Broadcom Q3 2026: $29.6B Revenue, Custom AI Chips Driving 86% Growth

Broadcom reported third-quarter fiscal 2026 revenue of $29.59 billion on September 2, beating the $29.36 billion Wall Street expected and marking 86% growth from the $15.95 billion it posted a year ago. The company's custom AI accelerator business is expanding rapidly, with chips shipping to Google, Meta, OpenAI, and Anthropic. Despite the strong quarter, guidance for Q4 came in at a $34.8 billion midpoint, short of analyst estimates, pulling the stock down about 2% in after-hours trading.

What happened

Metric	Value
Q3 2026 revenue	$29.59B (up 86% year-over-year)
Q3 2025 revenue (same period)	$15.95B
Adjusted EPS	$3.32 (vs. $3.24 expected)
Net income	$13.09B (up from $4.14B a year earlier)
Semiconductor unit revenue	$16.7B (vs. $15.2B estimate)
Infrastructure software revenue	$8.75B (vs. $8.82B estimate)
Q4 2026 revenue guidance (midpoint)	$34.8B (vs. $35.03B expected)
Market capitalization	Over $1.7 trillion
Stock performance, year to date	Up 6% (vs. S&P 500 up 12%)

Broadcom designs custom AI accelerator chips (processors built to a client’s specific requirements, rather than general-purpose GPUs) for several of the largest AI labs. That business is now the clearest driver of its semiconductor division’s revenue, which tripled from a year ago to $16.7 billion.

Who is buying Broadcom’s chips and what are they getting?

CEO Hock Tan detailed the company’s customer pipeline on a call with analysts. The roster covers some of the biggest names in AI infrastructure spending.

• OpenAI: Broadcom developed the Jalapeno custom processor for OpenAI. Shipments are ongoing, and OpenAI is already taping out a second-generation Jalapeno while engineers are working on concepts for a third. In 2027, Tan expects OpenAI to deploy about 1.3 gigawatts of Jalapeno capacity, with a path to more than 5 gigawatts beyond that.
• Google and Anthropic: Broadcom is accelerating delivery of Google’s Ironwood tensor processing units (TPUs, Google’s in-house AI chips) to both Google itself and to Anthropic, which licenses them. Tan said Broadcom expects to ship “tens of billions of dollars worth of processors” to Google per year for the next several years. Anthropic is projected to deploy about 5 gigawatts of Google’s TPU8i chips next year, with a path to an additional 10 gigawatts after that.
• Meta: Production shipments of Meta’s MTIA (Meta Training and Inference Accelerator) chip, optimized for inference and recommendation workloads, are also expected to continue.
• Apple: Apple committed additional investment toward Broadcom’s U.S. chip production operations during the quarter.

Why it matters

Broadcom’s numbers confirm that the custom silicon market is not a side bet. AI labs are moving aggressively away from general-purpose GPU clusters and toward chips designed specifically for their own architectures. For Broadcom, that shift is worth an 86% revenue jump in a single year.

Tan’s forward projections are striking. He expects Broadcom’s AI revenue to reach $115 billion in fiscal 2027 and then double again to $230 billion the year after. He also called for earnings of $30 per share in fiscal 2027. Those numbers, if correct, would mean AI chip revenue alone will grow larger than Broadcom’s entire company revenue was just two years ago.

The company also disclosed something unusual: it is considering providing “residual value guarantees” to two of its most strategic AI lab customers. These are contingent liabilities, meaning Broadcom would backstop some of the financial risk those labs take on when committing to large-scale chip infrastructure. CFO Amie Thuener framed it as bridging the gap between the labs’ current cash flow and the heavy upfront costs of building out AI compute. That is a significant shift from supplier to financial partner.

For anyone tracking the pace of AI infrastructure investment, Broadcom’s numbers are a useful ground-truth signal. The labs are not slowing spending. They are locking in multi-year chip supply chains with their foundry partners.

Our take

The headline growth is real and the customer list is credible. But there are a few things worth watching carefully here. First, Tan’s $115B and $230B AI revenue projections are management forecasts, not analyst consensus, and they assume sustained, accelerating deployment by a handful of very large clients. If any one of those clients hits a funding wall, delays a product, or shifts strategy, Broadcom’s numbers shift too. Concentration risk is high.

Second, the residual value guarantee angle is genuinely new and deserves attention. Broadcom acting as a de facto financial backstop for AI labs signals that those labs are stretching their balance sheets to compete on compute. It also means Broadcom is taking on exposure it never had as a pure chip designer. That is a different kind of company than it was two years ago.

For businesses that rely on AI APIs built on top of these chips (which is most businesses using AI tools today), the practical takeaway is that the underlying compute layer is expanding fast and is being funded aggressively. That should keep AI service capacity growing, and it likely keeps pricing competitive, at least in the near term. If you are thinking about integrating AI into your own workflows, the infrastructure to support it is being built at scale.

What to do about it

1. Watch Broadcom’s Q4 2026 earnings closely. If actual revenue clears $34.8 billion, Tan’s fiscal 2027 projections become more credible.
2. Track Anthropic’s TPU8i deployment timeline. Five gigawatts is a large commitment; any delays would signal broader cooling in AI infrastructure spend.
3. If you are evaluating AI tools or vendors, factor in that compute supply is expanding rapidly. Pricing pressure on AI APIs is likely to continue downward.
4. Keep an eye on whether Broadcom formalizes the residual value guarantees. If it does, read the structure carefully: it will tell you a lot about how confident both sides are in near-term AI demand.

The biggest risk here is not Broadcom’s execution. It is whether the AI labs it serves can keep translating compute into revenue fast enough to justify the scale of infrastructure they are committing to.

Source: Bing News · OpenAI