2026-06-09 — views $AVGO · Broadcom · Tomahawk / Jericho AI networking
Broadcom's networking quarter: AI silicon hits $10.8B as Tomahawk 6 passes a year of shipping and a 200-terabit switch tapes out
Broadcom's Q2 FY2026 put AI semiconductor revenue at $10.8B, up 143% YoY, with networking nearly 40% of it. Hock Tan said the 100-terabit Tomahawk 6 has shipped for over a year and a 200-terabit successor tapes out this quarter; Q3 AI guidance is $16B, up over 200% YoY.
The quarter in numbers
On June 3, 2026, Broadcom reported its second quarter of fiscal 2026, ended May 3, 2026, and the print was the kind that resets the baseline for everyone selling networking silicon. Consolidated revenue grew 48% year-over-year to a record $22.2 billion. The Semiconductor Solutions segment did $15.0 billion of that, up 79%, while Infrastructure Software added $7.2 billion, up 9%. Free cash flow came in at $10.3 billion — 46% of revenue — and adjusted EBITDA at $15.2 billion, or 69% of revenue.
The line that matters for this section: AI semiconductor revenue of $10.8 billion, up 143% year-over-year and above the company’s own forecast, “driven by increasing demand for custom AI accelerators and AI networking,” in CEO Hock Tan’s words. The guide is even louder. For Q3, Broadcom expects total revenue of roughly $29.4 billion and AI semiconductor revenue of $16.0 billion — growth of over 200% year-over-year for the AI line.
Networking is almost half the AI story
Buried in the call was the number connectivity watchers should anchor on: “Networking represented almost 40% of our Q2 AI revenue.” Run the arithmetic and that is roughly $4.3 billion of switch, fabric, SerDes, and optics silicon in a single quarter — a quarterly run-rate larger than most pure-play networking companies’ annual revenue. Tan did flag that the mix should normalize, with networking’s share of total AI revenue settling “closer to around 30%” — not because networking slows, but because custom accelerators are ramping even faster and inflating the denominator.
The product commentary was specific. Broadcom has been shipping “the industry’s only 100 terabit Ethernet switch, the Tomahawk 6” for over a year, and Tan confirmed the company “will now be taping out our next-generation 200-terabit switch this quarter.” On the routed-fabric side, Jericho 3 and Jericho 4 fabric solutions are enabling “the world’s largest deployments of multiple hyperscalers.” For scale-up inside the rack, Broadcom pointed to its 200G and 400G SerDes enabling co-packaged solutions, and management described the company as “the de facto standard in the industry” for co-packaged optics extending AI clusters across data centers.
The IP read
Three things stand out for anyone tracking the interconnect-IP landscape.
First, the switch cadence is holding. Tomahawk 5 (51.2T) to Tomahawk 6 (102.4T) took roughly the customary two years, and a 200-terabit-class tape-out this quarter keeps the doubling clock running. Every doubling at the switch raises the bar for the SerDes underneath it — and Broadcom owning both the switch ASIC and the 200G/400G SerDes lanes is precisely the vertical stack that standalone PHY and retimer vendors have to route around rather than through.
Second, this is the strongest quarterly evidence yet for the Ethernet-everywhere thesis. The competing scale-up stories — NVLink and its Fusion ecosystem, UALink’s promised open fabric — are architecture announcements and partner lists. Broadcom’s counterargument is a revenue line: roughly $4.3 billion of AI networking in thirteen weeks, shipped, recognized, and guided higher. Hyperscalers building custom accelerators overwhelmingly wire them together with Ethernet, and each new XPU program Broadcom wins pulls its own attach of Tomahawk, Jericho, and optics along with it.
Third, the accelerator-networking flywheel is now explicit in the guidance. Q3’s $16 billion AI number implies the custom-accelerator side is scaling faster than networking, but networking attach is what makes those XPU deals defensible. The fabric is the lock-in.
Practitioner note
If I were planning a 2027 cluster build, the actionable detail here is the 200-terabit tape-out timing. Tape-out this quarter typically means production silicon in volume roughly 12 to 18 months out — call it late 2027 for deployable systems. That gives a clean planning seam: current builds standardize on Tomahawk 6 at 102.4T with 200G SerDes lanes, and the next refresh window aligns with the 200T part, likely pushing per-lane rates and co-packaged optics harder. The question I would put to Broadcom field engineering now: is the 200T generation copper-first with CPO optional, or does the power math at that radix effectively mandate optics? The answer determines whether your next data hall needs a different cabling and service model, and it is worth asking before the rack layouts freeze.
Under-considered angle
The “normalization to 30%” comment is being read by some as networking deceleration. The more interesting reading is the opposite: it is a denominator effect that quietly confirms how large the custom-accelerator ramp is. If networking holds near a $4 billion-plus quarterly run-rate while falling to 30% of AI revenue, the implied accelerator number is enormous — and every one of those accelerators terminates into a fabric someone has to supply. The risk to watch is not demand; it is concentration. A networking line this size is increasingly tied to a handful of hyperscale XPU programs, and a single program slipping a quarter would move the whole segment in a way the old, diversified Tomahawk merchant business never experienced.