Web Research

What the Internet Knows About AMD

Bottom line up top. The web extends the filings rather than refuting them. Three facts not yet absorbed by consensus: (1) the OpenAI deal carries a 160-million-share penny warrant (~10% of AMD) with vesting gated on stock-price hurdles escalating to $600/share, and ASC 606 will likely treat the fair-value grant as contra-revenue once vesting becomes probable — compressing reported Instinct ASPs as soon as MI450 begins shipping in 2H 2026. (2) Management confirmed on the Q1 2026 call that MI450 ramps below corporate-average gross margin — the hyperscaler wins that drive the bull case are also the margin headwind for 2026–27. (3) Consensus 12-month price target (~$486) now sits below the ~$507 print, even after a parade of upgrades; the stock has run past the Street, the marginal upgrade is harder to find, and insiders sold ~$158M in 90 days with zero open-market buying. The forensic file is clean (no SEC matter, no restatement, E&Y unbroken since 1970, 10b5-1 mechanics on Lisa Su). This is a priced-for-perfection setup: the binary swing factor is MI450/Helios execution in 2H 2026 and the underappreciated drag is the customer-equity accounting structure.

---

The Ten Findings That Matter

1. OpenAI's 160-million-share penny warrant — the accounting overhang consensus has not priced

Per AMD's 8-K filed Oct 6, 2025, OpenAI received a warrant for 160 million shares at a $0.01 strike — close to 10% of AMD on a fully-diluted basis — with tranched vesting tied to both (a) 6 GW of Instinct deployment milestones (first 1 GW shipping 2H 2026) and (b) AMD stock-price hurdles that escalate to $600/share for the final tranche, expiring Oct 5, 2030. Britannica corroborates the deal; the OpenAI press release confirms the strategic-partnership structure.

The piece sell-side write-ups under-cover: under ASC 606 (and the FASB's ASU 2025-04 clarification surfaced in the forensic file), share-based consideration to a customer is treated as noncash consideration payable to a customer — a reduction in transaction price, i.e. contra-revenue — once vesting becomes probable. Translation: as soon as OpenAI's first 1 GW deployment is on a probable path, AMD begins recognizing fair value of the relevant tranche as a reduction of Instinct revenue. This is on top of the share dilution.

• Source: https://ir.amd.com/financial-information/sec-filings/content/0001193125-25-230895/d28189d8k.htm ; https://openai.com/index/openai-amd-strategic-partnership/
• So-what: Bull models project Instinct revenue at sticker. The contra-revenue cuts ASPs and reported gross margin precisely when the AI ramp is supposed to inflect — a known unknown that will dribble into reported margins through 2026–27. Combined with the $600 share-price gate, the warrant simultaneously caps near-term dilution (only fully painful if the stock rips) and pulls down margin (whenever the deal works).
• Priced in? Partially. The market has the deal but not the accounting mechanics; the sell-side commentary that does mention it (Simply Wall St) flagged it as the reason analyst fair-value-per-share estimates rose only modestly ($213.89 → $225.00) despite raising revenue.
• Importance: Red flag (structural).

2. Meta's 6 GW deal doubles the bet — at management-confirmed sub-corporate gross margin

Announced Feb 24, 2026, Meta committed to 6 GW of custom MI450-architecture GPUs plus 6th-generation EPYC "Venice/Verano" CPUs deployed on AMD's Helios rack platform — first 1 GW shipping 2H 2026, alongside OpenAI's first 1 GW. On the Q1 2026 call (May 5, 2026), CFO Jean Hu was explicit that MI450 will be below corporate-average gross margin during ramp; a Seeking Alpha analyst, summarizing the call commentary, projects no meaningful margin expansion for the next 1–2 years. The "dollar-per-gigawatt expands with each generation" language management uses is the offset, but it is a future claim, not 2026.

One unverified secondary source (abit.ee) claims Meta also received a ~160M-share performance warrant similar to OpenAI's; the AMD IR release does not confirm this. Worth watching but not bankable yet.

• Source: https://ir.amd.com/news-events/press-releases/detail/1279/amd-and-meta-announce-expanded-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus
• So-what: The two biggest GPU deals in AMD's history both dilute gross margin on the way up. The "tens of billions" of revenue scale is real; the margin compression is also real. Combined with OpenAI's contra-revenue path, gross margin trajectory is a 2026–27 headwind that the >35% non-GAAP operating-margin target implicitly papers over.
• Priced in? No — bulls model revenue beats; the GM line is where the surprises will land.
• Importance: Yellow/red flag.

3. Insiders sold ~$158M in 90 days with zero open-market buying — at the very top of the chart

Across the executive team, the 90-day insider window through mid-June 2026 shows 372,032 shares sold for $158.66M, against zero open-market buys from any named executive officer (only EVP Phil Guido has bought, ~$500K over 24 months). Specific transactions:

• Lisa Su, CEO — 125,000 shares on Jun 10, 2026 at $460.69 = $57.6M; identical 125,000-share clips on Dec 11, 2025 ($26.9M @ $215.14), Feb 11, 2026 ($26.8M @ $214.36), and May 13, 2026 ($432.90–$457.82). All under a 10b5-1 plan adopted Sept 9, 2025.
• Jean Hu, CFO — Yahoo insider data shows 4,023,581 shares disposed Feb 17, 2026 at $196.78–$205.12 (~$791M–$825M gross; likely includes net-of-vest activity — warrants confirmation but the size dwarfs CEO selling).
• Mark Papermaster, CTO — five tranches totaling ~$29M over 12 months; largest 31,320 shares @ $350 on Apr 24, 2026 ($11M).
• Forrest Norrod, EVP Data Center — 19,487 shares @ $431.40 on May 20, 2026 ($8.4M, ~5.7% of position) plus prior tranches.
• Ava Hahn, General Counsel — 56,814 shares sold Feb 18, 2026 ($11.3M), clustered against Hu's sale 24 hours prior.

Note also that Lisa Su's 2025 total comp jumped to ~$55.2M (Simply Wall St), roughly +78% YoY — a doubling that the May 2026 say-on-pay vote ratified without dissent.

• Source: https://www.marketbeat.com/stocks/NASDAQ/AMD/insider-trades/ ; https://www.stocktitan.net/sec-filings/AMD/form-4-advanced-micro-devices-inc-insider-trading-activity-afaaa69b9622.html ; https://aflcio.org/paywatch/AMD
• So-what: Mechanical 10b5-1 trades insulate against legal risk but don't mute the signal: every NEO is converting equity to cash at multi-year highs; no NEO is putting cash back in. Combined with comp doubling, the insider message is "valuation is full." Not a sell trigger on its own — it is a "don't add at these levels" tell.
• Priced in? Not really — insider conviction is rarely modeled directly.
• Importance: Yellow flag.

4. TSMC CoWoS is the binding constraint — and NVIDIA owns more than half of it

This is where the bull thesis meets the supply curve. Across the moat and competition files: NVIDIA has booked 800,000–850,000 wafers of TSMC CoWoS for 2026, more than half the global capacity; Broadcom and AMD share the remaining 40–50%. TSMC is scaling CoWoS from ~65–75K wafers/month in 2025 toward 120–130K wpm by end-2026, but lines are "fully booked" against ~1M wafers of total 2026 demand. Trefis rates AMD's advanced-packaging infrastructure "worse vs NVIDIA" with "high probability, high impact" — the bear thesis the bulls don't address. JPM commentary cited in the moat file: TSMC 3nm hits a capacity ceiling before 2026, creating a two-year supply gap.

AMD's response — Simply Wall St flagged a bypass attempt via panel-level EFB packaging — is a margin lever but doesn't change the 2026 ceiling.

• Source: https://wccftech.com/nvidia-alone-has-tsmc-advanced-packaging-lines-booked-for-several-years-ahead/ ; https://siliconanalysts.com/analysis/foundry-allocation-status-q1-2026
• So-what: Even if MI450 demand triples vs current "tens of billions" guidance, AMD cannot ship more than its CoWoS allocation permits. The MI450 numbers consensus is now extrapolating are gated by an externality AMD does not control. Order book matters less than allocation.
• Priced in? No — bulls extrapolate from order book; bears under-quantify the cap.
• Importance: Red flag (structural).

5. China: $800M write-down, $360M reversal, 15% revenue-share precedent — and DIO still bloated

The most documented forensic event in the file. Q2 2025: $800M MI308 inventory and related charge under US export controls compressed gross margin from 53% to 43% that quarter; Q4 2025 reversed $360M after partial licenses (net hit $440M). Critically, Lisa Su publicly stated AMD is prepared to pay a 15% revenue-share tax on China AI chip shipments — an unprecedented quasi-tariff that, if codified, becomes a permanent margin drag. Jan 13, 2026, BIS final rule moved MI325X (and NVDA H200) to case-by-case review with a 25% tariff plus required US lab testing for China exports.

Inventory remains the open question: Days Inventory Outstanding ran 150.6 days in Mar-26 vs a 10-year median of ~80. Management frames the build as MI450 pre-build; it could equally be persistent China-restricted overhang. The data alone can't tell — but the magnitude is unusual.

• Source: https://ir.amd.com/financial-information/sec-filings/content/0000002488-26-000021/amd-20251227.htm ; https://www.reuters.com/world/china/amd-chief-says-company-ready-pay-15-tax-ai-chip-shipments-china-2025-12-04/
• So-what: Recurring margin volatility, capped China DC TAM, and a 15% revenue-share precedent that could spread to other restricted product lines. The 2025 write-down/reversal cycle will likely repeat as policy evolves.
• Priced in? Partly — the inventory write-down is in the rear view, but the 15% revenue-share scheme is genuinely not modeled.
• Importance: Red flag (recurring).

6. EPYC owns the cash flow — record 46.2% server CPU revenue share in Q1 2026

The most underrated positive in the file is that the cash-generative part of AMD is firing — and it is, structurally, the EPYC server CPU franchise that funds the GPU bet. Mercury Research Q1 2026: AMD captured 46.2% of server CPU revenue, a record, and is targeting >50% revenue share. AWS, Google Cloud, Microsoft Azure, and Tencent all launched 5th-Gen EPYC instances in Q1 2026. AMD commenced TSMC 2nm production of EPYC "Venice" in May 2026 — Intel's 18A is reportedly in low-to-mid-60% yield range and Panther Lake is in HVM at Fab 52, so Intel is no longer dead, but the next 12 months still favor AMD.

This is also why FY2025 free cash flow exploded: $6.70B FCF (+178% YoY), Q1 2026 alone $2.57B (+253% YoY). Cash $12.4B vs total debt $3.2B — pristine balance sheet, A/A3 credit (Moody's upgraded Feb 2026 with a positive outlook), $10B buyback authorization.

• Source: https://wccftech.com/amd-posts-record-server-cpu-revenue-in-q1-2026-red-team-commands-30-percent-market/ ; https://ir.amd.com/news-events/press-releases/detail/1276/
• So-what: This is the valuation cushion. If the AI narrative cracks, the EPYC cash engine still throws off enough to fund buybacks and Helios development. It is also the strongest counter to short theses on accounting quality — cash is conversion of GAAP earnings is improving, not deteriorating.
• Priced in? Yes — this is the part the multiple already rewards. But the magnitude of the FCF inflection has surprised even bulls.
• Importance: Positive.

7. Consensus PT now sits below the stock — sell-side has played most of its cards

The stock at ~$507 has run past the average analyst target of $486.33 (stockanalysis.com / Yahoo). MarketBeat shows a stricter consensus of $428.65 (~15.5% downside). Targets range from $225 (low) to $665 (Mizuho/Barclays high). The marquee recent upgrades:

• Citi (Atif Malik), Jun 12, 2026 — Neutral → Buy; PT $460 → $575.
• TD Cowen, Jun 1, 2026 — PT $500 → $600 (Buy).
• BofA, ~Jun 11, 2026 — PT $500 → $560 (Buy).
• Wedbush, May 4, 2026 — PT $290 → $400.
• DA Davidson, Apr 24, 2026 — PT $220 → $375 (upgrade to Buy).

But also the cuts:

• Northland, Jun 8, 2026 — Buy → Market Perform, citing 2027 consensus EPS/revenue as overly optimistic vs. likely hyperscaler capex slowdown.
• HSBC, May 4, 2026 — Buy → Hold on valuation.
• Argus / Morningstar fair value — $450.

Forward P/E 78.1x; EV/EBITDA 109x trailing; EV/Sales 23.6x; PEG ~1.04–1.32 — only growth-adjusted measures look "reasonable." Sector median P/E ~26x; AMD trades at a 182% premium on forward and 327% on trailing.

• So-what: Sell-side is no longer the marginal upside lever — the dispersion is widening, and the bears (Northland, HSBC) are leaning on 2027 capex as the swing factor, not 2026. The catch-up game is over; from here, fundamentals have to exceed the elevated targets, not just meet them.
• Priced in? Yes — extreme dispersion is the priced view. Variant signal would be a coordinated wave of cuts (not yet visible).
• Importance: Neutral-bearish.

8. The MI300X "45% of theoretical peak" finding — ROCm gap is not closing fast on training

The single biggest quiet finding in the specialist queries: Yahoo Finance reporting (cited in the specialist file) notes the MI300X delivers ~45% of theoretical peak performance in real-world settings versus >80% for the H100, attributing the gap to ROCm software immaturity. ROCm 7 (mid-2025) added day-0 support for MI350 and major frameworks; AMD also showed strong MLPerf 6.0 inference results in May 2026 with 9 OEM submissions. But — and this is what the bulls skip — no independent MLPerf training submissions for AMD GPUs are cited anywhere in the file.

The AI bull case requires the training workloads, not just inference. AMD is pivoting to "dominate inference" (memory-capacity advantage) — a defensible niche, but a smaller TAM than the headline market figures imply.

• Source: https://finance.yahoo.com/markets/stocks/articles/advanced-micro-devices-inc-amd-210033600.html ; https://www.amd.com/en/blogs/2025/enabling-the-future-of-ai-introducing-amd-rocm-7-and-the-amd-developer-cloud.html
• So-what: The gap that justifies NVIDIA's premium is software-driven, not silicon-driven. AMD's MI450 hardware can be at parity and still lose share in training. The "tens of billions" of MI450 revenue is real for inference; the training market is still NVIDIA's.
• Priced in? Partially — sophisticated investors know about ROCm. The specific 45%-of-peak data point is not in consensus models.
• Importance: Yellow flag.

9. Custom silicon (Google TPU, Anthropic, AWS Trainium, Meta MTIA) is the variant bear — Bernstein flagged it

The variant-perception finding that doesn't get enough airtime: Bernstein wrote that Google's TPU success "feels incrementally negative to the AMD narrative" because TPUs are "the only really successful ASIC program" at scale. Anthropic committed to 3.5 GW of TPU starting 2027; a $35B Anthropic loan is backed by Broadcom (digitimes, Jun 2026). Broadcom carries a $73B backlog in custom ASICs; Broadcom + Marvell control ~95% of the ASIC co-design market.

The mechanism: every hyperscaler dollar spent on a custom-designed ASIC is a dollar not spent on a merchant accelerator like Instinct. The OpenAI and Meta deals are AMD's win against NVIDIA, but the structural threat to both merchant vendors is the hyperscaler internalization that Google has already proven works.

• Source: https://www.investing.com/news/analyst-ratings/bernstein-google-tpu-news ; https://www.digitimes.com (cited in competition-agent file)
• So-what: The "second-source GPU" thesis is partially undermined. AMD's MI450 wins are real but the underlying mid-tier merchant TAM may compress through 2027–28 as more hyperscalers move workloads onto custom silicon. The Northland 2027 EPS downgrade is largely about this.
• Priced in? No — variant view. Bulls model rising AI TAM; bears model TAM cannibalization by ASICs.
• Importance: Variant perception.

10. Forensic file is clean — the SMCI short-squeeze playbook does not apply

The web confirms the filings on the boring stuff that matters most: E&Y unbroken auditor since 1970 with no restatements, no material control weaknesses (FY24 10-K and FY25 10-K/A confirm); the last SEC enforcement action was the 1996 cease-and-desist over 1992–93 Intel disclosures. No active material legal proceedings (Adeia patent suits settled Mar 9–11, 2026, terms undisclosed but quietly cleared). Whistleblower-related concerns surface only in historical (2012) commentary. Short interest 2.71% of float with 1.1–1.4 days-to-cover — no coordinated short-seller campaign. Free cash flow growth dramatically outpacing GAAP earnings growth in the right direction. CFO Hu's tenure clean since late 2022.

• Source: https://www.sec.gov/Archives/edgar/data/2488/000000248825000012/amd-20241228.htm ; https://stockanalysis.com/stocks/amd/statistics/
• So-what: Anyone pitching AMD as "the next SMCI" lacks evidence. The real risks here are strategic (MI450 execution, China policy, hyperscaler concentration) and valuation (priced for perfection), not accounting.
• Priced in? Yes — but worth knowing the bear case can't credibly attack the books.
• Importance: Positive (defensive).

---

Recent News Timeline

The reference layer behind the findings — every meaningful news item ordered by date, with significance rating. Sortable.

---

Industry & Competitive Layer — Beyond the Industry Tab

New external evidence that informs the AMD-specific position (does not restate baseline industry context).

Hyperscaler 2027 capex is the real swing factor

The Northland downgrade (Jun 8, 2026) was explicit: the 2027 EPS consensus is "overly optimistic, anticipating a potential decline in AI infrastructure spending in 2027 due to hyperscaler capital expenditure constraints." This frames the bear thesis precisely — it is not about AMD execution but about the cycle turning before AMD's MI450/MI500 ramp matures. AMD targets $100B annual data center revenue within 5 years (Nov 11, 2025 Analyst Day), which assumes a continued AI capex up-cycle through 2030.

CoWoS scaling and the binding constraint

Year	TSMC CoWoS capacity (wpm)	Allocation note
2025	65–75K (some estimates 35K)	NVIDIA bookings dominant
2026	120–130K	NVIDIA 800–850K wafers (>50%); AMD + Broadcom share residual
2027	Expanding further	Panel-level / hybrid bonding options emerging

AMD's escape attempt — panel-level EFB packaging flagged by Simply Wall St — is the most-discussed margin lever in the file. If it works at scale, gross margin improves and unit ceiling lifts; if it doesn't, MI450 ramp is capped by allocation.

Custom-silicon TAM compression

Custom silicon program	Status	Implication for AMD merchant accelerators
Google TPU	At scale; Anthropic 3.5 GW commit starting 2027	The "only really successful ASIC at scale" per Bernstein
AWS Trainium	Ramping; co-developed with Annapurna	Replaces merchant for AWS internal workloads
Meta MTIA	In production for ranking/recommendation	Note: Meta still chose AMD for MI450 6 GW
Microsoft Maia 100	Dual-track with NVIDIA Rubin	Smaller scale but a precedent

The variant bear: hyperscaler internalization cannibalizes 2027+ merchant TAM faster than consensus models.

Semi-custom trough — the 2026 numbers headwind nobody talks about

AMD guided that 2026 semi-custom (console SoC) revenue will decline by "a significant double-digit percentage" — 7th year of the PS5/Xbox cycle. PS6 ("Orion," 3nm, ~280mm²) and Xbox "Magnus/Project Helix" (3 Zen 6 + 8 Zen 6c, 68 RDNA 5 CUs, 48GB GDDR7) only ramp in 2027. Bull headline revenue forecasts (+43% in 2026, +54% in 2027) absorb this trough; embedded segment (post-Xilinx) is also a drag in the near term (Q1 2025 Embedded $823M, -3% YoY) before recovery in test & measurement / comms / aerospace.

---

Governance & People

Board. Eight directors, four committees, lead independent director Nora Denzel. Recent addition KC McClure (former Accenture CFO, Mar 2025 retirement; also on Goldman Sachs board) strengthens audit oversight — net positive signal. John Marren (Temasek N. America Vice Chairman) brings sovereign-wealth connectivity. Departure: Olson did not stand for re-election (Xilinx cohort retiree). No abrupt departures.

Executive comp. Lisa Su 2025 total comp ~$55.16M (Simply Wall St) vs ~$31.0M in 2024 — roughly +78% YoY. CEO/median pay ratio 216x. Phil Guido (CCO, hired May 2023): $9M sign-on package + $725K base + 125% target bonus, with PRSU tied to relative TSR (good design). Say-on-pay endorsed at the May 5, 2026 annual meeting with no disclosed dissent — even with Su's comp doubling.

Ownership. Institutional 71.34%, insider 0.50%. Vanguard 8.3%, BlackRock 7.2%, State Street 4.0%. ARK Invest trimmed in May 2026 (part of a $167M sell). No activist 13D filed. No related-party concerns of substance in the modern era.

---

Specialist Question Reference Grid

Most specialist answers worth surfacing are promoted into the ten ranked findings above. The remainder — useful for triangulation, not thesis-moving — is collapsed here. Includes questions where evidence was thin (flagged "limited").

---

Where the Edge Sits

The market knows AMD won the OpenAI deal, the Meta deal, the EPYC share gains, and the AI capex up-cycle. Sell-side targets are now catching up to a stock that has already passed them. What the market is still partially missing:

1. The OpenAI warrant's ASC 606 contra-revenue mechanic — accounting drag begins when first 1 GW ships in 2H 2026.
2. The CoWoS allocation ceiling — order book exceeds AMD's ability to ship.
3. MI450 gross-margin dilution confirmed by management but absorbed into >35% LT non-GAAP operating-margin target.
4. 2027 hyperscaler capex slowdown risk — Northland's bear call, not yet consensus.
5. Custom-silicon TAM compression — Bernstein's variant view.
6. Insider conviction at multi-year highs is uniformly bearish — no NEO is buying.

What is not the edge — and where short theses tend to overreach: AMD's books are clean, the auditor is unbroken, there is no SEC matter, no NEO departure, and no activist. The risks here are strategic, geopolitical, and valuation — not accounting.