We value AMD at $245 a share. The report's $529.95 price (Oct 26, 2026) was 116% above fair value: Avoid.
Restated Oct 26, 2026 (method case-spread-2026-10-24); the report's original value was $245, Avoid.
- Five-year return
- -2.8%
- a year, at the report's price
What would change the view
- Data-centre AI revenue on a path to $180B or more in 2031 (for example 2027 above $45B with second gigawatts firm at two buyers): our value would reach the Avoid line ($432, 85% of the price) and the view becomes a Hold
- 2027 data-centre AI revenue below $20B, or OpenAI or Meta stop after their first gigawatt: the pessimistic path ($71)
- Data Center segment operating margin below 25% for two quarters in 2027 as Helios racks ship: AI margins are heading toward our pessimistic 12%, about -$40 a share at 22%
- Server processor revenue up less than 40% in 2027 (management says more than 70%): about -$38 a share at 10 points a year below our path
- AMD's data-centre AI revenue growing faster than NVIDIA's data-centre revenue for four straight quarters in 2027: share gains toward the Street's path, which gives $452-626 through our model
Catalysts
- Q3 2026 results after the close (date set by AMD, 2026-10-06), with the Q4 guide. We expect revenue $13.50B and a Q4 guide midpoint of $14.52B (preview); a quarter barely moves the value, the gigawatt timing does: each $1B of 2031 AI revenue is about $1.55 a share
- The MI450 and Helios ramp: AMD expects a step up in data-centre AI revenue in Q4 2026 and another in Q1 2027 (Q2 2026 call), and each warrant's first tranche vests on delivery of the buyer's first gigawatt (8-Ks 2025-10-06, 2026-02-24). A Q4 on schedule supports our 2027 AI revenue of $30B; a slip into late 2027 points to the pessimistic path
- Q4 2026 results (date not yet set by AMD; early February is its habit), with the first 2027 commentary on data-centre AI. Above $45B of AI revenue implied for 2027 would put AMD on our optimistic path; under $20B on the pessimistic
- Anthropic's first gigawatt, due in the first half of 2027 (Q2 2026 call); Meta's first gigawatt is binding (8-K 2026-02-24). A firm order for a second gigawatt from any of the three is the first sign the deals run beyond their floors
This text argues the original call. The current figures are at the top.
We keep AMD an Avoid at $245 a share, up $2.30 from $242.91, against $529.95. Anthropic filed its S-1 on 2026-10-26 and Meta its offering documents the same day. Together they show AMD paying for part of the AI demand it sells into. The filings change how AMD's cash is counted, not what AMD will sell, so the value barely moves.
What the filings show
Anthropic. Fact (S-1 p. 164): in July 2026 Anthropic signed a Hardware Supply Agreement with AMD for Helios racks built on MI450 GPUs. Under it, AMD committed to buy up to $5.0B of Anthropic's non-voting Class N stock. Anthropic can draw that money in steps after AMD's GPUs are delivered and deployed, the first step expected in 2027. After the listing, each purchase is priced at the 30-day volume-weighted average price of Anthropic's Class A stock. Fact (S-1 p. 172): Anthropic committed to an initial tranche of capacity in 2027, bought from approved rack makers or rented through service providers. The tranche is subject to conditions on deployment proposals, is non-cancelable by Anthropic except for AMD's uncured breach, and has no fixed dollar amount. If fully procured, Anthropic expects to spend more than $20.0B over the term, plus an option for about as much again. These amounts are not in Anthropic's commitments table.
AMD's own disclosure. Fact (10-Q for the quarter to 2026-06-27, filed 2026-08-04): after the quarter AMD "entered into investment commitments of up to $5.0 billion, subject to certain conditions, which are expected to be made through fiscal year 2028". It did not name the counterparty, and AMD filed no 8-K for the agreement (EDGAR filing index, read 2026-10-26). Inference: the S-1 identifies the commitment; amount, conditions and timing all match.
Meta. Fact (Meta free-writing prospectus, 2026-10-26): NVIDIA, AMD and Broadcom, "severally and not jointly", indicated interest in buying up to $5.0B of Meta Class A stock in total in Meta's $22.5B common offering, at the offering price. The indications are not binding. AMD's share is not stated.
What changes in our value, and what does not
| Step | Bear | Central | Bull |
|---|---|---|---|
| 2026-10-22 report, restated under the current method (case-spread-2026-10-24) | $56.26 | $242.91 | $522.58 |
| Valuation date moves to 2026-10-26 | +$0.07 | +$0.32 | +$0.69 |
| The $5.0B buys an Anthropic stake, held at exit | +$0.62 | +$1.45 | +$1.37 |
| The stake leaves operating capital, so it no longer lowers the return on new capital | +$0.00 | +$0.53 | +$0.61 |
| This revision | $56.96 | $245.21 | $525.25 |
Source: bridge.py, bridge.txt. The bear and bull shown in the 2026-10-22 report ($71, $415) were restated to $56 and $523 on 2026-10-23, when the shared method set the cases' exit multiples at the 25th and 75th percentiles of history. That was a method change, not evidence about AMD.
The stake. Our 2026-10-22 model already paid the 10-Q's $5.0B, $2.5B in each of 2027 and 2028, but counted nothing in return. That was right for an unknown commitment, but wrong for stock bought at market prices. It also counted the 2028 half as operating capital, which lowered the return on new capital our fade starts from. The stake now sits outside operating capital (+$0.53, found by the outside eval), and we hold it at exit (our assumption):
- Central and optimistic: at its $5.0B cost. That is a zero return over three to four years, below what the stock should be expected to earn. We use it because the stake is non-voting, can be sold only by converting to Class A on sale to an unrelated buyer (S-1 p. i, p. 118), and rises and falls with AMD's own AI revenue.
- Pessimistic: at 40% of cost. This is the wrong-way world. A slower-growing Anthropic draws the full $5.0B because it needs the cash most, its shares fall with its growth, and its AMD orders shrink at the same time. 40% is our judgment; at zero the pessimistic value is $0.6 lower.
Meta. We count none of it: the indication is non-binding and AMD's share is unknown. If AMD bought a third of the $5.0B at the offering price and the shares were worth the same at exit, the value would fall $0.42; if it bought all $5.0B, $1.26 (output.txt). Inference: buying at the public offering price is close to value-neutral. The cost is the cash it ties up, in a customer whose AI spending already drives AMD's warrant deal.
AI revenue: no change. The S-1 makes Anthropic's first tranche firmer than our 2026-10-22 report said ("planned, not binding"). Our central case already sells about 21 gigawatts of AI systems over 2026-31, more than all 14 signed gigawatts in full. The pessimistic case's $42B of AI revenue in 2027-28 already covers three first gigawatts at $12B each. Inference: Anthropic's "more than $20.0B" for its first tranche is its whole outlay, including memory, networking, power and the rack makers' and service providers' margins. It is not AMD's revenue. If the tranche is one gigawatt, our $12B of AMD revenue per gigawatt fits inside it. We keep $12B.
The loop: how much of AMD's growth it finances itself
AMD now finances, in some form, every customer behind its signed gigawatt deals (fact, except where marked):
| Customer | Signed | What AMD gives | Firm |
|---|---|---|---|
| OpenAI | up to 6 GW | warrants for up to 160M AMD shares, vesting as it buys and as AMD's price clears hurdles (8-K 2025-10-06) | first gigawatt |
| Meta | up to 6 GW | warrants for up to 160M AMD shares on the same pattern (8-K 2026-02-24); part of a $5.0B non-binding indication for Meta stock, shared with NVIDIA and Broadcom (FWP 2026-10-26) | first gigawatt |
| Anthropic | up to 2 GW (Q2 2026 call, management claim) | up to $5.0B of Anthropic stock bought at market as GPUs are delivered (S-1 p. 164) | initial 2027 tranche, subject to deployment proposals (S-1 p. 172) |
Inference (output.txt): at $12B a gigawatt the 14 signed gigawatts are $168B. That is at most 71% of our central 2027-31 AI revenue ($237B) and 32% of AMD's total 2027-31 revenue ($529B). It is an upper bound, since only the first tranches are firm. Anthropic alone is about $24B, 10% of our AI revenue, and the $5.0B stake is about a fifth of it.
Credit exposure if Anthropic's growth slips. AMD lends Anthropic nothing and holds no collateral under the S-1's terms. Unlike Broadcom, it is a shareholder, not a creditor. Its exposure is three things that fall together:
- the stake, at most $5.0B (about $2.7 a share at exit);
- its Anthropic revenue;
- its guarantees of partners' data-centre leases ($4.1B maximum, partners not named) and its own data-centre leases ($9.5B plus $4.5B), all in the 10-Q.
The guarantees and leases were already charged in the 2026-10-22 model. Losing Anthropic entirely, with AI revenue 10% lower every year (its about $24B of our $237B), no stake bought and OpenAI's and Meta's warrants as in the central case, takes about $11 off the central value, to $234 (model result, output.txt). Inference: the Anthropic loop alone is a small risk to AMD's value. A slip across OpenAI and Meta as well is what moves toward the pessimistic case.
How AMD will book it (inference; AMD has not said). AMD's 10-Q carries marketable equity at fair value through other income (expense), below operating income. Class N is not listed, but it converts to listed Class A on sale. We expect AMD to mark it to Anthropic's share price, so its swings reach GAAP earnings per share but not segment profit or our valuation's operating line. The Q3 10-Q, due with results on 2026-11-03, may say.
The view
At $529.95 the price is 2.2 times our value. To justify it, AMD's data-centre AI revenue would have to reach about $236B in 2031 (ours $60B, our optimistic case $104B). None of this week's filings moves that gap. They show that part of the demand behind the gigawatt deals is paid with AMD's own stock and cash. That makes the deals' second tranches harder to read as independent demand. In 13% of our simulated values AMD is worth the price (odds.json).
Market reconciliation
The price, $529.95, sits above our optimistic case ($525). The 2026-10-22 report's reconciliation and its section 4 still hold, because no operating driver moved. The Street's revenue and EPS by year (FMP, read 2026-10-26: 2030 revenue $232.4B from 29 analysts, EPS $43.74 from 9) are unchanged from the figures that report ran through the model. Run through our model, they give $452-796 a share. The gap is AI revenue: the Street's path implies about $169B in 2030, against our $55B. The market's world, with the signed gigawatts delivered in full and AI revenue near $168B in 2031 at a 40% margin, carries a 10% weight in odds.json and justifies the price in 47% of its draws. This week's filings add no evidence for that world. They show that some of the demand behind it is paid for with AMD's own stock and cash.
Next results (Q3 2026, 2026-11-03 after the close)
The Q3 forecasts committed in the 2026-10-22 record stay the graded set. This revision adds none. Three things to read on the day:
- whether the 10-Q names Anthropic as the $5.0B investment commitment;
- how AMD plans to measure the stake;
- whether the Q4 guide still has the step up in data-centre AI from Helios that the Q2 call promised.
What would change the call
The falsifiers in the 2026-10-22 report stand. One addition: if Anthropic orders the second tranche of about the same size, AMD's AI revenue path moves toward our optimistic case ($104B in 2031, a value near $525). The price would still sit above it.
Revision history
| Report | View | Pessimistic | Central | Optimistic | Central change | Price | Blinding |
|---|---|---|---|---|---|---|---|
| 2026-10-22-amd | avoid | 71 (restated 56) | 243 | 415 (restated 523) | first | 507.77 | partly unblinded |
| 2026-10-26-amd | avoid | 57 | 245 | 525 | +2 | 529.95 | unblinded |
Sources
- Anthropic, PBC, Form S-1, 2026-10-26, p. i, 118, 164, 172 (data/anthropic-s1-paged.txt; the fund's read: 2026-10-26-anthropic-s1).
- Meta Platforms free-writing prospectus, 2026-10-26 (data/meta-fwp-2026-10-26.htm).
- AMD 10-Q Q2 2026 (filed 2026-08-04); 8-Ks 2025-10-06, 2026-02-24; Q2 2026 call (data/, carried from the 2026-10-22 report); EDGAR 8-K index read 2026-10-26.
- Robinhood quote, 2026-10-26 12:38:50 ET.
- The 2026-10-22 report: report, restatement.
This work was done with Claude, a model made by Anthropic, a counterparty in the agreements discussed: a conflict of interest.
Details
- Report date
- Oct 26, 2026
- Price source
- Robinhood last trade 12:38:50 ET 2026-10-26 (regular session; official close 2026-10-23 $516.79), Oct 26, 2026, 9:38 AM PDT
- Bear / bull
- $56.96 / $525
- Exit price
- $458 (on Dec 27, 2031)
- Confidence
- 23%
- Price against our scenarios
- The price is justified in under 20% of our scenarios4,000 simulated outcomes, each drawing the uncertain drivers from ranges we chose: our judgement, not a measured chance
- Worlds
- 4 stories about the businessEach outcome first picks one of these stories by its weight, then draws that story's own driver ranges; the answer above is their mix.
- World: base (45%)
- the base drivers. The price is justified in under 20% of its scenarios; middle outcome $239, 10th to 90th $114 to $492Weight: Our judgment, matching the central case's 50% weight. Scaled by 0.90 on 2026-10-23 to make room for the market world.
- World: bear_lean (22%)
- the gigawatt deals stall after their first tranche and AI spending digests in 2029. The price is justified in under 20% of its scenarios; middle outcome $152, 10th to 90th $75.59 to $306Weight: Our judgment, matching the pessimistic case's 25% weight. Scaled by 0.90 on 2026-10-23 to make room for the market world.
- World: bull_lean (22%)
- the signed gigawatts arrive in full and AMD wins other buyers. The price is justified in under 20% of its scenarios; middle outcome $330, 10th to 90th $165 to $656Weight: Our judgment, matching the optimistic case's 25% weight. Scaled by 0.90 on 2026-10-23 to make room for the market world.
- World: market (10%)
- the market's path: the signed gigawatt deals arrive in full and AI revenue reaches about $168B in 2031 at a 40% margin. The price is justified in 20-50% of its scenarios; middle outcome $509, 10th to 90th $252 to $1,022Weight: Our weight, 10%, argued from evidence and labelled unblinded (2026-10-23): the Street's 2030 revenue is $232B (29 analysts, FMP 2026-10-23) against our bull's $167B, and it is the only support for this path. The signed frameworks (up to 14 GW with OpenAI, Meta and Anthropic, only the first gigawatts firm) carry $42-56B a year on the report's own arithmetic (14 GW at $12B each over three or four years), which is the bull-lean world, not $168B. The base rate puts the path beyond the 90th percentile of five-year growth for $20B+ companies (19.9% a year, odds_reference.json), so it gets under the bull's weight.
- What moves the answer
- the exit price (market outcome) 57%; data-centre AI revenue (Instinct GPUs and Helios racks) in 2031, US$ billions; the 2028-2030 path scales with it from 2027's $30B 29%; data-centre AI operating margin in 2031 (segment basis, before stock pay and amortization); 2028-2030 shift by the same points 23%; server CPU revenue growth 2027-2031, points a year added to the central path (+60%, +18%, +10%, +7%, +6%) 16%How much of the spread in value each driver explains on its own; drivers that move together share credit, so the shares need not add to 100%.
- Leaning cases
- 3 other views we ranThe same simulation with some ranges moved to another view, to show how much the answer depends on our ranges.
- Case: bear_lean
- the gigawatt deals stall after their first tranche and AI spending digests in 2029. The price is justified in under 20% of this case's scenarios; middle outcome $152, 10th to 90th $75.59 to $306
- Case: bull_lean
- the signed gigawatts arrive in full and AMD wins other buyers. The price is justified in under 20% of this case's scenarios; middle outcome $330, 10th to 90th $165 to $656
- Case: market
- the market's path: the signed gigawatt deals arrive in full and AI revenue reaches about $168B in 2031 at a 40% margin. The price is justified in 20-50% of this case's scenarios; middle outcome $509, 10th to 90th $252 to $1,022
- Cost of equity (original)
- 12.9%
- Original figures
- fair value $245 (bear $57, bull $525), exit $458; Avoid
- Street
- average $613, median $641, low $260, high $1,250; 17 targets in the last quarter; FMP price-target-consensus, Oct 26, 2026, 9:52 AM PDT
- Restated
- Oct 26, 2026, method case-spread-2026-10-24; forecast.json valued by method.py at market inputs 2026-10-01
- Cost of equity calibration
- +0.0 points
- Calibration source
- none: the 1-point cut set 2026-10-16 to pull the batch toward the Street was removed on 2026-10-18 (Dong); the cost of equity is risk-free plus beta times the premium
- View rule
- strong buy when the price is at or below 0.80x fair value; buy when at or below 0.90x fair value; avoid when fair value is below 0.85x the price; otherwise hold
- Blind step
- Unblinded; the model was committed in b7fb063 before the record
- Method
- As of Oct 26, 2026, built without: Mature return holds own capital
Original call, Oct 26, 2026 (before the Oct 26, 2026 restatement)
We keep AMD an Avoid at a fair value of $245 (was $243), against $529.95. Anthropic's S-1 names the $5.0B investment commitment AMD's 10-Q left unnamed: AMD will buy Anthropic Class N stock at the market's 30-day average price as its Helios racks are delivered, and Anthropic's first tranche (over $20B of spend if fully procured) is non-cancelable by Anthropic. Our earlier model already paid the $5.0B and counted nothing for it; holding the stake at cost adds $1.45 a share (and $0.53 more once it leaves operating capital), and $0.6 in the pessimistic case, where Anthropic draws it all and its shares fall with AMD's Anthropic revenue. AMD, NVIDIA and Broadcom also indicated, without commitment, up to $5.0B in aggregate for Meta's stock offering; all $5.0B by AMD would cost $1.26 a share. AMD now finances all three customers behind its signed gigawatt deals (warrants for OpenAI and Meta, stock for Anthropic), which are at most 71% of our 2027-31 AI revenue. No operating driver moves; the price still needs about four times our 2031 AI revenue.
As of 2026-10-26 · unblinded: a revision of the 2026-10-22 report, which read the price after its blind commit; this one changes only inputs dated by the 2026-10-26 filings. Model committed before this report (b7fb063). The 2026-10-22 report ($243, Avoid) stays on the record.
- Fair value today
- $245 per diluted share, at a 12.91% cost of equity. Weighting the pessimistic, central and optimistic cases 25/50/25 (our weights) gives $268
- Price
- $529.95, Robinhood last trade 2026-10-26 12:38:50 ET (official close 2026-10-23 $516.79) → 117% above fair value
- Five-year exit price
- $458 on 2031-12-27: the average of a long fade of returns toward the cost of capital ($271) and an outside view at the peers' 29.9 times operating profit ($645). Fading over 10 or 20 years gives $452 or $463
- Five-year return
- -2.8% a year from $529.95 to the central exit (no dividend)
- Bear / bull today
- $57 / $525: the pessimistic and optimistic cases, each one coherent world
- Confidence
- About 23%: the share of our simulated values within 20% of $245, interpolated from odds.json's percentiles (subjective, not calibrated)
- Moat
- One of two companies selling complete AI data-centre systems at scale, and the faster-growing x86 server-chip maker; no moat yet against NVIDIA in AI or the buyers' own chips
- Market expects
- Data-centre AI revenue of about $236B in 2031 (ours $60B), everything else as in our central case
- Price needs
- No single driver within our ranges: an AI margin of 55% gives $338; server growth 20 points a year faster gives $368 (odds.json)
- Odds
- P(price justified): 13%, the share of our simulated values at or above the price; eight in ten fall between $105 and $573. Subjective, not calibrated (odds.json)
- Key driver
- Data-centre AI revenue in 2031: each $1B is about $1.55 a share. The Anthropic stake this revision adds is worth $2
- Street
- Average 12-month target $613 (median $641, range $260-1,250); 3 targets in the last month averaging $675 (FMP, read 2026-10-26). An expectations reference, not validation