The Interconnect War
▶ See it in 3D The CXL Fabric region
NVLink vs UALink vs CXL — whoever wins decides which chipmakers get the design wins for years. Astera Labs is the pure-play that wins whichever way it breaks.
The fight over how GPUs are wired together inside a rack. NVIDIA's NVLink is the locked-in champion; a coalition called UALink wants to pry it open. Whoever wins decides which chipmakers get the design wins for years — and Astera Labs sells the connecting silicon either way, which is why it's the cleanest bet.
What it is
Inside a rack, GPUs must talk to each other and to pooled memory at terabytes per second. Three standards fight for it. NVIDIA's proprietary NVLink is shipping at 1.8 TB/s — the moat. The open UALink consortium (AMD, Broadcom, Intel, Google, Meta) published its 2.0 spec in April 2026 before any 1.0 silicon shipped; AMD's MI400 carries the first in H2 2026, with broad products in 2027. CXL pools coherent memory.
The business
The highest-stakes battleground in AI hardware. NVLink wins → NVIDIA's lock-in holds. UALink wins → Astera (switches), Broadcom (silicon), AMD (competitive GPUs at scale) all benefit. Most likely they coexist — and Astera wins either way as a multi-standard chip house. NVIDIA put ~$2B into Marvell to keep it inside the NVLink camp.
Bottleneck severity
Standards lock-in, not manufacturing: whoever wins NVLink vs UALink decides which chipmakers get the design wins for years. NVIDIA defends the lock (even paying ~$2B to keep Marvell in its camp); everyone else funds the challenger.
Who makes it
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.
BroadcomAVGO
Tomahawk switching + CPO (Davisson 102.4 Tbps) + custom XPUs for hyperscalers. Anchor of the UALink/merchant-silicon camp.
MarvellMRVL
~70% optical-DSP share + custom AI silicon + Celestial AI photonic memory fabric ($3.25B). Most complete optical stack in public equities.
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).
Montage Tech688008.SH
CXL memory controllers + DDR5 RCD/MRDIMM buffers. China's CXL silicon play.
RambusRMBS
Memory-interface IP + chips (CXL, MRDIMM, PCIe). Licensing royalties scale with every DDR5/CXL device.
Metrics that move this layer
- UALink siliconMI400 H2 2026 → products 2027; spec is ahead of silicon
- Astera Scorpio Xin volume since Q2 2026; switch TAM ~10× retimers
- NVLink Fusion breadthAstera building hybrid racks for 2027 — even the challenger sells into the moat
- CXL host CPUsIntel Diamond Rapids slipped to 2027; AMD Venice shipping