SAN JOSE — The number Broadcom (AVGO) needs to defend on September 2 is not $372. It is $16 billion.
That is the AI revenue figure Hock Tan guided for the fiscal third quarter ending in August — a 200%-plus increase year over year, a pace no other semiconductor company in this cycle is running at. Five days before earnings, AVGO trades at $372.64 on Friday, down slightly from Thursday’s close, while the options market is pricing a 12.4% move in either direction when results hit. For a stock sitting at a $1.77 trillion market capitalization, that is not a routine number.
The day’s broader context is not helping. Federal Reserve Chair Kevin Warsh delivered his first major policy address at Jackson Hole on Friday morning, and the tone was mildly hawkish — no rate cut signal, no pivot language. The 10-year Treasury yield held at 4.679%. The Nasdaq fell 0.5%. Broadcom, whose customers are largely flush corporate AI spenders rather than rate-sensitive retail borrowers, absorbed the macro backdrop without a dramatic move.
The structural story heading into earnings is straightforward: Broadcom built a durable advantage in custom AI silicon at the exact moment hyperscalers decided they did not want to be entirely dependent on Nvidia. The company’s XPUs — application-specific accelerators built to a customer’s exact training or inference architecture — now power AI workloads at Google, Meta, Apple, and, under the Jalapeño project, OpenAI. Anthropic has committed $21 billion toward Broadcom-made TPU capacity. That is not a partnership announcement; that is a capital allocation decision.
The fiscal Q2 results, reported in June, established the runway. Total revenue reached $22.2 billion, up 48% year over year. AI revenue came in at $10.8 billion — up 143% from the same period a year earlier. The Q3 guidance of $29.4 billion in total revenue, with $16 billion attributed to AI, implies the AI segment is now more than half of Broadcom’s total business.
The question heading into September 2 is whether Tan can convert the revenue guidance into a structural thesis about margin durability. Custom silicon is inherently a lower-margin business than standard GPUs — each design cycle requires significant non-recurring engineering investment, and the customer base is narrow by definition. Broadcom’s ability to price those engagements while maintaining the 60%-plus adjusted EBITDA margins the company has run for six quarters is the real earnings variable. Revenue beats without margin expansion will disappoint.
The competitive picture is evolving in ways that benefit Broadcom without eliminating risk. Nvidia remains the default choice for general-purpose AI compute — the H200 and GB200 NVLink racks are where most model training still runs. But the hyperscalers have made clear through their capital allocation that custom silicon is not an experiment. It is a procurement strategy. Broadcom controls roughly 70% of the custom AI ASIC market, a share that reflects both technical execution and the difficulty of building the design and packaging relationships that underpin a competitive XPU program.
Intel and AMD are both attempting to close the gap. Intel’s 18A process node is now yielding at rates above 60%, and the company has signaled it is open to foundry relationships with AI ASIC designers. AMD’s Instinct MI400 series targets inference workloads where cost-per-query matters more than peak throughput. Neither threat materializes before Broadcom’s September earnings call, but the competitive trajectory is worth watching across the next four to six quarters.
Wall Street has not pulled back on Broadcom despite the premium valuation. The analyst consensus sits at a Strong Buy, with an average price target of $511 — a 37% premium to Friday’s price. The gap reflects the degree to which the AI revenue ramp is being modeled as durable rather than cyclical. Whether that modeling survives a softer-than-guided Q3 print is the clearest near-term risk.
The Jackson Hole dynamic is worth a brief note. Broadcom’s customer base — Google, Meta, Microsoft, Amazon, Apple, OpenAI, Anthropic — is not interest-rate sensitive in the way that housing or consumer credit is. Their AI capital expenditure is driven by competitive positioning, not financing cost. A hawkish Fed does not halt a hyperscaler’s XPU procurement cycle. What it does do is compress multiples on growth-oriented equities, which is why AVGO sold off slightly on Friday even though nothing about the fundamental business changed.
The Micron earnings date sits two days later, on September 30. The semiconductor sector is running a dense earnings calendar into the back half of 2026, and Broadcom sets the tone for AI-infrastructure stocks when it reports on September 2. A print that confirms $16 billion in AI revenue and holds margins above 60% adjusted EBITDA would likely reprice the sector upward. A miss on either line would reset expectations across the XPU and custom silicon space more broadly.
What Broadcom does not yet know — and what no analyst model can reliably capture — is the pace at which the Jalapeño architecture scales beyond the initial OpenAI deployment. The project was disclosed in broad strokes; the chip count, the power envelope, and the timeline to full production volume remain outside the public record. That gap between disclosed partnership and disclosed shipment data is where the earnings call will either add clarity or defer it.
Related Posts
