Broadcom's $230 Billion AI Target: The Growth Is Real. The Question Is Whether the Stock Price Can Keep Up.

On September 2, Broadcom reported AI semiconductor revenue that had tripled to $16.7 billion and laid out a plan to reach $230 billion by 2028. Six named customers. "Simply hot" demand, in CEO Hock Tan's words. A $35 billion financing platform with Apollo and Blackstone to build the compute infrastructure behind it.

The stock fell 4% after hours.

The decline had nothing to do with that $230 billion figure. It was the Q4 revenue guide of about $34.8 billion, roughly $300 million below consensus expectations. On a company that large, $300 million is a rounding error. But the market reads these misses as signals, not arithmetic. And here is the signal: Broadcom is growing so fast that the tolerance for a quarterly shortfall is vanishing.

That is the tension this earnings report reveals. Not whether the long-term picture is real, but whether the near-term mechanics can carry the weight of what the market has already priced in.

Broadcom's custom-chip business has become a machine. In fiscal Q3, AI semiconductor revenue reached $16.7 billion, up 221% year over year. Custom accelerators alone made up 73% of total AI revenue — about $12.2 billion from six customers who design their own chips and hire Broadcom to build them. These are not generic GPU purchases like the ones Nvidia sells. Each chip is an application-specific integrated circuit, tailored to a specific company's model architecture, trained on their data, optimized for their workloads.

The economics of this arrangement are straightforward. A custom chip costs less than half of what an equivalent Nvidia GPU costs for the specific workload it was designed for, as Tan put it on the call. That is the entire value proposition for customers. If you are OpenAI running inference on GPT models, or Anthropic running Claude, a chip designed for exactly that job costs less in silicon, power, and rack space than a general-purpose GPU forced to do it. The more inference workloads dominate the AI cycle, the more the custom-chip model wins on unit economics.

Tan said AI demand is "simply hot" and guided to approximately $58 billion in fiscal 2026, roughly $115 billion in fiscal 2027, and approximately $230 billion in fiscal 2028. That is a fourfold increase from this year's total in two fiscal years. He described the FY2028 number as a "line of sight" target, not delivered revenue. The difference between those phrases matters: a line of sight means Broadcom has the orders and the supply chain to make it possible, but individual customer deployments can slip, accelerate, or change scope.

The customer list is where the picture gets sharper. Anthropic is on track to become Broadcom's largest XPU customer in 2027, deploying one gigawatt of Ironwood in 2026, five gigawatts of TPU v8i in 2027, and a line of sight to 10 additional gigawatts in 2028. OpenAI's custom Jalapeño chip began shipping in Q3, with a 1.3 gigawatt commitment in 2027 and over five gigawatts in 2028. Google is shipping Ironwood at high volume and signed a long-term agreement for "multi-tens of billions of dollars of TPUs annually" Meta has three generations of MTIA accelerators on track through 2028.

These are not orders in the traditional sense. They are deployment roadmaps that assume continued investment, sustained demand, and no major architectural pivot. Broadcom calls it "secured supply" — meaning it has already reserved the wafers, memory, and substrates to build the chips if the customers deliver on their end. The dual nature of supply commitments is always the question: they signal both demand strength and delivery risk.

And here is what the market was already watching on the day Broadcom reported. Less than 24 hours before the earnings call, Reuters reported that Anthropic had signed a $35 billion cloud-computing deal with Lambda, a data center provider backed by Nvidia. Nvidia holds the lease on the Texas data center that will power part of the arrangement. This was one of at least four massive Anthropic cloud deals totaling over $150 billion, alongside agreements with Nscale, neocloud Fluidstack, and SpaceX.

Tan did not acknowledge this on the call. He called Anthropic his coming largest customer. The Anthropic-Lambda deal does not directly compete with the Broadcom XPUs — it is cloud capacity, not silicon. But it signals something more important about the behavior of frontier AI labs. They are not betting on one supplier. They are building capacity everywhere, across every architecture, and they are doing it at a scale that dwarfs any single semiconductor company's capacity to deliver.

This is not a Broadcom-specific problem. Every chip company in the AI infrastructure space faces the same reality: the customers are writing checks to everyone, and the total amount they need is so large that no single provider could satisfy it even if they wanted monopoly terms. The custom-chip business grows precisely because the market is big enough to support multiple architectures in parallel.

But it also means Broadcom's revenue line is not a pipeline that belongs to Broadcom alone. It belongs to Anthropic, Google, OpenAI, and Meta — and any one of them could shift mix, slow deployment, or pursue internal silicon capability. Anthropic explored acquiring chip startup MatX for roughly $7 billion. Google signed a large custom-chip agreement with Marvell that could generate up to $120 billion through fiscal 2033. Broadcom's Tan acknowledged Marvell's emergence but said the revenue impact would not show until around fiscal 2029.

The point is not that Broadcom is losing customers. The point is that the revenue these roadmaps represent is conditional on customer decisions that Broadcom does not control. That conditionality is baked into every custom-chip business. The question is only whether the concentration is high enough to make it material.

Six customers producing 73% of AI revenue is not a diversified business by any standard reading. The fact that those six are the most aggressive AI companies in the world is the argument in Broadcom's favor. But the math is the math.

The stock was trading around $378 at the close, down from a 52-week high of $495 — a drawdown of roughly 24% from the peak. The P/E ratio sits around 60x on trailing earnings, a compression from the mid-70s range earlier in the year. Broadcom's net debt stands at roughly $45 billion, but the net debt-to-EBITDA ratio of about 1.1x is manageable for a company generating $13.7 billion in free cash flow in a single quarter.

The valuation picture tells its own story. Broadcom trades at a much higher multiple than Nvidia, which sits around 28x P/E on trailing earnings despite posting roughly $194 billion in annual data center revenue. Broadcom's premium reflects the VMware software business — which generates roughly $12 billion in recurring annual revenue with the stability of long-term enterprise contracts — and the expectation that the custom-chip machine will compound at a pace that justifies the spread.

Whether that premium is earned depends on what happens in the next 12 months, not 24. The $230 billion figure for FY2028 is meaningful as a ceiling, not as a floor. It sets the scale of the opportunity Broadcom has built. But the investment case does not live at the ceiling. It lives in the quarter-over-quarter evidence that the machine is shipping on schedule, at the margins Tan describes, with the customer base he names.

The Q4 miss of $300 million was small enough to be dismissed as noise. It was also small enough to show how thin the margin for error has become. When a company guides to fourfold growth over two years, every quarter is a referendum on whether that trajectory is real.

What changes how you view Broadcom depends on what you were seeing before the report. The earnings show a company with a real operating engine, not a narrative. AI semiconductor revenue more than tripled in one year. Operating margins hit a record 67.9%. Free cash flow was 46% of revenue. These are the financial characteristics of a business that has found a structural position in a growing market and is executing.

The $230 billion figure is less important as a prediction than as a map of the commitments Broadcom has secured. It tells you what the customers are planning, what the supply chain looks like, and what scale the company believes it can handle. Tan has a track record of setting aggressive targets and treating them as operational anchors rather than sales pitches. The question is not whether he is optimistic. It is whether the supply bottlenecks — leading-edge wafers, HBM memory, substrates, land, power, and shell capacity — will resolve fast enough to keep pace.

The risk is not a binary one. Broadcom is not going to lose six customers simultaneously and watch the AI revenue line flatline. The risk is incremental: a deployment that slips into the next quarter, a customer that diversifies a larger share of spend, a supply constraint that pushes delivery back by months. Each one reduces the growth rate from the headline projection, and at this valuation, the growth rate is what supports the multiple.

The story is not about whether Broadcom wins the AI chip market. It is about whether the growth it is delivering today — tripling year over year, with 67.9% operating margins and custom chips accounting for three-quarters of AI revenue — can continue at a pace that justifies a 60x P/E. That is the gap between what the business is doing and what the stock price assumes. The evidence from this quarter shows the business is real. The gap remains.