Broadcom told investors on Tuesday evening that its AI semiconductor business will roughly double to $115 billion in fiscal 2027 and double again to $230 billion in fiscal 2028, one of the largest forward revenue commitments any chipmaker has ever put on a slide. The stock did not celebrate. It slipped.
That reaction is the story, and almost nobody covering the print has explained it properly. The coverage has fixated on the $230 billion headline and treated the muted move as investor fatigue or profit-taking after a long run. It is neither. The answer is sitting in a different line of the same release: Broadcom’s infrastructure software segment grew revenue 29% year over year to $8.8 billion, while the annual recurring revenue underneath that segment grew 15%. Reported revenue is running at nearly double the rate of the recurring base that is supposed to produce it. That divergence is what a professional reads first, and it is the part the headline number is drowning out.
The Print Itself Was Enormous
There is no version of this quarter that is weak. Broadcom posted record third-quarter results: revenue of $29.6 billion, up 86% year over year, and operating income of $20.1 billion, up 92%, for an operating margin just under 68%. Semiconductor solutions revenue rose 127% to $20.8 billion. AI semiconductor revenue specifically came in at $16.7 billion, up 221% from a year earlier and up 54% sequentially, and management guided the fourth quarter to $21.7 billion of AI revenue inside $34.8 billion of total revenue.
For the full fiscal year, that puts AI semiconductor revenue near $58 billion, up roughly 186%. Chief executive Hock Tan then raised the fiscal 2027 target from the “in excess of $100 billion” he gave in June to about $115 billion, and said the company has line of sight to $230 billion in fiscal 2028. He attributed the confidence to six customers making multi-gigawatt deployment commitments, and said Broadcom has already secured wafer, high-bandwidth memory and substrate supply through 2028. Anthropic is expected to become the largest XPU customer in 2027. OpenAI could deploy more than five gigawatts of Broadcom’s next-generation Jalapeno part in 2028.
The Segment That Pays for All of It
Here is why the software line matters more than its size suggests. Broadcom’s AI roadmap is capital-hungry in a specific way: to promise $230 billion of XPU revenue in fiscal 2028, the company has to pre-commit to wafer allocation, HBM supply and advanced packaging years ahead of the orders converting. Tan said as much on the call. Securing supply through 2028 is not a talking point, it is a purchase obligation.
The thing that lets Broadcom write those cheques without stressing its balance sheet is VMware. Infrastructure software carries roughly 93% gross margins and an operating margin near 79%, and in the quarter the segment’s operating margin expanded about 650 basis points to around 84%. That is an annuity, and annuities are what fund cyclical bets. One of the few analyses to flag this, an Investing.com breakdown of the software segment, made the point that the AI surge is masking a deceleration in exactly the business that de-risks the AI surge.
The 29% versus 15% gap has a mundane explanation and an uncomfortable one, and both are probably true. The mundane one: Broadcom is still converting VMware’s legacy perpetual licences into subscriptions, and that conversion flatters reported revenue in the period it lands without adding proportionally to the recurring base. The uncomfortable one: when the conversion runway is finished, reported software growth converges down toward ARR growth, and 15% is a materially different funding profile than 29%.
Six Customers, and Not All of Them Are Self-Funding
The other thing worth saying plainly is that a $230 billion number resting on six buyers is a concentration story, whatever else it is. Google and Meta generate enough operating cash flow to fund multi-gigawatt commitments out of pocket. OpenAI and Anthropic do not. They are funding compute through capital raises and vendor arrangements, in a market where the cost of financing the AI buildout has been repricing all summer. Broadcom is not wrong that the demand is real. It is exposed to the question of whether the demand stays financeable.
That exposure runs in the opposite direction from a normal chip cycle. In a normal cycle, a chipmaker that overestimates demand cancels foundry bookings and eats a modest penalty. Broadcom has told the market it has locked supply through 2028 in order to make the forecast credible. Credibility purchased that way converts a demand miss into an inventory problem. The company also faces a sharpening competitive picture in custom silicon, where Marvell’s warrant arrangement with Google showed hyperscalers are actively cultivating a second source.
Where We Land
The $230 billion forecast is not a lie and it is not hype, but it is being reported as a demand signal when it is closer to a supply commitment. Broadcom has decided to guarantee capacity in order to win designs, which is a legitimate strategy and a real transfer of risk from its customers onto its own balance sheet. Investors are entitled to price that, and on Tuesday they did.
Our view is that the market got this one right and the coverage got it wrong. A forecast three years out from a company whose disclosure cadence is quarterly is worth less than a live deceleration in the segment that funds the forecast. Broadcom management should be pressed on infrastructure software ARR on the next call, not on Jalapeno gigawatts, and if the ARR number keeps growing at half the pace of reported software revenue, that is the metric that will move the stock. The $230 billion will still be three years away. The annuity is being spent now.
The right question for the fourth quarter is not whether AI revenue hits $21.7 billion. It almost certainly will. It is whether Broadcom is still able to fund its 2028 promises out of VMware’s cash flows, or whether it has to start funding them out of somebody else’s.