$230 Million Short, $51 Billion Gone: Broadcom Now Guarantees the Supply — and the Money That Pays for It

Broadcom Inc. – Broadcom: 230 Millionen zu wenig, 51 Milliarden weg

Broadcom reported the best quarter in its history after Wednesday’s close: revenue of $29.59 billion, up 86 percent, of which $16.7 billion came from AI semiconductors — up 221 percent year over year and 54 percent sequentially. Free cash flow was $13.67 billion, or 46 percent of revenue. On the call, chief executive Hock Tan added two numbers of a kind the company has never put on the table before: roughly $115 billion of AI revenue in fiscal 2027 and $230 billion in fiscal 2028.

On Thursday the stock fell. The Nasdaq Composite rose 1.4 percent that day to 26,584.06 and the Dow gained 1.2 percent to 53,686.11 — and Broadcom closed 2.94 percent lower at $356.45, after touching $342.44 intraday, a 6.75 percent decline. With roughly 4.74 billion shares outstanding, that is about $51 billion of market value gone between Wednesday’s close and Thursday’s. At the low it was $118 billion.

The trigger was a single figure. Broadcom guided fourth-quarter revenue to about $34.8 billion. Consensus stood at $35.03 billion. The gap is $230 million, or 0.66 percent. The market removed roughly $51 billion of capitalisation over a $230 million shortfall in one quarter — 223 times the number. At the intraday low it was 511 times. On the same call, the company put $345 billion of two-year AI revenue on the table and said the supply behind it was already secured.

What the numbers actually say

The third fiscal quarter ended on 2 August 2026. Semiconductor Solutions produced $20.84 billion of revenue, up 127 percent from $9.17 billion a year earlier. Infrastructure Software — essentially the acquired VMware franchise — contributed $8.75 billion, up 29 percent. The interesting part is the shift underneath: software is now only 30 percent of group revenue. A year ago it was materially more. Broadcom, which spent years deliberately positioning itself as half chip house and half software house, is turning back into a semiconductor company at speed.

Adjusted earnings came in at $3.32 per share against $1.69 a year earlier; consensus was around $3.24. Reported net income under US accounting rules jumped from $4.14 billion to $13.09 billion. Operating cash flow rose 98 percent to $14.20 billion. Capital expenditure was $532 million. That is the number that tells you what the business is: 1.8 percent of revenue. Broadcom does not build fabs. It designs, contracts out the manufacturing, and sells. Cash rose to $23.98 billion against total debt of $59.42 billion. The quarterly dividend of $0.65 is payable on 30 September.

The guidance changed its unit of measure

What is genuinely new about this quarter is not in the income statement but in how the outlook was phrased. Tan said, in as many words, that for 2027 the company had secured the supply to double AI revenue again, to approximately $115 billion. For 2028 he repeated the construction — the supply to meet the $230 billion outlook was secured as well. And he made clear that actual demand is running ahead of both figures.

That is a different species of guidance from the one investors are used to. Normally a company tells you how much it expects to sell, and the binding constraint is the customer. Broadcom is telling you how much it can ship, and the binding constraint is the wafer. Asked how he distinguishes real demand from inflated demand, Tan pointed explicitly to the availability of leading-edge wafers, substrates and high-bandwidth memory, alongside customer data-centre readiness. His summary was blunt: this is real demand.

Framing a forecast as a capacity reservation says two things. First, the bottleneck is no longer willingness to pay; it is physics. Second, and this is the part that tends to go unread: somebody has already committed. Secured supply means purchase obligations have been entered into with foundries and memory makers. If demand softens, the obligation does not.

Twenty-five gigawatts on a single conference call

More revealing than the dollar figures is the unit in which Broadcom now describes its customer commitments. Named on the call: Anthropic, with one gigawatt of the Ironwood generation this year, five gigawatts of the successor TPU v8i in 2027 and a further ten gigawatts in 2028. OpenAI, with 1.3 gigawatts of the accelerator developed under the Jalapeño codename for 2027 and more than five gigawatts in 2028. Meta, with three gigawatts across three generations of its MTIA part through 2028. Google, described as buying multiple tens of billions of dollars of tensor processing units annually over the next several years.

Add up only the explicitly stated power figures and you clear 25 gigawatts through 2028. For scale: average electricity demand across the entire United States runs around 480 gigawatts. One supplier, on one ninety-minute call, named data-centre projects amounting to roughly five percent of the average load of the American grid. When a chip company states its order book in gigawatts rather than units, the binding constraint has moved — away from the die and toward the interconnection queue, the transformer and the permit.

The margin falls because the business is winning

Chief financial officer Amie Thuener gave a number that was lost in the noise: gross margin was 75 percent in the quarter, down 210 basis points sequentially. For the fourth quarter she guided to roughly 73 percent. A year ago it was 78 percent. Broadcom is therefore giving up five points of gross margin in the year its revenue nearly doubles.

The cause is not price pressure but mix. Custom accelerators carry a great deal of bought-in memory. The more HBM stacks sit on a module, the larger the share of revenue that is, in substance, a purchase from Samsung, SK Hynix and Micron passed through. With every accelerator socket Broadcom wins, it also resells a piece of somebody else’s memory bill — at a markup, but only at a markup. What rescues the operating line is cost discipline: semiconductor operating expenses were $1.2 billion, or six percent of segment revenue. Adjusted operating margin therefore sits near 68 percent despite the falling gross margin, with roughly 66 percent guided for the fourth quarter.

$532 million of capex — and up to $100 billion of guarantees

This is where the balance sheet becomes more interesting than any growth rate. Broadcom spends $532 million a quarter and generates $13.67 billion of free cash flow. By every conventional measure this is a capital-light business. But the capital has not disappeared. It has changed line items.

Earlier this year Broadcom backstopped most of a $35 billion debt package under which investors including Apollo Global Management and Blackstone financed the purchase of custom AI chips to lease to Anthropic. A considerably larger structure is now under negotiation: more than $60 billion of senior secured debt routed through special-purpose vehicles, plus a junior tranche of roughly $30 billion. Taken together, as much as $100 billion could be in play, a portion of which Broadcom would guarantee. Credit markets have noticed. Five-year protection against a Broadcom default widened by 28 basis points during August — more than Oracle’s and more than SpaceX’s.

So a company now stands on both sides of its own order book. It guarantees the debt with which a special-purpose vehicle buys its chips, which are then leased to the customer, who pays the lease out of revenue it has yet to earn. The investment risk a conventional manufacturer carries in its plant appears nowhere in Broadcom’s capital intensity. It sits in guarantees — and those show up first in the price of default protection, not in the price-to-earnings ratio.

Six customers behind a $230 billion number

Broadcom serves six customers in the custom accelerator business. Estimates put well over 60 percent of the segment’s quarterly revenue with just three of them. Anthropic, the company says, is on track to become the largest customer in that category in 2027 and to remain so in 2028.

That is exactly where the circle closes. The largest future buyer is not a listed company with published accounts but a private one whose procurement runs through a financing structure the seller helps carry. Broadcom confirmed on the call that these structures are principally for Anthropic and OpenAI, while its other customers fund their deployments themselves. That is an important clarification — it separates the financed portion of the pipeline from the self-supporting portion. It also means the fastest-growing part of the outlook is the part standing on borrowed money.

The case against the bearish reading

None of this is the same as an uncovered promise, and it is worth being fair about it. The financing is senior secured, the collateral is real, scarce and currently coveted hardware, and the junior tranche absorbs first loss. Broadcom guarantees a portion, not the whole. Cash of $23.98 billion sits against $59.42 billion of debt, with free cash flow that on an annualised basis should clear $50 billion. A 28 basis point widening in default protection is a repricing, not distress.

The margin dilution is mix rather than erosion: gross profit dollars rose sharply, and the operating margin remains historically extraordinary. And the largest customers do not need the structure at all. Google and Meta self-fund; a multi-tens-of-billions annual commitment from an issuer of that credit quality is a different order of promise from a lease through a special-purpose vehicle. An investor looking at the stock near $345 after the drop, at roughly eighteen times expected fiscal 2027 earnings, can argue with a straight face that the market is currently paying very little for secured supply.

What this means for a US investor

For most American portfolios the exposure is already there and was never chosen. Broadcom is among the largest single weights in every broad US and global index; anyone holding an S&P 500 fund owns a meaningful slice of this outcome by default. The more useful question is which listed names sit on the other side of those 25 gigawatts, and the answer is rarely the chip itself. Micron supplies the high-bandwidth memory whose rising content is compressing Broadcom’s own gross margin. GE Vernova, Eaton and Vertiv sell the turbines, switchgear, busways and cooling that turn a gigawatt on a slide into a gigawatt on a site. Marvell is the other merchant custom-silicon house, and its results read as a control experiment for the same thesis.

On tax, the relevant mechanics are unglamorous but consequential. Shares held longer than one year are taxed at 0, 15 or 20 percent as long-term capital gains rather than at ordinary rates, with the 3.8 percent net investment income tax on top above the applicable thresholds. Broadcom’s $0.65 dividend qualifies for the same preferential rates only if the 61-day holding-period test around the ex-dividend date is met. And anyone tempted to harvest a loss into this drawdown should note the wash-sale rule: repurchasing a substantially identical position within 30 days before or after the sale disallows the loss.

What has to be measured from here

The tests over the coming months are uncomfortably concrete. First: does AI revenue in the fourth quarter, which ends on 1 November, reach the guided $21.7 billion? Second: does the $60 billion-plus financing actually close, at what size, at what spread, and how much of it does Broadcom ultimately guarantee? Third: does Anthropic secure the power, the land and the interconnection for five gigawatts in 2027? Fourth: where does gross margin settle as memory content keeps rising? And fifth, the hardest test: to reach $115 billion, fiscal 2027 AI revenue has to average just under $29 billion a quarter, against the $16.7 billion the company just reported as a record.

Thursday showed how the market currently weights all of that. A $230 million shortfall in the next thirteen weeks was worth $51 billion to it; two years of secured supply worth $345 billion was worth approximately nothing. You can call that short-sighted. You can also read it as a precise judgement: secured supply is not a booked order, and a guarantee is not revenue. The market still pays for shipments rather than for commitments — and it has begun reading the price of those commitments where it appears first, in the credit market rather than in the equity research.

PARTNER PICK

Try TradingView Free for 30 Days

Plus get a $15 discount on your first subscription through this link.

30 Days Free Trial
$15 Discount
Pro Charts & Tools
Start 30-Day Free Trial →
Affiliate link: we earn a commission if you subscribe through this link, at no extra cost to you.
Daniel Herzog
AUTHOR

Daniel Herzog

Founder of Butterfly Market Insider

More about Daniel →

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top