TSM · Forward model

Revenue by vertical, 20 quarters out

Model as of 2026-08-25

Each segment is projected from its own operating driver, rolled up into consolidated cash flow, and discounted back to a fair value per share. The assumptions below are editable — change them and every number on this page moves with them.

TSMC has no reportable segments in the Microsoft sense - it is a single foundry business and reports one consolidated profit and loss. What it publishes every quarter, on one slide of its earnings deck, is revenue BY PLATFORM as a percentage of net revenue: High Performance Computing, Smartphone, Internet of Things, Automotive, Digital Consumer Electronics and Others. Those six percentages sum to exactly 100 in all eight quarters here, so multiplying them by the reported dollar revenue gives six lines that reconcile to the consolidated total exactly. That is the whole basis of this model's verticals, and its limitation: the percentages are WHOLE NUMBERS, so every line carries up to half a point of rounding - about $200M a quarter on a $40.2B base, which is half the DCE line and a quarter of the Others line. Read the two small platforms as indicative and the three large ones as sound. What TSMC does NOT publish is any economics by platform. There is no gross margin, no operating margin, no capital expenditure and no capacity by platform - only consolidated figures - so all six verticals here carry the SAME margin and the SAME capital intensity, and the split does no economic work at all. It changes the revenue path and nothing else. The margin used is the disclosed 60.3% operating margin plus depreciation and amortisation added back at 15.14% of revenue, the first-half 2026 rate, giving 75.4%. There is no corporate overhead layer because there is nothing unallocated to put in it. One number in that add-back deserves attention: TSMC's depreciation was NT$359.5bn in the first half of 2026 against NT$359.0bn in the first half of 2025 - essentially FLAT - while capital expenditure ran at NT$496bn in the basis quarter alone. The fabs being built now are not yet in service, so the depreciation from this capital wave has barely started to arrive. Finally, the driver: TSMC publishes wafer shipments only in total (4,336 thousand 12-inch equivalents in the basis quarter) and gives no capacity or shipment figure per platform, so all six verticals run growth drivers. A per-platform wafer driver would have to assume every platform earns the same revenue per wafer, which is the one thing that is certainly false when HPC sits on 2-nanometer and IoT sits on mature nodes. A seventh vertical, Tesla AI5, is carried at ZERO in every reported quarter and starts in 2027. It is not a TSMC disclosure of any kind - TSMC has never named Tesla as a customer, never mentioned AI5 and publishes no wafer price - and it exists on this page to make one question answerable: what is a flagship automotive AI design win worth to a company running at $45bn a quarter? The arithmetic is two million cars a year at two half-reticle dies each, so four million chips or one million a quarter, at roughly $150 of TSMC revenue per die - about $22k a wafer over some 175 good dies, plus packaging. That is $153M in its first quarter, growing to $225M by 2031 as Optimus and training clusters lift the die count toward 1.8 million a quarter. It never exceeds 0.28% of revenue and it moves base fair value by one dollar, from $472.80 to $473.80. Note also that Tesla's EXISTING silicon is already inside the Automotive and HPC lines, which are projected from history that contains it, so this vertical must be read as incremental AI5 volume above the current run rate or it double-counts.

Where each case comes from

From cash flow to fair value

The projected path

Every row is projected. A year-over-year change is shown only where the quarter it compares against exists — an em dash means there is no comparable quarter, not a flat year.

Model revisions

Assumptions are marked to reality as each quarter prints. Every change is appended here, with the fair value the model produced at the time, so the model's own history stays visible.