Everything About AI
NextFin News - Broadcom Chief Executive Hock Tan told investors this week that the chipmaker's artificial intelligence revenue will double to about $115 billion in fiscal 2027 and double again to $230 billion in fiscal 2028, a forecast that puts a custom-chip specialist on a direct collision course with Nvidia's merchant-GPU empire. The market's reaction was less impressed: shares fell about 5% in extended trading after Broadcom's fiscal third-quarter report on Sept. 2, even though quarterly revenue rose 86% and AI chip sales more than tripled. The gap between a record quarter and a punished stock is the question this piece answers.
The Quarter: Records on Paper, Disappointment in the Guide
Broadcom reported fiscal third-quarter revenue of $29.59 billion, up 86% from $15.95 billion a year earlier and slightly above the $29.36 billion consensus compiled by LSEG. Net income more than tripled to $13.09 billion, or $2.68 a share on a GAAP basis, from $4.14 billion, or 85 cents, in the same period a year ago. Non-GAAP earnings came in at $3.32 a share, beating the $3.24 estimate. Free cash flow hit $13.7 billion, equal to 46% of revenue.
The AI engine inside those numbers is the part that matters. AI semiconductor revenue reached $16.7 billion in the quarter, up 221% from a year earlier and 54% from the prior quarter, Tan said in the company's release. That is more than triple the $5.2 billion of AI chip sales Broadcom recorded in the same period a year ago, when AI was still a secondary business rather than the center of gravity. Infrastructure software revenue of $8.75 billion came in just under the $8.82 billion consensus polled by StreetAccount.
Then came the guide. Broadcom expects fourth-quarter revenue of about $34.8 billion, a 93% year-over-year increase but below the $35.03 billion that analysts were expecting. The non-GAAP operating margin is forecast at 66%, a touch under the 66.5% consensus. The miss was small in percentage terms, but it was enough to send the stock lower because investors had been pricing perfection into a name that has already jumped more than sixfold since the end of 2022.
The real story, however, is not the quarter. It is the multiyear trajectory Tan sketched out on the earnings call: AI revenue of about $58 billion in fiscal 2026, $115 billion in fiscal 2027, and $230 billion in fiscal 2028, with earnings per share topping $30 in fiscal 2028. Those figures imply that Broadcom's AI business alone would grow roughly fourfold in two years from a base that barely existed three years ago. That is the claim worth dissecting.
Why Broadcom's Model Is Different From Nvidia's
Nvidia sells general-purpose graphics processors to thousands of customers, from cloud providers to enterprises to research labs. Broadcom sells custom accelerators to a handful of hyperscalers, designing silicon to one customer's exact workload and locking that design in for years. The difference sounds technical, but it is really a difference in revenue durability.
When Google, Meta, OpenAI, or Anthropic designs a custom chip with Broadcom, the engagement runs across multiple generations of silicon and is tied to gigawatt-scale data-center buildouts. Tan said on the call that Broadcom is speeding up shipments of Google's Ironwood tensor processing units to Anthropic and TPU 8i chips to Google. He pointed to a 1.3-gigawatt deployment of the OpenAI-designed Jalapeno chip in 2027, with a line of sight of more than 5 gigawatts for Jalapeno and its second generation. Meta is expecting 3 gigawatts of computing power through the end of 2028, with the first 1-gigawatt order beginning delivery in the second half of next year.
These are not spot orders. They are capacity commitments that map onto the physical buildout of AI data centers, and they are why Broadcom can guide to $115 billion with a straight face. A merchant GPU buyer can switch vendors between product cycles. A hyperscaler that has co-designed a chip, ported its models to it, and built a data center around its power envelope does not walk away lightly.
"Demand for our custom AI accelerators and networking continues to be very strong," Tan said. "Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter. In Q4 the momentum continues, and we expect AI semiconductor revenue to accelerate to $21.7 billion, up 236% year-over-year."
The networking business is the second half of the moat. AI clusters are only as fast as the links between their chips, and Broadcom's switching and optical networking gear is embedded in the same racks as its accelerators. Goldman Sachs has estimated the optical networking market could expand from $15 billion in 2026 to $154 billion in 2028. Broadcom does not break out networking revenue separately, but if roughly 30% of its AI revenue carries through networking, that is a business approaching $17 billion this year on a $56 billion to $58 billion AI base.
There is also a valuation asymmetry. Broadcom's market capitalization sits around $1.8 trillion. Nvidia's, at more than $5 trillion, is nearly three times larger. If Broadcom captures even a meaningful slice of the custom-silicon shift, the re-rating math favors the smaller name. That is the bullish case in one line: Broadcom is being paid like a supplier while behaving like a platform.
Is This Cyclical or Structural? The Share Shift Is the Durable Part
The cleanest way to think about Broadcom's boom is to separate two forces that are running at the same time. One is cyclical: the AI capital-expenditure wave itself, which will have quarters of acceleration and deceleration, inventory adjustments, and margin swings. The other is structural: the migration of frontier AI compute from merchant GPUs to custom ASICs inside the largest cloud builders.
The cyclical leg is mean-reverting by nature. Broadcom's own history proves it. In June, after guiding fiscal third-quarter AI revenue to $16 billion, the stock fell 14% because the number came in below the roughly $17.2 billion analysts had modeled. In September, the stock fell again on a guide that missed by about $230 million. Quarterly volatility around hyperscaler capex is not going away, and memory content — which Tan flagged as a reason margins are not stronger — will ebb and flow with commodity prices.
The structural leg is different. Three pieces of evidence point to a regime shift rather than a temporary swing. First, the customer concentration is a feature, not a bug: Broadcom works with roughly 7 to 10 hyperscaler customers on multi-year contracts, and those relationships are measured in gigawatts of committed power, not unit orders. Second, the design-in cycle is long and expensive; once a cloud provider has built its software stack and data-center architecture around a custom accelerator, switching costs compound with every model generation. Third, the economics favor custom silicon at scale: a hyperscaler running tens of thousands of identical inference workloads can extract more performance per watt from an ASIC than from a general-purpose GPU, and at frontier scale, watts are the binding constraint.
That last point is the mechanism. AI training and inference are ultimately physics problems — compute density, memory bandwidth, and power delivery. Merchant GPUs win when the customer base is fragmented and workloads are heterogeneous. Custom ASICs win when one buyer standardizes enough volume to amortize the design cost. The hyperscalers have now crossed that threshold. The shift is not a bet on AI demand continuing to rise in a straight line; it is a bet on who captures the silicon margin inside that demand.
The Second-Order Question the Market Is Not Asking
The first-order read of Tan's guide is obvious: AI demand is strong, so buy the AI supply chain. The second-order question is harder, and it cuts the other way. If custom accelerators are genuinely more efficient than merchant GPUs, what happens to Nvidia's pricing power — and to the total dollar value of the AI chip market — as the mix shifts?
Efficiency gains are deflationary for the buyer. Every dollar a hyperscaler saves by moving inference to a custom chip is a dollar it does not have to spend on Nvidia hardware. That is good for Broadcom's share of the pie and potentially bad for the size of Nvidia's slice. The cross-asset transmission runs like this: custom-silicon share gains compress the merchant GPU growth rate; a slower merchant growth rate forces a re-rating of any name priced for perpetual triple-digit expansion; and the capital freed up by cheaper inference either funds more model training, which is bullish for the long run, or flows back to shareholders, which is a sign the capex peak is near.
There is a third-order expectation gap hiding inside the $230 billion number. The market has priced Broadcom as an AI winner, but it has not fully priced the possibility that Broadcom wins by taking share from the very AI leader everyone already owns. If that thesis is right, the trade is not "long AI infrastructure." It is "long the substitution of AI infrastructure." The distinction matters because the two trades diverge the moment AI spending growth slows while ASIC penetration keeps rising.
Chief Financial Officer Amie Thuener's margin commentary is the tell. "We delivered non-GAAP operating income growth of 92% year-over-year, as consolidated revenue grew 86% year-over-year to $29.6 billion," she said. "Q4 consolidated revenue growth is forecasted to increase 93% year-over-year to $34.8 billion, and we expect to maintain our non-GAAP operating margin at 66%, flat from a year ago." Operating income growing faster than revenue, with margins held flat despite higher memory content, is the signature of a business with operating leverage that the headline multiple has not yet recognized.
The Strongest Case Against the Thesis
The bear case is not that AI demand is fake. It is that Broadcom's customers are the same entities deciding how much AI demand is economically justified, and they have both the incentive and the ability to squeeze their supplier. A hyperscaler that has learned to design chips with Broadcom can decide to design more of them in-house. Google already builds its own TPUs; the question is how much of the value chain it keeps with Broadcom versus internalizing over time. Nvidia's CUDA software ecosystem remains the deepest moat in the industry, and a software moat is harder to dislodge than a hardware one.
There is also a valuation-of-demand risk. Investors such as Jeremy Grantham and Michael Burry have warned that AI capital spending is running ahead of the revenue it can generate, and that a reckoning in hyperscaler budgets would hit the entire supply chain. If AI capex growth slows, Broadcom's gigawatt commitments do not disappear overnight, but new bookings would, and a bookings slowdown would show up in the numbers well before the $230 billion target date.
The answer to the bear case is that Broadcom's exposure is back-loaded and contractual, while Nvidia's is spot-market and cyclical. But that answer only holds if the commitments are real and binding. The falsifying signal is specific: if Broadcom's AI bookings-to-revenue ratio falls below roughly 1.5x for two consecutive quarters, or if hyperscaler AI capital-expenditure growth slows to less than 15% year over year for two consecutive quarters, the structural share-shift thesis should be downgraded to a cyclical bet. Either metric is observable and dated. Until then, the burden of proof sits with the bears.
What to Watch: Three Horizons
Short term (next quarter): The market will focus on whether the fiscal fourth-quarter guide of $34.8 billion is a one-time guide miss or the start of a pattern. Watch the Q4 AI semiconductor print of $21.7 billion against the $20.93 billion FactSet consensus, and whether the 66% operating margin holds despite higher memory content. A beat on AI revenue with a margin hold would repair the post-earnings damage quickly.
Medium term (fiscal 2027): The $115 billion AI revenue target is the number that separates believers from skeptics. Hitting it requires the Jalapeno ramp for OpenAI, continued TPU volume for Google and Anthropic, and the Meta 3-gigawatt program to stay on track. Any slippage in those three customer programs would be the first sign that the guide was aspirational rather than contracted.
Long term (fiscal 2028 and beyond): The $230 billion target implies Broadcom becomes one of the largest chip companies on earth by AI revenue alone. The structural thesis survives if custom ASICs keep taking share from merchant GPUs inside hyperscalers. It breaks if hyperscalers internalize more of the design work or if AI model demand plateaus before inference volumes justify another generation of custom silicon.
Base case: Broadcom delivers $115 billion in fiscal 2027 as hyperscaler capex keeps growing, and the stock re-rates toward a premium multiple as the market recognizes operating leverage. Upside case: custom-silicon penetration accelerates faster than expected, Nvidia's merchant growth slows, and Broadcom's market cap converges meaningfully toward Nvidia's. Downside case: hyperscaler capex growth decelerates below 15% year over year, bookings slow, and the $230 billion target gets pushed out, compressing the multiple back toward a conventional semiconductor peer.
The kicker: Broadcom is not betting that AI spending will keep rising forever. It is betting that however much the hyperscalers spend, an increasing share of every dollar will flow through its custom silicon — and that is a trade against Nvidia's dominance, not a bet alongside it.
Explore more exclusive insights at nextfin.ai.
