The $30 Billion Number Nobody Read: Why Broadcom's Sell-Off Misprices Tan's Backlog Math

John Ian Lau — Technology & AI infrastructure desk

Broadcom reported its fiscal second quarter after the bell on June 3 and the market did something it has not done with a custom-silicon name in over a year — it sold the print. Shares fell as much as 13 percent during the conference call before clawing back to roughly minus four percent by the after-hours close. The reaction was reflexive and, in our view, almost entirely wrong. Buried in Hock Tan's prepared remarks was a single number that recontextualizes the entire AI infrastructure cycle. It was barely discussed. It is the most important data point of the quarter.

The headline figures were strong by any reasonable standard. Total revenue came in at $22.19 billion, up 48 percent year-over-year. AI semiconductor revenue printed $10.8 billion, a 143 percent increase and a hair above the company's own $10.7 billion guide. Networking — the unglamorous, high-margin half of the AI silicon business — accounted for roughly 40 percent of that AI line. For the third quarter, Broadcom guided to $29.4 billion in total revenue, comfortably ahead of the $28.53 billion sell-side consensus, with the AI semiconductor component pointing to $16 billion, up roughly 200 percent year-over-year.

That last figure is what triggered the selling. The buy-side whisper had drifted up to $17.2 billion for AI semis in the third quarter, and the 2027 AI guidance — last set at "in excess of $100 billion" — was reiterated rather than raised. Long-only managers who had positioned for a beat-and-raise on the out-year framing took profits. Hedge funds covering shorts faded the move. The narrative congealed within an hour: the chip rally is exhausted, AI infrastructure is decelerating, sell the news.

That narrative ignores the number that matters.

The $30 billion line

Roughly fifteen minutes into the call, Tan delivered what should have been the headline of the entire print:

"Demand for XPUs and networking is simply insatiable. During the quarter, bookings for AI semiconductors were over $30 billion against the $10.8 billion we shipped."

Tan, the chief executive of Broadcom, was describing a 2.78x book-to-bill ratio in the company's AI semiconductor business — in a single quarter. To put that in perspective, the semiconductor industry's overall book-to-bill ratio rarely sustains above 1.2x even in cyclical expansions. A 2.78x reading is not a deceleration signal. It is the opposite. It is the signature of a supply chain that cannot physically fill the orders coming in.

The math implications cascade. Broadcom's full fiscal 2026 AI semiconductor revenue is now tracking toward roughly $56 billion, a 180 percent increase from fiscal 2025. To then post "in excess of $100 billion" in fiscal 2027 — Tan's reiterated framing — requires roughly 80 percent growth on top of that base. Against $30 billion of bookings per quarter, the $100 billion target is no longer aspirational. It is mathematically conservative. For Broadcom to miss it, bookings would have to collapse for three consecutive quarters, or hyperscaler customers would have to cancel firm purchase orders en masse. Neither is happening. To the contrary: the company disclosed during the call that its core XPU customer count has expanded to six, including Google, Anthropic, Meta and OpenAI, with two additional unnamed customers. OpenAI's contracted deployment alone is now 1.3 gigawatts in 2027, sitting inside a larger 10-gigawatt-by-2029 arrangement.

This is the third consecutive quarter in which Tan has guided AI semiconductor revenue below where it printed. Fiscal first-quarter AI guidance of $8.2 billion produced an actual print of $8.4 billion. Fiscal second-quarter guidance of $10.7 billion produced $10.8 billion. The pattern is not coincidence. It is calibrated under-promising, and the third-quarter $16 billion figure that just spooked the tape is almost certainly the floor, not the ceiling, of what the company will deliver.

Later in the same call, Tan added the framing that should have rallied the stock:

"Now we expect this momentum to continue into fiscal year 2027, and reiterate our AI semiconductor revenue guidance to be in excess of $100 billion."

"In excess of" is the operative phrase. Tan does not use language loosely. The company has spent the past eighteen months systematically lowering the visible bar in front of it and then stepping over the bar. The market — conditioned by two years of NVIDIA's "raise everything every quarter" cadence — has begun to mistake this discipline for caution. It is not caution. It is sandbagging in service of a multi-year compounding setup.

The Dell cross-check

Broadcom is not the only data point. Six days before the print, Dell Technologies reported its fiscal first quarter and the result was, by any historical reference, extraordinary. Revenue grew 88 percent year-over-year. AI server revenue alone reached $16.1 billion, a 757 percent annual increase. The company booked $24.4 billion of AI orders in the quarter and exited with a record $51.3 billion AI backlog. Adjusted earnings per share came in at $4.86 against a $2.94 consensus expectation. The stock closed up 32.76 percent the following session — the largest single-day gain in the company's public-market history.

Ben Reitzes, head of technology research at Melius Research and one of the most measured voices on the hardware desk, captured the moment in a single sentence widely circulated in the days after the print:

"I've never seen anything like it."

Reitzes has covered the technology hardware complex for more than two decades, through the dot-com cycle, the smartphone supercycle and the cloud build-out. His admission that the current cycle has no clean historical analog is, in itself, the analytical signal. When the leading custom-silicon supplier and the leading AI-server integrator both report results that exceed the cyclical bounds of every prior comparable period, the appropriate interpretation is not that the cycle has peaked. It is that the cycle is still establishing a baseline.

Where the pricing power compounds

The broader trade implication is more important than the Broadcom-specific call. The Broadcom and Dell prints, read together, describe a supply chain operating at capacity-constrained early-cycle conditions, not late-cycle exhaustion. Bookings outpace shipments by multiples. Backlogs are at record highs and still growing. New customers are being added rather than churning out. None of these are signatures of an over-supplied or saturated end market.

The differentiation now is not whether to be long AI infrastructure. The differentiation is where in the stack pricing power compounds. The supply-constrained inputs sit upstream of the integrators: high-bandwidth memory, advanced packaging capacity, optical interconnect and the specialty tooling that fabricates all three. These are the businesses that get pricing leverage as the cycle extends, because their physical capacity cannot be expanded inside a single fiscal year. Commodity server integrators that lack backlog visibility face the opposite dynamic — margin compression as components get scarce and pricier.

This is the same playbook that worked in the early years of the cloud build-out, when the marginal compounder was not the hyperscaler but the equipment vendor selling into the hyperscaler. The 2025—2027 analog is one layer deeper in the stack.

Our view

The Broadcom sell-off is buyable. The market has misread Tan's discipline as weakness in a quarter where the underlying bookings are the strongest in any company's history we cover. A 2.78x AI book-to-bill is a structural signal, not a quarterly anomaly, and the $100 billion 2027 AI guidance is now a floor rather than a ceiling. We would add on the post-earnings pullback rather than fade it.

More broadly, we would use this print to rotate within AI infrastructure exposure. Long the upstream suppliers — high-bandwidth memory, advanced packaging, optical — where capacity is the binding constraint. Underweight commodity integrators that lack the multi-year backlog visibility Dell and Broadcom now possess. Watch for the next Broadcom AI semiconductor guide in early September; if the third-quarter print materially exceeds the $16 billion bar — which the bookings math implies it will — the 2027 number will be raised, and the sandbagging trade will have completed its first full cycle.

This note is for informational purposes only and does not constitute investment advice. Solomon Grey Capital, its principals and affiliates may hold positions in the securities discussed.