L2 · CHOKEPOINT
Compute Hardware
This is the layer most people picture when they think about the AI supply chain: the chips, servers, networking gear and memory that actually do the computing. NVIDIA's GPU dominance is the headline, but the full picture includes custom accelerators built by Google and Amazon, high-bandwidth memory from SK Hynix and Micron that stacks inside each GPU package, and the optical interconnects that wire racks together at speed. Each of those components has its own tight supply chain.
WHY IT'S A CHOKEPOINT
The merchant GPU gets the headlines, but HBM memory and the lasers and optical components feeding the interconnect are the tighter constraints. Behind every GPU order sits the in-rack power-delivery chain, now moving to 48V and 800V.
Signals
- SK Hynix targets ~200k wafers/month of HBM by 2026, with output fully sold out (TrendForce).
- HBM4 prices ~29% above HBM3E and HBM4e ~61% higher, reflecting severe supply tightness (TrendForce).
- TSMC CoWoS reaches ~125-130k wafers/month by end-2026 (toward ~220k by 2028), yet a 15-20% supply gap persists (TrendForce).
- Marvell holds ~70% of 800G optical DSPs and ~50% of 1.6T, the interconnect feeding the GPU (TrendForce).
The investment angle
HBM is structurally supply-constrained through at least 2027, giving SK Hynix and Micron pricing power; in-rack power delivery (48V/800V) is the overlooked bottleneck, with exposure via Vertiv, Eaton and Infineon.
Inside this layer, node by node
The atlas data behind this layer: 52 nodes, 11 of them chokepoints. Every node links back into the network map; market figures carry their source.
Chips purpose-built for parallel matrix and tensor workloads in AI training and inference, spanning merchant GPUs, hyperscaler custom ASICs, and alternative architectures. These dominate both capital spend and power draw in AI data centres. NVIDIA captures ~80% of merchant GPU revenue; Broadcom and Marvell capture design-partner margins on custom ASICs.
General-purpose processors handling OS, orchestration, and non-tensor tasks in AI servers. Required to run accelerator platforms but not the primary compute element. Commoditised in AI servers; Intel under margin pressure from AMD, Arm server CPUs gaining share via hyperscaler custom designs.
PCIe cards connecting servers to network fabrics; SmartNICs and DPUs embed processors to offload storage, security, and networking tasks from host CPUs. Offload reduces CPU overhead in AI clusters running distributed training and inference. NVIDIA BlueField/DOCA SDK locks in cloud providers; Marvell OCTEON leads storage offload.
Memory hierarchy from HBM stacked on accelerators through server DRAM, CXL expanders, and experimental persistent memory. Bandwidth and capacity constraints directly throttle model size and training throughput. SK hynix captures highest HBM margin; Samsung and Micron compete on DRAM and next-gen HBM4.
SSDs, NAND controllers, and mechanical drives for dataset staging, checkpoints, model repositories, and archival. Training throughput depends on feeding GPUs fast enough from storage; checkpoints protect expensive compute state. NAND oligopoly controls flash pricing; controller IP held by specialist firms.
Hardware connecting AI accelerators within and between servers, racks, and pods. Bandwidth and latency here determine how large a model cluster can train as a single unit. NVIDIA controls GPU-to-GPU fabric and InfiniBand; Broadcom dominates Ethernet switch silicon.
Server and rack-level systems integrating accelerators, CPUs, memory, and networking. Physical density, thermal limits, and interconnect topology determine how AI capacity deploys. ODMs (Quanta, Foxconn, Wistron) capture assembly margin on GB200 NVL72 racks; hyperscalers are the primary buyers.
Power conversion and distribution from rack input down to silicon voltage rails. Efficiency losses here directly increase total datacenter power draw and operating cost. Infineon and Delta won the bulk of GB200 power delivery, with MPS, TI, Renesas and Vicor competing for sockets; MLCC shortage creates upstream pricing power.
Companies we track
Supply chain
Raw inputs
Key suppliers
Buyers