AMD: The Second Platform Claim

I. Investment Question and Initial View

Can AMD ($AMD) convert the Instinct AI accelerator cycle into a durable second-platform position in data centre infrastructure — and at a pace consistent with what the current share price requires?

Three conditions define the answer. The MI450 hardware cycle must execute on its 2H2026 delivery schedule. AMD's ROCm software platform must reduce customer migration costs to the point where hardware economics govern procurement. And the OpenAI infrastructure partnership must convert from a performance-gated equity commitment into recurring shipment revenue. The conditions are sequential: hardware slippage breaks the software and systems arguments before they can be tested.

The structural logic holds. AMD is the only credible large-scale alternative to NVIDIA ($NVDA) in high-performance AI accelerators, and the transition from a CPU-centric business to an AI infrastructure platform is now visible in the reported margin structure. The direction of travel is not in question. What the current share price asks is whether the pace and scale of that transition match the trajectory it is pricing — a question the filed results as of 1Q2026 support directionally but do not yet validate in magnitude.

Client and Gaming and Embedded not disaggregated in 1Q2026 quarterly disclosure. FY2025 FCF excludes $1.2B from discontinued ZT Manufacturing operations.

Three qualifications condition the reading.

The EPYC and Instinct composition.

The data centre segment houses two businesses with distinct competitive characters. EPYC, AMD's server CPU franchise, has established genuine competitive leadership — its trajectory is not in contention. Instinct, the AI accelerator franchise, has demonstrated volume shipment capability, but its supporting software platform, system integration, and hyperscaler penetration have not been established at the scale the current share price appreciation implies. The incremental case is Instinct.

ROCm maturity as the binding software condition.

ROCm is AMD's software platform for AI development (similar to NVIDIA's CUDA platform) , the mechanism through which customer migration costs fall to the point where hardware economics govern procurement. Management characterises its maturity positively but provides no standardised metrics against which that characterisation can be verified. Parity can only be tracked indirectly through customer win rates, hyperscaler GPU allocation patterns, and migration friction commentary.

The OpenAI warrant as success-contingent dilution.

AMD issued warrants to OpenAI covering 160 million shares at effectively zero cost, with vesting contingent on Instinct GPU purchase milestones and AMD stock price targets. None had vested as of FY2025 year-end. Dilution materialises only if AMD delivers the AI revenue that triggers the milestones — a structure that concentrates dilution pressure at the moment the bull case is confirmed.

The horizon for this primer is FY2026 through FY2028, with FY2029 as the sustained-scale validation point. FY2026 tests the MI450 production ramp and the first deployments under the OpenAI commitment. FY2027 tests whether purchase milestones are converting to observable warrant vesting and whether system-level integration is generating measurable margin differentiation. FY2028 tests the sustained-scale claim — whether AMD's data centre franchise has reached the earnings base required to validate the growth trajectory the current share price is pricing.

---

II. Business Architecture and Earnings Composition

AMD is a fabless semiconductor designer: it develops products and contracts all manufacturing to third parties, principally TSMC. Revenue is transactional — generated through the sale of processors, graphics chips, and adaptive computing devices to original equipment manufacturers, systems integrators, and distributors. There is no recurring subscription base; durability comes from annual product cadence and the compounding of customer relationships in the data centre.

Three reportable segments organise the business. The table below shows operating economics for FY2025 alongside the corporate line that reconciles segment results to the consolidated GAAP figure.

All Other includes $2.3B amortisation of acquisition-related intangibles, principally Xilinx, and $1.6B in stock-based compensation. FY2025.

The gap between segment aggregate margin and the consolidated GAAP figure is partly structural rather than purely operational. As Xilinx acquisition intangibles run off over the horizon, the GAAP margin should mechanically recover part of the gap to segment-level economics as the amortisation component rolls off, without requiring any improvement in underlying performance.

Data Center.

The segment houses two product families with distinct competitive characters. EPYC is AMD's server processor franchise, where AMD has established genuine competitive leadership: AMD management has stated that all ten of the world's largest hyperscalers now standardise on EPYC. Instinct is AMD's AI accelerator line — the growth claim. The MI300X cycle established AMD as a credible volume supplier of AI accelerators; the MI450 cycle beginning 2H2026 is the test of whether that position extends at scale. The FY2025 margin dip relative to FY2024 reflects ramp and R&D investment required to sustain the annual hardware cadence; the 1Q2026 recovery signals that investment is absorbing into the cost base.

Client and Gaming.

The segment combines two operationally distinct businesses whose trajectories diverge. Client — Ryzen desktop and mobile processors — recovered strongly as the post-pandemic inventory correction cleared and AMD gained premium market share. Gaming — semi-custom chips for PlayStation and Xbox consoles — is in structural decline as the current console cycle matures. AMD combined its reporting of these two businesses from 1Q2025, which obscures the Client growth story and the Gaming drag within the same line.

Embedded.

The direct output of the 2022 Xilinx acquisition: field-programmable gate arrays (FPGAs) and adaptive processors deployed in industrial, automotive, and telecommunications infrastructure. Customers pay a premium for logic that can be reconfigured without new silicon — a product characteristic that structurally supports the segment's margin profile. Embedded reached peak revenue in FY2023 following the acquisition before a sharp inventory correction; FY2025 represents stabilisation. At 36%, it is the highest-margin segment in the portfolio at any revenue level.

Operating model.

AMD's fabless structure concentrates capital in R&D rather than physical manufacturing. The key operating metrics at the current point in the ramp are summarised below.

R&D intensity reflects the cost of maintaining parallel development across multiple hardware generations simultaneously — the price of the annual cadence commitment. Purchase commitments represent management's confidence in the 2H2026 production ramp and, symmetrically, a fixed obligation if demand disappoints.

Geographic distribution.

Europe and Japan growth incorporates a reclassification of revenue previously reported in a consolidated "Other" category. Underlying growth in both regions reflects hyperscaler AI infrastructure deployment.

China is the primary geographic concentration and regulatory risk point. Management excludes China-based AI GPU revenue from formal guidance until export licence status is confirmed — a discipline that treats China as a potential upside variable rather than a baselined assumption.

ZT Systems.

In October 2025, AMD acquired ZT Systems and immediately divested its manufacturing operations to Sanmina Corporation, retaining only the design teams. At a net cost of approximately $800M, AMD acquired system design capability: engineers able to configure AMD CPUs, GPUs, and networking components into complete rack-scale AI infrastructure. The product output is Helios, AMD's forthcoming rack-scale platform for hyperscale deployment.

Disclosure constraints.

Instinct GPU revenue is not disaggregated from EPYC CPU revenue within the Data Center segment. The Client and Gaming segment reorganisation from 1Q2025 prevents historical comparison of the two constituent businesses. The ZT Design Business contribution is not reported at segment level. Customer concentration is disclosed only when an individual customer exceeds 10% of total revenue.

---

III. What Must Be True

Six conditions must hold for the investment case to pay out across the FY2026 to FY2028 horizon. Four are load-bearing — failure at any breaks the sequential mechanism the thesis depends on. Two are amplifying — they determine the margin and scale of the outcome but do not gate the underlying recovery. The matrix below identifies where each condition is first testable and where it must be resolved.

Condition 1 — MI450 production ramp.

The Instinct MI450 series must enter volume production and begin revenue recognition in the second half of 2026. Management has committed to a sharper ramp in 4Q2026 relative to 3Q2026 and positioned the first 1-gigawatt OpenAI deployment as an MI450-powered event. Failure is not a scheduling inconvenience: the foundry commitments AMD has made for FY2026 wafer capacity are unconditional, meaning a ramp delay produces stranded supply costs against a revenue shortfall simultaneously.

Condition 2 — Advanced node and HBM supply execution.

The condition has two parallel dimensions. The first is TSMC wafer yield: AMD must convert its FY2026 foundry commitments into high-yield, leadership-class silicon. If the transition to the node required for MI450 encounters yield difficulties, AMD pays the commitment cost without generating the corresponding product. The MI308 China charge demonstrated the write-down mechanics; yield failure is a more severe version of that mechanism because it cannot be resolved by regulatory relief.

The second dimension is HBM availability. Instinct AI accelerators incorporate High Bandwidth Memory — a specialised memory component packaged alongside the AMD compute die, sourced from SK Hynix, Samsung, and Micron. AMD could receive every committed TSMC wafer on schedule and still face MI450 production constraints if HBM allocation is insufficient. The two supply chains are independent: TSMC controls wafer yield; memory manufacturers control HBM capacity allocation. Both must resolve in AMD's favour for the MI450 ramp to proceed at the volume 2H2026 guidance requires. HBM supply concentration, pricing dynamics, and the competitive economics of the memory supply chain are examined in HBM Supply Economics and Margin Durability.

Condition 3 — ROCm software parity.

AMD's ROCm software platform must reduce customer migration costs to a level where hardware economics govern procurement. The condition cannot be tested directly from AMD's own disclosures — it can only be inferred from enterprise customer win rates, hyperscaler GPU allocation patterns, and commentary on migration friction. The failure mode is sustained migration friction: migration costs remain high enough that hardware-price-advantaged Instinct GPUs cannot displace incumbent silicon in workloads where software tooling is the binding constraint.

Condition 4 — OpenAI purchase milestone conversion.

The warrants AMD issued to OpenAI vest on undisclosed Instinct GPU purchase milestones and stock price targets. The first observable test is whether any tranche vests in the FY2026 annual filing. Vesting is a binary commercial signal — it confirms OpenAI has met an undisclosed purchase threshold even if the underlying quantity remains opaque. Non-vesting through FY2026 is not automatically a failure, given the warrant horizon extends to October 2030, but it shifts the burden of proof onto FY2027 deployment data.

Condition 5 — Helios rack-scale integration.

The Helios platform must reach production readiness for hyperscale deployment in FY2026 and generate measurable margin differentiation at the Data Center segment level by FY2027. Helios is the product test of whether the ZT Systems acquisition justifies its cost. Failure leaves AMD as a component supplier — capturing processor and accelerator margin but ceding the system design and integration premium that rack-scale deployment commands.

Condition 6 — Client market share threshold.

AMD must reach and sustain its stated target of greater than 40% unit share in the client processor market. The condition is amplifying rather than load-bearing: failure does not break the data centre thesis, but it removes the free cash flow contribution that funds parallel investment in the Instinct roadmap and reduces the financial resilience of the overall business if data centre ramp timing slips.

Condition interactions.

The four load-bearing conditions form a sequential dependency chain. Supply execution (2) gates hardware shipment (1). Hardware (1) then gates both software adoption (3) and commercial conversion (4) — ROCm parity is untestable at volume and OpenAI deployments cannot begin without MI450 in production. A yield failure at Condition 2 therefore cascades through all four load-bearing conditions before any can be independently tested. The amplifying conditions operate in parallel: Helios integration and Client market share improve the outcome of a successful chain but cannot compensate for failure within it.

---

IV. Capital Allocation and Balance Sheet Posture

AMD's fabless structure concentrates capital at the architectural layer rather than the manufacturing layer. The output is a balance sheet that carries net cash at a point in the cycle when Intel is leveraged and NVIDIA is returning capital at a rate that reflects structural monopoly margin. AMD operates as a capital-light business directing most of its financial advantage into the R&D intensity required to sustain the annual hardware cadence, with a buyback programme running as a secondary priority and no dividend.

R&D as the primary capital commitment.

R&D expenditure rose 25% in FY2025 to $8.1 billion — approximately 23% of revenue. The 1Q2026 run-rate of $2.4 billion per quarter implies a further step toward $9.6 billion annualised as the MI450 and next-generation development cycles run in parallel. This is the financial expression of the cadence commitment: maintaining annual product launches requires simultaneous investment across multiple hardware generations, and the cost of that parallelism sits entirely in the operating expense line rather than in a capital expenditure base AMD does not carry. R&D intensity has remained above 20% of revenue across the transition from CPU-centric to AI-infrastructure business, confirming that the cadence operates as a prior claim on the revenue growth it is simultaneously enabling, ahead of margin expansion.

Acquisition history.

Two acquisitions define AMD's balance sheet trajectory. The 2022 Xilinx integration introduced a $2.3 billion annual amortisation charge — the primary source of the GAAP-to-segment margin gap — alongside the Embedded business, which at 36% is the highest-margin segment in the portfolio. The ZT Systems acquisition in October 2025 was structured as an engineering transaction: AMD paid $3.2 billion, immediately divested the manufacturing operations to Sanmina for $2.4 billion, and retained the design teams at a net cost of approximately $800 million. The retained capability is the engineering input for Helios, AMD's rack-scale AI infrastructure platform. Both acquisitions were funded from the balance sheet without recourse to debt.

Balance sheet.

AMD closed FY2025 with $10.6 billion in cash and short-term investments against $3.2 billion in total debt, producing a net cash position of $7.4 billion and net leverage of negative 1.1 times. The liquidity position has expanded materially — FY2024 cash and investments totalled $5.1 billion — funded by the FCF step-up as Data Center margins recovered. The capital structure is a resource that management has committed to the ramp through the purchase commitment structure rather than through incremental debt.

Purchase commitments as the balance sheet risk.

AMD's $12.2 billion in unconditional purchase commitments at FY2025 year-end — $8.5 billion of which falls due in FY2026 — is the primary balance sheet feature requiring analytical attention. These are take-or-pay obligations: AMD owes the commitment regardless of whether demand materialises at the volume the ramp requires. The MI308 China charge demonstrated the write-down mechanics when inventory built ahead of uncertain demand; an $8.5 billion commitment against a demand environment that deteriorates is a more severe version of that mechanism, because the obligation cannot be resolved by regulatory relief or inventory reclassification. The commitment is simultaneously management's strongest signal of conviction in the 2H2026 ramp and the most significant financial downside exposure if Conditions 1 and 2 from the investment framework are not met on schedule.

The OpenAI warrant.

The 160 million warrants issued to OpenAI at a $0.01 exercise price represent potential dilution of approximately 9.7% of the current diluted share count. No tranches had vested as of FY2025 year-end; the warrants are not yet reflected in the diluted share count. The analytical distinction from standard SBC dilution is that the warrant is success-contingent: it materialises only if AMD delivers the Instinct GPU revenue that triggers the purchase milestones. The base case risk — where revenue grows but not at a pace that fully absorbs the dilution at the moment of vesting — is examined in the quantitative conditions section.

Shareholder returns.

AMD returned $1.3 billion through buybacks in FY2025, repurchasing 12.4 million shares against a $9.4 billion remaining authorisation. Stock-based compensation ran at $1.6 billion in FY2025 — 4.6% of revenue — producing modest net dilution on an annual basis. AMD does not pay a cash dividend; management's stated priority is reinvestment, with buybacks as a capital return mechanism. The buyback programme size relative to the R&D commitment confirms the hierarchy: the annual cadence is funded first, capital returns follow from what remains.

---

V. Competitive Context

AMD's competitive position does not map cleanly onto a single peer set. In server CPU architecture, AMD is the established market leader executing against an incumbent in structural difficulty. In AI accelerators, AMD is the challenger to a full-stack incumbent whose operating margin and software moat have no direct parallel in listed semiconductors. The two battlegrounds share a product roadmap and a balance sheet but operate under different competitive dynamics: CPU leadership was earned through sustained execution and is now AMD's to defend, while the accelerator position is still being built and remains contingent on closing a software gap that hardware performance alone cannot resolve.

FY2025 for all three. NVIDIA FY2025 ends January 2025; AMD and Intel end December 2025.

NVIDIA.

Jensen Huang's characterisation of NVIDIA as selling "AI factories" rather than chips is the most precise description of the competitive asymmetry AMD faces. CUDA — optimised across 4.7 million developers and more than 4,400 accelerated applications — is a software moat that compounds independently of hardware performance: every workload trained on NVIDIA silicon, every model optimised for CUDA, and every enterprise deployment built around the NVIDIA stack extends the switching cost that AMD's hardware must overcome at the point of sale. NVIDIA's acquisition of Mellanox extended the moat into the networking layer, giving NVIDIA margin capture across the full scale-up fabric — compute, memory bandwidth, and interconnect — that AMD must assemble from third-party components or through the ZT Systems design capability. The 62.4% GAAP operating margin is the financial expression of a position where procurement decisions are made on ecosystem compatibility rather than hardware total cost of ownership. AMD's multi-source thesis depends on hyperscalers prioritising supply chain diversification and negotiating leverage over that ecosystem inertia — a commercial dynamic that is real but structurally weaker than a hardware performance argument.

Intel.

Under new CEO Lip-Bu Tan, Intel's competitive position in FY2025 is defined by the capital requirements of its IDM 2.0 transition rather than by its product roadmap. A negative 4.2% GAAP operating margin and 0.97x net leverage reflect a manufacturing base whose fixed costs cannot currently be absorbed by the revenue the business generates. The 18A node transition — Intel's bid to reclaim process technology leadership through its owned fabs — is the mechanism on which the recovery thesis depends, and it is consuming capital that AMD, as a fabless operator, does not need to deploy. AMD has already claimed Intel's server CPU market share systematically: management states that all ten of the largest hyperscalers now standardise on EPYC. Intel is not a credible near-term threat in AI accelerators; its analytical relevance to this primer is as the operational template for what execution failure at scale produces in an integrated margin structure, and as the re-rating inversion reference examined in the quantitative conditions section.

Synthesis.

AMD's competitive differentiation is as much structural as it is technical. The open alternative thesis — that hyperscalers will allocate GPU infrastructure across multiple vendors to avoid single-source dependence and maintain pricing leverage — is a demand-pull argument that operates independently of ROCm's current maturity relative to CUDA. It provides AMD with a commercial tailwind that hardware performance parity alone would not generate. The binding question is whether that tailwind is durable at the scale the current share price requires, or whether it produces a volume floor without the margin trajectory needed to close the gap to NVIDIA's economics. AMD's CPU leadership in EPYC demonstrates that winning the open alternative argument in a software-defined market is achievable over a sufficiently long execution horizon. Whether the AI accelerator market compresses that horizon to within the FY2026 to FY2028 window is what the conditions framework is designed to test.

---

VI. Risk Architecture

AMD's risk profile is defined by a single structural dependency that propagates across multiple categories simultaneously. The fabless model concentrates every operational risk at the manufacturing interface — TSMC yield, node transition timing, and HBM supply allocation are all expressions of the same underlying exposure. Regulatory risk sits independently of this but is concentrated in one geography: China contributes 22% of revenue under a licensing regime that has already produced a material write-down and carries an uncodified government revenue-share expectation with no formal mechanism. The consequence is a risk architecture where the downside cases in the operational and regulatory categories can compound without any natural offset between them, and where the $8.5 billion FY2026 purchase commitment creates a fixed cost floor against which both can simultaneously apply pressure.

Foundry concentration.

AMD has no manufacturing of its own. The entire Instinct and EPYC product roadmap runs through TSMC's advanced nodes, and the $8.5 billion due in FY2026 purchase commitments are unconditional — AMD owes them whether or not the corresponding products ship. A disruption to TSMC capacity, whether through yield difficulties on the node required for MI450, geopolitical instability in Taiwan, or capacity allocation to NVIDIA at AMD's expense, produces a compounded outcome: stranded supply costs against a revenue shortfall, with no internal manufacturing base to draw on. Management acknowledges alternative foundry options in qualitative terms but leadership AI products remain TSMC-concentrated in every disclosed product plan through the primer horizon.

China regulatory exposure.

China contributed $7.8 billion — 22.4% of FY2025 revenue — and the regulatory environment governing that contribution is unstable. AMD recorded an initial $800 million charge on Instinct MI308 products in 2Q2025 following new Department of Commerce restrictions; partial licence grants later in the year allowed a partial reversal, producing a net charge of $440 million for the year. The more structurally significant element is the mechanism that has since emerged: the U.S. government's uncodified "expectation" of a 15% revenue share on licensed China GPU sales. No regulation has been published; no formal mechanism exists. AMD excludes China-based AI GPU revenue from formal guidance until licence status is confirmed — an appropriate discipline that also acknowledges the fragility of the revenue base. If the 15% expectation hardens into codified policy, it functions as a gross margin charge on a material revenue line without any corresponding cost reduction.

Annual cadence execution risk.

The annual AI GPU development cadence — a competitive requirement, not a self-set preference — concentrates execution risk at each release window. Parallel development of multiple hardware generations raises the probability of design errors or missed manufacturing timing that a sequential model would not produce. Falling behind by one quarter in the FY2026 window delivers disproportionate commercial consequences because hyperscaler procurement cycles are not paused to accommodate slippage.

Two interaction patterns require attention.

The first is the Advanced Node Dependency triad, linking foundry concentration, execution cadence, and inventory valuation. A TSMC yield failure on the node required for MI450 cascades into three simultaneous outcomes: stranded foundry commitment costs, a cadence slip that misses the FY2026 hyperscaler window, and an inventory write-down on silicon that cannot be deployed. The three conditions share the same trigger and propagate together before any can be tested in isolation.

The second interaction involves China exposure and the warrant vesting structure. A licence revocation or tightening of the uncodified revenue-share expectation removes revenue that currently funds parallel R&D investment and reduces AMD's operational margin at the exact moment the FY2026 purchase commitment obligations fall due. Margin compression in this scenario weakens the probability that Instinct GPU volumes reach the undisclosed purchase milestones required for OpenAI warrant vesting. The regulatory and commercial risk vectors converge on the same financial constraint.

---

VII. What the Observed Price Implies

The analysis in this section characterises the financial conditions that the current enterprise value implies. It does not constitute a price target or a view on fair value. The reverse DCF framework maps the territory; the watch conditions in Section VIII track whether the terrain is being navigated as the market has priced.

AMD's enterprise value of approximately $814 billion — derived from a share price near $500 against $3.2 billion in debt and $10.6 billion in cash — places a burden of proof on the forward earnings trajectory that the filed results as of 1Q2026 do not yet validate in magnitude. This section works backwards from that enterprise value to characterise what must be true in financial terms, identifies the discount rate and growth assumptions under which the current price is justifiable, and maps the load-bearing conditions to the financial exposure they represent in the current valuation.

The reverse DCF

Model inputs

The starting point is AMD's FY2025 normalised free cash flow of $5.52 billion. At $814 billion in enterprise value, AMD trades at approximately 147 times normalised free cash flow on the current earnings base. The market is pricing forward earnings, not the current base — the question is what forward earnings trajectory is required.

Working backwards from $814 billion at the fundamental WACC of 12.7% and a 3.5% terminal growth rate, the required terminal free cash flow — the level at which FCF stabilises into perpetuity — is approximately $79 billion. Against the FY2025 starting point, that is a 14-fold increase. If the terminal state is reached by FY2030, the implied five-year FCF compound annual growth rate exceeds 60%.

This required terminal FCF is distinct from the explicit FY2030 bull-case FCF used in the sensitivity table below. The sensitivity assumes AMD reaches $60 billion of FCF by FY2030 and then capitalises the terminal value from that base; the $79 billion figure is the stabilised terminal FCF level required for the current enterprise value to clear under the central WACC/TGR assumptions.

A cross-check against management's forward target establishes where the guidance scenario sits relative to the current price. AMD management has characterised a greater-than-$20 EPS level as a medium-term objective. At $20 earnings per share on a 1.65 billion diluted share count, net income would be approximately $33 billion; normalised free cash flow, which runs above net income for AMD given amortisation add-backs on the Xilinx intangibles, would be in the range of $35–38 billion. Discounted at 12.7% with a 3.5% terminal growth rate and treating FY2028 as the year the target is reached, that FCF level implies a share price in the range of $180–195. At the fundamental discount rate, that places the guidance-implied share price materially below the current market level.

The gap is not evidence of mispricing in isolation. It indicates that the observed price is consistent with either a higher probability weight on the bull-case FCF trajectory, or a lower risk premium than AMD’s fabless supply exposure and MI450 execution burden would normally imply.

Sensitivity

The table below maps implied share prices across combinations of discount rate and terminal growth rate using the bull case free cash flow trajectory — the scenario under which all six conditions from Section III execute on schedule and AMD achieves platform-level margin economics by FY2030. The bull case assumes FCF of $10 billion in FY2026, rising to $18 billion in FY2027, $30 billion in FY2028, $47 billion in FY2029, and $60 billion in FY2030. Net cash of $7.4 billion is added to the enterprise value before converting to per-share price on a 1.65 billion diluted share base, excluding the 160 million OpenAI warrants that remain unvested.

Bull case FCF trajectory: $10B (FY2026E) → $18B (FY2027E) → $30B (FY2028E) → $47B (FY2029E) → $60B (FY2030E). Net cash $7.4B added; 1.65B diluted shares (excludes 160M unvested OpenAI warrants; full vesting reduces every cell by ~8.8%). The market price cell (10% / 4.5%) is the only combination at which the bull case trajectory is consistent with the current ~$500 share price. The central case (12.7% WACC / 3.5% TGR) implies $297 per share.

In the entire feasible parameter space, only the combination of the lowest discount rate (10%) and the highest terminal growth rate (4.5%) produces a share price consistent with the current market. At the fundamental WACC of 12.7%, the bull case trajectory implies a share price in the range of $251–$329 across all TGR assumptions — materially below $500 even assuming the entire bull case executes. The central case (12.7% WACC / 3.5% TGR) produces $297 per share. Every one-percentage-point increase in the WACC from the 10% floor reduces the implied share price by approximately $50–80 per share across the full sensitivity space.

The OpenAI warrant creates an additional sensitivity not visible in the table. Full vesting of all 160 million warrants would raise the diluted share count to approximately 1.81 billion, reducing each cell by approximately 8.8%. The structure is success-contingent — dilution materialises at the same time the bull case FCF trajectory is confirmed — which means the per-share compression from warrant vesting coincides with the moment the narrative is strongest.

Conditions-to-price mapping

The sensitivity table establishes the outer bound: the current price is consistent with the complete bull case FCF trajectory only under a low execution-risk discount rate and extended terminal-growth assumption.

Conditions 1 and 2 — the MI450 hardware and supply execution chain — are the gatekeeping entries. A ramp failure in 2H2026 does not reduce the terminal FCF estimate modestly; it collapses the FY2026 and FY2027 FCF assumptions entirely. The $8.5 billion FY2026 supply commitment becomes stranded cost against a revenue shortfall, and the ramp deferral pushes the first confirmable Instinct volume data at least two to four quarters to the right. The bull case FCF of $10 billion in FY2026 and $18 billion in FY2027 would compress toward $4–6 billion and $8–10 billion respectively, removing approximately $15–20 billion in present-value FCF while the investor narrative simultaneously migrates from platform execution to ramp risk. Multiple compression applies pressure to the denominator in the sensitivity table independently of the FCF estimate.

Condition 3 — ROCm software parity — does not affect the FY2026 FCF assumptions directly but governs whether the FY2028–FY2029 trajectory is achievable. The enterprise segment of the AI GPU market, where migration friction is the binding procurement constraint, is the revenue tranche above the hyperscaler anchor. Failure on the software condition caps the addressable market at the hyperscaler layer and compresses the FY2028–FY2029 FCF assumptions from the $30–47 billion range toward a $15–25 billion range — conditions under which the $60 billion terminal FCF is not accessible within the FY2030 horizon.

Condition 4 — OpenAI warrant vesting — operates primarily as a multiple signal rather than an FCF estimate input. Non-vesting through FY2027 does not reduce the earnings model, but it removes the narrative anchor that supports the premium discount rate currently embedded in the market price. The sensitivity table shows that a two-percentage-point WACC increase from 10% to 12% reduces the implied share price by approximately $75–145 per share across TGR columns, without any change in the underlying FCF trajectory. The warrant not vesting is not that shock in one event; the correction would be gradual and conditional on subsequent management commentary. The analytic point is that the narrative premium is already embedded in the discount rate, not in an explicit FCF estimate, which makes it less visible to standard earnings-revision frameworks and more susceptible to an abrupt revision when the narrative catalyst does not arrive on schedule.

Inversion risk

NVIDIA's re-rating in 2023 — from approximately $150 per share in January to approximately $500 per share by December — provides the narrative structure AMD's current price references. The structural parallels are real: a hardware cycle with a confirmed hyperscaler anchor partnership, a platform narrative that transcends product-cycle frameworks, and a demand signal that the market concluded was structural rather than cyclical. The structural difference is material. When NVIDIA's re-rating began, the software moat — 4.7 million CUDA developers, more than 4,400 accelerated applications — had been compounding for a decade. That re-rating was pricing hardware acceleration of a software ecosystem that already existed. AMD's re-rating is pricing the simultaneous construction of the hardware ramp and the software ecosystem. The NVIDIA analog supports the scale of the addressable opportunity; it does not validate the assumption that AMD's ecosystem will compound at comparable velocity from a lower starting base.

The Intel inversion pattern provides the failure template. Intel's 10nm node delays from FY2019 did not produce immediate multiple compression commensurate with the competitive exposure. The sequence was: one quarter of disappointing data centre commentary, followed by analyst consensus reduction, followed by a modestly lower share price, followed by a second missed milestone that reclassified the first as a structural pattern rather than a scheduling anomaly. Multiple compression became non-linear because each missed milestone recalibrated the risk premium applied to subsequent guidance — the market stopped pricing management communications at face value, which compresses the implied WACC independently of the earnings estimate. AMD's FY2026 analogue is the MI450 ramp. A single quarter without confirmable MI450 revenue recognition initiates the same sequence: not a crash, but a recalibration of the probability weight assigned to subsequent guidance. Given that the current price is consistent with only the most favourable corner of the sensitivity space — 10% WACC, a rate that embeds near-zero execution risk — any upward recalibration of the risk premium moves the implied price materially.

---

VIII. Watch Conditions

The six conditions identified in Section III are the analytical instruments for tracking the investment case through each quarterly reporting cycle. Four are load-bearing — their trajectory determines whether the sequential mechanism the thesis depends on is functioning. Two are amplifying — they determine the margin and scale of the outcome. The accordion cards below contain the full signal framework for each.

Categorical condition — China regulatory resolution

The China regulatory environment does not lend itself to a trajectory-based watch framework: the condition is binary, and the mechanism is not an AMD execution variable. AMD's approach of excluding China-based AI GPU revenue from formal guidance until licence status is confirmed is the appropriate discipline, and it functions simultaneously as the monitoring framework: any guidance restatement that incorporates China AI GPU revenue signals a licence resolution, while any further write-down on China-facing inventory signals a tightening.

On track: AMD restates 2H2026 or FY2027 guidance to explicitly incorporate China AI GPU revenue, indicating that export licence conditions have stabilised and the 15% uncodified revenue-share expectation has not hardened into a codified gross margin charge. No further write-down charges on China-facing Instinct inventory are recorded.

Off track: AMD records an additional write-down on China-facing inventory in excess of the $440 million net charge recorded in FY2025, or management commentary confirms that the 15% revenue-share expectation has been formalised in a manner that AMD has incorporated into its cost structure. Either event would require a reassessment of the China revenue line in the base case financial model.

---

IX. Summary

AMD’s second-platform claim is structurally credible, but the observed share price already capitalises a demanding version of that outcome.

The case is governed by a four-condition sequential chain. Supply execution — TSMC advanced node yield and HBM allocation — gates the hardware ramp. The MI450 hardware ramp gates both ROCm software adoption and the OpenAI commercial conversion. Each condition depends on the one before it, which means a failure at the supply layer cascades through all four load-bearing conditions before any can be independently tested. EPYC's established CPU leadership is not in contention. The entire gap between the intrinsic value of management's guidance and the current share price belongs to Instinct.

FY2026 is the first resolution year: it tests the MI450 production ramp and supply execution, and generates the first evidence on whether the OpenAI infrastructure partnership is converting into confirmable volume. FY2027 tests whether ROCm software parity is reducing enterprise migration costs and whether warrant vesting signals that OpenAI purchase milestones have been crossed. FY2028 is the sustained-scale test — whether AMD's data centre earnings base has reached the level required to validate the trajectory the current price is pricing, rather than merely confirming the direction. FY2029 is the validation point for a durable second-platform position.

Even the complete bull case FCF trajectory implies approximately $251–$329 per share at the fundamental discount rate; the current price occupies the narrowest region of the sensitivity space and embeds near-zero execution risk. The MI450 ramp signal — Data Center segment revenue acceleration in 3Q2026 and 4Q2026, gross margin at or above 52.8%, no inventory write-down charges — is the earliest and most consequential data point. The ROCm parity and OpenAI warrant conditions are not observable until FY2027 and FY2028 reporting. The China regulatory condition adds a binary tail risk that operates independently of the execution chain.

The structural logic of the second-platform claim is sound. AMD is the only credible multi-source alternative in a market the company targets at over $400 billion in addressable value by 2027. EPYC demonstrates that winning the open alternative argument in a software-defined market is achievable over a sufficiently long execution horizon. The binding question over FY2026 to FY2028 is whether the AI accelerator market compresses that horizon to a timeframe consistent with the enterprise value the market has already assigned.