Broadcom vs SK Hynix

Custom Silicon vs. Memory: Two Sides of the AI Chip

Broadcom and SK Hynix are not direct competitors — they are complementary suppliers occupying different layers of the AI infrastructure stack. Broadcom designs the logic chips (custom AI accelerators and network switches), while SK Hynix manufactures the memory chips (HBM) that are physically stacked onto those accelerators. Every custom XPU that Broadcom designs for Google or Meta needs SK Hynix HBM to deliver its promised performance.

This symbiotic relationship means both companies benefit from the same macro trend — the massive buildout of AI training and inference clusters by hyperscale cloud providers. However, their economics differ sharply. Broadcom captures value through design IP and long-term contracts, earning high margins on fabless chip design. SK Hynix captures value through manufacturing scale and process technology leadership, earning margins that fluctuate with memory pricing cycles but have been historically strong during the current HBM supercycle.

The Hyperscaler Custom Silicon Wave

Broadcom's most strategically important business is designing custom AI accelerators for hyperscale customers who want alternatives to NVIDIA GPUs. Google's Tensor Processing Units (TPUs) are the most mature example — Broadcom has co-designed multiple generations of TPUs that power Google's AI training and inference workloads. In 2025, Broadcom expanded this model dramatically, signing multi-year deals with Meta, OpenAI, and reportedly Anthropic to design bespoke AI chips.

The scale of this opportunity is staggering: Broadcom's total AI-related order backlog reached $73 billion, with the OpenAI collaboration alone targeting 10 gigawatts of AI accelerator deployment. This custom silicon approach trades NVIDIA's general-purpose flexibility for workload-specific optimization, and Broadcom is the primary design partner making it happen.

SK Hynix benefits from this trend regardless of whether customers choose NVIDIA GPUs or Broadcom-designed custom chips — all of them need HBM. In fact, the proliferation of custom accelerators may increase total HBM demand beyond what NVIDIA alone would consume, making SK Hynix a picks-and-shovels play on the entire AI accelerator market.

HBM: The Chokepoint SK Hynix Controls

SK Hynix's dominance in High Bandwidth Memory represents one of the most consequential bottlenecks in the AI supply chain. With 62% of HBM shipments as of mid-2025, SK Hynix effectively controls the pacing of AI accelerator production worldwide. The company was first to market with HBM3E and first to complete HBM4 development in September 2025, maintaining a generational lead over Samsung and Micron.

HBM4 represents a significant architectural leap: 2,048 I/O channels (double the previous generation), over 40% improved power efficiency, and up to 48GB capacity in the 16-layer configuration showcased at CES 2026. SK Hynix has already delivered paid HBM4 samples to NVIDIA, with mass production contracts expected to finalize in early 2026. The company also plans ~20% price increases on HBM3E for 2026, reflecting sustained demand that continues to outstrip supply.

While Samsung is aggressively expanding HBM capacity and Micron has gained ground (overtaking Samsung in HBM market share), Goldman Sachs projects SK Hynix will maintain over 50% total HBM share through at least 2026. This structural advantage gives SK Hynix pricing power that is unusual in the historically commoditized memory industry.

Networking: Broadcom's Other AI Moat

Beyond custom accelerators, Broadcom dominates AI datacenter networking through its Memory Tomahawk switch silicon line. The Tomahawk 6, delivering 102.4 terabits per second of switching bandwidth, has become the standard for connecting thousands of GPUs in large-scale AI training clusters. Broadcom's AI switch backlog exceeds $10 billion, and the company's Ethernet-based networking solutions are increasingly competing with NVIDIA's InfiniBand for AI cluster interconnects.

This networking position is strategically important because AI cluster performance depends not just on accelerator compute power, but on the fabric connecting accelerators together. As clusters scale toward millions of GPUs, networking becomes a larger share of total infrastructure cost and a bigger determinant of training efficiency. SK Hynix has no presence in networking, making this a dimension where Broadcom's value proposition extends well beyond chip design.

Software and Diversification

Broadcom's $69 billion acquisition of VMware in 2023 transformed it into a hybrid semiconductor-software company. The infrastructure software segment generated $27 billion in fiscal 2025 revenue at 77% operating margins, providing a high-margin recurring revenue stream that insulates Broadcom from semiconductor cyclicality. VMware Cloud Foundation adoption is accelerating as enterprises modernize their cloud infrastructure.

SK Hynix lacks this kind of diversification buffer. While the company has expanded into enterprise SSDs and is developing computational storage and processing-in-memory technologies, its revenue remains overwhelmingly tied to memory chip pricing cycles. The current AI-driven supercycle has been extremely favorable — operating profit doubled year-over-year to KRW 47.2 trillion in 2025 — but memory markets are historically volatile, and SK Hynix's margins will compress when the cycle eventually turns.

Risk Profiles and Dependencies

Broadcom's primary risk is customer concentration. Its custom XPU business depends on a handful of hyperscalers, and losing a major design win (or a customer deciding to bring chip design in-house) would be significant. Broadcom also depends on TSMC for manufacturing, adding foundry capacity as a constraint. On the opportunity side, every new hyperscaler that decides to build custom silicon is a potential multi-billion-dollar customer.

SK Hynix faces competitive risk from Samsung's aggressive HBM capacity expansion (reportedly planning a 50% increase in 2026) and the cyclical nature of memory markets. However, the structural shift toward AI workloads has created a more durable demand profile for HBM than traditional DRAM, and SK Hynix's technological lead provides at least a 12–18 month buffer against competitors in each new generation.