Broadcom Reaffirms $100 Billion AI Forecast With Six Customers

Broadcom's AI chip revenue jumped 143 percent to $10.8 billion in fiscal Q2, and CEO Hock Tan reiterated that fiscal 2027 AI sales will top $100 billion, carried almost entirely by six hyperscale customers.

"Q2 semiconductor revenue from AI of $10.8 billion grew 143 percent year over year, above our forecast, driven by increasing demand for custom AI accelerators and AI networking," Tan said on the June 3 earnings call.

Total revenue reached $22.187 billion, up 47.9 percent, with non-GAAP EPS of $2.44 and free cash flow of $10.262 billion, equal to 46 percent of revenue. Q2 AI bookings ran over $30 billion against the $10.8 billion shipped, and Q3 guidance calls for about $16 billion in AI semiconductor revenue, a 200 percent-plus jump. Tan said visibility now extends into 2028.

The concentration risk is the counterweight. A handful of hyperscalers — Google, Meta, ByteDance, OpenAI and Anthropic among them — carry nearly all of the AI number, so a single pullback bends the growth curve. Broadcom shares fell 12.59 percent to $418.91 the day after the report before recovering, with Bank of America and Jefferies raising targets to $530 and $550.

The Custom-Silicon Wedge Into Nvidia's Turf

Broadcom's growth is the clearest evidence that hyperscalers no longer want to buy every AI chip from Nvidia. The company designs application-specific integrated circuits, or ASICs, built to a single customer's workload, in contrast to Nvidia's general-purpose GPUs that run on the CUDA software stack. Google's TPU line is the best-known chip family built through this partnership, and Alphabet has begun selling TPUs to outside customers, putting it in direct competition with Nvidia while enlarging Broadcom's addressable market.

The economics favor custom designs at hyperscale. A chip tuned to one company's model architecture strips out unused features and lowers the cost of each unit of compute, which matters for inference workloads run billions of times a day. The tradeoff is flexibility: Nvidia's CUDA platform carries over a decade of developer tooling, while a custom XPU locks a customer into a narrower software path. For a company running the same handful of models across a million-chip fleet, that cost is often worth paying.

Nvidia is not losing ground in absolute terms. Its data center segment generated $75.2 billion in fiscal Q1 2027, up 92 percent, roughly seven times Broadcom's $10.8 billion AI figure. But Broadcom's 143 percent growth rate, more than 50 percentage points faster than Nvidia's, shows where the incremental AI infrastructure dollar is heading. AMD's data center unit, at $5.8 billion, trails both.

Six Customers, One Growth Curve

The concentration question is the one analysts keep circling. Broadcom does not break out AI revenue by customer, and Tan has declined to confirm names on earnings calls, but reporting ties the custom-chip pipeline to at least five hyperscalers: Google, Meta, ByteDance, OpenAI and Anthropic. A large-scale compute agreement connected to Anthropic adds a fifth major name to a list that had centered on Google's TPU program, and the company describes "several" additional prospective customers without naming them.

The backlog softens the risk. Custom silicon orders lock in wafer capacity, advanced packaging slots and high-bandwidth memory more than a year ahead of delivery, making a $30 billion booking book harder to unwind than a quarterly sentiment shift. Tan has now repeated the $100 billion fiscal 2027 figure across back-to-back earnings calls, and the full-year 2026 AI target of $56 billion is already largely contracted.

Broadcom trades at roughly 23 times forward earnings against triple-digit AI growth, a multiple that looks reasonable if the backlog converts on schedule. The stock carries 44 Buy ratings against zero Sells with a consensus target near $528, versus the $418.91 close after the earnings selloff. The next test comes August 26, when Nvidia reports fiscal Q2 results; for Broadcom, the $16 billion Q3 AI guide is the tell. A beat would confirm the custom-silicon narrative; a miss would puncture it.

This article is for informational purposes only and does not constitute investment advice.