"Even as we ship the chips, are they going to be deployed on a timely basis?"
That was, reportedly, Hock Tan on the September 2 earnings call, and in this author's view it is the most important sentence Broadcom said all quarter — more important than the $230 billion number that grabbed every headline.
For the first half of 2026, the fear hanging over Broadcom was demand. The stock had fallen from roughly $495 to the mid-$360s, Macquarie had cut it to Neutral in June over Google concentration, and the whole trade rested on whether hyperscaler orders would hold.
Then Tan flipped the question on its head. Demand isn't the risk. Demand, he said, "actually exceeds this outlook" — and he said it about a target of $115 billion in fiscal 2027 AI revenue, roughly double this year's $58 billion.
He said it again about fiscal 2028's $230 billion. "Here again, we have secured the supply to meet this outlook."
Read that carefully. Tan isn't forecasting demand. He's rationing it. The $115 billion and $230 billion figures are what Broadcom thinks it can build and ship, not what its customers want to buy — which is more.
Tan isn't forecasting demand. He's rationing it. The figures are what Broadcom thinks it can build and ship, not what its customers want to buy — which is more.
That is the inversion this piece exists to price. The binding constraint on Broadcom's next two years is not orders. It is atoms — wafers, memory, substrates, and the gigawatts of power that turn a shipped chip into a working data center. And almost nobody spent 2026 watching that constraint.
Start with what "secured the supply" actually covers, because Tan was precise about it. On the call he named the pieces: leading-edge wafers, high-bandwidth memory, and substrates. Broadcom has been locking those up years in advance, down to building its own chip substrate plant in Singapore.
That is the part of the chain Broadcom controls, and it has, by its own account, contracted for it. Fine.
Now read the caveat, which is the part it doesn't control. "Even as we ship the chips, are they going to be deployed on a timely basis?" Land and power dictate when a customer's data center capacity turns on, Tan said, and any piece of the chain may become the bottleneck.
So the supply Broadcom secured stops at the loading dock. What happens after — energizing the site, wiring the power, standing up the building — belongs to the customer. The $230 billion assumes all of that lands on schedule.
Look at the physical scale of what "on schedule" means. $230 billion of AI revenue in fiscal 2028 averages nearly $58 billion a quarter. Broadcom's entire business — chips plus VMware software — just posted a record $29.6 billion quarter. Two years out, the AI line alone would need to run at almost double the whole company's current record.
The measure for that isn't dollars. It's gigawatts. And the customer ramps Tan laid out are staggering.
Anthropic: 1 gigawatt of Ironwood in 2026, 5 gigawatts of TPU v8i in 2027, with clear line of sight to another 10 gigawatts after that. That trajectory makes Anthropic Broadcom's largest XPU customer in 2027 — overtaking Google, the relationship the entire story was built on — and holding that spot through 2028.
OpenAI is running a parallel ramp on its "Jalapeño" chip: 1.3 gigawatts planned for 2027, with a next-generation successor approaching tape-out in 2028 and line of sight to over 5 gigawatts total, enough to make OpenAI the second-largest XPU customer.
Add Meta, whose latest MTIA accelerator begins production shipments in the fourth quarter. "Four of our six customers will grow to enormous scale," Tan said.
Six. That is the whole list underwriting a number that quadruples in two years. It is a short customer base for a figure this size — and every one of those six is racing to energize gigawatts of capacity on the same clock.
That is the concentration hiding inside the supply story. It isn't just that six names carry $230 billion. It's that all six are pulling on the same constrained resources — the same fabs, the same substrate lines, the same overloaded power grids — at the same moment.
If you want a second opinion on how fast that buildout actually converts to revenue, look at Marvell. In late August, Broadcom's closest custom-silicon competitor raised its own fiscal 2027 and 2028 outlook on broad Data Center strength — about $12 billion and $18 billion, respectively.
But Marvell disclosed something Broadcom's clean doubling narrative did not. Much of the financial benefit from its expanded hyperscaler partnerships — including Google — may arrive closer to 2029.
Hold those two reads side by side. Broadcom says the supply is secured and the doublings land in 2027 and 2028. Marvell, building for overlapping customers on the same physical infrastructure, is telling investors its own hyperscaler payoff slides toward 2029.
Those are competing views of the same buildout clock. One says the gigawatts energize on time; the other quietly budgets for slippage. They cannot both be describing the same timeline.
None of this means the backlog is fake. That is the trap in reading Broadcom as a demand story. The demand is real — AI semiconductor revenue tripled to $16.7 billion in the third quarter, up 221% year over year, and now makes up 56% of total revenue. The fourth-quarter guide calls for $21.7 billion, each quarter landing $5 billion to $6 billion above the last.
The orders are there. Tan's own framing makes that plain: the guidance is deliberately conservative, capped by what the supply chain can produce, not by what customers will take.
So the risk isn't a demand break. It's a deployment miss. Six customers all trying to bring gigawatts online simultaneously turns timing into the single point of failure — and timing is the one variable Tan explicitly said Broadcom cannot commit to.
"Sure, deployment timing could make a quarter or two look ordinary along the way," as one bullish read of the print put it. That is the honest shape of the downside: not a collapse, but a stutter, when land or power or a grid interconnect somewhere in one of six massive projects arrives late.
The market has been mispricing which risk it's carrying. It spent the first half of 2026 fearing the orders would stop. The orders didn't stop; they overflowed the guide.
What could still go wrong is more mundane and harder to model — whether a substrate line, a transformer, or a permit shows up when the schedule says it must. Tan told investors the outlook rests on "what data center sites, locations are ready [in] 2028 with respect to our customers."
That is the sentence to hold onto. The $230 billion is a bet on construction timelines and megawatt hookups across six of the most ambitious infrastructure programs on earth — a supply-and-power problem wearing a demand-story costume.
Price the constraint everyone stopped watching. That's the whole trade.