The case for a $2 trillion AMD is a case about earnings. At $633.91, the return depends almost as much on how many ways those earnings are divided: among shareholders, employees, sellers, and the customers AMD now pays in its own stock.
At $633.91, its close on October 2, AMD's price assumes that its earnings per share multiply. Value the company at thirty times earnings at the end of 2030, and the price offers 12% a year if AMD earns $34.19 a share that year and 15% if it earns $38.25: about eight and nine times the $4.17 it earned in 2025, on its own adjusted basis. Management's target is more than $20. Almost as much depends on how many ways the earnings are divided. A $2 trillion market value at the end of 2030 offers 16.3% a year shared among today's 1.66 billion diluted shares, and 11.3% shared among two billion, about the count that success would bring, because OpenAI and Meta can earn up to 320 million AMD shares, at a penny each, as they buy its accelerators. Three things will decide it, and none is in a filing yet: what the newest accelerators earn, how the warrants will reach earnings, and whether cash keeps pace.
Two trillion dollars turns up twice in the AMD story this autumn: as the size AMD says its markets will reach by 2030, and as the market value an analyst says the company could reach, about twice the $1.03 trillion its shares were worth on October 2, after a 28% rise since September 14. This note divides both.
The argument is short and unusually well documented. Artificial intelligence is moving from answering questions to doing work, and work done by software agents needs ordinary processors as well as accelerators. When Microsoft announced new Azure machines in July, built on nearly 500 of AMD's newest server cores, "agent coordination at scale" was on the list of jobs. Even software, it seems, needs middle management, and the management runs on CPUs. AMD sells both kinds of chip. On September 30, Ben Reitzes of Melius told clients, as CNBC reported, that this could carry AMD toward $2 trillion, and that it would likely need to show that "EPS power of $50+ is possible."
AMD's own case is more measured. Last November it set targets for three to five years out: revenue growth above 35% a year, an adjusted (non-GAAP) operating margin above 35%, and adjusted earnings per share above $20, which in August it said it now expects to exceed significantly. It expects Data Center revenue to more than double in 2027, and the markets for data center accelerators and server processors to reach about $1.4 trillion and $220 billion by 2030. OpenAI and Meta each intend to deploy up to six gigawatts of its accelerators, and Microsoft will run its Helios racks on Azure.
None of that is in dispute here. The question is the owner's: what does $633.91 already assume, and what is left for the person who pays it?
Several of Adrian Slywotzky's profit models describe a piece of AMD; one describes the way it makes money. Time profit, in The Profit Zone (1997), is the premium earned on what is new, which lasts only until a capable rival copies it, so the business lives by launching the next product before the copy arrives. Slywotzky's example was Intel under Andrew Grove (pp. 41 and 59); in the sequel, Profit Patterns (1999, pp. 197 to 199), AMD is the rival coming up underneath. AMD's latest 10-Q describes the same machine: "product pricing declines over the life of a specific product," new products are needed to hold the average selling price, and launch cycles are getting shorter.
A model that fits is half the question. The other half is whether the company is winning at it, which for time profit means launching faster than it is copied, with new products that hold the margin up rather than pull it down:
Here T is how long a capable rival takes to match a new generation, L the time between AMD's own launches (by its plan, now a year for its rack-scale accelerator systems), and m the gross margin: each business's margin mk, weighted by its share of revenue wk. A product with a below-average margin pulls the blend down as it grows. The second quarter, business by business:
| Business | Newest products | Price and margin on the newest | Winning? |
|---|---|---|---|
| Server processors (EPYC) | Fifth-generation Turin ramping; sixth-generation Venice in production | Units and average selling price both up by double digits; adds to the gross margin | Winning |
| Data center accelerators (Instinct) | MI355X shipping; MI450 in Helios racks from the third quarter | Gross margin "slightly below corporate average"; the first two gigawatt-scale buyers hold warrants for up to 320 million shares | Not yet: a challenger, paying to be adopted |
| AMD as a whole | All of the above | Adjusted gross margin about 54% a year earlier before the MI308 charge, 55% in the first quarter, 56% in the second; guided to about 56% | Holding |
From AMD's second-quarter 2026 release, 10-Q and earnings call. AMD reports one Data Center segment and does not disclose processor and accelerator margins separately.
In server processors AMD is winning. EPYC sales grew more than 70% in the quarter, AMD expects server revenue to grow more than 80% in the second half and more than 70% in 2027, and it says it gained share of x86 server revenue. In accelerators it is not winning yet, nor would anyone expect it to be: there AMD is the copy arriving, not the original being copied. Its 10-Q describes Nvidia using its position and proprietary software to influence AMD's customers, and AMD sells Helios on more inference output per dollar than the leading rival rack, which is a challenger's argument, not a leader's premium. Its bet is that processors and a recovering embedded business keep the blend steady while accelerators ramp.
Underneath the accelerators sits different mathematics, that of technologies worth more the more people use them (Arthur, 1989). Users bring software, tools and trained engineers, so markets tend to tip toward whichever platform gets ahead, and stay there. A buyer chooses AMD when
where a and b are what each platform is worth to that buyer on its merits, nAMD and nlead how widely each is already used, r what each additional user adds, and w anything AMD pays to be chosen.
In the language of the model, the warrants are w made visible. OpenAI and Meta can each earn up to 160 million shares as their purchases climb from the first gigawatt to six and as AMD's share price clears thresholds rising to $600 for the final tranche, with further conditions before exercise. The exercise price, one cent, is the only term nobody needs to model. The buyers also gain something the inequality leaves out: an outside option in every negotiation with the leader. Winning, in this model, means the w needed falls as nAMD grows, until the next buyer needs none, and the evidence so far points that way, faintly. Microsoft's plan to run Helios on Azure came, as AMD announced it, with no warrant, and the June-quarter 10-Q describes warrants for OpenAI and Meta alone.
This reminds me of Ecclesiastes 11, which holds both sides of these contracts in two verses. The first is the familiar counsel to send your bread out on the waters and trust it to come back, which is a fair description of a warrant: shares given now, to return as purchases. The second belongs to the buyers: "Divide your portion among seven, or even eight" (BSB). Meta calls it a portfolio approach. AMD is acting on the first verse, and its customers on the second.
We rebuilt the valuation behind this note line by line and turned it around, in the manner of expectations investing (Mauboussin, 2006): take AMD to the end of 2030, at a multiple of that year's earnings or at a market value shared among the shares then outstanding, and ask what annual return carries $633.91 there.
Here 4.2464 is the years from October 2, 2026 to December 31, 2030, EPS is on AMD's adjusted basis, and AMD pays no dividend. Vista does not publish what a share is worth. What the price offers on each case is the more useful question.
| The case | What it assumes | What $633.91 offers a year |
|---|---|---|
| Management's floor | $20 of 2030 EPS, the analyst-day target, at 30 times | −1.3% |
| Middle case, lower multiple | $35 of EPS at 25 times | 7.9% |
| The headline, widely shared | $2 trillion of market value, shared by 2.1 billion shares | 10.1% |
| The headline, success-case shares | $2 trillion shared by 2.0 billion shares: today's count plus both warrants and World Labs | 11.3% |
| Middle case | $35 of EPS at 30 times | 12.6% |
| The headline, today's shares | $2 trillion shared by 1.66 billion diluted shares | 16.3% |
| Bull case | $50 of EPS at 25 times | 17.3% |
Annual return at which the October 2 close compounds to each case at December 31, 2030. Inputs and method are at the end of this note.
The table says four things.
The earnings have to multiply. At thirty times, 12% a year needs 2030 EPS of $34.19 and 15% needs $38.25: 8.2 and 9.2 times 2025's $4.17, or growth of 52% and 56% a year for five years. (The 2025 figure carries about $440 million of net export-control charges.) The full grid:
| Return the reader requires | At 25 times | At 30 times | At 35 times |
|---|---|---|---|
| 10% a year | $38.01 | $31.67 | $27.15 |
| 12% a year | $41.03 | $34.19 | $29.31 |
| 15% a year | $45.90 | $38.25 | $32.79 |
2030 earnings per share, adjusted basis, at which $633.91 offers each return at each end-2030 multiple.
Chan, Karceski and Lakonishok (2003) found sustained high growth in earnings rare, and no more common than chance would produce; Fama and French (2000) found that unusual profitability reverts toward the average. At thirty times and 12%, the price assumes AMD is the exception.
On management's floor, the multiple carries everything. If AMD earns exactly $20 in 2030, the price offers 12% a year only if the market then pays 51 times earnings.
The two trillions meet. To be worth $2 trillion at thirty times, AMD must earn about $66.7 billion a year. At a 30% net margin, roughly what its target of an operating margin above 35% leaves after its 13% adjusted tax rate, that takes $222 billion of revenue, against $34.6 billion in 2025. For the market value to reach the first two trillion, AMD must sell about a ninth of the second.
The denominator matters as much as the numerator. The widest gaps in the table are between share counts, not between optimists and pessimists. The 2.0 billion is a bridge, not a forecast: today's 1.66 billion diluted shares, plus 320 million warrant shares at full vesting, plus about 13 million for World Labs at today's price, makes 1.99 billion. Full vesting alone, added to today's 1.63 billion shares outstanding, would cut earnings per share by 16.4% at the same earnings. The warrant shares arrive mainly when purchases do, so dilution and revenue move together, though not in proportion and not on a published timetable. Only the first gigawatt is binding for either customer, and as of June 27 no warrant share had vested. Timing matters too: the same $2 trillion over 2.0 billion shares offers 22.5% a year if it arrives at the end of 2028, and 7.6% at the end of 2032.
One more question hides in the word "earnings." AMD's $20 target, and every EPS figure here, is on its adjusted basis, which adds back stock compensation: $503 million in the second quarter, 30 cents of the $1.66, about 18%. Apply the thirty times after stock compensation at that share, and the $35 case offers about 7.5% a year, not 12.6%. That is a hypothetical, and a consistent one only because this model adds no future grants to the share count. The evidence favors charging it. Investors have priced stock compensation as a real expense (Aboody, Barth and Kasznik, 2004), analysts who left it out set estimates too high (Mohanram, White and Zhao, 2020), and expenses excluded from adjusted earnings have predicted lower future cash flow (Doyle, Lundholm and Soliman, 2003).
This reminds me of the question Warren Buffett put to Berkshire Hathaway's shareholders in 1992, about stock options: "If compensation isn't an expense, what is it?" AMD's contracts add a second clause. It now pays customers in shares as well as employees, and under the revenue standard, value given to a customer generally comes off revenue unless it buys something distinct. Once the first tranche vests, the question is whether the adjusted figure, the one the $20 target is written in, will count it.
The second quarter was a record. Revenue rose 50% to $11.5 billion. Data Center revenue more than doubled to $6.7 billion, 58% of the total, and the segment's operating margin rose from 27.7% in the first quarter to 31.3%, measured before the stock compensation and acquisition charges AMD reports outside its segments. Adjusted EPS was $1.66, against $1.38 reported. AMD guided third-quarter revenue to about $13 billion, plus or minus $300 million, at an adjusted gross margin of about 56%. Year-on-year comparisons flatter, because the second quarter of 2025 carried $800 million of charges for MI308 accelerators that export controls kept out of China.
The cash is more complicated. Free cash flow was $1.56 billion against $2.76 billion of adjusted net income, a conversion of 56%, against 82% for the half and 113% for the first quarter alone; operating cash flow fell $589 million and capital spending rose $419 million, to $808 million. Receivables, inventory, prepaid costs and accrued liabilities absorbed $3.0 billion of cash, and a $2.27 billion rise in payables, equal to 96% of the quarter's $2.37 billion of operating cash flow, gave most of it back, for a net use of $730 million ($274 million for the half). That ratio is arithmetic, not a counterfactual: payables could not vanish while purchases stayed the same. But earnings that run ahead of cash deserve attention (Sloan, 1996), and payables are the line to watch.
AMD is also committing money faster than its balance sheet shows. At June 27 it had $30.3 billion of unconditional purchase commitments, $17.4 billion of them due in the rest of 2026, mostly for wafers, components and cloud capacity; $4.5 billion of leases not yet commenced; and guarantees of up to $4.1 billion on partners' data-center leases, which the filing says may be issued in exchange for warrants. Since then it has added up to $5.0 billion of conditional investment commitments, and data-center leases with $9.5 billion of payments over as long as sixteen years. These are different obligations on different timetables, not to be added together or to debt, and all of them rest on one foundry: TSMC makes every AMD processor and GPU wafer at 7 nanometers and below.
In August AMD also borrowed $4.75 billion, in notes due from 2029 to 2036 at coupons of 4.6% to 5.5%, about $240 million of interest a year. The $875 million of notes due September 24 have since matured; no filing yet shows their repayment. At June 27, cash and short-term investments of $13.1 billion exceeded debt of $3.2 billion by $9.9 billion. The new notes raised cash and debt together, so net cash is broadly unchanged, but gross debt is now more than twice the June figure.
Over the first half, shares outstanding rose from 1,630 million to 1,632 million. Buybacks of $221 million and $341 million of shares withheld for employees' taxes, less $205 million from employee plans, came to a net $357 million, a cash subtotal rather than the cost of offsetting dilution, against $990 million of stock compensation expense. The equity plan has 107 million shares available, which is room to issue, not issuance. World Labs, an AI model lab AMD agreed in September to buy for about $8.2 billion in stock, will add shares priced just before closing, expected by year-end: about 12.9 million at October 2's price, as an illustration.
The research cuts both ways. The shares of companies whose share counts grow have tended to return less afterward (Pontiff and Woodgate, 2008), though shares given to customers for their purchases are a new kind of issuance that the research predates. Suppliers with concentrated customer bases have earned higher returns on their assets, through lower costs and better use of capital (Patatoukas, 2012). AMD had no customer above 10% of revenue in 2025 or 2024, but says it depends on a small number of customers and expects to go on doing so.
Here is the strongest case we can build without cheating. Management is running ahead of its plan: it says it is well ahead of last November's financial model, expects Data Center revenue to more than double in 2027, and told an analyst whose 2027 accelerator estimate was about $30 billion that his number was probably too low. The processor engine is real, adds to the margin, and is confirmed by customers: Microsoft's machines for agentic work, Meta validating sixth-generation EPYC platforms in its labs. The warrants pay out mainly where AMD is shipping gigawatts, so the dilution arrives with the revenue. Credit $35 of 2030 earnings at thirty times and the price offers 12.6% a year; $50 at twenty-five times, 17.3%.
Those are not poor returns, and the hurdle is the reader's. An investor who requires 12% will find that the middle case just clears it; one who requires 15% will find that only the bull case, and the headline shared among today's shares, does. Which hurdle is right is the reader's decision, not ours.
The case can fail without AI spending falling. AMD can win volume at thin margins. Agentic demand can go to Arm designs or to customers' own chips; Meta described its AMD deal as one part of a portfolio that includes its own accelerators. Customers can use AMD mainly as leverage against the leader. Commitments can absorb cash. Shares issued to employees, customers and sellers can outrun earnings. And even a good outcome can disappoint an owner if it arrives late, brings more dilution, or earns a lower multiple.
Likely, we can get some of it from primary research, which is what Vista is all about. The filings will report, a quarter at a time, whether the ramp arrived and what it earned. People who have bought these systems, run models on them and kept the books on contracts like these can say now what usually happens.
We would askWhat separated an announced gigawatt from capacity that was installed, used and reordered, and what made a second platform stick?
The answer that would change the view"Second platforms usually stall after the first commitment."
We would askWhere does the software gap still cost time or money: in training, in inference, or in the processor-side work of coordinating agents?
The answer that would change the view"Inference moves over cleanly; training does not, and will not for years."
We would askHow much of agentic AI's processor demand is new work rather than relabeled capacity, and how much will go to Arm designs or customers' own chips rather than x86?
The answer that would change the view"Most of the new agentic demand goes to the customers' own Arm chips."
We would askAs a matter of past practice, how do purchase commitments of this size behave when demand shifts, and which constraint binds first: wafers, memory, packaging or substrates?
The answer that would change the view"Commitments like these are rarely cut without paying for the capacity anyway."
We would askWhen warrants like these vest, how has the cost usually flowed through revenue, adjusted earnings and cash?
The answer that would change the view"Companies usually leave it out of adjusted earnings."
Every conversation would draw on past, general experience only: no current employees of AMD or of its customers and suppliers, nothing confidential, and every Advisor screened for conflicts before a word is said. For a client engagement, Vista would combine this note with those interviews: two or three Advisors, structured conversations led by a Research Director, and one written brief that says where they agree, where they split, and what would change the answer.
We will score this note after AMD reports its fourth quarter of 2026 and files its 10-K (it reported the fourth quarter of 2025 on February 3), and publish the result on the scorecard whether it flatters us or not. The tests are fixed today, so they cannot drift:
Four or five passes, and the first year of the path the price assumes has been walked. Two or fewer, and the price is ahead of the evidence. Three is too early to say. We will also report what the price then offers on each case, and whether any warrant tranche has vested. Some evidence comes sooner: Lisa Su gives an opening keynote at the OCP Global Summit on October 12, and AMD reports its third quarter on a date it had not announced on October 3.
Two trillion is a round number, and round numbers are easy to say. The work is in the division: by the earnings that must arrive, by the years they take, and by the shares that will be standing there when they do. The price has already assumed the first. The next two quarters will begin to settle the other two.
Annual return = (2030 EPS × end-2030 multiple, or market value ÷ shares, ÷ $633.91)1 / 4.2464 − 1. The horizon is the 1,551 days from October 2, 2026 to December 31, 2030, divided by 365.25. No dividend is assumed. The EPS the price needs = $633.91 × (1 + r)4.2464 ÷ the multiple. Growth from 2025 runs five years from AMD's 2025 adjusted EPS of $4.17. Revenue at a 30% net margin assumes 2.0 billion shares. The success-case share count adds both warrants at full vesting and the World Labs shares at an illustrative October 2 price to the 1.66 billion diluted shares AMD guided for the third quarter, for 1.993 billion, rounded to 2.0 billion; it is a scenario, not a forecast of issuance or a ceiling. The stock-compensation case applies the multiple to adjusted EPS less 18.1%, compensation's share of second-quarter adjusted EPS. Every figure in this note was recomputed in code from these inputs.
| Input | Value | Source |
|---|---|---|
| Price | $633.91 | Close, October 2, 2026 |
| Shares outstanding | 1,632,475,042 | 10-Q cover, July 29, 2026 |
| Diluted shares | about 1.66 billion | Third-quarter guidance, second-quarter call |
| Customer warrants | up to 320 million | 160 million each to OpenAI and Meta at $0.01; 8-Ks and 10-Q |
| World Labs | about $8.2 billion | All-stock; 12.94 million shares at $633.91, illustrative; 8-K |
| 2025 adjusted EPS, revenue | $4.17; $34.6 billion | Fourth-quarter and full-year 2025 results |
| 2030 market | about $2 trillion | AMD, July 23, 2026 |
| Net margin | 30% | Assumption; about a 35% operating margin after a 13% tax rate |
| Horizon | 4.2464 years | October 2, 2026 to December 31, 2030 |
This note is research, not investment advice. It states what a market price assumes under labeled assumptions; it is not a recommendation to buy, sell or hold any security, and the decision belongs to the reader. As of October 3, 2026, Russ Rosenzweig, Vista's founder, owns shares of AMD. The note began as a research question Russ Rosenzweig put to the Vista Research Desk, which challenged his draft over several rounds; the model and every figure were then rebuilt and checked against AMD's filings and releases.