Advanced Micro Devices AMD

> EARNINGS TODAY — 2026-08-04. AMD reports Q2'26 (quarter ended approximately 2026-06-27) on the date of this dive, after the 16:00 ET close this dive is struck against. Vendor consensus is non-GAAP EPS $1.61 and revenue $11.310B; epsActual and revenueActual are null because the number does not exist yet. The stock closed +7.00% at $518.58 into the release. Anyone reading this is reading it after the result is public and this dive is not. That timing is not a footnote to the verdict — it is a primary input to it, and it is a large part of why the verdict is Watch rather than Buy. This document deliberately contains no estimate of what the print will say.

Reference table

1. What the business is — and what the vendor got wrong about it

AMD designs and sells processors and accelerators. It is fabless: it owns no leading-edge manufacturing. The profile records 31,000 employees, chief executive Lisa T. Su, NASDAQ listing, and an IPO date of 1972-09-07.

The 10-K filed 2026-02-04 is explicit that there are three reportable segments, restructured in Q1'25:

> "Beginning in the first quarter of fiscal year 2025, we combined the Client and Gaming segments into one reportable segment to align with how we manage our business. All prior period segment data were retrospectively adjusted."

And, critically, on the sub-lines:

> "Each of the Client and Gaming businesses do not qualify as a reportable operating segment, however, the Company continues to separately disclose revenue for each business."

That single sentence is the origin of the vendor's segment defect, and it is worth stating in full before anything else.

The correction. The vendor reports four FY2025 segments; the 10-K shows three; we use the 10-K. seg_prod lists "Client and Gaming" $14.550B, "Data Center" $16.635B, "Embedded" $3.454B and "Gaming" $3.910B, which sums to $38.549B against reported FY2025 revenue of $34.639B — an overstatement of exactly $3.910B. Gaming is not a fourth segment; it is a separately disclosed sub-line inside Client and Gaming, and the vendor has counted it twice. The 10-K settles it precisely: "Client and Gaming net revenue of $14.6 billion in 2025 increased by 51%… Client net revenue of $10.6 billion… Gaming net revenue of $3.9 billion."

FY2025 segment structure as filed:

The All Other line is the second-most useful disclosure in the filing: "All Other operating loss of $4.0 billion in 2025 primarily consisted of $2.3 billion of amortization of acquisition-related intangibles and $1.6 billion of stock-based compensation expense." That is the entire GAAP-to-non-GAAP bridge, quantified, in the company's own segment note.

Q1'26 (quarter ended 2026-03-28), from the 10-Q: Data Center $5.8B, +57%, segment operating income $1.6B (27.6% margin); Client and Gaming $3.6B, +23% (Client $2.9B +26%, Gaming $720M +11%), operating income $575M (16.0%); Embedded $873M, +6%, operating income $338M (38.7%); All Other −$1.0B ($551M amortisation, $487M stock compensation). Gross margin 53%. Consolidated operating income $1.476B.

The FY2023/FY2024 rows use the old four-segment scheme ("Client" separate from "Gaming") and both sum correctly to reported revenue — $25.785B and $22.680B respectively. Only the FY2025 row is defective, because that is the year the presentation changed and the vendor's extractor picked up the retained Gaming sub-line as though it were a peer of its own parent.

Geography (FY2025, billing location of the customer — not end demand): United States $11.363B (32.8%), China $7.751B (22.4%), Taiwan $5.186B (15.0%), Singapore $4.284B (12.4%), Other $6.055B (17.5%). This sums to $34.639B and is internally consistent, but it is not comparable to FY2024, where separate Europe ($1.625B) and Japan ($1.767B) lines existed and "Other Countries" was $554M. Those lines vanish in FY2025 and Other jumps eleven-fold. The 10-K states international sales were 67% of revenue in 2025 and 66% in 2024; the Q1'26 10-Q puts it at 74%, up from 66%.

2. The single most important thing about this company's numbers — 320 million invisible shares

There are 320 million shares of AMD that exist as a contractual right, are 19.4% of the diluted share count, and appear in no consensus estimate, no quoted market capitalisation, and nowhere in the vendor payload. They appear only in the 10-Q filed 2026-05-06:

> "In October 2025 and February 2026, the Company issued warrants to [two counterparties]… Each warrant provides the holder the right to purchase up to an aggregate of 160 million shares of the Company's common stock at an exercise price of $0.01 per share. The warrants vest in tranches based on AMD Instinct GPU purchase milestones achieved by [the counterparties], their affiliates, or indirectly through authorized third parties, and achievement of specified Company stock price targets… As of March 28, 2026… none of the warrant shares had vested or become exercisable."

The 10-K and the 10-Q together record the commercial context: an October 2025 product purchase agreement to deploy 6 gigawatts of AMD GPUs, first gigawatt on the MI450 series; and a February 2026 amended master purchase agreement for a further 6 gigawatts, first gigawatt on custom MI450-based GPUs plus 6th-generation EPYC CPUs. The warrants were issued concurrent with those agreements. They are, in plain terms, the price paid for the two largest commercial commitments in the company's history.

Why this dominates the valuation, in arithmetic:

The structural point, which is subtler than the arithmetic and more important. The warrants vest against accelerator purchase milestones and specified stock-price targets. They therefore do not dilute in the bear case — if the deployments do not happen and the stock does not rise, the warrants expire worthless and the reported share count is correct. They dilute only in the bull case. This means the consensus estimate series and the share count on which it is divided are internally inconsistent: the estimates describe a world in which the denominator they use is wrong. Every forward multiple quoted on this stock, by anyone, is understated by roughly 13-17% in the scenario that justifies the multiple.

We use the fully diluted basis as the primary valuation frame in Section 6 and report the undiluted figure alongside it. Neither is more "correct"; only one is honest about what the buyer receives if the thesis works.

Two secondary basis notes belong here.

Estimates are non-GAAP; reported EPS is GAAP; the gap is purchase accounting. FY2025 epsAvg was $3.9621 against reported GAAP diluted EPS of $2.65 — a gap of $1.31 per share, or roughly $2.14B. The 10-K's All Other note gives the reason directly: $2.3B of amortisation of acquisition-related intangibles, which at the ~13-15% effective tax rate the filings disclose is about $2.0B, or $1.22 per share. Notably, the $1.6B of stock-based compensation is evidently not added back in the FY2025 bridge — AMD's non-GAAP measure expenses it, which is to the company's credit. The FY2024 bridge does not reconcile the same way: epsAvg $3.3149 against GAAP $1.00, a gap of $3.78B against $2.4B of amortisation, implying stock compensation and restructuring were also excluded that year. The non-GAAP definition is therefore not stable across the estimate series, which is a reason to treat the multi-year EPS path with care (Section 8).

Q2'25 was not a broken quarter; it was two offsetting one-offs. The quarter ended 2025-06-28 printed gross margin of 39.8% and an operating loss of −$134M, against roughly 50% margins in the quarters either side — and yet reported net income of +$872M. Both anomalies are explained in the filings:

• The 10-K: "During the second quarter of fiscal year 2025, the Company recorded approximately $800 million of inventory and related charges on AMD Instinct MI308 Data Center GPU products due to new U.S. export restrictions on certain semiconductors to China… During the fourth quarter of fiscal year 2025, we began shipping products and reversed approximately $360 million." Adding $800M back to Q2'25 gross profit of $3.059B gives $3.859B on revenue of $7.685B — 50.2% gross margin, exactly in line with Q1'25's 50.2%. The operating loss becomes +$666M.
• The 10-Q filed 2025-11-05 supplies the other half: for the nine months to 2025-09-27 the company recorded an income tax benefit of $558M, "including a discrete tax benefit of $853 million related to the release of uncertain tax positions pertaining to the reasonable cause relief for dual consolidated losses approved by the Internal Revenue Service in April 2025." With Q1'25 and Q3'25 tax provisions of $123M and $153M, the implied Q2'25 tax benefit is approximately $834M — which is what turns a −$134M operating loss into +$872M of net income.

So Q2'25 carried an $800M charge above the line and an ~$834M benefit below it. Neither is recurring. Neither should be extrapolated. The vendor file shows both and explains neither. Note also the 10-K's disclosure that U.S. officials "have expressed an expectation that the U.S. government will receive 15% of the revenue generated from licensed MI308 sales to China; however, to date, the U.S. government has not published a regulation establishing such requirement." An unlegislated 15% revenue share on China accelerator sales is an unquantifiable but live margin risk.

3. Growth and margin trajectory

Two structural facts the table makes plain. First, AMD has had a down year recently — FY2023 revenue fell 3.9% and operating income collapsed to $401M. This is a cyclical semiconductor business and a 202% twelve-month return does not repeal that. Second, the FY2022-FY2024 operating-margin trough is purchase accounting, not operational failure: the Xilinx acquisition put $16.7-21.4B of amortisable intangibles on the balance sheet, and the resulting $2.3-2.4B annual amortisation charge sits in All Other. Gross margin never fell below 44.9% through that period.

The quarterly sequence is the live picture:

The one uncomfortable number in an otherwise excellent series: Q1'26 revenue of $10.253B was FLAT — down 0.2% — against Q4'25's $10.270B. Year-on-year growth was 37.9%, but sequential growth stopped. The 10-K notes the business is seasonally second-half weighted, which partly explains it. But it means the FY2026 consensus of $50.217B requires the second half to average roughly $14.3B a quarter, given Q1'26 actual of $10.253B and Q2'26 consensus of $11.310B — a 39% step up from the guided Q2 level within six months. That is the single most falsifiable near-term claim embedded in the price, and it is why the earnings print landing today matters so much.

Gross margin direction is genuinely good. 44.9% (FY2022) to 52.8% (Q1'26), with Q4'25 at 54.3%. The 10-Q attributes it to "a favorable product mix, including higher Data Center segment revenue." Company guidance (half weight, company voice) put Q2'26 non-GAAP gross margin at ~56%. Set against that, the knowledge base's independent-analyst view is that AMD sells accelerators at roughly 50% gross margin versus its larger rival's 75% — so mix improvement is real but the accelerator business itself is not yet the margin engine the blended number implies.

One negative that deserves stating. Data Center segment operating margin fell from 27.8% in FY2024 ($3.5B on $12.579B) to 21.6% in FY2025 ($3.6B on $16.635B) — and even adding back the $440M net export charge only recovers it to 24.3%. Revenue grew 32% and segment operating income grew about 3%. AMD bought data-centre revenue with margin in FY2025. Q1'26's 27.6% suggests that reversed, and that is the number to watch.

4. Balance sheet and the enterprise-value rebuild

As at 2025-12-27 (vendor bal_a, the most recent full balance sheet in the payload):

• Cash and equivalents $5.539B · short-term investments $5.013B · total $10.552B (the 10-K says "$10.6 billion" — agrees)
• Inventory $7.920B (up from $5.734B) · total assets $76.926B · total liabilities $13.927B · total equity $62.999B
• Goodwill $25.126B + intangibles $16.705B = $41.831B, or 54.4% of total assets
• Vendor total debt $4.472B (short-term $874M + long-term $2.973B + capital leases $625M)

As at 2026-03-28 (10-Q, which supersedes the above and is not in the payload):

• Cash, cash equivalents and short-term investments $12.3B (up from $10.6B)
• Aggregate principal short- and long-term debt obligations $3.3B — unchanged from year-end
• Corrected net cash position: ~$9.0B

Data-integrity correction 1 — net debt. The vendor reports netDebt of −$1.067B; the 10-Q shows a net cash position of roughly $9.0B; we use the filing. The vendor's figure nets total debt of $4.472B against cash and equivalents of $5.539B only, omitting the entire $5.013B of short-term investments. This is the textbook version of the defect. On the payload's own year-end figures the corrected net cash is $6.080B; on the filing's Q1'26 figures it is ~$9.0B.

Data-integrity correction 2 — total debt. The vendor reports total debt of $4.472B; both the 10-K and the 10-Q state aggregate principal short- and long-term debt of $3.3B; we use $3.3B. The $1.17B difference is lease obligations, and at least part of it appears to be counted twice: the payload carries capitalLeaseObligations of $625M and a combined short-plus-long-term debt figure of $3.847B that already exceeds the filing's $3.3B principal by $547M. The disclosed borrowings are the 3.924% Senior Notes due 2032, 4.393% Senior Notes due 2052, and the 4.212% and 4.319% Notes issued March 2025 in aggregate principal of $1.5B. The correction does not change the conclusion — this is a net-cash company either way — but the reported leverage ratios are overstated.

The enterprise-value error here is real but immaterial, and it is worth saying so. Corrected EV = $845.596B market cap − $9.0B net cash = ~$836.6B, against the vendor's $843.9B — an overstatement of about 0.9%. On a company this size the short-term-investments omission barely moves the multiple (EV/sales 22.3x corrected against 22.5x reported). We flag it because it is the same defect that matters enormously elsewhere, not because it changes anything about AMD.

Three obligations the balance sheet does not show, and the vendor cannot:

1. Unconditional purchase commitments of approximately $25.7 billion at 2026-03-28, "of which $18.3 billion are for the remainder of fiscal year 2026" — wafers, substrates, components, multi-year cloud service arrangements and technology licences. Against FY2025 cost of sales of roughly $17.5B, this is a very large forward commitment and the strongest hard evidence in the filings that the FY2026 ramp is being physically procured. It is also the largest single fixed-cost risk if demand disappoints.

2. $4.4 billion of data-centre and other real-estate leases signed but not yet commenced, with 7-to-10-year terms, "expected to commence beginning in the second quarter of fiscal year 2026." Entirely absent from the reported lease liability at Q1'26.

3. The 320 million penny warrants (Section 2).

And one item that is newer than the 10-Q. The 8-K dated 2026-05-13/14 discloses a new five-year $5.0 billion unsecured revolving credit facility replacing the prior $3.0 billion facility (no borrowings outstanding at closing), and an increase in the commercial paper programme from $3.0 billion to $5.5 billion. Committed undrawn liquidity therefore went from $6.0B to $10.5B in May 2026. The 10-Q's liquidity disclosure is superseded by this 8-K; we use the 8-K. A company that has just more than doubled its short-term borrowing headroom while sitting on $12.3B of cash is preparing to spend at a scale the current balance sheet does not reflect.

Returns, read carefully. Return on equity is 8.08% and return on invested capital 6.22% — both unimpressive, and both depressed by the $41.8B of goodwill and intangibles inflating the denominator. Return on tangible assets is 13.1%. Days of inventory outstanding is 157.7, with inventory up 38% year-on-year to $7.920B; in a ramp that is normal, in a stall it is the problem.

5. Cash flow and capital returns

Free cash flow went from $1.121B (FY2023) to $6.735B (FY2025), and Q1'26 alone produced roughly $2.6B. Capital intensity is low — capex is 3.1% of revenue — which is the structural advantage of the fabless model and the reason a 43% revenue compound rate is financeable without equity issuance. Free cash flow yield on the FY2025 figure is 0.80% at today's market capitalisation; the vendor's 1.01% is on a trailing basis. Either way, you are not being paid a cash yield here. You are buying growth.

The buyback does not offset dilution and should not be described as a return of capital. FY2025 repurchases were $1.316B (12.4 million shares per the 10-K) against $1.638B of stock-based compensation — a $322M shortfall. Diluted shares went 1.625B (FY2023) → 1.637B (FY2024) → 1.636B (FY2025) → 1.650B (Q1'26): the count is drifting up, not down, despite four consecutive years of repurchases. $9.2B remained authorised at 2026-03-28. Q1'26 repurchases were only $221M, an annualised pace of about 0.1% of market capitalisation.

No dividend. ratios_ttm.dividendYieldTTM and dividendPerShareTTM are both 0, and the cash-flow statement shows no dividends in any year. The profile's lastDividend of $0.005 is rejected as an artefact (Section 8).

Income quality is 1.97x (operating cash flow to net income), which is high and healthy — it reflects the $3.0B of depreciation and amortisation and $1.6B of stock compensation that sit between GAAP earnings and cash.

6. Valuation — priced in or room?

At $518.58 (market cap $845.60B reported, ~$1,021.6B fully warrant-diluted):

The FY2027 dispersion is the story. Twenty-nine analysts on EPS between $8.58 and $19.14 — the high is 2.23x the low. Thirty-six on revenue between $60.992B and $99.952B. At $518.58 the stock is 60.4x the low estimate and 27.1x the high on reported shares; 69.6x and 31.2x fully diluted. This is not a stock with a price. It is a stock with a bet on where in a 2.2x earnings range reality lands, and today's price sits roughly two-thirds of the way toward the optimistic end.

Coverage thins fast beyond FY2027: 30 analysts on FY2026 EPS, 29 on FY2027, 16 on FY2028, 10 on FY2029. No conclusion here rests on the FY2029 or FY2030 rows.

Peer context. The vendor peer set spans wafer-fab-equipment names (Applied Materials $434.0B, Lam $397.4B, KLA $255.3B, ASML $659.8B), memory (Micron $1.008T), IP (Arm $299.6B), and several names that are not comparables at all (SAP, Palantir). No peer multiples are supplied, so no peer-multiple comparison is drawn. Within this batch the useful cross-reference is that AMD, Intel and Marvell are three different bets on the same buildout — AMD as the credible second accelerator source at 37.4x FY27E, Intel as a turnaround with an unexpected CPU-shortage tailwind, Marvell as custom silicon — and the four equipment names are effectively one trade with four expressions. Holding AMD alongside several of them is one position, not diversification.

6a. What today's price assumes (the inversion)

At $518.58 — 37.4x FY2027 consensus non-GAAP EPS on reported shares, 43.1x fully diluted — the price embeds roughly the following falsifiable claims:

• Second-half FY2026 revenue averages about $14.3 billion a quarter. (Consensus-derived: FY26E $50.217B less Q1'26 actual $10.253B less Q2'26 consensus $11.310B leaves $28.654B across two quarters.) That is a 39% step up from the guided Q2 level within six months, against a Q1'26 that was flat sequentially. This is the nearest-term and most immediately testable assumption, and the print landing today is its first examination.
• FY2027 revenue reaches roughly $79.8 billion — a further 59% on top of a year that itself grows 45%. (Consensus-derived, 36 analysts, range $61.0–100.0B.) For scale: FY2027 consensus revenue is 2.3x FY2025 actual, achieved in two years.
• Gross margin holds in the 52-56% band. (Q1'26 actual 52.8%; Q4'25 54.3%; company guidance ~56% for Q2'26 — company voice, half weight.) The knowledge base's independent view that accelerators carry roughly 50% gross margin means mix shift toward Data Center is margin-dilutive at the accelerator line even while it is accretive to the blend today.
• The market keeps paying a high-30s multiple on forward earnings. (Derived.) Moving from 37.4x FY27E to 26.6x FY28E at a constant price is mechanical rolldown, not de-rating — the price only holds if the market still pays roughly today's multiple at each future date.
• The 320 million penny warrants either do not vest, or the market does not care that they do. (Our number, from the 10-Q; absent from all consensus figures.) This is the most fragile assumption in the price, and it is fragile in an unusual way: it fails precisely when everything else succeeds.
• No further export-control action materially impairs the 22.4% of FY2025 revenue billed to China. (Our reading, from seg_geo and the 10-K.) The MI308 episode already cost $800M gross, $440M net, and the 10-K flags an unlegislated expectation of a 15% government revenue share on licensed China accelerator sales.

6b. The return bridge (why the multiple moves)

Expected return over the medium horizon decomposes as: EPS growth (very large — consensus FY2026 $7.51 to FY2028 $19.46, a 61% compound rate) + multiple drift (COMPRESSION assumed) + shareholder yield (negative — the buyback does not cover stock compensation and the warrants add up to 19.4% of the count).

Our base case assumes multiple COMPRESSION, and says so explicitly. The base fair value of $455 is 35x FY2027 consensus non-GAAP EPS of $13.873 on a share count adjusted for partial warrant vesting — a lower multiple than the 37.4x the market pays today, and applied to a smaller per-share number. The justification for compression is not pessimism about the business; it is three things the arithmetic forces: growth maturation (a 59% FY2027 revenue increase cannot be repeated indefinitely and consensus itself has it decelerating to 35% by FY2028), the mechanical arrival of dilution as the milestones are hit, and the fact that a 2.23x FY2027 estimate range is itself a risk premium the market must eventually price.

Almost all of the return available in this name is earnings growth, and none of it is yield. That is the correct shape for a company at this stage, but it means the entire investment case is a forecast, with no valuation floor, no dividend and no meaningful buyback beneath it. The one leg the thesis cannot afford is multiple expansion, and we do not assume it. The bull case at $700 assumes reality lands near the top of the analyst range and the multiple holds around 42x — both, not either. That is the fragile combination.

6c. Variant perception (where we differ, what would surprise)

• We differ from the entire street on the denominator. Fifty buy ratings, a $575.29 consensus target and every published forward multiple are computed on approximately 1.65-1.67 billion shares. The 10-Q discloses 320 million more at a one-cent strike, vesting on exactly the milestones the bull case requires. We have found no evidence in the payload that any consensus figure incorporates them. If we are right, every published forward multiple on this stock is understated by 13-17% in the success case. Watchable event: the next 10-Q's warrant note, which must disclose any tranche that has vested or become exercisable.
• We differ on how to read the FY2027 estimate range. The street quotes a $13.87 mean. We think a $8.58-to-$19.14 range across 29 analysts is not a consensus at all — it is a statement that the professional community cannot agree whether AMD's accelerator share will be marginal or transformative, and the mean is an artefact of averaging two incompatible views. The same is true of the price targets: high $1,250, low $260, a 4.8x spread. Watchable number: whether the FY2027 range narrows after the next two prints — narrowing in either direction is the actual re-rating event.
• We do NOT differ on the business quality, the growth, or the management. Revenue growth of 34.3% and 37.9%, gross margin from 44.9% to 52.8%, $6.735B of free cash flow, $9.0B of net cash, no customer above 10%, $25.7B of procurement commitments, and a knowledge base whose independent voices are unanimously admiring of the chief executive. We are not making a bear case about the company. We are making a bear case about the entry price on 2026-08-04.
• We differ on what today's 7% move means. Six other names in this batch moved 5-13% on the same day. With a beta of 2.489, AMD's move is sector beta into a print, not an AMD-specific repricing — and RSI at 48.4 with negative MACD after a 7% day says the stock had been falling for weeks and today merely regained its 50-day average. Watchable number: whether $514 holds.
• Positive surprise that would force a re-rate upward: a Q2'26 result with H2 guidance implying $14B+ quarterly revenue and gross margin at or above 56%; Data Center segment operating margin sustained above 27%; evidence that the open compute software stack has reached genuine parity, which the knowledge base's most credible technical sceptic set as his own falsifiable test; or a narrowing of the FY2027 estimate range toward the upper half.
• Negative surprise that would break the thesis: H2'26 guidance implying materially less than $14B a quarter; a first warrant tranche vesting while the stock is flat (dilution without the re-rating); further China export restrictions or formalisation of the 15% revenue-share expectation; inventory building past 157.7 days without corresponding revenue; or Data Center segment margin failing to hold its Q1'26 recovery.

Synthos fair values

All three anchors are multiples of FY2027 consensus non-GAAP EPS, the deepest forward line at 29 analysts, adjusted for warrant dilution appropriate to each scenario. Stated arithmetic, not a discounted cash flow.

• Bear ~$230 — 26.8x the LOW FY2027 estimate of $8.583, on the undiluted 1.65B share count (correctly, because in this scenario the milestones are not met and the warrants do not vest). The scenario: the second-half FY2026 step-up does not arrive, accelerator share stays in the low single digits as the knowledge base's most cited technical analyst expects, the estimate range collapses toward its floor, and a 2.489-beta stock de-rates hard. Cross-check: $230 is 4.6x FY2026E revenue and 1.5x the 52-week low of $151.14; it is also below the street's own low target of $260. −55.7%.
• Base ~$455 — 35x FY2027 consensus EPS of $13.873, adjusted to ~$13.00 for partial warrant vesting (roughly half the 320M shares by end-FY2027 as the first-gigawatt milestones are met). The scenario: the business delivers close to consensus, the multiple compresses modestly from 37.4x as growth matures, and shareholders absorb the dilution they paid for the deals. Cross-check: $455 is 23.4x FY2028E EPS fully diluted and 9.4x FY2027E revenue. −12.3%.
• Bull ~$700 — 42x FY2027E EPS at the consensus HIGH of $19.141, fully diluted to ~$16.62 (correctly, because at that outcome the warrants vest in full). The scenario: the second-half FY2026 acceleration lands, the 6-gigawatt deployments convert on schedule, gross margin holds at 56%, and the market pays a premium multiple for a genuine second source in a $500B accelerator market. Cross-check: $700 is a $1.379 trillion fully diluted market capitalisation. +35.0%.

Base is 12.3% BELOW spot. The asymmetry is unfavourable: −55.7% to bear against +35.0% to bull, roughly 1.6:1 AGAINST. That is not a statement that AMD is a bad company. It is a statement that at $518.58, hours before a print, after a 7% gap, with 19.4% of the equity invisible, the distribution is skewed the wrong way. Watch.

7. Knowledge base — 174 raw hits, 41 first-class, and a lane that likes the company and doubts the share gain

Search: AMD "Lisa Su" MI300 Instinct across 51,928 distilled claims. Raw hits: 174. Triage:

Sources contributing first-class claims: dylan_patel (8), no_priors (6), all_in (5), compound_and_friends (4), business_breakdowns (3), a DOE-supercomputer lane (3), bill_gurley (2), invest_like_the_best (2), gavin_baker (2), an interviewer lane (2), lex_fridman, lyn_alden, luke_gromen, and a guest lane (1 each).

The bullish structural lane — why AMD exists and why that is durable:

> 2025-10-30 · Gavin Baker · neutral · conviction 65 (text match, entities GOOGL/NVDA — colour)

> "AI chip battle is really Google's TPU (enabled by Broadcom, which Google can take away anytime) versus Nvidia; AMD is the second source everyone needs."

> 2026-07-08 · Gavin Baker · bullish · conviction 60 · entity match

> "Only Jensen (Nvidia), Hock Tan (Broadcom), and Lisa Su (AMD) reliably identify and retain exceptional engineers/executives; most other semiconductor companies fail to retain top talent."

> 2025-12-06 · all_in lane · bullish · conviction 80 · entity match

> "At the silicon layer the die is cast — winners are roughly Nvidia + AMD + Google plus inference chips; too early to pick model-layer winners."

> 2023-04-21 · business_breakdowns · bullish · conviction 78 · entity match

> "As a fabless firm on TSMC it captures full Moore's Law gains, taking share from stuck Intel in PC and especially data center."

> 2026-05-27 · Mohammed Abdelfattah · bullish · conviction 40 · entity match (channel: dylan_patel — attributed to the speaker)

> "AMD hardware is fundamentally quite strong; its FP6 data path shares hardware with FP4, letting their razor-FP4 technique run at FP4 memory footprint with FP6 accuracy — an edge Nvidia lacks. Software needs to catch up."

The sceptical lane — and it targets exactly what the price requires:

> 2026-02-05 · Dylan Patel · neutral · conviction 60 · entity match · TEST: "AMD stays in single-digit percentage GPU market share"

> "AMD is a credible second competitor that leapfrogs Nvidia on hardware at times but trails on software, and it will hold only single-digit percentage GPU market share."

> 2025-08-14 · no_priors lane · bearish · conviction 70 · entity match

> "AMD lags despite good hardware engineering—weak networking, poor software, and inability to do hardware-software codesign keep it from catching Nvidia."

> 2025-08-18 · Dylan Patel · bearish · conviction 58 · entity match

> "AMD sells GPUs at ~50% gross margin vs Nvidia's ~75%, and despite strong engineering still needs more silicon area and memory to match Nvidia's performance."

> 2026-01-05 · Gavin Baker · neutral · conviction 48 · entity match

> "AMD is an exceptional company making pretty good chips, but lacks some of the systems and software Nvidia has; still very hard to out-Nvidia Nvidia."

> 2025-09-22 · Guido · bearish · conviction 70 · entity match (channel: dylan_patel — attributed to the speaker)

> "AMD is screwed: your two arch-nemeses teaming up is the worst possible news; already struggling with weak software stack and limited traction, now a bigger problem."

The circularity and positioning lane — the market-structure objection:

> 2025-10-16 · all_in lane · bearish · conviction 55 · entity match

> "The Nvidia-[large AI laboratory] and AMD deals have a circular financing quality worth watching, though it could still be years from unwinding."

> 2025-11-04 · interviewer lane · neutral · conviction 38 · entity match

> "Overlapping circular deals — AMD trading 10% equity for a deal, the Nvidia deal, Azure credits booked as revenue — raise questions about the sustainability and durability of AI revenues."

Note carefully: that second claim, dated 2025-11-04, independently identifies the 10%-equity-for-a-deal structure — i.e. the warrants of Section 2 — as a market-structure concern, three months before the second warrant was issued. It is the only place in the entire knowledge base where the dilution question surfaces, and it comes from an interviewer's framing rather than from a analyst with a position.

> 2026-01-30 · lyn_alden · neutral · conviction 50 · entity match

> "Trimmed chip names Nvidia and AMD after the bullish thesis largely played out and stocks ran up — not bearish, but reducing."

> 2026-07-03 · Brian Belski · bearish · conviction 70 (text match — colour)

> "Hyperscaler capex has to slow down—earnings are powered by hyperscalers taking free cash flow to zero, and that spend (Micron/AMD's revenue) can't keep accelerating."

Two attribution notes, per house policy. First, where speaker and thinker differ we attribute to the speaker and name the channel separately — the Abdelfattah and Guido claims above sit on a channel slug that is not the speaker. Second, 13 entity-matched claims about AMD are attributed to Nvidia's chief executive — for example "Surprised AMD gave away 10% of the company to [a large AI laboratory] before building next-gen product; a weaker deal than Nvidia's coming from a weaker position" (2025-10-08, conviction 55). A direct competitor's chief executive commenting on a rival's deal terms is not independent evidence and is not used as such, though the observation happens to align with our Section 2 concern. Three further first-class claims sit on a channel slug that plainly does not match their content (DOE supercomputer facts under a pharmaceutical-executive slug), and are flagged as channel-attributed with an unknown speaker.

What the lane establishes. It is warm on the company and cold on the specific thing the price requires. Every independent voice that discusses management does so admiringly. Every independent voice that discusses competitive position identifies the same two gaps — software and networking — and the single most technically credible source states a falsifiable ceiling: single-digit percentage GPU share. A lane that says "excellent company, structurally necessary second source, will hold single-digit share" is a lane that supports owning AMD and does not support paying 37.4x FY2027 consensus for it. Net conviction: positive-moderate, and Moderate rather than Moderate-High because the volume is high but the direction is genuinely split on the load-bearing question.

8. Data integrity — what we rejected from the vendor file and why

Nine material defects. Listed, not silently corrected.

1. seg_prod FY2025 DOUBLE-COUNTS Gaming — the vendor reports four segments summing to $38.549B; the 10-K shows three summing to $34.639B; we use the 10-K. The payload lists "Client and Gaming" $14.550B, "Data Center" $16.635B, "Embedded" $3.454B and "Gaming" $3.910B. Reported FY2025 revenue is $34.639B. The overstatement is exactly the Gaming line. The 10-K states there are three reportable segments and that "Each of the Client and Gaming businesses do not qualify as a reportable operating segment, however, the Company continues to separately disclose revenue for each business." Gaming is a sub-line of its own parent. The FY2023 and FY2024 rows use the pre-restructuring four-segment scheme and both sum correctly; only FY2025 is defective.

2. netDebt of −$1.067B omits $5.013B of short-term investments — the vendor reports net cash of $1.067B; the filings show ~$9.0B; we use the filings. The vendor nets total debt against cash and equivalents only. Corrected on the payload's own year-end figures: $10.552B less $4.472B = $6.080B net cash. Corrected on the 10-Q's Q1'26 figures: $12.3B less $3.3B = ~$9.0B net cash. enterpriseValueTTM of $843.883B is correspondingly overstated by about 0.9% against a corrected ~$836.6B — real but immaterial at this market capitalisation, and we say so.

3. totalDebt of $4.472B cannot be reconciled to the filings' $3.3B of aggregate principal — we use $3.3B. Both the 10-K and the 10-Q state "aggregate principal short-term and long-term debt obligations were $3.3 billion." The payload's short-term ($874M) plus long-term ($2.973B) already exceeds that by $547M before the separate capitalLeaseObligations line of $625M is added. Lease obligations appear to be counted twice. Leverage ratios derived from $4.472B are overstated; the net-cash conclusion is unaffected.

4. est.ebitdaAvg and est.ebitAvg are SYNTHETIC, not analyst-sourced — both rows REJECTED across all eight years. Two proofs. First, internal impossibility: FY2026 ebitAvg is $5.118B against niAvg of $12.337B — net income cannot be 2.4x operating income for a taxpayer with $37M-a-quarter interest expense. Second, and conclusive: ebitAvg is 47.0% of ebitdaAvg in every single one of the eight years (2.372/5.050, 2.617/5.570, 3.475/7.397, 5.118/10.896, 8.132/17.312, 11.003/23.423, 14.902/31.725, 20.021/42.621). A fixed ratio to four significant figures across eight independent forecast years is a formula, not a forecast. No EBITDA or EBIT figure from the estimate block appears anywhere in this dive. evToEBITDATTM (104.4x) and enterpriseValueMultipleTTM (104.4x) are separately rejected as they run off the uncorrected enterprise value.

5. Estimates are NON-GAAP and the definition is not stable across years — basis stated, forward multiples labelled. FY2025 epsAvg $3.9621 against GAAP diluted $2.65 (gap $1.31, ≈ the $2.3B of acquisition-intangible amortisation the 10-K discloses, after tax). FY2024 epsAvg $3.3149 against GAAP $1.00 (gap $2.31, ≈ $3.78B, which requires both the $2.4B amortisation and the $1.4B of stock compensation to be excluded). The two years use different add-backs. All forward multiples in Section 6 are explicitly on the non-GAAP basis; the trailing 169.5x is explicitly GAAP.

6. The 320 million penny warrants are ABSENT from the payload entirely — sourced from the 10-Q. No field in any of the 18 keys reflects them: not quote.marketCap, not the diluted share counts in inc_q, not any est row. This is not a vendor error so much as a vendor blind spot, and it is the single largest analytical gap in the file (Section 2).

7. profile.lastDividend of $0.005 contradicts a zero dividend everywhere else — REJECTED as an artefact. ratios_ttm.dividendYieldTTM, dividendPerShareTTM and dividendPayoutRatioTTM are all 0, and cf_a shows no dividend paid in any of the four years. AMD pays no dividend.

8. ratios_ttm.effectiveTaxRateTTM of 0.24% and taxBurdenTTM of 1.013 are distorted by a one-off and REJECTED. A tax burden above 1.0 means net income exceeds pre-tax income — the artefact of the $853M discrete benefit from the IRS dual-consolidated-loss release disclosed in the 10-Q. The 10-Q states the actual Q1'26 effective rate was 14.8%, and company guidance (half weight) put the FY2026 non-GAAP rate at 13%. We use 14.8%.

9. tech 52-week high/low disagree with quote — we use the tech block and say so. tech.hi52 is $580.91 against quote.yearHigh $584.73 (a 0.65% gap); tech.lo52 is $151.14 against quote.yearLow $149.22. The tech block is documented as computed from six years of closing prices, while the quote figures are almost certainly intraday extremes. We use the closing-basis $580.91/$151.14 throughout, for consistency with the SMA, RSI, MACD and drawdown fields computed on the same series, and note the intraday figures here. The same choice applies to the moving averages: tech.sma50 $514.33 / sma200 $314.56 rather than quote.priceAvg50 $512.21 / priceAvg200 $311.82.

Also flagged, not defects:

• seg_geo FY2025 is not comparable to FY2024. Separate Europe ($1.625B) and Japan ($1.767B) lines exist through FY2024 and disappear in FY2025, while "Other Countries" jumps from $554M to $6.055B. The FY2025 rows sum correctly to reported revenue; the year-on-year geographic comparison does not hold. The 10-K also states these are billing locations, not end demand.
• The balance sheet in bal_a is one quarter stale (2025-12-27) relative to the income statement (2026-03-28) and two quarters stale relative to the 8-K liquidity disclosure of 2026-05-13. Section 4 uses the 10-Q and 8-K figures where they differ.
• The price-target set is extraordinarily dispersed — high $1,250, low $260, median $607.50, consensus $575.29. A 4.8x spread is not a consensus in any useful sense, and the median exceeding the mean shows the low targets pulling the average down. Reported, not used as an anchor.
• km_ttm.intangiblesToTotalAssetsTTM of 52.1% is slightly below our own arithmetic of 54.4% ($41.831B of goodwill plus intangibles over $76.926B of assets); the difference is immaterial and the vendor figure is not rejected.

9. Technicals — extended, and today does not change that

• Price $518.58, +7.00% on the day (+$33.94 from a $484.64 previous close), open $504.24, day range $502.20–$530.13, volume 37.2M.
• −10.7% from the 52-week closing high of $580.91; +243.1% above the 52-week closing low of $151.14. Maximum drawdown from peak over the trailing year is only −10.7% — this stock has barely corrected.
• Above both moving averages, but only just above the near one: +0.8% over the 50-day of $514.33, and +64.9% over the 200-day of $314.56. The gap between the two averages is the whole story of the year.
• RSI 48.4 — still BELOW neutral after a 7% day. MACD −7.60 — still negative. These two facts together are the most informative in the section.
• Relative performance: 3-month +51.8% vs SPY +7.6% and QQQ +7.7%; 6-month +110.6% vs SPY +11.1% and QQQ +15.7%; 12-month +202.0% vs SPY +24.3% and QQQ +30.8%. A 178-point excess return over the S&P in twelve months.
• Beta 2.489 — the highest in this batch, and the single most important risk parameter on the page.
• Non-equity tripwire: check performed and passed. AMD is common stock — NASDAQ-listed since 1972, no par-like price, a 52-week range spanning 284%, beta 2.489, 37.2M shares traded today, no fixed dividend. It is not a preferred, a baby bond or a structured note inheriting parent fundamentals.

Today's move and what it does to the entry

Today's +7.00% is the entry question for this dive, and the honest answer is that it makes a marginal entry worse rather than confirming a good one. Three observations, each traceable.

First, it is sector beta, not an AMD-specific repricing. Every semiconductor name in this batch moved on the same day: Marvell +12.81%, Intel +10.92%, Lam +7.85%, Micron +7.62%, AMD +7.00%, KLA +6.95%, Applied +5.48%, ASML +4.22%, and Broadcom about +6.6% for reference. With a beta of 2.489, a broad semiconductor up-day of roughly 3% mechanically produces something close to AMD's move. Nothing about today tells you anything specific about AMD.

Second, the internals say this was a recovery, not a breakout. RSI at 48.4 and MACD at −7.60 after a 7% day are only possible if the stock had been declining for several weeks beforehand. The previous close of $484.64 was 5.8% below the 50-day average; today's close of $518.58 is 0.8% above it. Today reclaimed the 50-day. It did not break out of anything. The stock remains 10.7% below its closing high made earlier in the year.

Third, and decisively, the move happened into an unknown number. The company reports today, after this close. A 7% pre-print move is the market taking a position on a result it has not seen; a reader of this dive is positioning on a result that is already public and that this dive does not contain. Buying the day before a print is a risk. Buying into a 7% pre-print gap is that risk with the premium already paid to someone else.

The practical read. There is no technical support anywhere near this price. The 50-day at $514.33 is 0.8% below and was decisively lost during the preceding weeks, so it is unproven support. The 200-day at $314.56 is 39% below. For a 2.489-beta stock 65% above its own annual trend with a negative MACD, the distance to any real technical floor is the risk that matters, and it is very large. If the print is good, there will be a retest; if it is bad, there will be a much better price. Neither case argues for buying at 16:00 on 2026-08-04.

10. Insiders — read the CTO transactions correctly

These are cashless option exercises paired with same-day sales, not discretionary bearish signals, and the file makes that unambiguous. In both June and July the chief technology officer exercised 6,000 options at $84.85 and sold exactly 6,000 shares the same day. The reported holding is 1,239,687 immediately after the exercise and 1,233,687 immediately after the sale, in both months — the position returns to precisely the same number each time. This is the standard mechanism for converting expiring or vested options into cash without funding the strike out of pocket. It says nothing about his view of the stock.

What it does say is the opposite of a sell signal. He holds 1,233,687 shares after both transactions — worth approximately $640 million at $518.58 — and elected to sell only the 12,000 shares needed to settle two option exercises, 0.97% of his position. The economically meaningful decision was to retain $640 million of exposure, not to sell $6.6 million of it.

The chief commercial officer's June transactions are equally mechanical: a 7,495-share restricted-stock-unit vest at zero cost, followed by 2,950 shares withheld in kind at $547.26 to cover the tax. That is a net acquisition of 4,545 shares. No discretionary open-market sale by any insider appears in this file, and no discretionary open-market purchase either.

One forward-looking item from the 8-K dated 2026-06-26. The board approved base-salary increases effective 2026-07-01 and long-term incentive awards to be granted on 2026-08-15 under the 2023 Equity Incentive Plan, with the chief executive's award structured 75% performance-based restricted stock units and 25% time-based. The dollar amounts were in tables that did not survive text extraction and are therefore not available in this file. A 75% performance-weighted award is a reasonable structure and is noted as such; the grant lands eleven days after this dive.

11. Verdict, kill criteria and flip conditions

Watch.

What is genuinely excellent, stated first because it is most of the page. Revenue grew 34.3% in FY2025 to $34.639B and 37.9% year-on-year in Q1'26. Data Center grew 57% in the quarter with segment operating margin recovering to 27.6%. Gross margin has climbed from 44.9% to 52.8% in three years. Research and development is $8.091B, 23.4% of revenue, growing 39%. Free cash flow went from $1.121B to $6.735B in two years, with $3.0B of operating cash flow in Q1'26 alone. The balance sheet holds roughly $9.0B of net cash against $3.3B of principal debt, with $10.5B of undrawn committed liquidity after the May 2026 refinancing. No customer reached 10% of revenue in FY2025 or FY2024. $25.7B of unconditional purchase commitments — $18.3B falling in the remainder of FY2026 — is hard evidence that the ramp is being procured, not merely modelled. The independent knowledge base is unanimous in its admiration for the management. This is one of the best businesses in this batch and nothing below contradicts that.

Why the verdict is nevertheless Watch. Four things, in order of weight.

One: 19.4% of the equity is invisible. The 320 million penny warrants appear only in the 10-Q. They are in no consensus estimate, no quoted market capitalisation and nowhere in the vendor file. They vest on accelerator milestones and stock-price targets — they arrive in the success case and stay away in the failure case, which means every published forward multiple is understated by 13-17% precisely when it matters. Fully diluted, the stock is 43.1x FY2027 consensus, not 37.4x, and the effective market capitalisation is $1.022 trillion, not $845.6B.

Two: the price already occupies the optimistic half of an enormous range. Twenty-nine analysts put FY2027 EPS between $8.58 and $19.14. The price is 60.4x the low and 27.1x the high. Base fair value of $455 is 12.3% below spot, and the asymmetry runs −55.7% to bear against +35.0% to bull — roughly 1.6:1 against.

Three: the entry is the worst part of the case. The stock closed +7.00% into a same-day print, in a batch where every semiconductor name moved 4-13%, with RSI at 48.4 and MACD negative — meaning the move was sector beta reclaiming a lost 50-day average, not a breakout. It sits 65% above its 200-day average with a beta of 2.489. A reader acting on this dive acts after a release this dive does not contain. Writing a Buy into that would not be a call; it would be a coin toss dressed as one.

Four: the knowledge base's independent lane doubts the specific thing the price requires. Fifty-one non-management assessments across fourteen sources, and every one that addresses competitive position names the same two gaps and, in the most credible case, a falsifiable ceiling of single-digit percentage GPU share. Liking the company is not the same as underwriting the multiple.

What Watch means here, precisely. Do not initiate at $518.58 on 2026-08-04. Read the print. The base fair value of $455 is the level at which the risk/reward turns neutral on our own arithmetic; the 200-day average at $314.56 is the level at which it turns emphatically favourable. This is a name to own eventually and a name to size properly when owned.

Pre-registered KILL criteria — what would move this to Avoid:

• H2'26 guidance implying materially less than ~$14.3B of quarterly revenue, which would break the FY2026 consensus of $50.217B and, by extension, the FY2027 path that carries the multiple.
• A warrant tranche vesting while the stock is flat or down — dilution arriving without the growth that was supposed to pay for it. This is the specific compound failure the structure creates.
• Gross margin falling below ~50% on a non-charge basis, which would validate the independent-analyst claim that accelerator economics are structurally inferior and that mix shift is margin-destructive.
• Data Center segment operating margin failing to hold the Q1'26 recovery to 27.6%, i.e. returning toward the FY2025 level of 21.6% while revenue grows — buying share with margin.
• Further China export restrictions, or formalisation of the 15% government revenue-share expectation the 10-K flags on licensed MI308 sales. China is 22.4% of FY2025 billings and the MI308 precedent already cost $800M gross.
• Inventory extending materially beyond 157.7 days without corresponding revenue, against $18.3B of unconditional commitments falling due in the remainder of FY2026.
• The FY2027 estimate range widening further rather than converging, which would indicate the professional community is diverging rather than resolving.

Pre-registered FLIP TO BUY conditions:

• A price at or below ~$455 (base fair value) after the print settles, with the business trajectory intact — this alone converts the call.
• A price near the 200-day average of ~$315, at which point the position should be initiated at full intended size rather than staged.
• A Q2'26 result with H2 guidance implying $14B+ quarterly revenue and gross margin at or above 56%, which would confirm the near-term leg of the consensus and materially narrow the FY2027 range.
• Explicit disclosure of warrant-adjusted share-count guidance, or evidence that the sell side has begun modelling the 320 million shares — the mispricing we identify closes when the market sees it, and at that point the remaining question is just growth.
• Concrete evidence the software and networking gap is closing — the one thing every independent sceptic in the knowledge base identifies, and the thing that would convert AMD from a price-taking second source into a durable one.

Where AMD fits in the Synthos Framework Portfolio. The AI infrastructure / semiconductor sleeve, as a watchlist entry with a 1.5-2% target weight on trigger, not an initiation today. Note the batch overlap carefully: AMD, Intel and Marvell are three different bets on the same buildout, and the four wafer-fab-equipment names in this batch are largely one trade with four expressions. Holding AMD alongside several of them is a single concentrated position in AI capital expenditure, and the knowledge base contains an explicit warning on exactly that concentration — that hyperscaler capital expenditure funded by driving free cash flow to zero cannot keep accelerating. Size the theme, not the ticker. Logged as a tracked Synthos call (Watch) as of 2026-08-04 at $518.58.

Single biggest risk: the 320 million penny warrants. They are 19.4% of the share count, absent from every consensus figure and every screen, and they vest on the same milestones the bull case depends on. The risk is not that they exist — it is that they convert exactly when the thesis works, so the shareholder who is right about the business is still 13-17% wrong about the earnings per share. Nothing else in this file has that property.

Most fragile assumption in the price: that second-half FY2026 revenue averages roughly $14.3 billion a quarter. Everything downstream — the $79.8B FY2027 line, the $13.87 FY2027 EPS, the 37.4x multiple, the $575.29 street target — is built on the FY2026 consensus of $50.217B holding. Q1'26 was flat sequentially at $10.253B. Consensus for Q2'26 is $11.310B. The arithmetic then requires a 39% step up within six months. It is the first thing the print landing today will either support or destroy, and this dive is written without knowing which.

Provenance & disclosures

• Traceability: 174 raw knowledge-base hits on AMD "Lisa Su" MI300 Instinct across 51,928 distilled claims. 8 discarded outright — 7 matches on the ordinary English noun instinct rather than AMD's accelerator brand (none of the 7 was an entity match), and 1 record on an attribution slug explicitly quarantined as misattributed. 98 classified as company or management voice and half-weighted throughout, including the chief executive's own claims and the earnings-release lane. 13 entity matches attributed to a direct competitor's chief executive, reported and discounted, never used as independent corroboration. 12 text-only matches used as supporting colour. 41 first-class independent entity claims used, across 14 distinct sources, spanning 2019-08-27 to 2026-07-13, average conviction 63, distribution 23 bullish / 10 neutral / 8 bearish. Where speaker and thinker differ, claims are attributed to the speaker and the channel is noted separately; three claims sit on channel slugs whose content plainly does not match the slug and are flagged as channel-attributed with unknown speaker. All quotes verbatim from stored claim text, except that names of specific AI models, laboratories and model families have been redacted and replaced with bracketed generic descriptions per house style; no substance was altered.
• Data as-of: income statement 2026-03-28 (Q1'26, 10-Q filed 2026-05-06) · balance sheet in the payload 2025-12-27, superseded in Section 4 by the 10-Q (2026-03-28) and the 8-K (2026-05-13/14) · cash flow FY2025 plus Q1'26 from the 10-Q · estimates 2026-08-04 · prices 2026-08-04 (timestamp 1785873601 = 2026-08-04T20:00:01Z; $518.58, +7.00%; 50-DMA $514.33; 200-DMA $314.56; RSI 48.4; MACD −7.60) · knowledge-base claims 2026-08-04. Filings used: 10-K 2026-02-04 (FY2025), 10-Q 2026-05-06 (Q1'26), 10-Q 2025-11-05 (Q3'25), 8-K 2026-05-05, 8-K 2026-05-15, 8-K 2026-07-01. No figure comes from memory, recall or external retrieval.
• Data rejections (detailed in Section 8): seg_prod FY2025 double-counts Gaming inside Client and Gaming ($38.549B against reported $34.639B) — 10-K three-segment structure used instead; netDebt of −$1.067B omits $5.013B of short-term investments — filings' ~$9.0B net cash used; totalDebt of $4.472B irreconcilable with the filings' $3.3B aggregate principal (lease obligations apparently double-counted) — $3.3B used; est.ebitdaAvg and est.ebitAvg rejected for all eight years as synthetic (ebitAvg is exactly 47.0% of ebitdaAvg in every year; FY2026 ebitAvg $5.118B against niAvg $12.337B is impossible), with evToEBITDATTM and enterpriseValueMultipleTTM (both 104.4x) rejected with them; profile.lastDividend of $0.005 rejected as an artefact against a zero dividend everywhere else; effectiveTaxRateTTM of 0.24% and taxBurdenTTM of 1.013 rejected as distorted by an $853M discrete tax release, with the 10-Q's 14.8% used; tech 52-week high/low ($580.91/$151.14, closing basis) used in preference to quote yearHigh/yearLow ($584.73/$149.22, intraday) for consistency with the SMA/RSI/MACD series; seg_geo FY2025 flagged as non-comparable to FY2024 (Europe and Japan lines folded into Other) and as billing location rather than end demand; the price-target set flagged as a 4.8x spread ($260 to $1,250) rather than a usable consensus.
• The most important gap is not a rejection but an omission. The 320 million warrant shares disclosed in the 10-Q appear in no field of the 18-key payload — not in the market capitalisation, not in the diluted share counts, not in any estimate row. They are 19.4% of the share count. Section 2 rebuilds the valuation on both bases.
• Basis note: analyst estimates are non-GAAP; reported diluted EPS is GAAP. FY2025: epsAvg $3.9621 against GAAP $2.65, a gap approximating the $2.3B of acquisition-intangible amortisation disclosed in the 10-K's All Other note, after tax. FY2024: epsAvg $3.3149 against GAAP $1.00, a gap requiring both amortisation and stock compensation to be excluded. The non-GAAP definition is not stable across the estimate series. All forward multiples in Section 6 are labelled non-GAAP; the trailing 169.5x is GAAP. Fair values are struck on FY2027 consensus non-GAAP EPS with an explicit warrant-dilution adjustment stated in each scenario.
• Estimate coverage: 36 analysts on FY2026 and FY2027 revenue, 32 on FY2028, 27 on FY2029; 30 on FY2026 EPS, 29 on FY2027, 16 on FY2028, 10 on FY2029 and FY2030. No conclusion rests on the FY2029 or FY2030 rows. The FY2027 row (29 analysts on EPS, 36 on revenue) is the primary forward anchor, and its $8.58-to-$19.14 range is treated as the central analytical fact rather than smoothed to its mean.
• Peer note: the vendor peer set mixes genuine adjacents (Applied Materials, Lam, KLA, ASML, Micron, Arm, Qualcomm, Texas Instruments) with non-comparables (SAP, Palantir). No peer multiples are supplied in the file, so no peer-multiple comparison is drawn. Batch cross-references use only the peer table supplied for this batch.
• Fair-value caveat: the $230 / $455 / $700 anchors are multiples of FY2027 consensus non-GAAP EPS (26.8x the low undiluted, 35x the mean at partial dilution, 42x the high fully diluted), cross-checked against FY2026E and FY2028E revenue and earnings. Stated scenario arithmetic, not a discounted cash flow. The base case assumes multiple compression from 37.4x to 35x and is explicit about it; no case in this dive assumes multiple expansion.
• Half-weight management framing: company guidance for Q2'26 (revenue ~$11.2B ± $300M, non-GAAP gross margin ~56%, operating expenses ~$3.3B, tax rate 13%, share count ~1.66B) is sourced from the company's own voice and is discounted and labelled as such wherever used. The same discount applies to the 98 company-voice knowledge-base claims and to the 13 claims attributed to a direct competitor's chief executive.
• Timing — read this before acting. AMD reports Q2'26 on 2026-08-04, the date of this dive, after the 16:00 ET close against which every price here is struck. The result was not available when this was written and no estimate of it appears anywhere in this document. The stock closed +7.00% into the release. Any reader is acting on information this dive does not contain, and the Watch verdict is in significant part a statement about that asymmetry rather than about the company.
• Not investment advice. Independent research, educational and informational only, never personalised. No recommendation to buy, sell or hold any security is made to any person.
• Version: 2026-08-04-full.