Citi Upgrades AMD to Buy: Dual Drivers from GPU and CPU, Target Price $575

Overview

This report upgrades AMD's rating from 'Neutral' to 'Buy' with a target price set at $575. Citi believes the current market primarily views AMD as a CPU company, pricing in only about a 60% probability for its GPU business to achieve over $50 billion in sales revenue by 2028, which severely undervalues its worth as a 'legitimate second source' outside of NVIDIA. The core catalysts are twofold: first, Meta will significantly increase procurement of AMD's custom AI GPUs (MI450); second, the explosion of Agentic AI is reshaping the server CPU market landscape, with AMD emerging as the biggest winner due to its architectural advantages. Based on SOTP (Sum of the Parts) segment valuation, we assign higher valuation multiples to the data center GPU/CPU businesses, deriving a target price of $575.

Core views

I. Significant Upside Expectation Gap in GPU Business Currently, Wall Street generally views AMD as a CPU manufacturer, expecting low market share in the AI GPU market. Citi believes that as AMD establishes itself as a reliable alternative supplier to NVIDIA, its share among cloud providers like Meta will explode. Specifically, AMD has signed a four-year, total 6 gigawatt (6GW) supply agreement for custom MI450 GPUs with Meta, with the first 1GW of capacity ramping up in the second half of 2026 and throughout 2027. Citi estimates that each 1GW of deployment will bring approximately $15 billion in revenue to AMD. We have upgraded our AI sales forecasts, expecting AI sales to reach $33 billion in 2027 (up 127%-137% YoY) and $50.8 billion in 2028 (up 54% YoY). This means there is a very high probability that AMD will achieve over $50 billion in GPU sales revenue by 2028. II. Agentic AI Ignites New Growth in Server CPUs Besides GPUs, the server CPU market is also seeing structural opportunities. Traditionally, CPUs were mainly used for regular workloads or as 'Head Nodes' for AI applications to handle simple request management. However, as AI evolves from simple chatbot inference to Agentic AI (autonomous agent workflows, such as Claude Code), the role of the CPU changes fundamentally. In Agentic AI, CPUs need to undertake high-complexity tasks such as orchestration, inference loops, data processing, and security components, causing the CPU-to-GPU ratio to rise sharply from 1:8 in model training and 1:4 in inference to 1:1 or even higher. This makes CPUs the new performance bottleneck, triggering strong restocking and upgrade demands. III. CPU Market Landscape and AMD's Competitive Advantages Citi updated its CPU TAM (Total Addressable Market) model post-Computex, significantly raising the 2030 CPU TAM from the previous $131.5 billion to $136.7 billion (CAGR of 36%), which is higher than AMD's own forecast of $120 billion. Among this, general-purpose CPUs account for 39%, AI head nodes for 16%, and Agentic CPU applications account for a high 45%. Regarding the competitive landscape, Citi believes AMD will be the biggest beneficiary of this CPU revival. Comparing competitors: 1. AMD: Its next-generation Venice CPU supports 8 to 256 cores, features an X86 Instruction Set Architecture (ISA), supports multi-threading and single-threading, has a wide SKU coverage, and offers better software compatibility and security than ARM architecture. 2. Intel: Its Diamond Rapids CPU does not support multi-threading and has fewer cores, so performance will lag behind Venice; the next-generation Coral Rapids restores multi-threading but is not expected until 2028. 3. Nvidia: Its Vera CPU uses proprietary spatial multi-threading technology and ARM ISA. While powerful, it faces software migration costs, and due to limited wafer capacity, Nvidia will prioritize high-margin GPU shipments, limiting CPU output. Accordingly, Citi predicts that by 2030, AMD's server CPU market share will rise from 25% in 2025 to 35%, surpassing Intel (46%) and ARM/Others (20%). IV. Valuation and Financial Forecasts Based on the above optimistic assumptions, Citi has upgraded AMD's earnings forecasts, with EPS for 2026/27/28 being 12%/13%/12% higher than market consensus. Using the SOTP (Sum of the Parts) valuation method, we assign the following valuations to each business segment (based on 2028 PE): - Data Center GPU Business: $281 (35x PE) - Data Center CPU Business: $204 (30x PE) - Client Business: $30 (15x PE) - Gaming Business: $2 (8x PE) - Embedded Business: $24 (20x PE) - Net Cash Per Share: Approx. $35 The combined target price is $575, implying an upside of approximately 17.7%.

Analysis framework

The logical derivation of this report follows the path of 'Macro Trend Identification -> Sub-market Structural Changes -> Quantitative Competitive Landscape -> Financial Forecast Revision -> Valuation Anchoring'. First, analysts observed the rapid penetration of Agentic AI technology through visits to Silicon Valley and underlying data, identifying this as a structural inflection point changing hardware demand (shifting from sole reliance on GPUs to equal emphasis on CPU/GPU, with extreme CPU demand elasticity). Second, at the micro customer level, by breaking down Meta and AMD's specific order agreements (6GW MI450), vague 'AI growth' is transformed into specific, quantifiable revenue increments ($15 billion per GW). Third, in competitive analysis, rather than stopping at qualitative descriptions, it delves into technical details such as instruction set architectures (X86 vs ARM), microarchitecture characteristics (multi-threading support, max core count), and supply chain priorities (Foundry capacity allocation), thereby demonstrating why AMD can seize share from Intel and Nvidia. Finally, using the SOTP model, differentiated P/E multiples are assigned to departments at different growth stages and with different business attributes, converting future market share and volume forecasts into current stock price targets.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply and Demand Framework

  Volume-Price Analysis and TAM Expansion under Supply and Demand Framework

  The research report not only analyzes the expansion of the overall market size (TAM) of CPUs/GPUs but further breaks down the growth rate differences across different application scenarios (general-purpose, AI head nodes, Agentic) and deduces the supply-demand mismatch and share transfer among various technological routes (AMD/Intel/Nvidia) in the coming years.

• Valuation MethodSOTP Segment Valuation

  SOTP (Sum of the Parts) Segment Valuation Method

  For diversified tech companies, it is inappropriate to simply use an overall PE for valuation. This method splits the company's various business segments (e.g., high-growth GPUs, mature CPUs, low-growth gaming) separately, prices them according to comparable company valuation multiples in their respective fields, sums them up, and deducts/adds net cash to derive a more accurate intrinsic value.

• Event Game Theory and Behavioral FinanceExpectation Gap/Expectation Management

  Finding Market Pricing Deviations (Expectation Gaps)

  The core argument of the report is built on 'insufficient market expectation'. Analysts point out that the market only priced in a 60% success probability for AMD's GPU business, while institutions believe the actual probability is much higher through research. Capturing this gap between 'consensus expectations' and 'fundamental truth' is the main source of excess returns.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AMD (AMD.US)

  Core beneficiary. Benefiting from the server CPU replacement wave brought by Agentic AI, and gaining significant share in Meta's custom GPU orders.

  Strengths

  Leading Venice CPU architecture (supports multi-threading, multi-core), excellent X86 ecosystem compatibility, TCO (Total Cost of Ownership) lower than NVIDIA solutions.

  Weaknesses

  Gaming and client businesses face cyclical pressures; compared to NVIDIA, its software ecosystem CUDA moat is still being caught up to.

  Comparison

  Compared to Intel, AMD leads in multi-core performance and energy efficiency, with a clearer next-generation product roadmap; compared to Nvidia, AMD is more cost-effective in the CPU field, but still lags in absolute GPU performance.

  Risks

  Agentic AI implementation falling short of expectations; delays in Meta order execution; intense price wars compressing gross margins.

• Intel (INTC.O)

  Indirectly mentioned competitor. Facing risk of share loss in the high-end server CPU market.

  Strengths

  Possesses massive manufacturing capacity and long-term customer relationships.

  Weaknesses

  Current flagship product Diamond Rapids does not support multi-threading and has fewer cores than AMD Venice, making it difficult to pose a strong challenge in the short term.

  Comparison

  Citi explicitly judges that Intel's next-generation products will return to multi-threading only in 2028, meaning AMD will occupy a clear performance generational advantage for the next two to three years.

  Risks

  Slow iteration of process technology; erosion of high-end market share by AMD and ARM architecture.

• NVIDIA (NVDA.US)

  Indirectly mentioned competitor. Dominant in the GPU market but constrained in the CPU field due to capacity and architecture reasons.

  Strengths

  Extremely high barriers in CUDA software ecosystem, strongest absolute GPU performance.

  Weaknesses

  Constrained by foundry capacity, needing to prioritize GPU shipments; Vera CPU adopts ARM architecture, facing software migration costs and ecosystem compatibility issues.

  Comparison

  In Agentic AI scenarios, AMD provides better TCO凭借 (by virtue of) X86 and full-stack capabilities, thereby carving out part of the CPU demand originally belonging to the Nvidia ecosystem.

  Risks

  Facing fierce competition from ASICs and AMD customized solutions.

Key data

• Target Price$575Raised from $460, based on SOTP valuation
• 2027 AI Sales Forecast$33 BillionSignificant YoY growth of 127%-137%
• 2028 AI Sales Forecast$50.8 BillionSignificant YoY growth of 54%
• 2030 CPU Total Addressable Market (TAM)$136.7 BillionCAGR of 36%, previously forecast at $131.5 Billion
• 2030 AMD Server CPU Share35%Significant improvement from 25% in 2025
• Estimated Value of Single Meta Order~$15 Billion per GWInvolves 6GW MI450 custom GPU agreement
• EPS Upgrade Magnitude12%/13%/12%2026/27/28 EPS higher than market consensus respectively

Impact & implications

For AMD, this means its valuation logic is shifting from a traditional cyclical PC/server chip company to a high-growth core AI infrastructure target. If Agentic AI adoption continues to exceed expectations, or if Meta's deployment schedule accelerates, AMD's actual profitability could far exceed current forecasts. For the industry, this indicates that the investment focus of AI computing power is spreading from single GPU training/inference to the CPU side involving complex logical processing, and the X86 architecture still possesses strong vitality in high-performance computing.

Risks

• Competitive Risk: AMD competes directly with Intel in the microprocessor market and directly with Nvidia in the graphics and AI GPU market. Any fluctuation in market share may affect our estimates.
• PC Terminal Market: Approximately 30% of AMD's sales come from the PC industry, which is highly dependent on IT spending. Major fluctuations in IT spending could lead to upward or downward revisions in estimates and ratings.
• Customer Risk: Approximately 15% of AMD's revenue comes from Sony and Microsoft. Major increases or decreases in orders from these key customers will have positive or negative impacts on our estimates.
• Valuation Risk: Although AMD's current trading price is slightly above its historical range, any revision to consensus revenue estimates could lead to high volatility in AMD stock.
• Takeover Bid: Any public or private takeover bid for all or part of AMD's business could drive the stock price up and lead us to re-evaluate the rating.

What to watch

• The actual adoption rate and implementation speed of Agentic AI in enterprises (there are already signs that adoption rates surged in February).
• Whether the first batch of 1GW capacity for Meta's custom MI450 GPU will ramp up as scheduled in the second half of 2026.
• The actual mass production yield and market feedback of AMD's next-generation Venice CPU.
• The release schedule and performance of Intel's Diamond Rapids and subsequent Coral Rapids products.
• Trends in enterprise IT spending budgets, especially willingness for capital expenditure on AI-related hardware.