Broadcom Posts a Record Quarter and Sees AI Chip Revenue Reaching $230 Billion by Fiscal 2028

Broadcom reported record revenue for its fiscal third quarter on September 2, 2026, and delivered a forecast that puts the company on a path to $230 billion in annual AI chip revenue by fiscal 2028, a number that would make its custom-silicon business larger than the entire semiconductor industry was a decade ago. In its earnings release on Sep 2, the company said total revenue reached about $29.6 billion for the quarter ended August 2, up 86 percent year over year, while AI semiconductor revenue hit $16.7 billion, up 221 percent from the same period last year and up 54 percent sequentially, now representing roughly 56 percent of total revenue. Non-GAAP earnings per share rose 96 percent to $3.32, according to the earnings release, extending a run in which Broadcom has become the primary beneficiary of a structural shift in AI computing toward custom chips designed for specific customers rather than general-purpose GPUs.

The quarter is the clearest evidence yet that the AI infrastructure buildout has a second pillar alongside Nvidia. Broadcom designs custom accelerators, which it calls XPUs, for a small group of hyperscale customers, and its disclosures on Sep 2 showed those shipments grew more than 3.5 times year over year and now account for about 73 percent of the company's AI semiconductor revenue, with AI networking revenue more than doubling. Yahoo Finance's Sep 2 transcript of the earnings call recorded management naming Google's Alphabet, Anthropic, OpenAI and Meta among the customers driving demand, with two additional XPU customers left unnamed, and executives describing a supply-constrained environment in which orders exceed what the company can ship. Broadcom generated $13.7 billion of free cash flow in the quarter, a figure that underscores how profitable the custom-silicon model has become once design costs are amortized across large committed volumes.

Key Facts

The forward numbers are where the report becomes extraordinary. Broadcom raised its fiscal 2026 AI semiconductor revenue guidance to roughly $58 billion, up 186 percent year over year, and then laid out a two-year arc that implies the business roughly doubles each year: approximately $115 billion in fiscal 2027 and approximately $230 billion in fiscal 2028, according to the company's Sep 2 release. For the current quarter, the company guided to total revenue of about $34.8 billion, including about $21.7 billion of AI semiconductor revenue, up 236 percent year over year. Management tied the fiscal 2027 figure to secured supply commitments rather than forecasts, and MarketBeat's Sep 2 earnings-call coverage recorded the company describing constraints across high-bandwidth memory, leading-edge wafers, substrates, land, power and data-center readiness.

The capacity math helps explain why the numbers are so large. Broadcom executives said on the Sep 2 call that total customer deployments should reach roughly 10 gigawatts of AI infrastructure in fiscal 2027 and roughly 20 gigawatts in fiscal 2028, with the dollar content per gigawatt staying in a relatively stable range. If those deployment figures hold, the implied revenue per gigawatt is consistent with a market in which each gigawatt of AI data center capacity requires several billion dollars of custom silicon, networking and packaging. Investors.com reported on Sep 3 that the scale of the forecast surprised even bullish analysts, several of whom had modeled fiscal 2028 AI revenue closer to $180 billion before the company raised the target to $230 billion.

The report also clarified the shape of Broadcom's customer base. Unlike Nvidia, which sells the same accelerator to thousands of buyers, Broadcom's model depends on a handful of customers each willing to commit to enormous, multi-year design and supply programs. The company said its two unnamed XPU customers are progressing through design cycles, and that the mix of demand now spans both the largest US cloud operators and the AI labs that have become the fastest-growing buyers of compute. The result is a business with high concentration risk but also high visibility, because revenue is contracted years in advance through design wins that are expensive for customers to abandon.

Analysis

What this really means is that the AI chip market is bifurcating faster than most investors expected, and Broadcom now owns the non-Nvidia half of it almost outright. The conventional narrative of the past two years was that Nvidia's GPUs were the default engine of AI and that custom silicon was a marginal experiment for a few cost-conscious hyperscalers. The $230 billion fiscal 2028 forecast inverts that story: it says the largest AI buyers, the ones with the most demanding workloads and the most price sensitivity, are systematically moving to custom accelerators that they co-design, and that this segment will be large enough to support a second supplier with a market cap to match. For Nvidia, the implication is that its most important customers are becoming its competitors by proxy, designing chips with Broadcom that replace the GPUs they would otherwise have bought.

The bigger picture here is about who actually captures the value of the AI buildout. Nvidia's model captures margin from every AI deployment because it sells a standardized product to everyone. Broadcom's model captures margin differently, by selling a small number of customers a product that is more efficient for their specific workloads but that required years of joint design investment. The custom approach only makes sense at enormous scale, which is why Broadcom's customer list reads like a roll call of the companies building the largest AI clusters on earth. The 10-gigawatt and 20-gigawatt deployment figures are the real headline of this report, because they imply that the AI infrastructure buildout has not peaked, that the demand is still accelerating, and that the bottleneck is not orders but the physical capacity to produce chips, package them and power them.

There is a reason the stock reaction to the report was muted despite the record numbers. Investors have learned to treat AI chip forecasts with skepticism after two years of aggressive guidance revisions, and Broadcom's own history includes quarters where supply constraints pushed revenue into later periods. The credibility of the $230 billion figure now rests on whether the two unnamed customers and the existing hyperscalers actually deploy the gigawatts they have signaled, and whether the supply chain, from TSMC's leading-edge capacity to high-bandwidth memory from SK Hynix, Micron and Samsung, can scale in step. If even one major customer slips its deployment schedule, the doubling trajectory breaks, and the downside to a stock priced for flawless execution would be sharp.

Why It Matters

For the semiconductor industry, the forecast resets the planning horizon for the entire supply chain, since $230 billion of custom AI silicon by 2028 implies commensurate demand for leading-edge wafer capacity, advanced packaging, high-bandwidth memory and networking silicon, and every supplier from TSMC to the memory makers will make investment decisions based on this trajectory. For Nvidia, the report is a competitive warning that its largest customers are voting with their design dollars for alternatives, which will put pressure on Nvidia to make its own accelerators cheaper and faster to retain share of the most lucrative workloads. For AI labs such as Anthropic and OpenAI, custom silicon offers a path to lower cost per token and more control over their infrastructure, which strengthens their economics as they prepare for public markets. For investors, the distinction between the two AI chip models, standardized GPUs versus custom XPUs, is becoming the central valuation question in the semiconductor sector.

Next Up

In the coming months, watch for the two unnamed XPU customers to be identified, because their design progress will determine whether the fiscal 2028 target is credible, and for any announcements about new design wins that would extend Broadcom's customer list beyond the current six. Watch also for the supply-side response, particularly capacity commitments from TSMC and the memory makers, since the 20-gigawatt fiscal 2028 deployment figure depends on the industry adding physical capacity at a pace it has never achieved. The key near-term test will be Broadcom's fiscal fourth-quarter report, when investors will look for evidence that the $21.7 billion AI revenue guide was met and that the growth trajectory into fiscal 2027 remains intact.