AMD has become the second source for AI compute. My conviction is that it grows into today’s price. My model says the price already assumes the plan works.
Two numbers, read together. The headline is my conviction call, valued at $624 a share on five stated beliefs. My probability-weighted model value is $313, and the rule I apply to every report reads Bearish at $528.75. Both are shown throughout, so you can judge the bet for yourself.
All “up to” commitments. The deals and the warrants
This is a conviction call. The headline, Bullish on a 3-5 year view, is my judgment, not the output of my model. It rests on a conviction case worth $624 a share, which needs five things to go right, listed in the conviction case. My base case is worth $311, well below today’s $528.75, and even my bull case ($486) is below it. The probability-weighted value is $313, so my rating rule reads Bearish. For a reader, that makes this a higher-risk position with an asymmetric outcome: about 18% upside to the conviction case, against about 40% downside if the base case is right.
The Street is enthusiastic: 47 Buy and 10 Hold ratings with no Sells, and a median price target of $635.50 (55 analysts). My conviction case, $624, sits close to that. My model, $313, sits far below it, for three deliberate reasons.
AMD’s shares move about 2.3 times as much as the market (measured two-year beta 2.31). Even after shading that toward 1.0, which is standard for a ten-year forecast, the cost of equity is 14.6%. Many Street models use 9% to 11%. That one input roughly doubles or halves the answer.
Consensus earnings per share use today’s share count. The OpenAI and Meta warrants add up to 320 million shares exactly when the gigawatt revenue arrives, so a model that counts the revenue but not the shares overstates value per share by up to 19%.
AMD paid $1.6 billion in stock-based compensation in 2025, 4.7% of revenue, and excludes it from its non-GAAP earnings. I treat it as a real expense, because it is paid in dilution.
AMD designs chips and outsources their manufacture, mostly to TSMC. It reports four segments. Data Center holds the EPYC server processors and the Instinct AI accelerators, now sold as complete Helios racks with 72 GPUs, EPYC CPUs and Pensando networking. Client is Ryzen PC processors. Gaming is mainly the custom chips inside Sony and Microsoft consoles, plus Radeon graphics cards. Embedded is largely the former Xilinx FPGA business, bought in 2022.
AMD does not split Data Center revenue between CPUs and GPUs. My estimate for 2026 is about $17 billion of EPYC and $14 billion of Instinct, with the GPU share rising sharply once MI450 ships at the end of Q3.
Reported results, USD billions
| Line item | FY2023 | FY2024 | FY2025 | Q2 2025 | Q1 2026 | Q2 2026 |
|---|---|---|---|---|---|---|
| Data Center | 6.50 | 12.58 | 16.64 | 3.24 | 5.78 | 6.72 |
| Client | 4.65 | 7.05 | 10.64 | 2.50 | 2.89 | 3.06 |
| Gaming | 6.21 | 2.60 | 3.91 | 1.12 | 0.72 | 0.78 |
| Embedded | 5.32 | 3.56 | 3.45 | 0.82 | 0.87 | 0.98 |
| Total revenue | 22.68 | 25.79 | 34.64 | 7.69 | 10.25 | 11.54 |
| Non-GAAP gross margin | 50% | 53% | 52% | 43% | 55% | 56% |
| Non-GAAP operating income | 4.85 | 6.14 | 7.77 | 0.90 | 2.54 | 3.09 |
| GAAP operating income | 0.40 | 1.90 | 3.69 | (0.13) | 1.48 | 1.99 |
| Free cash flow | 1.12 | 2.41 | 5.52 | n/a | 2.57 | 1.56 |
| Non-GAAP diluted EPS ($) | 2.65 | 3.31 | 4.17 | 0.48 | 1.37 | 1.66 |
Sources: AMD Q4 2023 release (Jan 30, 2024), Q4 and FY2025 release (Feb 3, 2026), FY2025 10-K, Q1 2026 release (May 5, 2026), Q2 2026 release, slides and 10-Q (Aug 4, 2026). Q2 2025 includes an $800 million charge for MI308 inventory blocked by US export controls; Q2 2025 client and gaming are derived from the annual totals. Free cash flow is AMD’s definition (operating cash flow less capital spending), before stock compensation.
In twelve months AMD signed the three largest customer commitments in its history. Each is measured in gigawatts of data center power, which is how AI labs now plan capacity. Two of them came with warrants that let the customer buy AMD shares for a cent each as the gigawatts ship.
| Customer | Commitment | Timing | Terms |
|---|---|---|---|
| OpenAI (Oct 2025) | Up to 6 GW, multi-generation | First 1 GW of MI450 from 2H 2026 | Warrant for up to 160M shares at $0.01; binding commitment only for the first 1 GW. AMD’s CFO expects “tens of billions of dollars” of revenue. |
| Meta (Feb 2026) | Up to 6 GW | First 1 GW, a custom MI450, from 2H 2026 | Warrant for up to 160M shares at $0.01. Also a lead customer for 6th-generation EPYC CPUs. |
| Anthropic (Jul 2026) | Up to 2 GW in Helios racks | First 1 GW from 1H 2027 | No warrant. AMD commits up to $5B of equity investment in Anthropic, released on milestones. |
| Oracle, Microsoft, others | Oracle about 50,000 MI450 GPUs; Azure Helios at scale | From Q3 2026 | No gigawatt figures disclosed. |
Value per share with and without the warrants (each scenario uses its own share count)
| Scenario | Warrant shares that vest | Without warrants | With warrants | Cost per share |
|---|
Sources: OpenAI warrant, 8-K and Exhibit 4.1 (Oct 6, 2025); AMD and Meta press release (Feb 24, 2026); Q2 2026 10-Q, Note 12 and MD&A; Q2 2026 release (Anthropic); CRN Asia and Bloomberg Government on the Anthropic terms (Jul 2026). The warrant schedule and exercise conditions are redacted; vesting in proportion to gigawatts is my assumption.
MI450 and Helios begin shipping at the end of Q3 2026 and ramp through Q4 and 2027. My base case has AMD shipping 2.3 GW in 2027 and 5.6 GW in 2030. Each quarter of delay moves revenue out and, because the warrants expire, also changes how many shares vest.
Lisa Su says each gigawatt brings “double digit billions” of revenue. I use $15 billion through 2029, easing to $13 billion by 2033 as competition matures. This is the single most important number in the model: the reverse DCF says today’s price needs about $30 billion.
AI agents run their workflows on CPUs, which lifts server CPU demand alongside GPUs. AMD raised its 2030 server CPU market estimate from about $60 billion to over $120 billion in May. My base case takes EPYC from about $17 billion in 2026 to $52 billion in 2030.
Non-GAAP gross margin is 56% today, and AMD targets 55% to 58%. MI450 racks earn below the average. Against that, up to 320 million warrant shares and about $2 billion a year of stock pay spread the profit across more shares.
The model builds revenue by segment. Data Center AI is gigawatts shipped times revenue per gigawatt; EPYC, Client, Gaming and Embedded are forecast in dollars. Each has its own gross margin. Operating costs and stock pay are in dollars, cash taxes are 14% to 15%, and capital spending and working capital grow with revenue. I treat stock pay as a cost and leave out the amortization of acquired intangibles, which is not cash.
Base case, USD billions. 2026E is my estimate, consistent with guidance; the valuation starts from H2 2026 free cash flow of $3.0 billion after stock pay.
For comparison, consensus expects 2027 revenue of $89 billion and non-GAAP EPS of $15.76. My base case has $84 billion; on AMD’s own non-GAAP basis and today’s share count it earns about $15.
AMD designs chips and outsources manufacturing, so its own capital spending is small: 2.8% of revenue in 2025. But to win the AI deals it has started to share the financing burden with customers. These commitments sit mostly outside the balance sheet, and they are the part of the story I watch most closely.
| Commitment, mid-2026 | Amount | What it is |
|---|---|---|
| Purchase commitments | $30.3B | Wafers, substrates and components, $17.4B of it in the rest of 2026 (up from $12.2B at year end) |
| Data center lease guarantees | $4.1B | Guarantees of partners’ data center leases, up to 15 years, “may be issued in exchange for warrants” |
| Investment commitments | Up to $5.0B | Signed after June, through 2028; consistent with the Anthropic equity commitment |
| AMD’s own data center leases | $9.5B | Signed after June, up to 16 years, starting 2027 to 2028 |
| Capital spending | $1.2B in H1 2026 | 5.5% of revenue, double the 2025 rate |
My reading: AMD is buying share partly with its balance sheet. The amounts are small against $13.1 billion of cash and $5.5 billion of 2025 free cash flow. But they tie AMD’s fortunes to the financing health of AI labs and cloud builders, and the warrants transfer equity to two of them. My model charges capital spending at 3% to 4% of revenue and working capital at 15% of each year’s revenue growth; it does not assume the guarantees are called.
Source: AMD Q2 2026 10-Q, Liquidity and Capital Resources, and Note 10 (Lease Guarantees); FY2025 10-K. The counterparties of the guarantees are not disclosed.
AMD is profitable, so an earnings multiple would work for today. It would not capture what decides the stock: how many gigawatts ship, at what price, and how many warrant shares that triggers. A cash-flow model built from those drivers can, and it charges the working capital, capital spending and stock pay that fast growth needs.
Gigawatts × revenue per gigawatt for AI; dollars for EPYC, Client, Gaming and Embedded; a gross margin for each.
Why: it ties value to the numbers each quarter updates.
Operating profit after stock pay and cash tax, plus depreciation, less capital spending and working capital.
Why: ten years reaches a mature growth rate.
Cost of equity 14.6%: beta 2.31 measured, adjusted to 1.87. WACC 14.5%.
Why: the risk charge comes from how the stock trades.
16 times 2035 EBITDA at the end of 2035, cross-checked against 3% perpetual growth.
Why: 16x is where mature chip companies trade today.
Each scenario’s share count adds the warrant shares its gigawatts vest.
Why: dilution arrives with the revenue that triggers it.
A 3-5 year rule, fixed in advance, against the cost of equity. The headline is my conviction, shown next to the rule.
Why: the reader sees both the bet and the discipline.
| Method | How I used it | Reason |
|---|---|---|
| Driver-based DCF, exit multiple, three scenarios | Primary | Values the gigawatt ramp, the price per gigawatt and the warrant dilution together. |
| DCF with perpetual growth | Cross-check | Gives $206 for the base case. The 16x exit implies 9.2% perpetual growth at a 14.5% discount rate, which is generous. |
| Reverse DCF | Cross-check | Today’s price needs about twice my base-case revenue per gigawatt. |
| P/E and EV/sales against peers | Context | AMD trades at 70x 2026E consensus EPS and 34x 2027E. Peer multiples range widely; they show what the market pays, not what AMD is worth. |
| Sum of the parts | Not used | AMD does not split Data Center into CPUs and GPUs, so the parts cannot be valued on reported cash flow. |
I discount unlevered free cash flow for the second half of 2026 and for 2027–2035 at a 14.5% WACC, using mid-year discounting, and capitalize 2035 EBITDA at 16 times as a value at the end of 2035. The cost of equity is 14.6%: a 5.2% risk-free rate plus a beta of 1.87 times a 5.0% equity risk premium. The beta starts from 2.31, measured from two years of daily returns against SPY, and is adjusted two-thirds toward the market’s 1.0 (the Blume method), because betas drift toward 1 over a ten-year forecast. On the unadjusted beta the cost of equity would be 16.75% and the model value about $270. Debt of about $7.1 billion, including the August 2026 notes, is 1% of capital. Net cash of $9.9 billion and today’s diluted share count, plus about 15 million shares for the World Labs acquisition, complete the bridge.
The solid line is the model value ($313.29). The dashed line is the price. The outlined bar is the conviction case, valued at discount rates from 11% to 13%.
Shaded cells are below today’s price. The outlined cell is the base case.
| Company | EV / sales | EV / EBITDA | Fwd P/E |
|---|---|---|---|
| AMD | 21.3x | 92.1x | 50.6x |
| NVIDIA | 17.2x | 25.9x | 17.7x |
| Broadcom | 18.4x | 31.4x | 19.0x |
| TSMC | 11.8x | 16.2x | 18.8x |
| Marvell | 21.7x | 71.8x | 40.5x |
| Arm | 48.1x | 233x | 102.6x |
| Intel | 8.8x | 29.8x | 55.7x |
| Qualcomm | 4.4x | 16.2x | 18.7x |
Source: StockAnalysis statistics pages, Oct 16, 2026 (trailing EV/sales and EV/EBITDA; forward P/E on each company’s fiscal year). Context, not inputs.
Holding every other base assumption fixed, $528.75 needs revenue per gigawatt about 2.0 times my base path: roughly $30 billion per gigawatt instead of $15 billion. Or it needs about 2.1 times the gigawatts, close to 17 GW a year by 2035. Changing the discount rate alone would need a beta near zero. The price makes sense only with both a lower discount rate and a higher exit multiple, for example 10% and about 25 times 2035 EBITDA. The market is paying for the plan to work and for AMD to be valued like a much less risky company.
My headline call rests on this case, not on the probability-weighted value. It uses the same model and the same bull-case operating path, with two changes a bull would make: a 12% discount rate (a beta of about 1.36, as AMD’s risk falls toward that of other $1 trillion companies) and 18 times 2035 EBITDA (between Qualcomm and TSMC at 16 times and Broadcom at 31 times today). It is worth $624.05 a share, 18% above today’s price. It needs all five of these to go right:
| Conviction belief | My base case | Today’s evidence |
|---|
What this means for a reader. The conviction case is not the most likely outcome in my model; my base case is. Holding AMD at $528.75 is a bet that the bull path happens and that the market keeps paying a premium for it. If both go right, the upside is moderate. If the base case is right, the downside is large.
Each slider applies to all three scenarios, and warrant dilution moves with the gigawatt slider. The model value and the rule update instantly; the published call does not change.
AMD has had the same chief executive since 2014, and the turnaround she led is the reason the AI customers take its roadmap seriously. Her pay is now tied heavily to the share price.
Chair, President and Chief Executive Officer, since 2014
What it means: her rewards depend on the share price, as an owner’s would. Planned selling is routine at this size.
Chief Financial Officer and Treasurer, since 2022
What it means: her stated margin range is the anchor for my gross margin assumptions.
EVP and General Manager, Data Center Solutions
What it means: the Helios ramp in Q4 2026 is his execution test.
EVP and Chief Technology Officer
What it means: an annual product cadence against NVIDIA depends on his team hitting every date.
| Holder | Shares | Stake | Note |
|---|---|---|---|
| Vanguard | 142.4M | 8.8% | Index and active funds; figure from early-2024 filings, as reported in the 2026 proxy |
| BlackRock | 124.9M | 7.7% | Same basis as Vanguard |
| OpenAI and Meta (potential) | Up to 320M | Up to about 16% after issue | If both warrants vest in full; none had vested by June 2026 |
| All directors and officers | 8.1M | Under 1% | 16 people, Mar 19, 2026 |
Sources: AMD 2026 proxy statement (DEF 14A, Mar 27, 2026): ownership and the 2025 Summary Compensation Table; 8-K of Feb 17, 2026 (CEO award) and Jul 1, 2026 (2026 grants); Q2 2026 10-Q, Item 5 (trading plans); MarketBeat insider trades (Form 4 summary, to Sep 15, 2026).
| When | Event | What I am watching |
|---|---|---|
| Nov 3, 2026 | Q3 2026 results | Revenue against the $13.0 billion guide; first Helios revenue; the Q4 guide against about $16 billion implied by consensus. |
| Q4 2026 | Helios ramp; first OpenAI and Meta gigawatts | First warrant tranches vest. Any disclosure of how the warrants hit reported revenue. |
| By end of 2026 | World Labs acquisition closes | About $8.2 billion in shares; Fei-Fei Li joins as chief scientist. |
| Early Feb 2027 | Q4 results and 2027 outlook | Confirmation that Data Center more than doubles in 2027 and server revenue grows over 70%. |
| 1H 2027 | Anthropic’s first gigawatt | Tests whether demand broadens beyond the two warrant holders. |
| 2027 | MI500 launch; Venice volume | Keeping an annual cadence against NVIDIA’s Rubin generation. |
Here are my estimates for the third quarter, so the call can be checked against the result.
Download the Excel model. It contains live formulas for the inputs, three scenario blocks, the conviction case, the valuation summary with both calls, a sensitivity grid and reported history. Every hardcoded input carries a source note, and it reproduces this page’s numbers exactly.
This is personal research by Phat Ngo, prepared as a work sample. It is not investment advice or a recommendation to buy or sell any security. The headline rating is the author’s stated conviction; the probability-weighted model value and the rule it produces are shown alongside it and read Bearish at the publication price. The author’s personal portfolio holds AMD (9% target weight) and NVIDIA (10%), from the October 16, 2026 close (10% and 12% before). Forecasts, scenario probabilities, the conviction case and the rating are the author’s judgments and may be wrong. Figures are as reported in public filings and news coverage on the dates cited.