AMDAdvanced Micro Devices, Inc.

AMD — dossier, 2026-09-21

Slot names-ai-capex-1, name pass. Drafted unattended by the cloud research routine following .claude/skills/stock-sourcing-pipeline/SKILL.md. Nothing here is human-verified. Every figure carries the document it was read from; anything from background knowledge is marked [background]. Never edited after today — corrections go in scorecard.md's log.

Price reference: $559.82 close 2026-09-18 (ROIC.ai). Shares outstanding 1,632,475,042 as of 2026-07-29 (EDGAR dei:EntityCommonStockSharesOutstanding, Q2 FY26 10-Q cover). Implied market capitalisation ≈ $914bn; ≈ $1.09tn if the 320m warrant shares are treated as outstanding.

---

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), cluster ai_semis, coverage: sec_domestic, cik 2488. No Tier 0 promotion: queue.json promoted was empty at the start of this slot.

---

S2 — Kill test

EDGAR XBRL companyconcept. Fiscal years end late December; FY2025 ended 2025-12-27.

Share count over five years. 1.229bn → 1.636bn, +33%. Almost all of that is the 2022 Xilinx acquisition [background: the step from 1.229 to 1.571 coincides with FY2022 and is acquisition-driven; the filings read this pass do not restate it]. Since 2022 the count has been flat to slightly down — AMD has been holding its share count, not growing it.

That is the fact about to change, and it is the whole risk. From the Q2 FY26 10-Q:

In October 2025 and February 2026, the Company issued warrants to OpenAI OpCo, LLC (OpenAI) and Meta Platforms, Inc. (Meta), each entitling the holder the right to purchase up to 160 million shares of the Company's common stock at an exercise price of $0.01 per share (the OpenAI Warrant and the Meta Warrant, respectively).

320 million shares at one cent, against 1,659m diluted in Q2 FY26 — ~19% dilution, and that is before the equity plan. AMD amended its 2023 plan on 2026-05-13, raising authorised shares by 65 million to 153 million, with 107 million still available at 2026-06-27. Total identified overhang: 427 million shares, ~26% of the current diluted count.

The warrants vest on Instinct purchase milestones plus stock price targets, and none had vested at 2026-06-27 — so nothing has hit the P&L yet. They are classified as liabilities until the conditions for equity classification are met.

Cash, debt and the maturity wall. Cash, cash equivalents and short-term investments were $10.6bn at FY25-end (from $5.1bn a year earlier) and $13.1bn at 2026-06-27 (Q2 call). Aggregate principal debt was $3.3bn at FY25-end (from $1.8bn). Then, in August 2026, AMD registered a new shelf (S-3ASR, 2026-08-13) and drew $4.75bn of senior notes off it the same day. Pro-forma debt ≈ $8.05bn against ≈ $17.8bn of cash and short-term investments — still net cash of roughly $9.8bn. Net debt/EBITDA is negative; there is no maturity wall.

Cash runway: not applicable. AMD is profitable and generated $6.7bn of FCF in FY25.

Analyst coverage count: not captured this pass. Thin.

Verdict — survives or dies: AMD survives comfortably. Revenue +34% in FY25 and +50% y/y in Q2 FY26, free cash flow $6.7bn, net cash, no going-concern issue. GAAP operating margin of 10.7% in FY25 is the weak number — an order of magnitude below NVIDIA's 60.4% — but it is rising and reached 17.2% in Q2 FY26. The kill risk here is not insolvency, it is that the equity is being handed to the customers: 320m warrant shares at $0.01 is, at today's price, roughly $179bn of value contingently transferred to two buyers in exchange for their purchase commitments.

---

S3 — Filings, at gate-minimum depth

Read this pass: FY25 10-K (Item 1A risk factors and Item 7 MD&A only), Q2 FY26 10-Q (statements, warrant note, segment MD&A), DEF 14A 2026-03-27 (ownership), the August-2026 424(b)(5), and the Q2 FY26 earnings call. One transcript, not two — the Equibles budget for this run was 25 calls across three names. Not read: the FY24 10-K, older 10-Qs, comment letters, critical audit matters, auditor tenure and fees. Note also a 10-K/A filed the same day as the 10-K (2026-02-04); the amendment itself was not read this pass — flagged for the next.

The three biggest risks management itself discloses

Management's own words, located by exact substring and copied with their locator.

1. Customer concentration — stated as a dependency, deliberately unquantified on revenue. 10-K FY25, Item 1A:

We depend on a small number of customers for a substantial portion of our business and we expect that a small number of customers will continue to account for a significant part of our revenue and receivables in the future.

The only number AMD actually gives is on receivables, in Note 11:

One customer accounted for approximately 11 % and another customer accounted for 24 % of the total consolidated accounts receivable balance as of December 27, 2025 and December 28, 2024, respectively.

No single customer above 10% of FY25 revenue is disclosed. That makes AMD materially less concentrated than either of the other names in this slot — NVDA discloses 22% and 14% of revenue in two customers, AVGO 32% in one. It is the one structural advantage AMD has here, and the warrants are in the process of trading it away: the whole point of the OpenAI and Meta agreements is to concentrate.

2. Export control, with a cost already booked. 10-K FY25, Item 1A:

In April 2025, the U.S. government implemented a new license requirement for the export of certain semiconductor products to a D5 Country, and to companies headquartered in, or with an ultimate parent located in such D5 Country. This restriction impacts our AMD Instinct™ MI308 products. As a result of the restriction, we incurred approximately $800 million in inventory and related charges in the second quarter of 2025.

Roughly $360m of that was reversed in Q4 2025 once licences were granted, for a net ~$440m — which is the entire difference between AMD's reported 50% FY25 gross margin and what it would otherwise have been.

3. Dependence on third-party manufacturing, now compounded by its own divestiture. 10-K FY25, Item 1A Risk Factors Summary:

We rely on third parties to manufacture our products, and if they are unable to do so on a timely basis in sufficient quantities and using competitive technologies, our business could be materially adversely affected.

AMD is fabless [background], and in the same year it bought ZT Systems for $3.2bn plus 8.3m shares and then sold ZT's manufacturing arm to Sanmina for $2.4bn plus 1.2m Sanmina shares and up to $450m contingent — keeping the designers, selling the factories, and naming Sanmina its preferred partner. The rack-scale ambition (Helios) rests on a contract manufacturer.

Latest quarter — Q2 FY26 (ended 2026-06-27)

The Q2 FY25 comparison flatters: it carried the ~$800m MI308 charge. Segments: Data Center $6.7bn, +107% y/y, 58% of revenue (42% a year ago), operating income $2.1bn against a $155m loss; Client and Gaming $3.8bn (+6%), within which Gaming fell 31% to $779m; Embedded $977m (+19%). Stock-based compensation was $990m in the first half, versus $733m.

Guidance given on the call (machine transcript, wording mangled in places): Q3 FY26 revenue ~$13.0bn ±$300m (+41% y/y), non-GAAP gross margin ~56%, opex ~$3.65bn, tax 13%, diluted share count ~1.66bn. Beyond that: server revenue +>80% y/y in H2 2026 and +>70% in 2027; Data Center segment revenue to more than double year-over-year in 2027; and an explicit statement that AMD expects to "significantly exceed our $20 annual EPS target within our strategic timeframe."

New customer announcements on the same call: Anthropic for up to 2 GW of MI450 in Helios, first gigawatt in H1 2027; Microsoft deploying Helios at scale on Azure. Neither is in a filing read this pass.

Proxy skim (DEF 14A filed 2026-03-27)

Based on 1,630,338,779 shares of our common stock outstanding as of March 19, 2026.

All 16 current directors and executive officers together: 8,107,566 shares, marked "*" — less than one percent. Insider alignment at AMD is through compensation, not ownership. That is a genuine contrast with NVIDIA (3.94%) and Broadcom (1.9%) and it matters more here, because AMD's board is the body approving warrants that hand ~19% of the company to two customers. Compensation structure and what the bonus is paid on: not extracted this pass — thin.

Insider pattern (12 months to 2026-09-21)

Zero open-market purchases by any insider (Equibles GetInsiderTransactions AMD, transactionType: Buy, fromDate: 2025-09-21 → "No insider transactions found"). 612 Form 4s on the current EDGAR submissions page, newest 2026-09-17; none of it is conviction buying.

Red-flag sweep

• Going concern: no flag. Newest examined filing is the 10-Q filed 2026-08-05; coverage there starts 2026-05-06, so earlier filings are unexamined.
• Executive changes (8-K Item 5.02, via Equibles): board only. Tim Ryan appointed 2026-08-19; Joe Householder retired the same day, two days after giving notice. Eight directors elected at the 2026-05-13 annual meeting. No CEO or CFO change — Lisa Su and Jean Hu are both in seat.
• Auditor change / restatement / NT 10-K: not run this pass — the efts.sec.gov full-text sweep was dropped for time. Thin; flagged for the next pass. The 10-K/A of 2026-02-04 should be read then.
• Leases not yet commenced: $1.3bn of data center and other real estate leases signed but not commenced at FY25-end, expected to start in 2026 with terms of 1 to 10 years. AMD is taking on data center property of its own — small against the balance sheet, new in character.

---

S5 — Ownership

Who is on the other side: short interest is 2.6% with 2.5 days to cover — the highest of the three names in this slot, but still not a short case. The more interesting answer is that OpenAI and Meta are on the other side of the equity, holding 320m shares' worth of one-cent warrants. They are not sellers, but they are the counterparties whose exercise sets the share count.

---

S8 — Valuation and sizing (model; both decisions are ⏳ OPEN (user))

AMD guides on a non-GAAP basis and the gap to GAAP is wide (Q2 FY26: 56% non-GAAP gross margin versus 53.8% GAAP; 27% non-GAAP operating margin versus 17.2% GAAP), almost all of it acquisition-intangible amortisation and stock-based compensation. Both are modelled below.

EV: market cap ~$914bn less ~$9.8bn pro-forma net cash → EV ≈ $904bn, ≈ 18x FY2026E revenue and ≈ 11x FY2027E revenue.

• Bear: Data Center does not double in 2027 — it grows 50% — while Gaming keeps shrinking and memory cost holds gross margin at 54%. FY27 revenue ~$63bn, non-GAAP EPS ~$9. At 30x that is ~$270, roughly half today's price. And the warrants still dilute on the stock-price tranches.
• Base: guidance is met. ~$13.8 FY27 EPS at ~41x today.
• Bull: the $20 EPS target lands early on Helios and MI500, and 45x on $20 is ~$900.

The honest summary of the valuation: AMD is the expensive one. NVIDIA trades at ~14x model FY28 EPS and Broadcom at ~12x management's own FY28 EPS target; AMD needs its Data Center segment to double in a single year merely to get to ~41x. It is priced as the option on the second-source thesis, and the warrants mean that if the thesis works, holders own ~19% less of it.

Liquidity: average daily volume 16,772,687 shares (FINRA, 2026-08-31 settlement) at ~$560 ≈ $9.4bn traded per day. No sizing constraint.

⏳ Buy price: OPEN (user). Size: OPEN (user). Both are owner-only under the skill.

---

S9 — Monitoring

• Next earnings: Q3 FY2026 ends ~2026-09-26; AMD has reported that quarter in the first week of November in recent years. Derived from the reporting cadence, not from a sourced calendar — ROIC.ai's earnings calendar is behind a paid plan. Thin.
• Filing feed: Tier 0 EDGAR sweep (watch_edgar.py, 06:40 UTC) covers cik 2488.
• Leading indicator between quarters: the Warrants note in each 10-Q. It states, in one sentence, how many warrant shares have vested and become exercisable. Vesting keyed to purchase milestones is a direct, filed read on whether OpenAI and Meta are actually buying — better than any revenue commentary, because it is the customers' own behaviour reported by AMD. Second: Data Center segment revenue against the "more than double in 2027" guide.
• Trigger that converts this to a buy: owner's call. What this routine can say is the condition that would resolve the central objection — warrant tranches vesting on purchase milestones (not on stock-price targets) while Data Center revenue tracks the doubling guide. That would mean the dilution bought real volume. The adverse case is tranches vesting on price targets while revenue lags, which is dilution paid for nothing.

---

Access and connector notes for this run

Same as the NVDA dossier of this date: SEC EDGAR reachable with the required User-Agent; Equibles reachable, 22 calls spent across the slot against a 25 cap; ROIC.ai reachable for prices but its earnings calendar is gated; EDINET DB and eu_data not needed for sec_domestic names and not called; Alpha Vantage, Bigdata.com and FMP not called, per the skill.

Process support, not investment advice.