Broadcom (AVGO) targets $230 billion in AI revenue by 2028, and the call turned a 4.6% after-hours drop into a gain

AI chips account for all of the growth

AI semiconductors generated 56% of quarterly revenue, up from 49% one quarter earlier. That percentage is set to rise again. Chief executive Hock Tan said in the release, "In Q4 the momentum continues, and we expect AI semiconductor revenue to accelerate to $21.7 billion, up 236% year-over-year." The forecast equals 62% of Broadcom's $34.8 billion fourth-quarter revenue guide. It also implies approximately $13.1 billion from the rest of the company, little changed from $12.9 billion in the third quarter. In other words, AI is providing essentially all of Broadcom's growth.

The semiconductor solutions segment, which contains the AI chips, posted $20.8 billion of revenue, up 127%. Infrastructure software, which is mostly VMware, added $8.8 billion, up 29%. Free cash flow was $13.7 billion, or 46% of revenue, and Broadcom ended the quarter with $24.0 billion of cash. The company declared its regular quarterly dividend of $0.65 a share, payable September 30 to holders of record September 21.

Chief financial officer Amie Thuener said in the release that Broadcom "achieved record revenue, operating profit and free cash flow in Q3," and that it expects to hold its adjusted operating margin at 66% in the fourth quarter, flat from a year ago. Operating income of $20.1 billion on an adjusted basis grew 92%.

Broadcom extends its AI targets through 2028

On the call, which began at 5 p.m. Eastern, Tan said custom accelerator shipments, the chips Broadcom calls XPUs, rose 3.5 times from a year ago and made up 73% of AI revenue in the quarter, with networking the rest. He raised Broadcom's fiscal 2026 AI revenue target to $58 billion, from the $56 billion given in June. Then he went two years further than he has before: about $115 billion of AI revenue in fiscal 2027, and about $230 billion in fiscal 2028. In June the fiscal 2027 figure was "in excess of $100 billion," with no 2028 number at all. The new targets double each year and land at four times this year's figure in two years.

Tan said supply is already secured for both 2027 and 2028, and that demand currently exceeds the 2027 outlook. Broadcom also set a target of more than $30 a share in adjusted earnings for fiscal 2028. The quarter just reported earned $3.32, so the target implies earnings more than doubling from the current run rate over two years.

Asked how the targets square with the gigawatt figures, Tan said the six customers' plans add up to about 30 gigawatts across fiscal 2027 and 2028, and that Broadcom judges the number conservatively because not all of it will be in production within the two years. He said the simpler way to read it is that the $115 billion and $230 billion add up to about $350 billion of AI semiconductors shipped to those customers over the two years, a figure he said Broadcom holds "with a pretty high degree of confidence." He drew the line between shipping and deploying himself: "It doesn't necessarily mean all 30GW need to have been deployed within that period."

Alphabet, Anthropic, OpenAI and Meta

Tan then walked through the customer list by name, which he has been doing more of each quarter:

• Alphabet (GOOGL): Broadcom began production shipments of Google's next-generation TPU v8i, Tan said, ahead of the v8t variant that MediaTek is building.
• Anthropic: shipments of Ironwood, the current Google TPU generation Anthropic is deploying, accelerate in the fourth quarter. Tan said Anthropic is expected to deploy 5 gigawatts of TPU v8i in 2027 and an incremental 10 gigawatts in 2028, which would make it Broadcom's largest XPU customer in 2027 and keep it there in 2028.
• OpenAI: Broadcom shipped Jalapeño, OpenAI's first custom accelerator, and is on track for 1.3 gigawatts of it in 2027, with more than 5 gigawatts of Jalapeño and successor chips expected in 2028. Broadcom is already developing OpenAI's third-generation XPU.
• Meta (META): Broadcom expects production shipments of Meta's MTIA accelerator at scale in the fourth quarter. It will deliver three generations of MTIA through the end of 2027, with line of sight to 3 gigawatts deployed through 2028.

Add those up and Broadcom has put numbers on more than 24 gigawatts from three customers alone, most of the roughly 30 gigawatts Tan cited across all six. On the networking side, Tan said Broadcom has just taped out Tomahawk 7, which it calls the industry's first 200 terabit per second Ethernet switch.

"AI demand is simply hot, and we're just getting started," Tan said.

Jalapeño versus Nvidia

On Jalapeño, Tan said the chip beats Nvidia's (NVDA) Grace Blackwell Ultra on performance per watt, latency, throughput and power, is comparable to Nvidia's next-generation Vera Rubin on OpenAI's workloads, and that custom XPUs can deliver that at roughly half the cost of a GPU. That is Broadcom claiming, on behalf of a customer, that a chip it physically designed matches Nvidia's best at the workload where most of the money is going. OpenAI published its own benchmarks at Hot Chips on August 25, and SemiAnalysis, which watched the runs, called Jalapeño and Vera Rubin a dead heat on cost per token. We laid out the shift behind this in August: for the first time, more capital is going into running models than training them, and inference is where custom silicon has the best case against GPUs. Nvidia shares were little changed after hours.

The call did what the release could not

What Wednesday did not resolve is the gap between what Broadcom ships and what it books. The company has been reportedly seeking more than $60 billion in financing tied to one AI transaction, and the OpenAI agreement alone calls for 10 gigawatts of accelerators by the end of 2029. A 54% sequential jump in AI revenue and a beat on every line could not clear the expectations already embedded in Broadcom's stock. A $230 billion target for 2028 did, at least as of 5:39 p.m. Eastern.