumer DDR4 toward higher-margin HBM. The crowding-out produced an absurd inversion: spot prices for consumer DDR4 at one point hit $2.10 per gigabit, a 5.5x surge in six months, briefly exceeding the price of the more advanced HBM3E itself.⁶

We believe that inversion powered the memory stocks' first-half rally. But a market this sensitive to marginal change, carrying this much leverage, cuts both ways. Since mid-June, the semiconductor sector has shed roughly $1.5 trillion in market value, nearly $350 billion of it from Micron alone.⁷ We also believe that none of this changes the underlying supply-demand picture, which is forecasted to stay constricted into 2028, or the earnings stability that a stack of long-term agreements provides. Capacity is sold out through 2027; contracts run through 2028.⁸ And on the valuation side, July's deleveraging compressed Micron's forward earnings multiple to six times.⁹

For our AI ETF AGIX, the KraneShares Public-Private AI & Technology ETF, the latest portfolio rebalance reflects this shift: we added KIOXIA, a Japanese semiconductor memory and data-storage company, and lifted the combined weight of the fund's memory names to roughly 6.6%, positioning AGIX for a steadier, more durable advance once valuations normalize.

The Next AI Bottleneck: Interconnect

In the Internet era, switch speeds doubled about every three years, a gentle curve.¹⁰ In the AI era, the data throughput of compute networks is growing at something closer to ten times a year.¹¹ Electrical interconnect bandwidth, by contrast, has improved only about 30x over the past two decades, while floating-point operations per second (FLOPs), a measure of computation amount, have grown 60,000x.¹² The gap is not closing. It is compounding.

The structure of traffic has changed, too. East-west (datacenter-to-datacenter) has overtaken north-south (datacenter-to-consumers). Data from Cloudflare, a U.S.-based platform used by roughly 25% of all websites to speed up and protect online applications, shows that as of June 2026, bot and AI-agent traffic has formally surpassed human traffic on the internet, 57.4% to 42.6%, with agent traffic alone surging nearly 8,000% over the course of 2025.¹³ Behind that statistic is the crowding of north-south traffic (users talking to data centers) by east-west traffic inside the data center: GPU to GPU, GPU to memory, rack to rack. A single human request now triggers far more machine-to-machine communication than it ever did.

Inside the rack, copper is approaching its physical edge. As rack scale extends to 72 accelerators and beyond, a single cabinet can require more than a thousand high-speed links, with total copper runs accumulating to several kilometers. Traditional copper wiring over such a long distance can't maintain signal integrity, which means the structure needs to switch to active copper cables that use built-in microchips to boost and clean up signals or jump straight to optics.

Optical interconnect, on the other hand, is following a clean path of value migration, from pluggable modules, to near-packaged optics (NPO), to co-packaged optics (CPO), each generation pushing the optical engine one step closer to the chip itself. The interesting thing about this migration is that the further we place the bet, the more we believe value accrues as the whole industry gradually feels the pain of the bottleneck unfold.

Penetration remains early. CPO accounts for only about 0.5% of AI data-center interconnect in 2026, with projections reaching roughly 35% by 2030.¹⁴ The AI optical transceiver market should exceed $10 billion in 2026, nearly double its 2024 size, yet, relative to total AI data-center capex, interconnect remains a far smaller line item than memory.¹⁵ Networking runs about 12% of AI data-center capital spending; memory now accounts for 40–70% of server bill-of-materials cost.¹⁶ Put differently: the memory bottleneck has already been cashed in, through earnings and through prices. The optical value-release curve is still at the base of its climb, somewhere between 0.5% and 35%.¹⁷ The repricing of this lane may only be beginning.

For our AI ETF AGIX's latest rebalancing, the fund allocated nearly 6.4% to optical interconnects, including names like Applied Optoelectronics, Corning, Lumentum, Coherent, and Marvell. KraneShares also recently launched the LUMA ETF, a fund dedicated to photonics and optics. Lumentum's FY26 Q4 revenue came in at $1.01 billion, up 109% year over year.¹⁸ Coherent posted $2.05 billion, up 34%. Corning Optical Communications reached $2.07 billion, up 32%, with Enterprise Networks up 65% and generative AI-related products growing even faster.¹⁹ Across all four company earnings reports, the same phrase kept surfacing: not enough capacity. That unanimity from the supply side is the main source of our conviction in this line.

Back to Insull. The men who bought boilers in the 1920s made the first money. The ones who pulled copper wire across state lines defined the shape of the power industry for the next fifty years. As agents continue to evolve, the scale and structure of data centers, and their demands on interconnect, memory, and everything else inside them, will keep shifting. We will continue tracking those shifts and aim to reflect them in AGIX's rebalancing.

---

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. NVIDIA Developer Blog, "Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era," as of 8/22/2025.
2. Silicon Analysts, "HBM Market Analysis: Exascale Inference Report 2025-2027," retrieved 8/24/2026.
3. LuminixAI, "Research publicly available data on DRAM and HBM pricing trends, SK Hynix/Samsung/Micron capacity expansion announcements, and...," as of 3/14/2026.
4. SK Hynix, "Five things you need to know about SK Hynix and the future of AI memory," as of 6/23/2026.
5. Yahoo Finance, "AI Data Centers Will Consume 70% of All Memory Chips in 2026. Here Are the Only 2 Stocks That Matter," as of 6/25/2026.
6. NVIDIA Developer Blog, "Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era," as of 8/22/2025.
7. Yahoo Finance, "Micron, Samsung, SK Hynix just dragged memory stocks into a bear market," as of 7/7/2026.
8. IGN, "Memory Shortage Sees 2027 Production Reportedly Sold Out, as Demand Far Outstrips Supply," as of 8/8/2026.
9. Data from Bloomberg as of 7/31/2026.
10. Agora Research, "The Optoelectronic Shadow War: Decoding the Physical Limits and Packaging Battles of AI Infrastructure," as of 3/16/2026.
11. Datacenters.com, "How AI Networking Is Becoming the Bottleneck Inside Modern Data Centers," as of 5/8/2026.
12. University of California, "AI and Memory Wall," retrieved 8/24/2026.
13. Cloudflare, "Content Independence Day, one year on: building the business model for the agentic Internet," as of 7/1/2026.
14. TrendForce, "NVIDIA Compute Architecture Paves the Way for Scale-Up Optical Interconnects; CPO Penetration in AI Data Centers Expected to Rise Steadily, Says TrendForce," as of 3/11/2026.
15. LightTrends Newsletter, "AI creates a new wave in demand for optical transceivers and accelerates LPO/CPO adoption," as of January 2025.
16. IOT Analytics, "Data center infrastructure market: AI-driven CapEx pushing IT and facility equipment spending toward $1 trillion by 2030," as of 11/12/2025.
17. TrendForce, "TrendForce Consulting: CPO's penetration rate in AI data center optical communication modules is expected to reach 35% by 2030," as of 3/11/2026.
18. Lumentum, "Lumentum Announces Fourth Quarter and Full Fiscal Year 2026 Results," as of 8/11/2026.
19. Corning, "Corning's Strong Second-Quarter 2026 Financial Results(1) Demonstrate Progress on Recently Upgraded Springboard Plan," as of 7/28/2026.

Term Definitions:

Capital expenditures (capex): Corporate spending on long-lived physical or technology assets such as data centers, network equipment, and manufacturing facilities.

Marginal tightness: The degree of shortage in the next incremental unit of supply relative to demand.

Interconnect: The hardware and links that move data among chips, servers, racks, and data centers.

High-bandwidth memory (HBM): A stacked DRAM technology placed near an AI accelerator to deliver much higher memory bandwidth than conventional server memory

Accelerator: A specialized processor, such as a GPU, designed to speed up computationally intensive workloads, including AI.

Graphics processing unit (GPU): A highly parallel processor widely used for AI model training and inference.

Dynamic random access memory (DRAM): The mainstream volatile memory used in PCs, servers, and many electronic devices.

Central processing unit (CPU): The general-purpose processor that manages many server tasks and coordinates system operations.

Agentic applications: AI software that can plan, make decisions, use tools, retain context, and execute multistep tasks with limited human direction.

Memory cache: A fast storage layer that keeps frequently needed data readily available to reduce retrieval delays.

Inference: Using a trained AI model to generate outputs, predictions, or decisions from new inputs.

Wafer: A thin circular disk of silicon on which semiconductor chips are manufactured

Yield: The percentage of usable, defect-free chips produced from a manufacturing batch.

Forward price-to-earnings multiple (forward P/E): A valuation ratio that divides a company's current share price by its expected earnings per share over the next 12 months.

Floating-point operations per second (FLOPs): A measure of computational processing capability based on the number of floating-point calculations completed each second.

Optical interconnect: A data connection that uses optical components and fiber to transmit information as light.

Bill of materials (BOM): The complete list and aggregate component cost to manufacture a product or system.