AVGO · Forward model · AI Semiconductors · Bear case
What has to happen in AI Semiconductors
Model as of
This page changes AI Semiconductors inside the complete AVGO model. Every other vertical, corporate programme and valuation assumption stays in the selected portfolio case.
Shares this vertical and portfolio case. Slider and horizon edits stay in your browser.
AI Semiconductors
Basis quarter$16.70B
Final quarter$13.22B
Implied CAGR−5%
Final revenue mix62%
Custom accelerators designed with a handful of hyperscalers and frontier labs, plus the Ethernet fabric that wires them together. Half the company already, and the only line whose slope decides whether the stock is cheap or expensive.
Last four quarters
2025 Q4
$6.44B
Estimated
2026 Q1
$8.40B
Reported
2026 Q2
$10.80B
Reported
2026 Q3
$16.70B
Reported
Custom AI accelerators (XPUs) for hyperscalers and frontier labsAI networking: Tomahawk and Jericho switching siliconOptical DSPs, retimers and PCIe/Ethernet interconnectCo-packaged optics and scale-up fabric
Megawatts energised
759 MW
at the basis quarter
759 MW of AI content shipped in the basis quarter: $16.7B at the $25M per MW management now discloses, at 88% utilisation. Derived.
Megawatts added
222 MW/qtr
changing −26.0% per quarter
222 MW more in the first projected quarter, which is what the guided $21.7B of Q4 AI revenue requires.
Utilisation
88%
gliding toward 93%
88% of shipped content deployed on schedule. Tan names timely customer deployment, not demand, as the risk to the outlook.
Revenue per MW
$25M/qtr
drifting 0.0% per quarter
$25M of Broadcom content per MW - the midpoint of the $20-30B per gigawatt Tan disclosed on the Q3 call.
AI Semiconductors
Latest: $13.22B (2031Q3E)
| Period | Value |
|---|---|
| 2025Q2 | $4.44B |
| 2025Q3 | $5.20B |
| 2025Q4 | $6.44B |
| 2026Q1 | $8.40B |
| 2026Q2 | $10.80B |
| 2026Q3 | $16.70B |
| 2026Q4E | $20.83B |
| 2027Q1E | $23.41B |
| 2027Q2E | $24.87B |
| 2027Q3E | $25.53B |
| 2027Q4E | $25.62B |
| 2028Q1E | $25.32B |
| 2028Q2E | $24.75B |
| 2028Q3E | $23.99B |
| 2028Q4E | $23.12B |
| 2029Q1E | $22.19B |
| 2029Q2E | $21.22B |
| 2029Q3E | $20.24B |
| 2029Q4E | $19.26B |
| 2030Q1E | $18.30B |
| 2030Q2E | $17.37B |
| 2030Q3E | $16.47B |
| 2030Q4E | $15.61B |
| 2031Q1E | $14.77B |
| 2031Q2E | $13.98B |
| 2031Q3E | $13.22B |
Assumptions & reasoning
- Modelled as megawatts of AI compute Broadcom's silicon ships into each quarter, times the silicon content per megawatt, rather than as a growth rate. That is deliberate: the customer commitments are quoted in gigawatts, the constraint is advanced packaging and HBM allocation, and a single growth number would hide both. The megawatt count is BACKED OUT of revenue at $18M of content per megawatt, not disclosed — Broadcom publishes no units, no wafer starts and no gigawatts shipped.
- The calibration is checkable against management's own numbers. The first two projected quarters land at $16.0B and $20.4B against a guided $16.0B for 2026 Q3 and the $20.8B implied by a $56B fiscal 2026, and the four quarters of fiscal 2027 total $110B against a stated 'more than $100 billion'. The model clears the guide rather than matching it; that gap is the model being slightly generous in the guided window, not a forecast of a beat.
- Content per megawatt FALLS toward 1.75% a quarter once the drift engages, from $18.0M to $14.3M by 2031. This is the assumption doing the most work and the one most likely to be wrong in either direction. Custom silicon exists because it is cheaper per unit of work than a merchant GPU, and that logic does not stop applying to Broadcom's own pricing once a second ASIC vendor is in the room. If you think Broadcom holds content per megawatt flat, that slider adds roughly a third to the terminal AI line.
- Build rate decays 17% a quarter, so megawatts energised converge on about 2,560 rather than compounding forever. This is where the model refuses to follow the trend line: at the fiscal 2027 exit rate held flat, AI alone would pass $120B a year, and a model with no asymptote arrives at a number larger than the whole accelerator market. The decay is the discipline, and the Bull case is the argument that it is wrong.
- Concentration is the risk that no slider on this page expresses. Roughly six customers carry this line — Google, Meta, Anthropic, OpenAI, Apple and one more — and any single programme moving in-house, slipping a node or being cancelled takes a visible bite out of a business now worth half of Broadcom. Bookings above $30B in the basis quarter against $10.8B shipped is the offsetting fact: the backlog is real and it is contracted.
- Margin starts at 74% and glides to 68%. AI silicon is a high-margin business today because the design win is sticky and the customer pays for the engineering, but the terminal number assumes competition and customer bargaining power eventually cost Broadcom six points.