AMD

IThe business

AMD designs processors and AI accelerators and owns no factories. TSMC makes every AMD wafer at 7 nm or smaller, and packaging and test run through partners such as Tongfu, SPIL and KYEC.⁶ It reports three segments. Data Center, 58% of revenue in the June 2026 quarter, holds EPYC server CPUs, Instinct GPUs, networking chips and FPGAs. Client and Gaming holds Ryzen PC chips, Radeon graphics and the chips in game consoles. Embedded holds the adaptive chips that came with Xilinx.¹⁷⁶

In the twelve months to June 2026, gross margin was 53.2%, research and development was 22.7% of revenue and capital spending 4.1%. Operating margin was 15.7% on the GAAP basis and 26.0% on AMD’s own non-GAAP basis C (calculated).^6417192021

Revenue grew from $4.3B in 2016 to $34.6B in 2025, 26.0% a year.^{1413121110986} The annual figures tie to the SEC’s XBRL facts.¹⁵ AMD reported net losses in 2016 and 2017.¹⁴⁶ Revenue fell in 2023, from $23.6B to $22.7B, and the Embedded segment fell 33% in 2024.⁹⁸

Data Center revenue more than doubled year-over-year

What it sells

• Data Center: EPYC CPUs, Instinct GPUs, networking chips and FPGAs for servers. Revenue was $6.72B in Q2 2026, up 107%, with a 31.3% segment operating margin. In the TTM it was $22.2B, up from $3.7B in 2021 C (calculated).¹⁷⁶¹⁰
• Client and Gaming: Ryzen chips for PCs, Radeon graphics and semi-custom chips for consoles. Q2 revenue was $3.84B: Client $3.06B, up 23%, and Gaming $779M, down 31%. AMD guides a significant double-digit decline in Gaming for Q3.¹⁷⁴¹⁴²
• Embedded: adaptive and embedded processors built on Xilinx. Q2 revenue was $977M, up 19%, which management calls its fastest growth in more than three years. It was $5.3B in 2023 and $3.5B in 2025.¹⁷⁴¹⁹⁶

Who pays

No customer was 10% of revenue in 2025 or 2024, and one Client and Gaming customer was 18% in 2023.⁶ On the 4 August call management named OpenAI, Meta and Anthropic as the anchor customers for MI450. I found no official transcript; I read two unofficial ones that agree, and cite them as the call.⁴¹⁴²⁴³

At $13B of revenue a gigawatt, my estimate, one gigawatt equals 15% of my year-2 base revenue E (estimate). I use gigawatts as a cross-check on the base case and not as a driver: revenue in the model grows by segment, and the gigawatts test whether the growth is plausible. Management says only “double-digit billions” a gigawatt.⁴¹⁴² The NVIDIA page uses $25B to $40B a gigawatt for its own platform. I cannot say how much of that gap is price and how much is what each company counts, which is why the cross-check in Valuation runs from $10B to $20B around management’s “double-digit billions”.

By billing location in 2025, the United States was $11.4B, China including Hong Kong $7.8B (22%), Taiwan $5.2B, Singapore $4.3B and other countries $6.1B, which sum to $34.6B before rounding C (calculated).⁶ Export controls reach into that number: AMD took $800M of charges on its MI308 chip in Q2 2025, released $360M in Q4 and shipped about $390M of MI308 to China in that quarter.⁶²⁰

The commitments below are evidence for part of the growth in my base case, not most of it: OpenAI and Meta cover at most 29% of the Data Center increase and Anthropic takes the total to 39% (table in Valuation) E (estimate).^2224354 They also differ in how binding they are.

For OpenAI and Meta, only the first gigawatt is binding. Anthropic’s S-1 commits it to a first block in calendar 2027, on conditions.⁶¹ Everything after that is “intends to deploy” or “up to”.²²²⁴⁴ The first gigawatt of each deployment is powered by MI450.⁴¹⁴

The cycle

Chip revenue is cyclical: AMD’s fell 4% in 2023, from $23.6B to $22.7B, and the 10-K says the industry has had severe downturns.⁶⁹ The stock fell 65%, from a close of $161.91 on 29 November 2021 to $55.94 on 14 October 2022, and has since risen 9.0 times C (calculated).¹ The newer cycle risk is the buyers’ own funding. The 10-K warns that customers may lack the capital for their AI infrastructure and may ask vendors for alternative financing or deferred payment.⁶

IIThe moat

I think AMD has a narrow moat, and it is two different things. In server CPUs it is a design and execution lead on a standard that Intel also builds to; it is widening in share and is not locked in. In AI accelerators it is a second-source position: the largest buyers want an alternative to NVIDIA, and AMD has issued them warrants for up to 320M shares to be it. That position is widening in volume and narrowing in price.

Server CPUs: a lead that has to be re-earned

Mercury Research, as reported by Tom’s Hardware (I did not read Mercury’s report), puts AMD’s x86 server unit share at 34.5% in Q2 2026, from 27.3% a year earlier. Its share of server revenue is 46.2%, for Q1 2026; I found no Q2 revenue figure.⁴⁴⁴⁵⁴⁶ Against Intel’s Xeon alone, EPYC held 46.4% of units in Q2.⁴⁴ A customer can swap an EPYC for a Xeon without rewriting software, because both run x86, so AMD can lose server share as easily as it won it. AMD’s lead is a design cadence: the sixth generation, “Venice”, is in production on a 2 nm process, and management says demand is stronger than for any prior EPYC generation.¹⁷⁴¹⁴²

Accelerators: a second source, paid for in shares

OpenAI and Meta each signed a binding order for the first gigawatt of MI450 and say they intend to deploy up to six. Anthropic plans up to two, and its S-1 commits it to a first block in calendar 2027, on conditions.⁶¹ OpenAI (October 2025) and Meta (February 2026) each hold a warrant for 160M AMD shares at one cent.^2224354

This is an important step for Meta as we diversify our compute.

Set the warrant beside the order. For all six gigawatts a customer can earn 160M shares, worth $80.8B at today’s price and $96.0B at $600. At $13B of revenue a gigawatt, six gigawatts is $78B of revenue E (estimate). The shares are worth 104% of that revenue at today’s price and 67% if a gigawatt brings $20B C (calculated).^2224411 The vesting schedule is redacted and the shares vest only as the share price clears thresholds that escalate to $600, so the real cost depends on the stock. If they vest evenly, each gigawatt hands over shares worth $13.5B against $13B of revenue.²²²⁴⁴

I count vested warrant shares as a cost to shareholders, whichever way AMD books them. My base counts 160M of the 320M, which lowers its value from $213 with none to $194; the bull counts all 320M and the bear 53M.⁴ Vesting depends on the market price clearing the thresholds, and only the top one, $600, is public; it does not depend on my value. The reverse DCF below counts no warrant shares, so it flatters the price. AMD’s chief financial officer said the OpenAI deal was expected to be highly accretive to non-GAAP earnings per share.²³ I cannot check that. The 10-Q does not say how a vested tranche will be booked, and as a matter of general accounting practice a payment to a customer usually reduces revenue, so reported revenue and margin may come out below my non-GAAP cases.⁴

AMD has issued warrants for up to 320M shares, 19.6% of the 1,632M outstanding, to win its place as the second source.

Evidence it is widening

Data Center against Intel. AMD’s Data Center segment has been larger than Intel’s Data Center and AI group in each quarter since Q3 2025, after trailing it in each quarter from Q3 2024 to Q2 2025 C (calculated).^{1719202147484950}

Product and software. Helios puts MI450 GPUs, “Venice” CPUs, Pensando networking and ROCm in one Open Compute rack, and the first shipments are guided for late Q3 2026.¹⁷⁴¹ AMD has also announced Helion, a rack for agentic AI work, with configurations planned for November 2026.³⁸ Anthropic will use its Claude model to optimize ROCm.³⁵ Management says open-source contributions to ROCm rose more than tenfold in a year.⁴¹⁴²⁴³

Evidence it is narrowing

Custom silicon. Broadcom’s AI semiconductor revenue grew 221% to $16.7B in the quarter to 2 August, 2.5 times AMD’s whole Data Center segment C (calculated).⁵²¹⁷ Broadcom’s line includes AI networking and its release does not split it out, so 2.5 times is an upper bound for custom accelerators alone. It shows what a buyer can do instead of buying from AMD or NVIDIA.

Price. Every gigawatt carries shares. A second source that has to be paid in equity has a weaker moat than one that is simply chosen.

Supply. AMD depends on one foundry for every advanced wafer, and its unconditional purchase commitments rose from $12.2B to $30.3B in six months, all non-cancellable.⁶⁵⁴

The moat is widest where AMD’s rival is Intel and thinnest where its rival is NVIDIA.

IIIThe competitive advantage

AMD’s advantage over Intel is growth and a higher whole-company margin, 17.3% against 11.1% on a GAAP basis in the June quarter; in data centers Intel’s segment margin is higher, 39.5% against 31.3%. Over NVIDIA and Broadcom AMD has no advantage in scale or margin; what it has is a place as the second source C (calculated).^17475152

Against Intel

AMD’s Data Center segment was $6.72B in Q2 2026 against $6.26B for Intel’s DCAI, up 107% against 59%. Intel’s segment earns the higher margin, 39.5% against 31.3%.¹⁷⁴⁷ Both are growing fast, so AMD’s gains come from a rising market as well as from Intel’s share. Mercury says Intel’s CPU supply rose in Q2 2026, especially in mobile client, and Tom’s Hardware reports that Mercury expects server CPU supply constraints could limit gains in Q3, with Q4 stronger; it does not say whose supply.⁴⁴ AMD’s 10-K says Intel uses its market position to price against it.⁶

In PCs AMD’s unit share is at records: 30.3% of client, 34.9% of desktop and 28.9% of mobile in Q2 2026.⁴⁴ Client revenue was up 23% in the quarter, but management plans for a softer PC market in the second half on memory costs.¹⁷⁴¹ In 2025 Client unit shipments rose 15% and the average selling price 31%, as the corrected 10-K/A states.⁷

Against NVIDIA and custom chips

NVIDIA’s Data Center revenue was $89.0B in the quarter to 26 July 2026, up 117%. AMD’s segment, which also holds CPUs, was 7.5% of that C (calculated).⁵¹¹⁷ AMD does not report Instinct revenue, so there is no like-for-like share, and I use none.

NVIDIA earned a 75.0% gross margin and a 66.2% GAAP operating margin in the quarter. AMD earned 54% and 17.3%, or 26.8% on its own non-GAAP basis C (calculated).⁵¹¹⁷ NVIDIA spent $7.05B on research and development in the quarter, AMD $2.53B, or 36% as much C (calculated).⁵¹¹⁷ Management says Data Center AI gross margin is slightly below the company average, and two customers hold warrants for up to 320M of AMD’s shares.⁴¹⁴

AMD’s 10-K says NVIDIA leans on its position in data-center GPUs and its software, and flags NVIDIA’s September 2025 investment in Intel as a competitive risk.⁶ It also warns that customers develop their own chips for AI work.⁶ Broadcom guided $21.7B of AI revenue for its next quarter.⁵²

Where AMD wins

Server CPUs. The share data above. Management expects server CPU revenue to grow more than 80% in the second half of 2026 and more than 70% in 2027, and targets more than 50% of server CPU revenue.⁴¹⁴²⁴⁰

Openness. Helios follows the Open Compute Project design, and buyers who want a second source get a rack built to a shared standard.¹⁷ Management claims up to 15% more throughput at the same rack power and up to 30% more tokens per dollar than “the competition”; I found no independent test.⁴¹

Where it loses

Software. ROCm is catching up to a standard it did not set. I treat it as the weak flank.

Scale. AMD’s research and development is 36% of NVIDIA’s C (calculated). The 10-K flags that competitors may have stronger market positions.⁶⁵¹¹⁷

Gaming. The console cycle has turned. Gaming revenue fell 31% in Q2 and AMD guides a significant double-digit decline in Q3.¹⁷⁴¹

AMD gains share against Intel at a 31.3% segment margin and pays for its place against NVIDIA with warrants for up to 320M shares.

IVManagement

Lisa Su joined AMD in January 2012, became chief executive in October 2014 and chair in February 2022. She is 56.¹⁶ The stock closed at $7.005 on 24 October 2016 and at $504.80 in Q4 2026, 72.1 times as much C (calculated).¹ That window starts two years after she became chief executive. Revenue went from $4.3B to $34.6B in nine years.¹⁴⁶ The proxy’s comparison is less kind: $100 invested at the close of 24 December 2020 was worth $234 at the end of 2025, against $547 for the S&P 500 Semiconductor & Semiconductor Equipment index.¹⁶

Pay and incentives

Su’s 2025 pay was $55.2M, 341 times the median employee’s $161,780, of which $41.7M was stock and $9.0M options. Pay “actually paid” under SEC rules, which marks awards to market, was $146.7M.¹⁶ The pay plan tracks relative return: the 2022 performance units paid out at 96.84% of target because AMD’s three-year return trailed the S&P 500 by 1.58 points. Bonuses use revenue, adjusted net income and adjusted free cash flow, with 30% on strategic milestones.¹⁶

In February 2026 the board approved a special award with a $75M target value. It can pay from 0% to 200% of the target in four tranches, and a tranche is earned when the 90-trading-day average closing price reaches its hurdle at any time to March 2031: 10%, 15% and 17.5% a year over five years from a baseline price, and $600 for the top tranche. The baseline is the average close over the 60 trading days to the 15 March 2026 grant date. The board set the hurdles with the all-time high of $267.08 at the grant date in view.²⁸

100% performance-based, tied to four stock price hurdles

By my arithmetic on those rules the baseline is about $215, which puts the first three hurdles near $345, $435 and $480. The 90-day average close passed them between 30 June and 1 September and was $530.64 in Q4 2026, so only the $600 tranche is left C (calculated).²⁸¹ I did not find the company’s own baseline or hurdles, so these are my estimates. With a conversion price near $207, the 30-day average to the grant date, the target is about 363,000 units, and the 544,000 units of the first three tranches are worth about $275M at today’s price C (calculated).²⁸¹ Earned units vest in March 2029 and March 2031 if she is still chief executive or executive chair.²⁸

The top hurdle, $600, is also the top threshold of both customer warrants.²⁸²²²⁴ A rising share price pays Su and the two customers at the same time. In August she received her annual award, a $36M target, 75% in units that pay on three-year relative return, and her base salary rose to $1,375,000 from 1 July.²⁹ Say-on-pay support was about 85% in 2025 and 92.6% of votes cast in 2026.¹⁶³⁰

Su owned 4,305,973 shares at 19 March 2026, including options and RSUs vesting within 60 days. That is 0.26% of the shares outstanding and about $2.2B at today’s price before option exercise costs C (calculated).¹⁶ On 8 June 2026 she adopted a Rule 10b5-1 trading plan for up to 450,000 shares, 10.5% of that holding, to run to 16 June 2027, about $227M at today’s price C (calculated).⁴

Put the incentives together. Su is paid for a higher share price over five years, and the top hurdle of her special award, $600, is the same line as the top of both customer warrants²⁸²²²⁴ That pays her to push the stock and the customer ramp up together. Her stake is 0.26%, but the proxy count includes options and units, so I cannot tell how much of it she bought with her own money¹⁶ Over 2021 to 2025 the capital went to Xilinx, which earns 2.5% to 2.8% a year on the price paid before amortization, to buybacks that did little more than offset stock pay, and since October 2025 to shares and guarantees for customers C (calculated).¹⁰⁶⁴

The bench and the board

Jean Hu has been chief financial officer since January 2023 and was Marvell’s before that. Mark Papermaster has been chief technology officer since October 2011. Forrest Norrod, who ran Dell’s server business, has led Data Center since January 2023. Darren Grasby became chief sales officer in January 2025. Emily Ellis became chief accounting officer on 15 December 2025; Hu had been acting.¹⁶³³ Fei-Fei Li is to become chief scientist when the World Labs purchase closes.³⁶

The board has eight directors, all independent except Su, who is both chair and chief executive. Nora Denzel has been lead independent director since November 2022 and a director since 2014. Joseph Householder, the audit chair, retired on 19 August 2026 after more than 11 years; KC McClure, formerly Accenture’s chief financial officer, joined in January 2026 and now chairs audit, and Tim Ryan, formerly the US chair of PwC, joined on 19 August.¹⁶³¹³²

Capital allocation

Capital allocation has three items that matter: Xilinx, bought for $48.8B in stock; ZT Systems, recorded at $4.4B and mostly sold on; and, since October 2025, financing for customers.¹⁰⁶⁴

Xilinx cost $48.8B in stock, 429M shares at $113.18. Those shares are worth $217B at today’s price C (calculated).¹ Embedded earned $1.36B of operating income in the TTM against $2.63B in 2023, and AMD books $2.2B a year of acquisition amortization on $25.5B of goodwill and $15.6B of intangibles, mostly from this deal.¹⁰¹⁷⁹⁶ Embedded’s operating income, before that amortization, is 2.5% to 2.8% a year on the $48.8B paid ($1.24B in 2025, $1.36B in the TTM) C (calculated). Xilinx technology also sits inside Data Center and Helios, which the segment line does not show.

Buybacks were $8.6B for about 81M shares from 2021 to 2025, about $107 a share. The net share count rose 425M, from 1,207M to 1,632M, and Xilinx alone issued 429M and ZT Systems 9M. Without the two deals the count fell about 13M, so buybacks covered what stock pay issued and a little more C (calculated).⁶¹⁰⁴ In the first half of 2026 AMD bought 1.1M shares for $221M, about $201 a share, close to my base value of $194 C (calculated); $9.2B of the authorisation remains.⁴ AMD pays no dividend.

On ZT Systems AMD recorded $4.4B in March 2025: $3.2B in cash, $0.86B in stock and $0.36B of contingent payments. When it sold the manufacturing arm to Sanmina in October 2025 it received $1.4B in cash net of cash divested and adjustments, 1.15M Sanmina shares worth $154M and up to $450M more by 2028, and kept the rack design and customer-enablement teams.⁶³⁴ Silo AI cost $665M, and AMD bought Taalas and MEXT for undisclosed prices in 2026.⁶³⁷⁴

AMD has begun financing demand. It issued warrants for 320M shares in October 2025 and February 2026, guaranteed up to $4.1B of a partner’s data-center leases in 2026, agreed to acquire World Labs for about $8.2B in stock, and committed to invest up to $5B in Anthropic. It has $4.5B of data-center and other real estate leases not yet started at 27 June, up from $1.3B in December, and signed $9.5B more after quarter end. It also sold $4.75B of senior notes on 17 August.^462224353626

Since October 2025 AMD has committed to give its biggest customers up to 320M shares ($162B at today’s price if all vest), guarantees of up to $4.1B, up to $5B of equity and $1B for Meta’s equity offering; cash on hand is $13.1B.

VFinancials

Free cash flow of $7.7B was 120% of net income in the TTM, and $2.2B of first-half cash came from payables.⁴¹⁷ From 2025 to the TTM, revenue rose 19%, GAAP operating income 76% and free cash flow 40% C (calculated).⁶⁴¹⁷

GAAP gross margin went from 49.5% in 2025 to 53.2% in the TTM. GAAP operating margin went from 10.7% to 15.7%, and was 17.3% in the June quarter C (calculated).⁶⁴¹⁷ The difference between the GAAP and non-GAAP operating margins is acquisition amortization and stock pay, together 9.9 points of the 10.3-point gap C (calculated).⁶⁴¹⁷ On AMD’s non-GAAP basis the TTM gross margin was 55.7%, operating spending 29.8% and operating margin 26.0% of revenue; stock pay was 4.6% and free cash flow after it 14.1% C (calculated).^17192021

Earnings quality

Three items flatter or depress the year-on-year comparison. A tax release of $853M raised 2025 net income and drops out of the TTM. An $800M inventory charge on MI308 cut 2025 gross profit, about $440M net after the Q4 release. And $549M of gains on long-term investments sit inside the first half of 2026 other income.⁶¹⁷²⁰

Operating cash flow of $5.3B in the first half of 2026 includes $2.2B from payables, which the 10-Q attributes to the timing of payments. Working capital was a net use of $0.27B in the half: the rise in payables paid for rises in receivables ($0.97B), inventory ($0.55B) and prepaid expenses and other assets ($0.99B), the last mostly prepayments under supply agreements.⁴ Suppliers are not carrying the ramp on net, but payables are the line that reverses if revenue stalls.

The balance sheet

Cash and short-term investments were $13.1B at 27 June 2026 against $3.25B of senior notes, with a $5.0B revolver and a $5.5B commercial-paper programme undrawn.⁴ AMD sold $4.75B of notes on 17 August at coupons of 4.600% to 5.500% and has $875M of 4.212% notes due on 24 September; I could not verify the repayment.²⁶⁴ The new notes add cash and debt together and raise interest by about $0.24B a year C (calculated).²⁶ The 10-Q does not say what AMD’s own data-center leases are for. It does say AMD expects to use its multi-year cloud compute capacity in its operations or assign it to third parties.⁴

VIValuation

Values below are as of 27 June 2026, the last balance sheet, as on the NVIDIA page. Dated Q4 2026 they would be about 3% higher, and discounting cash flows at mid-year instead of at year end would add about 5% more: together my base would be $209, my bull $469 and my bear $66 E (estimate).⁴¹⁷ The price is above the bull case on every one of these bases, and no conclusion below depends on the date.

Before the numbers, I ask how an owner loses money here. I see three ways. The customers do not fund their orders, the build-out stops, and AMD is left with $30.3B of purchase commitments and a plant sized for demand that left⁴. The warrants hand two customers 320M new shares, about a fifth of today’s count, so growth is shared before it reaches me C (calculated).⁴¹ And the price already pays for the best outcome, so a good result that falls short of my bull case still loses money.

What an owner can take out is smaller than the headline. Free cash flow in the TTM was $7.7B, and it is already after capital spending. Take off $1.9B of stock pay and $5.84B is left, about $3.52 a share, against non-GAAP earnings of $5.76 C (calculated).¹⁷⁴ That is 61% of the headline figure, and the price is 143 times it C (calculated).¹ Part of the first-half cash also came from suppliers not yet paid⁴.

Two things I cannot know. Nobody discloses what a gigawatt earns AMD, so I show $10B to $20B instead of one figure. And I cannot tell which customers will fund what they have said they intend. So the answer on this page is a range, $61 to $434, and the $221 weighted value is only an average of three stories.

What the price implies

Start with earnings. At $504.80 AMD trades at $837.5B on 1.659B diluted shares, 87.6 times its TTM non-GAAP earnings of $5.76 and 106 times after stock pay C (calculated).^117192021 The company targets non-GAAP earnings above $20 a share within three to five years of November 2025, and in August management said it expects to exceed that target significantly.⁴⁰⁴¹⁴² The price is 25.2 times $20, or 30 times if the same profit is spread over the 1.98B shares that exist when all the warrants vest C (calculated).⁴⁰¹

The return comes first. At $504.80 my cash flows return 3.8% a year in the bear case, 6.1% in the base and 9.2% in the bull, against 5.00% on the ten-year Treasury E (estimate).¹⁵³ The base case beats the risk-free rate by about one point, and the price is 8.3 times my bear value.

One constant rate is a blunt tool for a business that has just started to compound, so the table reads the price several ways. Solved as one rate, from TTM free cash flow of $5.84B after stock pay and with no warrant shares counted, the price needs 33.2% a year for ten years, to about $103B in year 10 C (calculated).⁴⁶¹⁷ AMD’s free cash flow after stock pay grew 8.1% a year from 2021 to 2025 C (calculated).⁶¹¹ Read as revenue, the price needs $410B in year 10 on my base margins, 2.7 times my base. Read as a return, a holder at 25 times my base year-5 earnings loses 2.1% a year.

At $504.80 the price needs free cash flow after stock pay to reach about $103B in year 10; my bull case reaches $83B.

What the market is paying for

There is a case for the price, and it is about the size of the market. Management gives about $1.4 trillion for data-center AI accelerators in 2030, growing above 45% a year, and about $220B for server CPUs, where it targets more than 50% of revenue.⁴¹⁴²⁴⁰ My base case has AMD’s Data Center segment, which includes CPUs, at 6.8% of the $1.4 trillion in year 5, my bull at 13.5%, and the price at about 16% C (calculated). AMD’s segment is about 7% of what NVIDIA’s and AMD’s data-center segments report today C (calculated).¹⁷⁵¹ A 50% share of $220B would be $110B on its own, more than my base year-5 Data Center revenue of $95B C (calculated). If the $1.4 trillion arrives, the price needs about 16% of it, a little above my bull’s 13.5% and 2.3 times my base’s 6.8%.

Put as one set of assumptions, the price needs my bull case with revenue about 1.2 times as high in every year: $252B in year 5 (my bull $216B, my base $119B), $353B in year 10, my bull margins and all 320M warrant shares. Year-5 Data Center revenue of $221B would be 10.2 gigawatts a year at $13B and 60% accelerators, against 4.4 in my base and 8.7 in my bull E (estimate). I give that case under 25% weight because my bull already has the two warrant holders buying every gigawatt they named, and delivers management’s August outlook in full.⁴¹

The price needs three things at once: the two customers take all 12 gigawatts, the operating margin passes 35%, and revenue runs 1.2 times my bull path. If I give each one chance in two, all three together are one in eight C (calculated). I weight my bull case under 25%, and the price asks for more than that.

Seven months of price and filings

The stock closed at $190.95 on 3 March 2026, almost exactly my base value, and at $649.42 on 6 October. The table lists AMD’s 8-Ks in that span and the Anthropic, Helion and Meta releases, with the close before and after each. I read AMD’s list of press releases only from 22 July onward, so releases between March and July that were not 8-Ks may be missing.¹⁵⁶

Four sessions carry a listed event and net to $37 of the $458 rise; the rest came on sessions with none of them. My model uses the filings through Q4 2026. I did not run it event by event, so I cannot say how much each event changed my value, and I offer no cause for the rest of the move C (calculated).¹⁵⁶

The discount rate

I use 10%, the rate I use for a large business with net cash, which is also what the capital asset pricing model gives for a beta of about 1.2 C (calculated). The model with AMD’s measured betas gives more: the 5.00% ten-year Treasury yield plus Damodaran’s 4.20% equity risk premium, times betas of 2.24 over ten years of monthly returns and 2.48 over five, gives 14.4% and 15.4% C (calculated).^5354551 Ten percent is generous for a stock with a ten-year beta of 2.24, and it raises every value: at 12% my base is $147, my bear $49 and my bull $314, and at 14.4% the base is $113 E (estimate). I keep 10% so the page compares with the others, and put the cyclical risk into the cash flows, where it can be seen. The bear case is that risk.

Three cases

I build each case from the income statement down: revenue by segment, non-GAAP gross margin, operating spending, stock pay as a cost, tax of 15% rising to 17%, capital spending of 3.0% to 3.5% of revenue against depreciation of 1.6% to 2.8%, and working capital at 15% of revenue C (calculated).⁶⁴¹⁷ Each case divides by its own share count: today’s 1.659B diluted shares plus the warrant shares that vest in that case. The base counts 160M of them, which lowers it from $213 to $194.

Growth after year 10 is 3.0%, 3.5% and 4.0% in the three cases. With 3.5% in all three the values are $64, $194 and $411. After year 10 comes 54% of the bear value, 59% of the base and 66% of the bull. In my base the ten explicit years are worth $77 a share, the terminal value $110 and net cash $7 E (estimate).⁴

The base case takes the company’s near-term outlook and then assumes a normal cycle.

• Year 1 revenue of $62.7B, about what $13.0B in Q3 and growth of 12.7% a quarter after it gives; Data Center up 93% E (estimate).¹⁷¹⁸
• Year 2 revenue of $87.3B and year 3 of $108.2B, with Data Center up 55% and 30%. My interpolation puts Data Center up 86% in calendar 2027, a little under management’s “more than double” E (estimate).⁴¹⁴²
• A flat year 4, with Data Center unchanged, as customer spending meets customer funding E (estimate). Then Data Center up 10% and 8% falling to 4.5% by year 10.
• Non-GAAP gross margin from 55.0% in year 1 to 50.0% in year 10, as rack-scale systems carry less margin than CPUs and rivals respond. Operating spending falls from 25.7% to 20.5% of revenue, so the non-GAAP operating margin peaks at 32.5% and ends at 29.5%. AMD targets above 35%.⁴⁰
• The customers take 6 of the 12 gigawatts, so 160M warrant shares vest. The table in Valuation shows what that explains: 29% of the Data Center growth, or 39% with Anthropic, and the rest is server CPUs and buyers not named E (estimate).

Year-5 Data Center revenue in the base is 4.4 gigawatts a year at $13B each if three-fifths of it is accelerators and racks; at $10B it is 5.7 and at $20B 2.9. Over five years the base needs 17 gigawatts at $13B, more than the 14 the three customers named, so buyers not named supply part of it or revenue per gigawatt is higher than I assume E (estimate). My base calendar-2030 revenue is about $114B, 27% a year from 2025, below the 35% a year of the company’s long-term model C (calculated).⁴⁰⁶

My base-case non-GAAP earnings reach $20.0 a share in year 5 on 1.66B shares, and $18.3 on the 1.82B shares that include the warrants. The company’s target is above $20 within three to five years of November 2025, so my base case reaches it about seven months after that window closes, and management said in August that it expects to exceed it significantly.⁴⁰⁴¹⁴² A buyer at $504.80 pays 25 times $20. At 25 times my base year-5 earnings of $18.26 on 1.82B shares the share would be worth $457 in mid-2031, which is 2.1% a year below today’s price C (calculated).⁴⁰¹

The bear case is stranded capacity. The customers take only the first gigawatt each is bound to purchase. Data Center grows 65% in year 1 on the current ramp, then 5%, then falls 15% as funding and power stop the build-out, and AMD books a $6B write-down, 20% of its $30.3B of purchase commitments E (estimate).⁴ Gross margin falls to 48% and the non-GAAP operating margin to 21%, the range AMD earned from 2023 to 2025 C (calculated).⁶⁹⁸ Revenue is $65B in year 10 and free cash flow $8.7B. The bear value of $61 is 9% above the $55.94 close of October 2022.¹

The bull case is AMD delivering its August outlook in full. Data Center more than doubles in 2027 (up 116% in calendar 2027 on my interpolation) and nearly doubles again in year 2, the two warrant holders take all 12 gigawatts they named and the 320M warrant shares vest, and the operating margin reaches 37%, above the 35% target.⁴⁰⁴¹ Calendar 2030 revenue of about $205B is 43% a year from 2025 and a third above the $155B that 35% a year gives; management said in August that revenue would grow substantially above 35% C (calculated).⁴⁰⁶⁴¹ Year-5 earnings are $35 a share on 1.98B shares and $42 on 1.66B, well above the target of $20 C (calculated).⁴⁰ Weighted 25%, 50% and 25%, the three cases give $221 a share E (estimate), against a price of $504.80.¹

What closes the gap

No single change gets there. The largest lever is the discount rate: my base case is $230 at 9%, $283 at 8% and $366 at 7% E (estimate). The largest operating lever is the bull revenue path, which takes the base to $324 even after the 320M warrant shares E (estimate). Bull margins on base revenue give $244. Skipping the flat year adds $18, to $212 E (estimate).¹

The price is my bull case at a 9.2% discount rate C (calculated). It is also year-10 revenue of $410B on my base margins, or $353B on my bull margins and shares C (calculated). NVIDIA’s June quarter, annualised, is $385B.⁵¹

The same dollar has other uses. On my NVIDIA page the base case returns 9.3% a year, and here it returns −1.6% C (calculated). How good the business is and how good the price is are separate questions. Nine years of 26% growth answer the first; a price above my bull case answers the second.

What the model leaves out

The model counts the dilution that comes with customer revenue, and none of the cash flows of World Labs, Taalas or MEXT. It treats the stakes and the guarantees as small. The table prices each in dollars a share. The largest item is AMD’s own data-center leases: $14.0B of aggregate future payments with no stated purpose, or about $1.0B to $1.3B a year of rent by my arithmetic. If that rent is not already in my operating spending it costs $6 to $8 a share, 3% to 4% of my base E (estimate).⁴

The margin of safety

• Price$504.80Close, Q4 2026 [1]
• Base value$194My estimate E (estimate)
• Margin of safety below$81set from my bear case C (calculated)

I set the margin of safety from the bear case, because the same customers and the same cycle drive this business. Below $81, my bear value of $61 is at least three-quarters of the price paid, and my base case would return about 19% a year E (estimate). That is 9 times my base earnings for the next year. At $81 the bear case still loses 24%. The cushion is the 58% discount to my base value, and it has to cover what the model does not hold: a customer failing to fund its order, a supply shock at TSMC, the loss of the chief executive, and my own error. The margin-of-safety price is 45% above the October 2022 low close of $55.94 and 58% below the March 2026 low close of $190.95; the price is 6.2 times it and 8.3 times my bear value C (calculated).¹

Where this could be in 5 and 10 years

For my base to be right at five years, revenue has to reach $119B by June 2031 and free cash flow after stock pay $26.4B, a 22% margin C (calculated). It is right at ten years if revenue is $153B at a 21% margin, $32.5B E (estimate). The value climbs from $194 today because the cash comes nearer, at 10% a year, and not because I add anything.

The bear and the bull differ on growth and margin. In the bear, revenue stays near $55B to $65B and free cash flow is 13% of it. In the bull, revenue reaches $302B at a 27.5% margin. The discount rate is 10% in all three, so the gap between them is the cash. At 12% my base is $147 today.

The most likely way my base is wrong is the build-out. If the large customers cannot fund their orders, Data Center stalls early, and the five-year value moves toward the bear’s $76. The price is above my base value at both dates.

These are estimates of value, not forecasts of the share price. The market price can sit above or below value for years.

VIIThe case for the price

What most people miss

AMD’s July release did not say whether Anthropic’s order was binding.³⁵ Anthropic’s S-1 of 26 October does. Anthropic committed to a first block of Helios capacity in calendar 2027, subject to deployment conditions. It cannot cancel unless AMD breaches the deal and fails to fix it.⁶¹ There is no fixed dollar amount, but if it takes the full block it expects to spend more than $20.0B. That money goes through manufacturers or third parties, so it is not all AMD revenue. No warrant appears. AMD instead commits to purchase up to $5.0B of Anthropic stock as delivery milestones are met.⁶¹

Intel’s 10-Q of 30 October shows server revenue up $2.5B in Q3. Volume rose 10%, and the average price rose 54%, mostly on a richer mix. Demand for some data-center products is above supply, and Intel expects the shortage to last into next year.⁶² That could limit AMD too.

The bull case

The first engine is the server CPU. On 6 October Citi’s Atif Malik raised his target to $800, TIKR reported. His reason is AI agents such as Meta’s Muse. He sees the CPU market going from $29B in 2025 to $300B in 2030, with AMD “the key beneficiary”.⁶⁷ AMD says about $220B for server CPUs, and expects its server revenue up more than 70% in 2027.⁴¹

The second engine is AMD as the other supplier of AI systems. BNP Paribas’s Karl Ackerman went to $960 the same day, 24/7 Wall St reported. He sees AMD taking at least 8% of a GPU market above $1 trillion by 2030.⁶⁶ I read both reports, not the notes. Frank Downing of ARK wrote on 25 March that AMD’s GPUs are “competitive with Nvidia on small model inference” for the cost, while Nvidia keeps a big lead on large models.⁶³ ARK Innovation’s AMD weight fell from 4.99% at 30 June to 1.82% in its 30 October file.⁶⁴⁶⁵ Lisa Su in August: Data Center growth “well over 100%” in 2027, and earnings that “significantly exceed” her $20-a-share target.⁴¹

Against reality

If AMD delivers that outlook in full, OpenAI and Meta take all 12 gigawatts, and the non-GAAP operating margin reaches 37%, the price is possible. That is my bull case: $434 at my 10% rate, or 9.2% a year from $504.80 E (estimate). To earn 10% from here, revenue also has to run about 1.2 times that path, to $252B in year 5 E (estimate).

Today is a long way from that. Revenue was $41.3B in the twelve months to June, so $252B takes about 44% a year for five years C (calculated). The past nine years gave 26%. Data Center was $22.2B and has to reach $221B E (estimate). The non-GAAP operating margin was 26.0%.¹⁷ No warrant share had vested by 27 June.⁴

In October 2025 Brad Gerstner of Altimeter called the OpenAI deal “a bet-the-farm bet by Lisa Su” on the All-In podcast, Benzinga reported. He said AMD’s last chip could not compete with Nvidia’s and that AMD is out if MI450 is not adopted. If it works, he said, OpenAI alone could bring $150B.⁶⁸ That test is still open: first shipments were guided for late Q3.⁴¹

Q3 is guided at $13.0B. To reach June 2027, the next three quarters must average $16.6B on my base, $18.0B on my bull and about $21.7B on the price’s path C (calculated).

What they have to hit

The thresholds match my signposts below.

Where I come down

The S-1 moved me on one point. Anthropic’s order is firmer than the release showed, and AMD takes its stock instead of giving away its own, so the warrants OpenAI and Meta got may not be the going rate. My $61, $194 and $434 did not move: the price still needs more than my bull, and one filing says nothing about the other two customers. Data Center more than doubling in calendar 2027, where my base has 86%, with orders past the first gigawatt on no new warrant, guarantee or equity, would raise my values.

VIIIRisks, and what would change my view

1. Customer funding and concentration. The 10-K warns that customers may lack the capital for their AI infrastructure and may ask for deferred payment.⁶ One gigawatt at $13B is 15% of my year-2 base revenue E (estimate). OpenAI’s and Meta’s orders are binding for one gigawatt each, Anthropic has committed to a first block on conditions, and the rest is intent.²²²⁴⁴
2. Financing the buyers. AMD gives warrants, guarantees up to $4.1B of a partner’s leases that are payable if the partner defaults, and has committed up to $5B to Anthropic. On 27 October it said it will invest $1B in Meta’s equity offering.⁴³⁵⁵⁷ The size is small against $13.1B of cash and large if several things go wrong together.
3. Dilution. The warrants cover up to 320M shares, 19.6% of the shares outstanding. Add 37M unvested awards, about 16.2M shares for World Labs at today’s price, and a larger equity plan with 107M shares left.^46273036 My base counts 160M of the warrant shares.
4. Supply and cost. TSMC makes every advanced wafer, the 10-K reports an industry-wide memory shortage, and $30.3B of purchase commitments cannot be cancelled. An analyst put Helios at almost 50% more high-bandwidth memory than the competing rack; Su called the larger memory footprint an advantage and said AMD can trim it.⁶⁴⁴¹
5. Execution. MI450 and Helios ship in volume from the fourth quarter of 2026, MI500 follows in 2027, and AMD is adding a new rack-scale product each year.¹⁷⁴¹³⁸ A one-quarter slip moves the first-gigawatt revenue into the next year.
6. NVIDIA, Broadcom and Intel. NVIDIA’s cadence, custom chips from Broadcom’s customers, Intel’s supply recovery and pricing, and Arm-based CPUs all aim at the same budgets.⁶⁵²⁴⁴
7. China and export controls. China including Hong Kong was 22% of 2025 revenue by billing location, and the MI308 charges showed how fast a licence can change.⁶²⁰
8. Accounting. Goodwill of $25.5B and acquisition intangibles of $15.6B could be impaired, the warrants are liabilities until equity-classification conditions are met, and the 10-Q does not say how vested tranches will be booked; a payment to a customer usually reduces revenue.¹⁷⁴
9. One leader. Su is chair and chief executive and, in the board’s words, the principal architect of AMD’s strategic plan. The $75M award vests only while she remains chief executive or executive chair.²⁸
10. A volatile share price. The stock rose 240% in seven months to a close of $649.42 on 6 October 2026 and fell 22.3% from that close to $504.80. It fell 10.9% on 30 October alone, the day after Intel forecast Q4 revenue of $15.0B to $16.0B against $16.5B in Q3, in a sector fall: the PHLX Semiconductor index lost 7.2% and the Nasdaq Composite 2.9%. AMD filed and released nothing that day, and I offer no other cause for the move that a filing does not show.^156585960

What would raise my base value: Q3 and Q4 deliveries on the first-gigawatt schedule, Data Center margin above 33%, orders beyond the first gigawatt with no new warrant, guarantee or equity commitment, and server revenue share above 50%.^17404445 What would lower it: any new warrant or guarantee, a Q3 revenue figure below $12.7B, purchase commitments above $40B with flat guidance, a gross margin under 52%, or payables unwinding.¹⁷⁴⁵ AMD’s $1B for Meta’s equity offering is a new equity commitment, so that test has already gone against it. It costs my base $0.55 a share if it were worth nothing, so $194 stands until the Q3 report.⁵⁷

The first test comes on 3 November 2026, when AMD reports Q3. It guided revenue of about $13.0B, plus or minus $0.3B, a non-GAAP gross margin of about 56% and operating expenses of about $3.65B, which implies non-GAAP earnings of about $1.93 a share C (calculated).¹⁷¹⁸³⁹

What I could not verify

• Instinct revenue is not disclosed in any filing, so I use no share of the AI-accelerator market.¹⁷
• The warrant vesting schedule and exercise conditions are redacted, and how a vested tranche will be booked is not stated.²²²⁴⁴
• The 10-Q’s $5.0B investment commitment names no counterparty. AMD’s release names Anthropic for up to $5B, and Anthropic’s S-1 describes the same size from July 2026. They are probably one item, and I count it once.⁴³⁵⁶¹
• What AMD’s own data-center leases are for, and their annual rent. The rent range is my arithmetic on the aggregate payments and lease terms.⁴
• Repayment of the $875M of notes due on 24 September.⁴
• Mercury Research’s report itself, and its Q2 2026 revenue shares; I used Q1 2026 for revenue and Q2 for units.⁴⁴⁴⁵⁴⁶
• AMD’s own call transcript. The server growth rates, the $1.4 trillion and $220B markets, “double-digit billions” a gigawatt, the Data Center AI margin remark, the memory exchange and the outlook above the $20 target come from two unofficial transcripts and a summary that agree. They anchor my year-1 and year-2 growth: without “more than double” for 2027 I would have no company statement past Q3 guidance for year 2. The filed release and slides do not carry them.⁴¹⁴²⁴³
• The accelerator share of Data Center revenue (60% in the cross-check) and revenue per gigawatt ($13B) are my estimates; neither is disclosed.⁴¹
• The CEO award’s baseline and hurdles, which are my arithmetic on the 8-K’s rules, and the company’s own count of tranches earned.²⁸
• Taalas and MEXT prices, and the holdings of Vanguard and BlackRock after early 2024.³⁷¹⁶
• The close. The price rests on Nasdaq’s 4:00 pm closing print, and Yahoo’s chart feed shows the same close. Nasdaq’s daily history had not posted the row when I checked.¹²³
• The beta. It uses the S&P 500 monthly series I pulled for my Microsoft analysis, because Yahoo refused my own requests.⁵⁵