Broadcom told investors its AI chip revenue will double in each of the next two fiscal years, reaching roughly $115 billion in fiscal 2027 and $230 billion in fiscal 2028, as custom accelerator deals with Anthropic, OpenAI, Google and Meta move from promise to committed capacity. The forecast, delivered on the company’s fiscal third-quarter earnings call, came with a caveat that soured the market slightly: fourth-quarter revenue guidance of about $34.8 billion fell below Wall Street’s $35.03 billion consensus, nudging shares lower in overnight trading.
The near-term numbers were strong. Third-quarter AI semiconductor revenue more than tripled year over year to $16.7 billion, lifting total company revenue 86 percent to $29.59 billion, ahead of the $29.36 billion analysts expected. Adjusted earnings per share came in at $3.32 against a $3.24 estimate. Bookings for AI chips alone topped $30 billion last quarter, and fourth-quarter AI semiconductor revenue is expected to accelerate to $21.7 billion, up 236 percent year over year.
Committed capacity, not aspiration
CEO Hock Tan went out of his way to frame the 2027 and 2028 numbers as backed by signed supply. “In 2027, we have secured the supply to again double AI revenue to approximately $115 billion,” he said, adding that demand actually exceeds that outlook and that the company will work to improve supply. For 2028, the company has what Tan called “line of sight” to another doubling to $230 billion, with supply secured to meet it.
Tan gave specific deployment visibility: more than 10 gigawatts of Broadcom-designed chips for Anthropic through 2028, over 5 gigawatts for OpenAI, and 3 gigawatts for Meta. Anthropic, he said, is on track to become Broadcom’s largest XPU customer in 2027 and sustain that position in 2028, planning to deploy another 5 gigawatts of TPU v8i chips in 2027 and an additional 10 gigawatts in 2028. OpenAI has a path to more than 5 gigawatts of Broadcom’s next-generation Jalapeno accelerator in 2028.
“That is committed capacity, not aspiration, and it closes most of the gap to what the market wanted.” – Patrick Moorhead, CEO, Moor Insights & Strategy
The framing matters because Broadcom’s custom accelerator business, built on so-called XPUs designed for individual customers, depends on a small number of very large buyers. Tan said the vast majority of AI compute demand still comes from the concentrated group of labs training frontier models, and that their appetite for compute infrastructure is set to inflect even more in 2027 and 2028. Google, OpenAI and Meta are already shipping products on Broadcom-designed silicon, and each new gigawatt commitment locks in years of revenue across design wins, manufacturing and follow-on networking equipment.
Analysts on the call spent considerable time on Tomahawk 6, Broadcom’s high-bandwidth Ethernet switch chip aimed at AI and high-performance data-center networking. Networking rides the same buildout as compute, which is why Broadcom benefits whether a given cluster runs on its accelerators or a competitor’s chips. J.P. Morgan’s Harlan Sur was among the analysts who pressed management on the Tomahawk pipeline during the question session.
The second supplier problem for Nvidia
The forecast underlines how AI infrastructure spending is broadening beyond Nvidia. Broadcom’s accelerators are custom silicon: a customer like Anthropic co-designs a chip tuned to its own models and software stack, then buys it at scale instead of renting Nvidia GPUs. The pitch is cost and power efficiency at enormous volume. Nvidia still dominates the merchant market, and its August quarter showed it: $96.2 billion in quarterly revenue, more than double a year earlier, with $89 billion of that from data centers. But the labs writing the biggest checks increasingly want a second source, both for pricing leverage and for supply-chain insurance.
That dynamic reshapes the semiconductor landscape. A few years ago the question was whether custom chips could match Nvidia’s performance. Now the question is when, not whether, frontier labs deploy them at gigawatt scale, and the answer has moved from 2028 to 2027 for Anthropic. Every hyperscaler running an AI business has a custom silicon program either shipping or in development, and Broadcom is the design partner behind most of them. The company’s financials already reflect the shift: AI semiconductors, once a side business to its networking and broadband lines, are becoming the core of its growth story.
Why the stock shrugged anyway
Despite the headline numbers, the stock dipped because the market has been trained on supernormal expectations. A quarter showing 86 percent revenue growth still missed the consensus for next quarter’s guide, and at Broadcom’s valuation, anything short of perfection gets punished. Analysts also flagged unresolved risks: concentration among a handful of customers, the durability of multi-year commitments from companies whose own revenue is still maturing, and whether the gigawatt figures translate into scheduled orders.
Those risks are real but so far hypothetical. Anthropic’s computing commitments this year alone exceed $135 billion across Lambda, Nscale, AWS and others, and OpenAI has struck similarly sized deals with multiple suppliers including Broadcom and Nvidia. The chip industry is effectively financing the AI buildout alongside the labs, on the assumption that frontier model demand keeps compounding through the decade. If AI revenue growth at the labs slows, these commitments become the kind of stranded capacity the telecom industry lived through in the dot-com era, a comparison skeptics of the buildout raise on every earnings call.
Tan said the company is on target to produce over $30 in earnings per share, well above the LSEG fiscal 2028 consensus of $25.86. Whether that lands depends on customers whose own business plans assume the same exponential curve. For now, both sides of those contracts are betting on each other, and the bets keep getting larger.