A Compute Host
▶ See it in 3D The CXL Fabric region
One server doing the actual work — a GPU box or a CPU box. The catch: the memory inside it is bolted in and can't be shared, so one machine starves while its neighbour wastes hundreds of idle gigabytes. CXL is the plumbing that lets the starving one borrow. Every box that runs short of memory is, ultimately, a customer for everything else on this map.
What it is
A host is one server doing the work — either a GPU box (8 accelerators, fast but memory-starved) or a CPU box (lots of slow DRAM). The problem: each one's memory is locked inside its own chassis. A GPU server can run out of HBM mid-job while the CPU box next to it has hundreds of idle gigabytes — stranded and unreachable.
The business
Every memory-starved host is a customer for this whole atlas. It buys more HBM (SK Hynix/Micron) when it can, and reaches through a CXL switch (Astera) to borrow from a shared pool when it can't. The demand signal that pulls the entire supply chain starts here — at the box that needs more memory than it owns.
Bottleneck severity
Who makes it
NVIDIANVDA
Buys the most HBM, NVLink and optics on earth and pre-buys supply to lock rivals out: ~60% of CoWoS, $2B each into Coherent and Lumentum (Mar 2026), $2B into Marvell. B300 (288 GB HBM3E) is the shipping reference GPU; Rubin (288 GB HBM4, 22 TB/s) samples Q4 2026, volume Q1 2027.
AMDAMD
MI-series GPUs with leading HBM capacity; the anchor of UALink. Zen 6 EPYC "Venice" launched July 2026 with roughly a year before Intel answers; MI400 carries the first UALink production silicon (H2 2026).
Astera LabsALAB
The only large-cap pure-play on AI connectivity — Aries retimers, Leo CXL controllers, Scorpio fabric switches. Q2 2026 revenue $392M (+104% YoY) as Scorpio X entered volume; Q3 guided $540–560M. Building NVLink Fusion hybrid racks for 2027. Risk: one customer is most of revenue.
SK Hynix000660.KS
#1 HBM maker (~62% share). Qualified HBM4 at NVIDIA first and holds an estimated 60–70% of Rubin's HBM4 allocation, with a TSMC-made logic base die. Sold out through 2027; its CEO calls 2027 potentially the tightest supply year the industry has seen. The bellwether of the whole cycle.
MicronMU
#3 HBM, the only US-HQ DRAM maker. Qualified on Rubin HBM4; sold out through 2027 and meeting only ~50–65% of what customers ask for. New Idaho capacity doesn't land until 2027–28 — so the squeeze is structural, not seasonal.
Metrics that move this layer
- memory per GPUHBM4 = 288 GB, still not enough
- pooling attach% of racks wired for CXL
- stranded DRAMthe waste CXL recovers