← TSM forward model

TSM · Forward model · Tesla AI5 · Bear case

What has to happen in Tesla AI5

Model as of

This page changes Tesla AI5 inside the complete TSM model. Every other vertical, corporate programme and valuation assumption stays in the selected portfolio case.

TSM forward model
Horizon
Consolidated fair value $201.80 all other verticals held in this portfolio case
Final-quarter revenue $122M 0% of company revenue
Explicit segment contribution $697M EBITDA less segment capex, before corporate items

The depreciation arrives. TSMC spent NT$496bn on capital expenditure in the basis quarter alone, and its depreciation was FLAT year over year in the first half - NT$359.5bn against NT$359.0bn - because the fabs are not in service yet. That is a timing gap, not a saving, and it closes. Management already guides gross margin down to 65-67% next quarter from 67.7%, citing the 2-nanometer ramp and overseas fabs. This case runs HPC decelerating from 20% sequential while the cost of the capacity built to serve it lands, and prices TSMC as what it is on the balance sheet: a business converting 39% of revenue into plant every quarter.

Tesla AI5

Basis quarter$0
Final quarter$122M
Final revenue mix0%

A line that does not exist yet, carried at zero across every reported quarter and starting in 2027. Tesla's AI5 is a roughly half-reticle inference die that Musk has said will be built at TSMC, two per car, against a stated ambition of two million cars a year. TSMC has never named Tesla as a customer, never mentioned AI5, and publishes no wafer price, so every number on this line is an outside estimate rather than a disclosure. It is here to make one question answerable on the page: what would a flagship automotive AI win actually be worth to a company running at $45bn a quarter? The answer, on the arithmetic below, is about half a percent of revenue - which is itself the finding.

Last four quarters
2025 Q3 $0 Reported
2025 Q4 $0 Reported
2026 Q1 $0 Reported
2026 Q2 $0 Reported
AI5 inference die for vehiclesAI5 for OptimusAI5 for training clusters
Units 1000000/qtr growing +3.0% per quarter 1,000,000 dies a quarter: two million cars a year at two AI5 dies each. Tesla's ambition, not a TSMC number.
Price per unit $150 drifting −1.0% per quarter $150 a die: roughly $22k a wafer over ~175 good half-reticle dies, plus packaging and test. An estimate.
Tesla AI5

Latest: $122M (2031Q2E)

Period Value
2024Q3 $0.00
2024Q4 $0.00
2025Q1 $0.00
2025Q2 $0.00
2025Q3 $0.00
2025Q4 $0.00
2026Q1 $0.00
2026Q2 $0.00
2026Q3E $0.00
2026Q4E $0.00
2027Q1E $140M
2027Q2E $138M
2027Q3E $137M
2027Q4E $136M
2028Q1E $134M
2028Q2E $133M
2028Q3E $132M
2028Q4E $131M
2029Q1E $130M
2029Q2E $129M
2029Q3E $128M
2029Q4E $127M
2030Q1E $126M
2030Q2E $125M
2030Q3E $125M
2030Q4E $124M
2031Q1E $123M
2031Q2E $122M

Assumptions & reasoning

  • Zero in every reported quarter, by construction. Tesla already buys silicon from TSMC and that revenue is inside the Automotive and HPC lines already; this vertical is the INCREMENTAL AI5 ramp above the current run rate, and reading it any other way double-counts.
  • The arithmetic: two million cars a year at two dies each is four million chips, or one million a quarter. At roughly $150 of TSMC revenue per chip - about $22k a wafer over some 175 good half-reticle dies, plus packaging - that is about $150M a quarter, or $600M a year.
  • The chip count grows toward 1.8 million a quarter, which is where Optimus and training clusters would take it. At $150 a chip that is $270M a quarter. A reader who wants the $250M a quarter that circulates for this programme is implicitly assuming roughly six to eight million AI5 dies a year, which is cars plus everything else, not cars alone.
  • At its peak this line is about half a percent of TSMC revenue and moves fair value by cents. That is the point of carrying it: a flagship automotive design win at a company this size is a rounding error, and the model says so explicitly rather than leaving it to intuition.
TSM model map

Explore another vertical