Broadcom: Consensus Models the Guidance. The Order Book Models Something Else.

Sell side estimates for AVGO run through 2028. We rebuilt them line by line, found out exactly what they assume, and measured what happens if the order book is telling the truth instead.

Broadcom reported its fiscal Q2 on June 3. Revenue of 22.2 billion dollars, up 48 percent. AI semiconductor revenue of 10.8 billion, up 143 percent and above the company’s own forecast. Q3 guided to 16 billion in AI revenue, growth of over 200 percent year over year. All of that sits in the 8-K, primary filing, no interpretation required.

What caught my attention was the distance between two other numbers from the call. Broadcom shipped 10.8 billion of AI silicon in the quarter and booked over 30 billion. A book to bill near three should reorganize a forecast. So I pulled the consensus estimates through 2028 and checked whether it did.

I own Broadcom, the following is NFA.

What we actually did

Consensus gives you one revenue line per quarter. Broadcom is three businesses wearing one income statement: AI semiconductors growing triple digits, non-AI semiconductors in a slow cyclical recovery, and infrastructure software compounding in the teens. A single number through 2028 tells you nothing until you take it apart.

So we took it apart. We held software and non-AI semiconductors to the company’s Q3 guidance, 8.9 billion and 4.5 billion, and grew them at modest steady rates from there. Subtract those two from each consensus quarter and the residual is the AI revenue the street is implicitly modeling. On sourcing: the quarterly totals are third party estimates from Fiscal.ai, the segment anchors are company guidance from the call, and the implied AI line is our own derivation.

Here is what falls out, fiscal years, in billions:

Implied consensus AI revenue FY2026: 56

Implied consensus AI revenue FY2027: 113

Implied consensus AI revenue FY2028: 162

Compare that to the call. Hock Tan guided FY2026 AI revenue to 56 billion. He reiterated FY2027 in excess of 100 billion, and in the Q&A said 2027 would be “in the range of 2x” 2026, which is 112.

The street has transcribed the guidance. FY2026 matches the guide to the decimal. FY2027 matches the stronger Q&A framing, with nothing added and nothing subtracted. Which collapses the question of whether estimates are too low into a different question: is the guidance itself conservative, and if so, where does the error live?

The error lives in the shape of 2028

Follow the implied AI line quarter by quarter and watch what happens after October 2027. Through fiscal 2027 consensus has AI revenue compounding at double digit sequential rates, back half loaded, exactly as management described the 10 gigawatts of 2027 shipments. The moment the fiscal year turns, sequential growth falls to between 2 and 5 percent and stays there through all of fiscal 2028.

Nothing in the source material motivates that kink. Analysts extend guided years faithfully and let unguided years fade, because nobody gets fired for modeling guidance plus a little. The fade carries exactly as much information as the silence it is built on.

The order book, meanwhile, carries plenty. Four data points from the quarter, each one traceable:

Bookings of over 30 billion dollars against 10.8 billion shipped, per Hock Tan on the call. Purchase orders totaling 6 billion from the two customers Broadcom has not yet named, shipping from late 2026. Inventory days up from 68 to 86 in a single quarter, per the CFO, an explicit build ahead of the second half ramp. That one I weight heavily because it is the balance sheet voting. And visibility. Three months ago management said it ran to 2027. Now they say it runs to 2028. That sentence moved more than any guidance number on the call.

One quarter of bookings does not fund a fiscal year two years out, so let me be precise about what the 30 billion does in our math. Bookings mix program orders, reservations placed ahead of power availability, and orders with acceptance conditions, so they convert unevenly. Even so: shipping 10.8 against 30 booked means the backlog grew by roughly 19 billion in one quarter. The entire gap between our fiscal 2028 and the street’s is 42 billion of AI revenue. A little over two quarters at the current net bookings pace covers it. And the Q2 10-Q discloses roughly 164.6 billion of remaining performance obligations under firmly committed contracts, customers without termination rights, with about 30 percent expected to convert to revenue over the next twelve months. For the consensus fade to hold, either that inflow dries up or the conversion of orders already placed slips past 2028.

Behind the bookings sit multi gigawatt committed programs at Anthropic, OpenAI and Meta, with deliveries that management says make 2027 back half loaded and bring “a lot more gigawatts” in 2028. I use those gigawatt figures only as evidence of commitment. Content per gigawatt varies by generation and architecture, so the revenue line in our model is built from the company’s dollar guidance and the order flow, never from gigawatt arithmetic.

A back half loaded 2027 produces a steep exit run rate. The whole disagreement between us and the street is what happens to that run rate on January 1. Consensus parks it. We let it run.

The scenario, and what it produces

We built the alternative in the same framework, same software and non-AI assumptions, only the AI line changes. A modest Q3 beat at 16.5 billion against the 16.0 guide, consistent with the pattern of recent quarters. The back half loaded 2027 management described. Then sequential AI growth holds around 10 percent per quarter through fiscal 2028 instead of fading. Ten percent is the scenario the order book forces me to test, no more than that. If you reject it, you should be able to say which leg you reject: the order flow, the delivery schedule, or the rising content per generation.

The output, fiscal years, in billions:

Scenario AI revenue FY2026: 57

Scenario AI revenue FY2027: 124

Scenario AI revenue FY2028: 204

Against consensus, fiscal 2026 and 2027 barely move. The entire gap is fiscal 2028: 204 against 162 on the AI line, roughly 272 against 229 on total revenue. Consensus understates the year by about 18 percent and the exit quarter by 26 percent, precisely where the screen stops scrolling.

The bear case that deserves an answer

The serious case against our curve is about sequencing, and parts of it come from Hock Tan himself. Asked why bookings so far outrun shipments, he explained that customers order years early because the chips are no longer the constraint, the power is. They need the power shell and the grid connection before anything can deliver. Read backwards, a book to bill of three is partly a symptom of lengthening lead times. An order placed today for 2028 delivery moves zero revenue into 2028 if the substation slips into 2029.

Supply carries the same asymmetry. Asked directly about wafers and HBM, management said supply is secured for 2026 and 2027, and that they are “working on 2028 and 2029 right now.” That cuts against us. Fiscal 2026 and 2027 rest on contracted capacity. Fiscal 2028, the exact year where our scenario diverges, rests on negotiations still in progress.

Two more legs. Google has a new long term agreement that management called very substantial in dollars while acknowledging there will be “some diversity of sources” at that customer. A fixed dollar commitment with a shrinking share of incremental architecture can produce exactly the fading profile consensus draws, without Broadcom losing a single program. And the financing layer behind part of the 2028 demand, the Apollo and Blackstone compute platform, has a first tranche that is being launched rather than closed. Until the capital is committed and the investors named, I weight it below a hyperscaler purchase order.

So consensus does not need a demand problem to be right about 2028. A sequencing problem is enough, gigawatts arriving a year late.

What being right is worth at 385

Suppose the scenario lands and consensus proves 18 percent light on fiscal 2028 revenue. That error never survives until 2028. It gets corrected as a staircase of beat and raise quarters through 2027, and since markets price twelve to eighteen months forward, most of the gap’s value gets realized during 2027.

Rough arithmetic, our derivation: take each fiscal 2028 revenue path, apply the segment operating margins Broadcom actually prints, roughly 62 percent in semiconductors and 79 percent in software, deduct the interest line, tax at the guided 16 percent, divide by the diluted count of about 4.94 billion shares. That segment method reproduces Q2’s actual operating income within a fraction of a percent. The consensus path produces something near 25 dollars of fiscal 2028 earnings. The scenario path produces something near 29. At 385.57 the stock trades around 15 times the consensus version of that year and around 13 times ours. On forward sales, the same screen that shows 8.0 times consensus revenue at the start of calendar 2028 shows 6.8 times on the scenario. One basis note: the screen’s 1.83 trillion market cap implies the basic share count, while the earnings math uses diluted. On the diluted basis the cap runs closer to 1.90 trillion and every sales multiple here sits about four percent higher.

Two caveats. The surprise would be XPU heavy, and custom accelerators carry the lowest gross margins in the semiconductor segment, so the earnings upside runs behind the revenue upside, and the bear narrative will rotate from growth to margins the moment the revenue argument is lost. The CFO preempted this by telling investors to model the two segments’ margins separately, which is what the math above does and what the attached model does on its own tab. And being right about the numbers does not settle the return. If the market holds today’s forward multiple as estimates catch up, the two year outcome is fine but unspectacular. Business quality and price remain separate axes, and this piece argues only about the first one.

The read through may matter more than the stock. Management said networking ran almost 40 percent of AI revenue this quarter and should normalize near 30. Thirty percent of a 204 billion AI year is over 60 billion of networking silicon in fiscal 2028, switches, DSPs, optics, retimers, the entire connectivity layer this publication spends its weeks grading. If the scenario is right, it is right for the whole scorecard universe at once. That is also the risk stated plainly: a basket built on this thesis wins together and loses together.

What I am watching

Four checkpoints between now and the September 3 report, each one falsifiable:

Does Q3 print at or above 16 billion in AI revenue, and does the full year number move above 56, or merely get reiterated? Does another quarter of bookings land anywhere near 30 billion, or was Q2 a one time wave of customers ordering early around the power constraint? Does management ever say about 2028 supply what it now says about 2026 and 2027, that it is secured? Does the first 35 billion Apollo tranche close with named investors?

Three or more yes, and the consensus fade in fiscal 2028 is a modeling artifact waiting to be repriced. Three or more no, and our 10 percent line is wishful thinking dressed up in a spreadsheet. A split decision is an answer too: the order book has not yet earned the curve, and the scenario stays a scenario for another quarter.

Consensus and the order book can both be right about fiscal 2028 in exactly one scenario: the orders are real and the gigawatts slip a year. Every other path breaks one of them. Which one would you rather be short?