✦ The Future BetsNarrow moat

Arista's future bets all answer the same question: after winning the argument that AI clusters should run on Ethernet, how much of the AI data center can one Ethernet company own? The bets march outward and inward at once — across buildings, into the rack, and out to the branch office.

The near bet is already guided: an AI-networking book targeted at roughly $3.6 billion for 2026¹, with a new domain — 'scale-across,' the fabrics that lash separate AI data centers into one training machine — expected to carry nearly a third of it². The iron bet shipped in June 2026: the 7060XE7 series, Arista's first 1.6-terabit-per-port generation, built on Broadcom's Tomahawk 6 and offered, notably, with EOS or an open operating system³. The long bet is scale-up — the network inside the rack, today owned by Nvidia's NVLink — where Arista joined an industry consortium (ESUN) to drag even that domain onto Ethernet⁴. And the sideways bet is VeloCloud, the SD-WAN business bought from Broadcom in mid-2025, which points the company at the one network it never touched: the branch office⁵.

The pattern is confidence expressed as geography. Every bet extends the same two assets — EOS and the merchant-silicon partnership — into an adjacent room, on the theory that the protocol war is over and Ethernet won. That theory's strongest evidence is the enemy's behavior: even Nvidia now sells Ethernet and sits inside ESUN.

Grade each room separately. The $3.6 billion AI book grades against its own target every quarter. The 1.6T generation grades on shipments from Q4 2026. Scale-up grades slowly — it is a 2027-and-beyond market, and the incumbent's lock is silicon, not sentiment. VeloCloud grades on whether an AI-era WAN becomes a real revenue line or a strategic hedge. The common falsifier is concentration: the same handful of titans fund every one of these futures, and what two customers giveth, two customers can re-architect away.