AI Infrastructure in 2026: Memory Bottleneck & Optical Interconnect Shift

The evolution of infrastructure has always been pivotal in transforming industries. Much like how Samuel Insull's efforts in the 1890s shifted the focus from coal to electricity distribution, today, the world of Artificial Intelligence (AI) is experiencing a similar shift in its core infrastructure demands. This modern transformation is driven by the increasing need for advanced memory and interconnect technology to support AI's growing computational demands.

By the 1920s, American industry shifted its capital spending from boilers to the expansive reach of high-voltage transmission lines. Now, in the AI realm, the limitations of memory are being addressed, but a new challenge emerges: the interconnect. As AI continues to advance, the supply chain adjusts, sensing a persistent shortage but with marginal easing. This transition highlights a new focus area: the connectivity between powerful server clusters and specialized chips.

Memory Constraints and Price Dynamics in AI

The surge of high-bandwidth memory (HBM) has become a crucial element in AI chips, particularly with graphics processing units (GPUs). Nvidia's progression from 80GB HBM in the H100 to a staggering 288GB in the Blackwell Ultra exemplifies this trend, marking a 3.6x increase over two generations.¹ As capacity grows, so does cost, with companies like Samsung and SK Hynix raising HBM3E prices by nearly 20%. Market forecasts predict that sixth-generation HBM4 stacks could surpass $500-600 per unit, significantly up from their predecessors.² This rapid growth is expected to propel the HBM market from $35 billion in 2025 to $54.6 billion in 2026, potentially reaching $100 billion by 2028.³

Dynamic random access memory (DRAM) also plays a parallel role, essential for AI servers that require high quantities for various computational tasks. As AI applications expand, the demand for DDR5 memory—used for preprocessing and coordination—surges by 50% quarter over quarter.⁴ However, the shift towards more valuable HBM has led to the crowding out of standard DRAM, with each gigabyte of HBM consuming significant manufacturing capacity.⁵ This competition has led to unexpected price inversions, where consumer DDR4 prices briefly exceeded that of HBM3E.⁶

The volatility in memory prices influenced the stock market, with significant fluctuations in semiconductor stock values. Despite these changes, the long-term supply-demand dynamics are projected to remain tight, with capacity booked through 2027 and contracts extending into 2028.⁸ This outlook has influenced KraneShares' AI ETF AGIX to adjust its portfolio to capitalize on these shifts, including the addition of KIOXIA to fortify its position in memory stocks.

The Interconnect Challenge in AI Infrastructure

As AI technology advances, the demand for efficient interconnect solutions becomes evident. Unlike previous eras where network speeds doubled every few years, AI networks require a tenfold increase annually.¹¹ However, the growth in electrical interconnect bandwidth falls short, lagging behind advancements in computational power measured in floating-point operations per second (FLOPs).¹² This growing gap emphasizes the necessity for improved data throughput in AI networks.

The nature of data traffic has shifted significantly, with east-west (datacenter-to-datacenter) traffic surpassing north-south (datacenter-to-consumer) exchanges. By June 2026, bot and AI-agent traffic accounted for 57.4% of internet activity, marking a significant rise from the previous year.¹³ The intensive machine-to-machine communication in datacenters emphasizes the need for robust interconnect solutions.

The physical limitations of copper wiring are becoming apparent as rack scales grow, necessitating a transition to active copper cables or optical interconnects that leverage light for data transmission. This shift is set to transform the structure of AI data centers, with projections indicating a rise in co-packaged optics (CPO) penetration from 0.5% in 2026 to about 35% by 2030.¹⁴ Though the optical interconnect market is growing, it remains a smaller component of AI data-center expenditure compared to memory.¹⁶

In response, KraneShares' AI ETF AGIX has reallocated funds towards optical interconnects, including companies like Applied Optoelectronics and Marvell. Additionally, KraneShares introduced the LUMA ETF, focusing on photonics and optics, as demand for optical solutions continues to grow. Recent corporate earnings highlight the sector's potential, with companies like Lumentum and Corning reporting substantial revenue increases driven by AI-related products.¹⁸ This trend suggests a promising future for optical interconnect technology in meeting AI's evolving demands.

The evolution of data center infrastructure continues to mirror historical shifts seen in other industries, with the potential to define the future landscape of AI technology. As KraneShares monitors these developments, its investment strategies aim to align with the changing dynamics in AI infrastructure.

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.