AMD in a DCF: separating AI revenue, margin and capital-spending assumptions

In March 2022, a Hacker News investor described a rough AMD DCF using a 6% discount rate and backward-looking revenue fitting, producing a $42–$121 per-share range alongside a then-current market price of $108.41. These are historical discussion figures, not a current valuation or this article’s price target. A market price inside that range does not validate the model: price is not intrinsic value, and no retrospective backtest of the original model is provided here.

The poster identified the discount rate and assumptions for revenue, operating income or net income as influential over five-, ten- and twenty-year horizons. That is the actual question addressed here: which assumptions should be tested separately as the business mix changes? Revenue assumptions feed profit, cash flow and terminal value. Separating inputs helps identify sensitivities; it does not eliminate forecast error.

AMD disclosed Q2’26 Data Center revenue of $6.718 billion versus $3.240 billion in Q2’25, up 107%. Client and Gaming rose 6%, from $3.621 billion to $3.841 billion, with gaming revenue down 31%. Total revenue rose 50%, from $7.685 billion to $11.536 billion. Data Center’s share increased from about 42% to about 58%. These are same-quarter year-over-year revenue-mix observations, not evidence that every business grew 50%.

Consolidated growth reflects both business growth rates and their weights. Extending a 50% rate does not mathematically guarantee a valuation error, but can conceal assumptions about individual businesses. Segment modelling makes alternative paths for Data Center, Client and Gaming and other businesses explicit before adding them together. Those paths may accelerate, slow or offset one another; a single quarter’s year-over-year growth does not determine them.

The first assumption group is revenue: test the conditions for Data Center and Client and Gaming growth separately rather than mechanically extending 107% or 50%. The second is gross margin: reported GAAP margin was about 40% in Q2’25 and 54% in Q2’26, with the earlier quarter including $800 million of MI308 export-control-related inventory and other charges. The 14-percentage-point difference cannot all be treated as recurring mix improvement, and does not establish future segment margins.

The third group is operating expenses. They rose from $3.193 billion to $4.213 billion, more slowly than revenue, reducing the expense-to-revenue ratio from about 41.5% to about 36.5%. These are two observed quarterly outcomes, not proof of superior Data Center revenue per R&D dollar, future expense growth or permanent operating leverage. The fourth is capex, which rose from $282 million in Q2’25 to $808 million in Q2’26, or about 3.7% to 7.0% of revenue. These are consolidated amounts; without an allocation disclosure, the increase cannot all be attributed to AI accelerators or rack construction.

The 55-analyst consensus aggregated by stockanalysis.com shows FY2026 revenue of $50.95 billion and FY2027 revenue of $89.05 billion, implying approximately 75% growth between the forecasts. FY2027 would average roughly $22.26 billion a quarter, about 1.93 times Q2’26 actual revenue of $11.536 billion. These conversions expose revenue hurdles. They are not company commitments and do not disclose analysts’ actual segment models, so they cannot establish that only one modelling method can produce the consensus.

Treating the $50.95 billion FY2026 consensus as an assumption, subtracting $21.789 billion of H1 actual revenue and the $13 billion Q3 guidance midpoint leaves roughly $16.16 billion for Q4, about 24% above that Q3 midpoint. This is a conditional bridge combining actuals, company guidance and external forecasts, not a Q4 forecast or signed-order total. Different business combinations can produce the same aggregate; the disclosures do not establish that Data Center must maintain 107% growth or that other segments must remain flat.

The loaded references to up to two gigawatts of Anthropic MI450-series capacity, Microsoft Azure Helios deployment and other customer names come from AMD’s own announcements. They are business-direction evidence, but these materials do not disclose the contract amounts, minimum purchase obligations, complete delivery schedules or pricing needed to reconcile FY2027 consensus customer by customer. “Up to” is not a guaranteed purchase floor. Missing public terms do not prove an absence of orders; they limit independent verification of the forecast.

The strongest counterargument is that a more detailed segment model can merely express unverified judgments more precisely. The loaded materials do not provide independent market share, third-party deployment verification, complete analyst modelling methods or capex allocation by segment. Subsequent results can test revenue mix, margins and expenses, cash flow and delivery against guidance. Until then, compare assumptions and retain ranges. Current growth neither proves a high valuation justified nor proves consensus unattainable.