1Q26 AI Channel Checks: Pt 3 - NWing & Optics

研报英文原文证据摘录

1Q26 AI Channel Checks: Pt 3 - NWing & Optics

implemented through copper, CPO, NPO, other optical approaches – affects far more than

the link itself

⎯ Decision flows into accelerator tray design, switch placement, cable/optical-engine

location, thermal design, service model, compiler assumptions, collective-

communications libraries, vendor qualification, and multi-year supply commitments

⎯ Thus hyperscaler decisions over the next 12-24mos matter: once the platform is

validated, tooled, qualified and reflected in SW assumptions, switching later becomes

expensive and operationally risky

◼ In our view, this explains why the market may not converge quickly…hyperscalers may each

prefer a “good enough” proprietary or semi-open path over a cleaner but less strategically

favorable industry standard

Protocol Battle: NVLink Real, Ethernet Durable, UALink Still Ecosystem Bet

◼ Three scale-up protocols may be too many

◼ NVLink is already real, deployed, and tightly optimized with NVDA systems, which gives it

the strongest near-term credibility

◼ Ethernet has unquestioned survivability and benefits from large installed base of endpoints

and fact that AVGO has silicon for which SUE protocol can be integrated into

◼ UALink remains attractive because it is open, and backed by AMD & INTC, but needs

credible switch suppliers (multiple)

⎯ This is bit of catch-22, as accelerator vendors need to believe switches will exist while

switch vendors need to believe accelerators will use it

⎯ NVDA’s partnership & investment in MRVL seen as an attempt to shift MRVL’s strategy

towards NVLink ecosystem and by default lower hyperscaler interest in building towards

UALink

CPO vs NPO Debate Rages On

◼ CPO/NPO debate is one that is not purely technical, it is also a debate about system-control