AMD › analyze
AMD · Analyze
A portfolio-specific passage was removed from the public build.
A portfolio-specific passage was removed from the public build.
⚠️ Verdict up front
WATCH · conviction 5.0 · fair value $380–480 (central ~$430) · spot $476.67 sits AT/ABOVE the top of it.
Three things need saying before any number:
- The watchlist multiple was wrong by ~2x. AMD does not trade at 32.2x forward. It trades at
63x FY2026 non-GAAP, ~43x on a blended next-twelve-months basis, and 30.9x only on
FY2027 — which is what Yahoo's
forwardPEfield actually contains. - The
Trim $360+instruction was not merely stale — it was pointing the wrong way. $360 is below fair value. Acting on it would have sold a position 25% under the analysis's own floor. The correct fix was always to re-derive it upward, and nobody did, for months. - The largest fact about this company is not in any published multiple. AMD has issued 320 million warrants at a $0.01 strike to OpenAI and Meta — 16.2% of the post-issue company, ~$152B of value at spot — and none of it has touched the financial statements.
1 · Fundamentals
Headline scorecard
| Metric | Value | Read |
|---|---|---|
| TTM revenue | $41.30B | Q2-26 alone +50% YoY |
| TTM operating income | $6.49B (15.7% margin) | |
| TTM GAAP diluted EPS | $3.89 | vs FY2025 $2.65 |
| TTM FCF | $8.40B (20.3% margin) | |
| Revenue 3y CAGR | 13.6% | FY2022–25; understates the current ramp |
| Net income 3y CAGR | 48.6% | off a depressed FY2023 base |
| FCF 3y CAGR | 29.3% | lumpy, not a trend line |
| ROIC (FY2025) | 4.5% | below any sane cost of capital |
| ROIC (TTM) | ~8.1% | improving, still weak |
| ROE / ROA | 10% / 5% | on 11.6x book |
| Debt/Assets | 5.0% | net cash $1.69B — balance sheet is not the issue |
| Goodwill | $25.13B = 32.7% of assets | Xilinx |
| Goodwill + acq. intangibles | $41.1B = 49% of assets | amortising $544M/quarter |
| Diluted shares | 1,630M → 1,659M YoY (+1.8%) | still diluting |
The two things the fundamentals actually say
(a) AMD earns a poor return on the capital it employs. ROIC of 4.5% in FY2025 and ~8% TTM is the single most under-discussed number on this name. The denominator is inflated by the Xilinx purchase — $41.1B of goodwill and intangibles, 49% of total assets — and the market is paying 11.6x book for an asset base earning single-digit returns. Contrast ANET in the same report at ~30% ROIC on almost no capital. AMD is not a high-return business that happens to be expensive; it is a moderate-return business priced as a high-return one.
(b) Buybacks do not cover dilution. TTM SBC $1.90B against TTM buybacks $1.18B — the repurchase programme retires 62% of what compensation issues, and the diluted count still rose 1.8% YoY. Per §1.4 of the framework this is a standing headwind to per-share value, before the warrants are considered.
🚩 Quality of earnings — a tripwire fired
Q2 2026 GAAP net margin 19.9% exceeded operating margin 17.3%. That is arithmetically impossible from operations, and it is the exact signature in [[pattern-net-margin-above-operating-margin-is-a-tripwire]].
| Q2-26 | $M |
|---|---|
| Revenue | 11,536 |
| Operating income | 1,990 (17.3%) |
| Other non-operating income | +598 |
| Net income | 2,297 (19.9%) |
At the quarter's 9.8% tax rate that $598M is $0.33/share — 24% of the reported $1.38 GAAP EPS.
Operating-only GAAP EPS was ~$1.05. The cash-flow statement shows Gain Loss On Investment
Securities −$483M, consistent with an equity mark under ASU 2016-01 —
[[pitfall-unrealized-equity-marks-break-headline-pe]]. Do not read Q2's GAAP EPS as an operating
result.
A second base effect compounds it: GAAP gross margin went 40% → 54% YoY, but the prior-year quarter carried the $800M MI308 export-control inventory charge. On a comparable basis non-GAAP gross margin moved 54% → 56%, +200bps, not +1,300bps. The headline margin expansion is an absent charge, not operating leverage.
Capital allocation — the AI build is now visible
| TTM | $B | Note |
|---|---|---|
| Operating cash flow | 10.08 | |
| Capex | 1.68 | Q2-25 $282M → Q2-26 $808M, +187% YoY |
| Free cash flow | 8.40 | 20.3% margin |
| R&D | 9.39 | 22.7% of revenue |
| Buybacks | 1.18 | |
| SBC | 1.90 |
Capex tripled year-over-year and H1-2026 alone ($1,197M) already exceeds all of FY2025 ($974M). Alongside $30.3B of unconditional purchase commitments (of which $17.4B falls in the rest of FY2026), rising inventory ($8,468M, WIP $5,265M) and ~$1.0B of new supply prepayments, this is the correct behaviour in a shortage and a large operating-leverage problem if AI demand air-pockets. Receivables grew +21% q/q against +13% revenue — worth watching, not yet a flag.
Per-share / owner's-eye view
| Reported | Fully diluted for warrants | |
|---|---|---|
| Revenue/share | $24.89 | $20.87 |
| FCF/share | $5.07 | $4.25 |
| FCF yield | 1.08% | 0.90% |
| P/FCF | 92.6x | 112x |
2 · The forward-EPS resolution — the pitfall fired, in the direction that costs money
Yahoo's payload carries both fields simultaneously and settles it without ambiguity:
epsCurrentYear : 7.56794 priceEpsCurrentYear : 62.99 <- FY2026
epsForward : 15.45073 forwardPE : 30.85 <- FY2027
trailingEps : 3.94 trailingPE : 120.98 <- TTM GAAP
This is the offset-year mechanism in [[pitfall-vendor-forward-eps-is-the-wrong-fiscal-year]] — the vendor "forward" points at the next fiscal year, ~16 months out. The note's own tripwire fires cleanly: implied forward EPS $15.45 against TTM $3.94 is +292%, an order of magnitude past the ~25% threshold.
Reconstruction from reported quarters + the company's own guide:
| Revenue | GAAP dil. EPS | Non-GAAP dil. EPS | |
|---|---|---|---|
| Q1 2026 reported | $10,253M (+37.8%) | $0.84 | $1.37 |
| Q2 2026 reported | $11,536M (+50%) | $1.38 | $1.66 |
| Q3 2026 guide | ~$13.0B ±$300M | — | ~$1.93 (derived; street $1.93) |
| FY2026E | ~$50.8B | ~$6.00 (low confidence) | $7.57 |
| FY2025 actual | $34.6B | $2.65 | $4.17 |
The corrected multiple ladder
| Basis | EPS | P/E | Fully diluted | P/E f.d. |
|---|---|---|---|---|
| TTM GAAP | $3.89 | 122.5x | $3.26 | 146.2x |
| FY2026 GAAP (derived) | ~$6.00 | 79.4x | $5.03 | 94.8x |
| FY2026 non-GAAP | $7.57 | 63.0x | $6.35 | 75.1x |
| NTM blended | $11.01 | 43.3x | $9.23 | 51.6x |
| FY2027 non-GAAP (what Yahoo shows) | $15.45 | 30.9x | $12.95 | 36.8x |
Corroborated at stockanalysis ($7.57 / 62.99x) and Bernstein (2027E $14.60).
⚠️ New failure mode for the playbook. stockanalysis.com published two different forward P/Es on two of its own pages on the same day at the same price — 62.99x (Forecast, FY2026) and 43.30x (Statistics, NTM), 46% apart. The existing guard "cross-check one independent vendor" is insufficient. The guard must be: check which year each number is on, including within a single vendor.
Own-band test — per [[pattern-ai-levered-fields-trade-above-own-band]]. AMD's GAAP P/E is unusable (the $544M/qtr amortisation swamps it), so use P/S:
| Period | Revenue | Mkt cap | P/S |
|---|---|---|---|
| FY2022 | $23.60B | $121B | 5.1x |
| FY2023 | $22.68B | $238B | 10.5x |
| FY2024 | $25.79B | $188B | 7.3x |
| FY2025 | $34.64B | $386B | 11.1x |
| TTM now | $41.30B | $777B | 18.8x |
18.8x is 70% above AMD's own highest prior reading. The stock is −18.5% off its 52-week high and +219% off its 52-week low — the drawdown is far smaller than the run-up, exactly the shape that note predicts.
3 · Moat
Evergreen rating: 6.0 / 10. (Moat Analyst)
The blunt finding: this is two companies with opposite moat profiles bolted together, and the market is paying for the weaker one.
Segment reality, Q2 2026
| Segment | Revenue | YoY | Op income | Op margin | % of revenue |
|---|---|---|---|---|---|
| Data Center (EPYC + Instinct + Pensando) | $6,718M | +107% | $2,103M | 31.3% | 58.2% |
| Client & Gaming | $3,841M | +6% | $582M | 15.2% (was 21.2%) | 33.3% |
| — Client | $3,062M | +23% | |||
| — Gaming | $779M | −31% | |||
| Embedded (Xilinx) | $977M | +19% | $386M | 39.5% | 8.5% |
"All Other" drag: −$1,081M/quarter ($544M intangible amortisation + $503M SBC) — 35% of segment operating income erased below the segment line.
Moat sources, named
| Source | Verdict |
|---|---|
| Network effects | No in GPU — CUDA has it; ROCm is a follower ecosystem AMD funds itself. Borrowed in x86 — the 10-Q states AMD's ability to innovate "depends partially on Microsoft designing and developing its operating systems to run on or support our x86-based microprocessors." That is a moat AMD is a tenant in. |
| Switching costs | High in Embedded/FPGA (Vivado/Vitis lock, 10–20yr programme lives). Moderate in server CPU — but an x86→x86 port is trivial, which is precisely why AMD could take Intel's share; the door swings both ways. Negative in AI GPU — AMD is paying customers equity to switch to it. |
| Cost advantage | No. Same foundry as Nvidia, worse allocation (~11% vs ~60% of CoWoS), less HBM buying power. Scoreboard: 56% gross margin vs Nvidia's ~73–75%. |
| Intangibles | Yes — the most underrated. The x86-64 cross-licence with Intel is irreplaceable and non-transferable. Plus the Xilinx IP estate. |
| Efficient scale | No in CPU/GPU; arguably yes in FPGA. |
AMD's genuine moat is IP-licence + x86 duopoly + FPGA switching costs. The AI GPU business that carries essentially the entire $778B market cap has no durable moat today — it has design wins that were bought with equity.
🚨 The warrants — the largest fact on this name
Verbatim, Q2 2026 10-Q Note 12:
"In October 2025 and February 2026, the Company issued warrants to OpenAI OpCo, LLC and Meta Platforms, Inc., 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 warrants vest in tranches based on specified AMD Instinct GPU purchase milestones … and achievement of specified stock price targets." "The warrants will be classified as liabilities until certain conditions for equity classification are met." "As of June 27, 2026, no warrant shares had vested or become exercisable, and the warrants had no impact on the Condensed Consolidated Financial Statements."
| Warrants | 320M shares @ $0.01 (OpenAI 160M + Meta 160M) |
| Diluted shares today | 1,659M — the warrant shares are in none of them |
| Fully diluted | 1,979M → 16.2% of the post-issue company |
| Value transferred at spot | ~$152.5B (19.6% of market cap) |
| Plus | $4.1B of contingent guarantees on customer data-centre leases, "may be issued in exchange for warrants" |
| P&L impact to date | Zero |
Plus the reflexive part, and it is the whole point: the final Meta tranche carries a reported ~$600 share-price target. The dilution triggers on the thesis working. You cannot own the upside without funding the giveaway — the 320M shares land precisely when the multiple is being justified.
This is the large-cap instance of [[pattern-customer-warrants-are-contra-revenue-that-buys-the-backlog]], with one difference that matters: CBRS is already amortising its customer warrants through revenue; AMD's cost is entirely prospective. The 10-Q is silent on how the charge gets recognised once a tranche vests. Under ASC 606 a milestone-vested warrant issued to a customer is the textbook case for a reduction of revenue. When the first tranche vests — plausibly Q4-26 or Q1-27, since Meta's first-gigawatt shipments start 2H26 — reported revenue growth and gross margin may both step down. This is the single highest-value item to re-check at the Q3 print.
The one-line test: a company that pays its customers in equity to take the product does not have pricing power, whatever the growth rate says.
Adversarial stress test
vs Nvidia. The revenue comparison is the honest framing: Nvidia Data Center $89.0B (quarter ended 2026-07-26, +117%) against AMD Data Center $6.7B — and AMD's number includes EPYC. ~13:1 on total DC; roughly 30:1 on accelerators alone. Secondary estimates put AMD at ~5–7% of AI accelerator revenue. Nvidia leads on scale-up fabric (NVLink 1.8 TB/s vs UALink 800 GB/s, systems "in earnest" 2027), on rack-scale production experience, and decisively on supply.
Inference is genuinely the more winnable market — 60–66% of AI compute in 2026, memory-capacity bound (MI450 carries 432GB HBM4), ROCm's gap 5–10% there vs 20–30% in training. The caveat that matters: frontier inference is migrating to rack-scale disaggregated serving (Nvidia's Rubin CPX for long-context prefill, end-2026), which re-imports the scale-up networking advantage into the market AMD was supposed to win — and inference is also where custom ASICs are strongest. AMD wins the easier market against more competitors at worse prices.
vs Intel — AMD is winning, and this updates our own playbook. [[pattern-shortage-pricing-masks-share-loss]] was written from Intel's side. Mercury Research Q2 2026 has now fired that note's own falsification condition, in AMD's favour:
| Segment (units) | Q1'26 | Q2'26 | YoY |
|---|---|---|---|
| Overall x86 | 32.6% | 34.1% | +4.7pt |
| Client | 29.6% | 30.3% (first time >30%) | — |
| Desktop | 33.1% | 34.9% | +2.7pt |
| Mobile | — | 28.9% | +8.4pt |
| Server (ex-IoT) | 33.2% | 34.5% | +7.3pt |
Intel added millions of units of mobile capacity in Q2 after two supply-constrained quarters and still lost 8.4 points of mobile share YoY. Supply relief did not restore share — so AMD's gains have a genuine product component that Intel's revenue growth did not. AMD's gains are also concentrated in the high-ASP datacenter socket (~46% on the narrow Xeon-SP-vs-EPYC basis against 34.5% broad), meaning Intel is retaining the cheap sockets.
⚠️ Basis conflation to correct in our own notes. Our playbook records 46.2% as Q1-26 server revenue share. The 46.4% now quoted for Q2-26 is Mercury's narrow-basis unit share (Xeon SP vs EPYC only). Press coverage blurs the two. Do not carry "46.2% revenue share" forward as a Q2 number; 34.5% server units is the figure to stand behind.
But the x86 duopoly itself is under attack, and not by Intel. Mercury Q2-26: Arm at 13.6% of server units and 15.3% of client, both records. Every Nvidia rack ships an Arm CPU and every AMD rack ships an EPYC — AMD's CPU franchise is partly hostage to its own GPU share. It is taking share of a pool shrinking as a fraction of datacentre compute.
vs custom ASICs. One number frames it: Broadcom's single quarter of AI semiconductor revenue ($10.8B, +143%) exceeds AMD's entire Data Center segment ($6.7B), CPUs included. Broadcom guides
$100B AI revenue in FY2027. Custom ASIC shipments +44.6% YoY vs merchant GPU +16.1%. Meta and OpenAI are simultaneously AMD's two warrant-holding anchor customers and two of Broadcom's six XPU customers — they are dual-tracking: use AMD to break Nvidia's pricing while migrating steady-state inference onto their own silicon.
Read: the bear case is stronger on economics, the bull case on volume. AMD will very likely grow Instinct revenue several-fold and is much less likely to earn Nvidia-like margins on it. AMD can win the units and lose the economics — and the warrant structure is management encoding exactly that trade.**
Dependency risk
- TSMC, verbatim from the 10-Q: "We rely on TSMC for the production of all wafers for microprocessor and GPU products at 7nm or smaller." No long-term capacity commitment with some suppliers — purchase-order basis.
- CoWoS: AMD ~11% vs Nvidia ~60% (>70% of CoWoS-L). AMD's Instinct ceiling is set by an allocation it does not control on a line whose largest customer is its principal competitor. Winning more orders does not produce more chips.
- HBM: all three suppliers sold out through 2026 on non-cancellable contracts; SK hynix supplies ~2/3 of Nvidia's HBM4. This is the mechanical reason AMD guides 56% gross margin — and why "flat 56%" spooked the market more than the revenue beat pleased it.
- China ATMP, under-discussed. The 10-Q: "The majority of our ATMP services are provided by the ATMP JVs" — two joint ventures with Tongfu Microelectronics affiliates. Most of AMD's assembly/test sits inside a China-based JV structure. A distinct geopolitical exposure from Taiwan, rarely modelled.
- China/MI308: licences granted; AMD remits 15% of China AI-chip revenue to the US Treasury; MI325 licences granted but "we do not yet know whether any imports … will be allowed into China," subject to US inspection and a 25% tariff. Treat China as an option, not a line item.
4 · Sentiment
The Q2 print — and a correction to the watchlist's own account of it
The watchlist recorded that AMD fell 7% because "the Q3 revenue guide missed an aggressive bar." The revenue guide did not miss — it beat. ~$13.0B ±$300M against a $12.5B sell-side consensus. What actually happened:
| Actual | Consensus | ||
|---|---|---|---|
| Revenue | $11,536M (+50%) | ~$11.3B | beat |
| Non-GAAP EPS | $1.66 | $1.61–1.62 | beat |
| Non-GAAP GM | 56% | ~56% | in line |
| Q3 revenue guide | ~$13.0B | $12.5B | beat (buy-side whisper ~$14B) |
| Q3 GM guide | ~56%, flat q/q | expansion expected | this is the miss |
The stock had run +20.7% in the four sessions into the print ($429.56 → $518.58). It fell on a flat gross-margin guide — CFO Jean Hu attributing it to datacentre AI margins running "slightly below company average" as Helios/MI450 ramps. The de-rating was about margin and cash conversion, not growth.
⚠️ Several secondary outlets reported a "gross margin miss, 54% vs 56%." That is a GAAP number compared against a non-GAAP consensus. Do not propagate it.
Guidance that matters
Data Center to "more than double year-over-year in 2027"; server CPU +>70%; data-centre AI "well above 100%"; "tens of billions of dollars in annual Data Center AI revenue in 2027." When a Cantor analyst floated ~$30B of 2027 Instinct revenue, Lisa Su said that was too low.
Helios is "now in production with initial shipments on track to begin later this quarter."
AMD does not disclose Instinct revenue separately. The only visible line is Data Center, which bundles EPYC + Instinct + Pensando. Every "AMD AI revenue" figure in circulation is an analyst construct, not a disclosure.
Analyst positioning — crowded long, and the tell is in the ratings, not the targets
49 analysts, mean rating 1.48 (Strong Buy), mean target $613.84 (Yahoo) — but MarketBeat shows $553.72 on the same day. Use the range.
Post-print, every firm raised numbers and almost nobody upgraded. Morgan Stanley (Equal-Weight $465 — below spot), JP Morgan (Neutral, target +43% to $550) and RBC (Sector Perform $540) each marked their models to the print and kept the neutral rating. That is the tell. Only Raymond James upgraded. Short interest is 2.45% of shares out. Institutional ownership 75.4%. Zero sell-rated coverage of substance.
Insiders — unambiguous
Zero open-market purchases by any insider in the last 12 months, against ~$261M of open-market sales in calendar 2026. Lisa Su ~$157M (125k/quarter — textbook 10b5-1; she sold 210,000 shares in Feb/Mar at $197–219 and left enormous money on the table, which is what a pre-set plan looks like, so weight her cadence as noise).
What is worth noting is the August cluster: five officers sold ~$41.8M in the 14 sessions after the print, including Forrest Norrod, EVP of the Data Center business, selling into $453–468 on 08-24, and CTO Mark Papermaster 28,811 shares at $466–475 on 08-20. Whether these were 10b5-1 is not verifiable from the available feed. Treat as weak-negative, not signal.
ARK has been a persistent mechanical seller through all of 2026, rotating into NVDA, AVGO and CBRS — not exiting the theme. One manager's rotation; no information about AMD's fundamentals. (Worth the note: ARK is selling AMD, whose customer warrants have not yet hit the P&L, to buy CBRS, whose customer warrants already run through the revenue line.)
5 · Valuation
DYT and DDM are N/A — AMD pays no dividend. Stated rather than faked, per §3.
Graham IV = √(22.5 × 3.89 × 41.19) = $60.04; on tangible-ish BVPS of $23.23, $45.09. Price is 7.9–10.6x Graham. Graham is the wrong tool for a fabless IP business with no book anchor — reported for completeness, carries ~zero weight.
Reverse DCF (10% discount, 3% terminal, TTM FCF $8.40B): to justify $778B, FCF must compound 28.1%/yr for ten years, reaching $99.6B in year 10 — more free cash flow than Nvidia generates today. That is the bar the current price sets.
Bogle expected return, on FY2026 GAAP ~$5.20 base with no dividend:
| Growth | Exit 40x | Exit 55x |
|---|---|---|
| 20%/yr | +1.7%/yr | +8.3%/yr |
| 25%/yr | +5.9%/yr | +12.9%/yr |
| 30%/yr | +10.1%/yr | +17.4%/yr |
Fair value triangulation on FY2027 non-GAAP, fully diluted for the warrants ($15.45 × 1,659/1,979 = $12.95):
| Multiple | Value | Discounted 1yr @10% |
|---|---|---|
| 30x | $389 | $353 |
| 35x | $453 | $412 |
| 40x | $518 | $471 |
| 45x | $583 | $530 |
Cross-check on FY2026 non-GAAP fully diluted ($6.35): 50x = $317, 60x = $381, 70x = $444.
Fair value $380–480 · central ~$430 · spot $476.67 is at/above the top.
| Level | ||
|---|---|---|
| Entry | $330–380 | a real discount to central FV |
| Strong buy | <$300 | |
| Trim | $545 | ≈42x FY2027 non-GAAP fully diluted |
⚠️ The trim is written as a dollar deliberately — do not convert it to a multiple.
site.pycomputesEPS = price ÷ forwardPE, so aTrim NNx fwdon AMD would use $476.67 ÷ 30.85 = $15.45 — which is (a) the FY2027 estimate, not the current year, and (b) excludes all 320M warrant shares. Both defects push the rendered level too high:Trim 40x fwdwould render $618,45xwould render $695. Worse, it would faithfully capitalise the exact number this analysis doubts — the failure mode named in [[pitfall-multiple-trim-inverts-on-peak-cycle-cyclicals]] and [[pitfall-multiple-trim-inherits-the-broken-vendor-field]]. Same call as MU and CBRS. This dollar trim must be re-set at the next/analyze; that is the cost of not using the multiple form.
6 · Tensions surfaced (Phase 3)
Fundamentals vs Moat — resolved in Moat's favour. Fundamentals sees 50% revenue growth, 107% Data Center growth, record share in every x86 segment, and a clean balance sheet. Moat sees ROIC of 4.5–8%, a 56% gross margin against Nvidia's 73%, an 11% CoWoS allocation, and anchor customers bought with 16% of the equity. Both are true and they describe different halves of the company. The reconciliation: the CPU + FPGA core (~55–60% of revenue) is genuinely evergreen and rates ~8/10 standalone; the AI GPU business that carries the entire multiple rates 3–4/10. The capital, the R&D, the management attention and the whole valuation are allocated to the weak side, while the best-moat segment (Embedded, 39.5% margins) is 8.5% of revenue and shrinking as a share of mix.
Sentiment vs Valuation — resolved in Valuation's favour. Sentiment reports a genuine beat-and-raise, a 1.48 Strong Buy consensus and a $613 mean target. Valuation observes that the same consensus contains three of the largest houses declining to upgrade, that P/S is 70% above AMD's own prior peak, and that the price requires 28%/yr FCF compounding for a decade. A raised target from a firm that did not raise its rating is a mark-to-market, not a call.
A portfolio-specific passage was removed from the public build.
8 · Verdict
WATCH · conviction 5.0 · fair value $380–480 · spot $476.67
Not a trim at $360 — that instruction was pointing the wrong way and should be retired, not executed. Not an add either: the stock is at the top of a fair-value band that already assumes the FY2027 bull case, and the one asset the whole valuation rests on has no moat and a 16% equity bill that has not been paid yet.
The bull case, stated fairly. EPYC is a genuine, structural, high-ASP share-taker that just passed the supply-relief test with a decade of runway. Embedded is a high-margin annuity with >$50B of cumulative design wins barely converted. ZT Systems bought AMD the system-integration competence it structurally lacked, cheaply and well. Management guides Data Center to more than double in 2027, and Lisa Su called a $30B Instinct estimate too low. In a >$1T compute TAM, even 7% share is an enormous business.
The bear case, and why it wins on price. AMD holds ~5–7% of AI accelerators, ~11% of the CoWoS capacity that physically caps how many it can build, a residual HBM allocation, and a 56% gross margin against Nvidia's 73%. Its two largest AI customers are simultaneously building competing silicon with Broadcom. To secure them AMD issued 320 million shares' worth of warrants at a penny and guaranteed $4.1B of their leases — and none of it has touched the financial statements, because vesting is tied to milestones and to a stock price not yet reached. The threat is reflexive: the more the AI thesis works, the more of the company transfers to the customers who made it work, and the dilution arrives exactly when the multiple is being justified.
Conviction 5.0. A good business paying for a great story. The price already contains the outcome management is guiding to, and does not contain the cost of achieving it.
Key risks
- Warrant recognition. How the cost lands on first vesting — contra-revenue would hit reported revenue growth and gross margin simultaneously. Unknown; the 10-Q does not say.
- Margin, not growth. Flat 56% gross margin is the constraint; it is set by HBM and CoWoS AMD does not control.
- Custom ASIC absorption of the volume tail by AMD's own anchor customers.
- Arm shrinking the x86 pool faster than AMD gains share of it.
- $30.3B of unconditional commitments against demand financed by circular AI capital.
- Taiwan — non-diversifiable, and China ATMP as a second, separate exposure.
Open questions — explicitly unresolved, do not fill in
- AMD's actual Instinct revenue. Not disclosed.
- How warrant cost is recognised on vesting. The 10-Q is silent. Highest-value open question.
- FY2026 GAAP EPS consensus — the ~$6.00 used here is a derivation, not a published consensus.
- Whether the August insider sales were 10b5-1 — Form 4 footnotes not retrieved.
- MI308 China revenue recognised in FY2026 — never quantified.
- Whether the 46% server figure is units or revenue — sources conflict.
- Reported Nvidia Kyber delay to 2028 — unconfirmed, and the highest-value item to monitor if true, since it would give AMD its only scale-up topology window this decade.
Sources
AMD Q2 2026 10-Q (period ended 2026-06-27) · AMD Q2 2026 earnings release and slides (ir.amd.com) ·
Q2 2026 earnings call transcript · Mercury Research Q2 2026 x86 share (via The Register,
2026-08-21) · Nvidia Q2 FY2027 results · Broadcom Q2 FY2026 results · Yahoo Finance
get_stock_info / get_financial_statement / get_recommendations / get_holder_info ·
stockanalysis.com · MarketBeat.
Playbook notes applied: [[pitfall-vendor-forward-eps-is-the-wrong-fiscal-year]] · [[pattern-net-margin-above-operating-margin-is-a-tripwire]] · [[pitfall-unrealized-equity-marks-break-headline-pe]] · [[pattern-customer-warrants-are-contra-revenue-that-buys-the-backlog]] · [[pattern-ai-levered-fields-trade-above-own-band]] · [[pattern-shortage-pricing-masks-share-loss]] · [[pitfall-stale-entry-zone-suppresses-a-name]] · [[pitfall-multiple-trim-inverts-on-peak-cycle-cyclicals]] · [[pitfall-multiple-trim-inherits-the-broken-vendor-field]] · [[pattern-ai-build-inflates-earnings-while-destroying-fcf]]