Broadcom Guides AI Revenue to $230 Billion by Fiscal 2028, Says Anthropic Will Overtake Google as Largest Custom Chip Customer
Q3 fiscal 2026 earnings call, September 2, 2026
Broadcom used its fiscal third-quarter earnings call to lay out a multiyear AI revenue trajectory that few semiconductor companies have dared to publish, telling investors it expects AI semiconductor revenue to double to approximately $115 billion in fiscal 2027 and double again to $230 billion in fiscal 2028. CEO Hock Tan was explicit that the company is not simply extrapolating trend lines: "This is real demand, we believe, based on what data center sites, locations are ready... against the supply chain we have in leading-edge wafers, substrates and HBM memory. This is, again, a carefully structured outlook that we believe we can achieve." The disclosure marks one of the most concrete multiyear AI capex commitments given by any semiconductor supplier to date, effectively pre-announcing two years of growth well ahead of when investors would otherwise model it.
Fiscal 2026 numbers already running ahead of plan
The forward guidance came on top of a quarter that beat its own targets. Third-quarter revenue hit $29.6 billion, up 86% year-on-year, with AI semiconductor revenue tripling to $16.7 billion. Operating margin reached a record 68% of revenue, and free cash flow hit $13.7 billion, or 46% of revenue. Management now expects fiscal 2026 AI revenue of $58 billion, above the $56 billion figure it had guided to previously, with the fourth quarter alone expected to deliver $21.7 billion in AI revenue, up 236% year-on-year, on total company revenue of $34.8 billion, up 93%. Tan reiterated that the company remains "very much on target to exceed $30 in earnings per share in fiscal 2028," a level that would represent a near-tripling of current annualized EPS.
Anthropic overtakes Google as the customer to watch
The most consequential new disclosure was the customer-by-customer roadmap Tan laid out for the six XPU customers Broadcom serves. Anthropic, which is deploying 1 gigawatt of Ironwood TPUs this year, is expected to scale to 5 gigawatts of the next-generation TPU v8i in 2027 and add another 10 gigawatts in 2028. Tan said flatly that "Anthropic is on track to become our largest XPU customer in 2027 and sustain that in 2028" — a notable shift given Google has been Broadcom's flagship TPU partner for a decade. OpenAI, meanwhile, is ramping its first custom chip, Jalapeno, toward 1.3 gigawatts of deployment in 2027, and Tan said Broadcom has visibility to more than 5 gigawatts of Jalapeno and its successor by 2028, which would make OpenAI the company's second-largest XPU customer. A third-generation chip is already in development with OpenAI, while the second generation is "approaching tape-out."
On the underlying chip economics, Tan made a pointed comparison to merchant GPUs, noting that Jalapeno "outperforms the Grace Blackwell Ultra in performance per watt, latency, throughput and power" while running at "half the cost of a GPU." He framed this as evidence of a broader thesis: "When you co-develop a chip that is optimized for your particular LLM workloads, you will outperform any GPU." Meta's MTIA program remains on schedule too, with three generations of accelerators arriving through 2027 and cumulative deployment of 3 gigawatts by 2028. Google's relationship, Tan said, has "never been stronger," anchored by a new long-term agreement to supply "multi-tens of billions of dollars of TPUs annually" in coming years, with the newest chip, TPU v8i, already shipping in production ahead of MediaTek's competing v8t design — a detail Tan highlighted as evidence of Broadcom's execution edge in a increasingly contested custom-silicon market.
Revenue per gigawatt lands below bullish estimates, but Tan says that's by design
Bernstein's Stacy Rasgon pressed management on unit economics, noting that Broadcom's implied revenue per gigawatt of roughly $11-12 billion looked low next to competitor commentary suggesting figures near $40 billion. Tan and Charlie Kawwas, president of the Semiconductor Solutions Group, clarified that Broadcom's own content estimate is actually $20-30 billion per gigawatt, and explained why it should stay roughly there even as chips get more powerful: as each new XPU generation draws more power per unit, fewer chips fit into a gigawatt of capacity, offsetting rising per-chip ASPs. The bigger driver of revenue growth, Tan said, is the sheer number of gigawatts being deployed, not price per unit. He also cautioned that the 30 gigawatts of cumulative demand visible across customers through 2028 will not fully convert into shipped revenue in that window, since land, power and data center shell construction — not chip availability — increasingly dictate the pace of deployment. Broadcom's $115 billion and $230 billion figures, he said, reflect a "judged" and deliberately conservative view of what will actually ship rather than the full addressable demand.
Gross margin dilution is a feature of the mix shift, not a red flag
CFO Amie O'Toole flagged that consolidated gross margin is expected to fall to approximately 73% in the fourth quarter, down from 78% a year earlier, as XPUs with higher memory content become a larger share of revenue. Operating margin, however, is expected to hold flat at roughly 66% because revenue growth is far outpacing operating expense growth. Tan was blunt about how he wants investors to read this: "Stop focusing on gross margin... look at where it matters, operating margin at the bottom." Semiconductor segment gross margin came in near 67% in the quarter, with Tan and O'Toole both signaling that further mix shift toward XPUs will keep pressuring the headline gross margin metric even as profitability at the operating line continues to expand.
Networking and optical components emerge as a second growth engine
Beyond XPUs, Broadcom detailed rapid traction in AI networking, with revenue up over 2.5 times year-on-year and expected to grow at a similar pace to XPU revenue over the next several years. The company's Tomahawk 6 switch, according to Kawwas, has seen the fastest ramp of any switching family in Broadcom's history and is deployed across nearly all AI hyperscalers building custom silicon, including those not using Broadcom XPUs. The newer Tomahawk Ultra platform, which enables GPU/XPU scale-up over Ethernet rather than proprietary interconnects, is seeing adoption that "has surprised us," per Kawwas, with deployments starting this quarter. Broadcom also disclosed it has taped out Tomahawk 7, a 200-terabit-per-second switch, positioning it ahead of the field in scale-out networking silicon. On the optical side, Tan said demand for EML and CW lasers is "far surpassing supply" industry-wide, prompting the company to more than triple capacity at its indium phosphide fabs in the U.S. and Singapore over the next two years — a capacity buildout that shows up directly in the quarter's step-up in capital spending.
Financing structure for Anthropic and OpenAI stays opportunistic, not formulaic
Analysts pushed for more detail on the AI XPV platform, the vehicle Broadcom established with Apollo and Blackstone in June to fund more than 20 gigawatts of compute for OpenAI and Anthropic through 2028, with a first $35 billion tranche already closed for Anthropic's 1-gigawatt buildout. O'Toole declined to specify how much of the remaining 15 gigawatts across both customers would be financed this way, saying only that "any that we do in the future is going to have unique features and be tailored specifically to the lab and to the investor needs." Tan added context on why Broadcom is willing to extend financial support to just two of its six XPU customers: Anthropic and OpenAI are, in his words, akin to "two geniuses in the middle of Outer Mongolia" who need help reaching scale, while the other four customers are financially self-sufficient. He characterized the economics as attractive on their own terms, noting each gigawatt of compute those labs deploy can generate roughly $30 billion of annualized revenue for them, calling it "a hell of a business model" worth backing.
Land, power and shell — not chips — are the real bottleneck heading into 2027
Multiple analysts, including Goldman Sachs' Jim Schneider, asked what could actually constrain the multiyear guidance. Tan was candid that land, power and shell availability — the physical data center infrastructure needed to house the chips — is now a bigger constraint than silicon supply itself, saying it "dictates specific timing of when this capacity gets deployed and be available." He described the planning process as a close collaboration with customers to validate real-world construction timelines before those volumes are baked into guidance. Substrate capacity is another pressure point Broadcom is addressing directly, with a new substrate fab in Singapore set to begin production in fiscal 2027 to relieve what Tan called "a key part of our supply bottlenecks." Taken together, management's message was that chip design and manufacturing are no longer the limiting factor in AI infrastructure buildouts — physical infrastructure and power availability now are, a dynamic likely to shape capital spending and site-selection decisions across the hyperscaler and AI lab ecosystem well beyond Broadcom's own supply chain.