Advanced Micro Devices

Phase A — Understand the business

Company Overview

AMD designs CPUs, GPUs, FPGAs, DPUs, and adaptive SoCs — fabless, outsourcing manufacturing to TSMC (leading-edge logic) with packaging through TSMC CoWoS/SoIC and OSATs. It is the only credible second source to the incumbent on two of the most valuable franchises in computing simultaneously: x86 server/PC CPUs (vs. Intel) and data-center AI GPUs (vs. NVIDIA). That dual-front position is the whole story.

Four reportable segments (FY2025, full year):

• Data Center — $16.6B (32% YoY), server CPUs (EPYC), data-center GPUs (Instinct), plus Pensando DPUs and Xilinx data-center products. The growth engine.
• Client and Gaming — record $14.6B (+51% YoY): Client (Ryzen) $10.6B (+51%), Gaming (semi-custom consoles + Radeon) $3.9B (+51%).
• Embedded — $3.5B (−3% YoY), the Xilinx FPGA/adaptive-SoC franchise, soft on customer inventory digestion earlier in the year.
• Total FY2025 revenue $34.6B (+34% YoY) vs. $25.8B in FY2024.

Customers. Hyperscalers and OEMs. On the GPU side the marquee names are now signed, not aspirational: Microsoft Azure (MI300X in production), Meta (MI300X deployed; committed to up to 6 GW of Instinct across generations including custom MI450 in the Helios rack), Oracle/OCI (MI355X rack-scale; a 50,000-unit MI450 cluster; zettascale clusters up to 131,072 MI355X), and OpenAI (the 6 GW strategic partnership, MI450 first 1 GW H2 2026). On CPU, the EPYC franchise sells into all major clouds and enterprise OEMs. Console semi-custom = Sony/Microsoft.

Suppliers / contract structure. TSMC is the binding upstream relationship (wafers + CoWoSCoWoSTSMC’s method of packing a processor and its memory onto one carrier so they sit close together. Supply of it has been a hard limit on how many AI chips can be built. allocation); HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips. from SK Hynix / Samsung / Micron. Revenue is largely transactional product sales, not take-or-pay — there is no recurring/subscription floor. The OpenAI deal is a commercial commitment to purchase tied to a warrant (below), not a guaranteed revenue contract; volumes vest milestones, not minimums. ``.

---

Supply Chain

Map, with named stakeholders (the chain ; AMD-specific reads ):

Upstream inputs → AMD → end customer

1. EUV lithography — ASML (monopoly). Single vendor for the scanners TSMC needs at N3/N2. Scanner delivery cycle 18–24 months; High-NA EUV delayed. AMD is exposed only through TSMC, but it is a hard ceiling on leading-node supply.
2. Leading-edge foundry — TSMC (with Samsung/Intel as distant alternatives). AMD's MI350/MI400 and EPYC are TSMC N3/N2-class. AMD shares this queue with Apple and NVIDIA — TSMC's top-3 (Apple, NVIDIA, AMD) are ~50% of TSMC revenue. AMD is third in that priority line behind Apple and NVIDIA — a structural disadvantage in a shortage.
3. HBM memory — SK Hynix > Samsung > Micron. MI300/MI350 carry large HBM3E stacks; MI400 moves to HBM4 (432 GB/GPU). HBM is a named severity-9 bottleneck: 2026 capacity reportedly pre-booked by NVIDIA + OpenAI; SK Hynix + Samsung raised HBM3E prices ~20% into 2026. AMD's MI-series differentiation has historically been its HBM capacity advantage — so HBM allocation is both its weapon and its risk.
4. Advanced packaging — TSMC CoWoS (overflow to ASE CoWoP, Amkor, Intel). Severity-9 chokepoint, binding through 2026: ~35K wafers/mo (late-2024) → 75K (end-2025) → 130K target (late-2026); NVIDIA secures ~60% of capacity. Easing late-2026 disproportionately benefits the number-two GPU vendor — i.e., AMD. This is the single most important supply-side swing factor in AMD's favor.
5. AMD (fabless design) → integrates into the Helios rack-scale system (72 MI400 accelerators linked by UALink/Ultra Accelerator Link).
6. End buyers — hyperscalers (Microsoft, Meta, Google, Amazon, Oracle) + OpenAI + sovereign-AI + enterprise.

Chokepoints / single-source dependencies: EUV (ASML, no substitute); CoWoS (TSMC, limited substitutability); HBM4 (SK Hynix-led, multi-quarter lead times). AMD's deepest single-source dependency is TSMC — for both wafers and packaging — where it ranks behind Apple and NVIDIA for allocation. Names present; lens passes.

---

Competitive Advantages (moats)

for the matrix and moat/vulnerability framing; AMD-specific reads.

Where AMD is genuinely advantaged:

• Only credible x86 server second source. EPYC has taken durable share from Intel; FY2025 Data Center CPU strength is real and Intel's server roadmap keeps slipping. This is a structural moat (x86 duopoly, validated installed base, multi-year design wins) and the most defensible part of AMD — and it is not the part the multiple is paying for.
• Memory/capacity edge on Instinct. MI355X claimed ~30% faster inference than NVIDIA B200 on Llama 3.1 405B and ~40% better tokens-per-dollar; MI400 = 432 GB HBM4 vs. less on Vera Rubin-class parts; Helios claims ~50% higher memory capacity/bandwidth vs. VR-NVL144. For inference (memory-bound), this is a real wedge — and it aligns with the house thesis that AI compute is substantially a memory-movement problem ``.
• Second-source / open-ecosystem position. Every hyperscaler wants a credible NVIDIA alternative to discipline pricing. AMD is the only merchant-silicon answer (vs. captive ASICs). UALink (open) vs. NVLink (proprietary) is AMD's coalition play.

Where the moat is thin — the central bear fact:

• ROCm vs. CUDA. NVIDIA still holds ~86% of data-center GPU revenue (down from ~90% in 2024); AMD ~5–7% of AI-GPU share. CUDA is a ~20-year software moat with system-level lock-in (cuDNN, TensorRT-LLM, NCCL). ROCm 7 closed the gap (≈within 10–30% of CUDA on most workloads; PyTorch first-class, JAX full support, OpenAI Triton generating AMD kernels) — but "10–30% behind" is still behind, and switching cost NVIDIA→AMD is rated High ``. The moat AMD needs is software, and software is exactly where it is weakest.

Bargaining power: Over customers — low to moderate (hyperscalers are giants buying a second source partly to gain leverage). Over suppliers — weak on the binding inputs (third in line at TSMC; price-taker on HBM). AMD needs TSMC and HBM more than they need AMD. The OpenAI warrant is, in part, AMD paying for demand it could not otherwise lock.

---

Segments

our figures is empty → all figures ``. (Flagging: research-layer requirement unmet because the CSV has no rows.)

The shape of the story: Data Center is now ~48% of revenue and the only segment that matters to the valuation. Within DC, the mix is shifting from CPU (steady share-gain compounding) to GPU (the hyper-growth, lumpy, customer-concentrated piece). Q1-2026's DC +57% vs. FY25's +32% says the GPU ramp is inflecting. Caveat to model: AMD does not cleanly disclose Instinct-GPU revenue separately from EPYC within Data Center — a forensic flag (Lens 10) and the single biggest gap for sizing the AI franchise.

---

Phase B — Measure performance

Earnings Result (Q1 2026, reported ~2026-05-05)

``

• Revenue $10.3B, +38% YoY — beat consensus ~$9.84B.
• Non-GAAP EPS $1.37, +43% YoY — beat consensus ~$1.29.
• Non-GAAP gross margin 55%, +170 bps YoY.
• Segment drivers: Data Center $5.8B (+57%) — EPYC + Instinct; Client & Gaming +23% to $3.6B; Embedded +6% to $873M.
• Cash flow / balance sheet: record FCFFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. $2.6B (~25% margin); cash + ST investments $12.3B.
• Q2-2026 guide: revenue ~$11.2B ±$300M (~+46% YoY, +9% QoQ at midpoint), non-GAAP GM ~56%. But management flagged consumer/gaming revenue declining QoQ on memory/component cost inflation — a margin/mix watch-item.
• Market reaction: stock +16–18% after hours — the beat-and-raise + DC acceleration was rewarded hard, confirming the market is trading AMD as an AI-GPU story, not a diversified-semi story.

Unusual vs. own history: DC growth re-accelerated (57% vs. FY25's 32%) — the Instinct ramp is the cause. The offset is the explicit gaming-revenue-down guide on input-cost inflation (HBM/memory pricing is now a cost headwind to AMD's own non-DC products, the flip side of the supply tightness that helps its GPU competitive position).

---

Earnings Calls (sentiment trend)

transcripts/ is empty → ``.

• Tone trajectory (last ~3–4 calls): rising conviction, increasingly specific. Through 2025 management moved from "we have a competitive roadmap" to hard commitments: "strong and increasing confidence in our ability to deliver tens of billions of dollars in annual data-center AI revenue in 2027" (Lisa Su). The signing of OpenAI (Oct-2025) and the Meta 6 GW expansion converted the narrative from hope to backlog.
• Recurring phrases: "open ecosystem," "rack-scale" (Helios), "tens of billions … 2027," "annual cadence" (MI350→MI400→MI500), "data-center AI."
• What they stopped saying: the older hedged "ROCm is maturing / multi-year journey" language has been replaced by win-citations ("seven of the top-10 AI model builders run on Instinct"). The software gap is now framed as closing rather than as the central risk.
• Read: sentiment is at a multi-year high and management has staked credibility on a specific 2027 number — which makes any 2027 stumble a sharp de-rate catalyst. Confidence is a double-edged provenance: it's a real signal and it's now priced.

---

Comps

Peer set: the AI-accelerator + large-cap logic cohort. **Multiples are with source/date; where not sourced → `n/a`.** Pulled.

The damning line: AMD at ~58x forward trades at ~2.4× NVIDIA's forward multiple (23.8x) and ~2.3× Broadcom's (25x) — i.e. the second-place franchise is priced richer than the runaway leader and the ASIC king. PEG ~1.0–1.26 only "rescues" the multiple if the ~50–60% forward EPS growth actually compounds for years. EV/EBITDA 110x and trailing P/E 169x are nosebleed; the bull must lean entirely on forward numbers, which is exactly where dilutionDilutionIssuing new shares, so each existing share owns a smaller slice of the same company. (Lens 10/12) bites.

---

Stock-Price Catalysts (>5% moves, ~5-yr pattern)

``. AMD is a high-beta name; 12 largest up-days over 3 years averaged +12.5%, largest single up-day +23.8% (2025-04-09).

What actually moves the stock (the pattern):

1. AI-GPU commercial milestones — the biggest discrete catalysts. AMD rose ~24% on the OpenAI 6 GW deal (2025-10-06); MI300/MI350 ramp news; Oracle/Meta deal headlines. This is now the dominant driver.
2. Earnings beat-and-raise on Data Center — Q1-2026 +16–18% AH. The market reacts to DC growth rate and the 2027 AI number, far more than to Client/Gaming.
3. NVIDIA read-through & sector beta — AMD trades as the NVIDIA-derivative/second-source; Broadcom's −14% on soft AI guide (2025-06-04) dragged AMD with it.
4. Macro / rates / China export-control headlines — export controls are a named severity-7 regulatory bottleneck ``; MI-series China-SKU news moves the stock.

What it reveals: the market prices AMD almost entirely off the AI data-center trajectory and the NVIDIA-share-gain narrative. CPU share gains and the diversified base barely register. That concentration of attention cuts both ways — a single DC disappointment or an OpenAI-deal wobble is a disproportionate de-rate risk.

---

Phase C — Judge people & books

Management

``.

1. Track record — elite, quantified. Lisa Su (CEO since Oct-2014, also Chair) executed one of the great corporate turnarounds: AMD from ~$1.61 stock / sub-$2B market capMarket capitalisationThe share price multiplied by the number of shares. What the market says the equity is worth. / ~$2.5B debt in 2014 to a $850B company in 2026 — stock +~9,000% over the decade. The Zen architecture bet and the disciplined EPYC server assault against Intel are textbook. This is a top-quartile semiconductor operator. Credit the bench too — Mark Papermaster (CTO), Jean Hu (CFO), Forrest Norrod (DC).
2. Tenure & skin in the game. ~11 years; Su owns ~3.6–4M shares (net worth ~$1.1–1.3B, almost entirely AMD). Aligned, though her ownership percentage of a $850B cap is small (founder-archetype incentives via equity comp, not a founder stake).
3. Capital allocation — mixed, improving. The $49B all-stock Xilinx acquisition (Feb-2022) — largest chip deal ever — and $1.9B Pensando (2022) built the Embedded + DPU franchises. Xilinx is strategically sound but Embedded was down 3% in FY25 and the deal was struck at a cyclical peak in stock terms; the jury on per-share value creation is still out. AMD also runs buybacks. ROE is only ~8.1% — depressed by the huge Xilinx goodwill/intangible base and stock-based comp, a genuine blemish on an otherwise stellar operating record.
4. Red flags (governance): the OpenAI warrant (160M shares at $0.01, up to ~10% of the company) and the Meta warrant arrangement together create ~320M shares of dilution to "buy" demand — a related-party-adjacent, shareholder-cost structure that flatters reported revenue while diluting per-share value (see Lens 10/12). Not fraud — but it is management paying equity for a customer, and bulls quoting EPS must net it.
5. Archetype: professional manager-engineer operating with founder-level conviction and a multi-decade horizon. The right profile for this stage — the risk is over-promising on the 2027 AI number to sustain the multiple.

---

Forensic Red Flags

our figures empty + no filings/ (wave constraint) → income-statement/balance-sheet specifics are ``; flag accordingly.

• Instinct-GPU revenue is not separately disclosed. AMD reports Data Center as one line (EPYC + Instinct + Pensando + Xilinx-DC). Investors cannot independently verify the AI-GPU run-rate vs. the CPU base — the single most important number is bundled. This is a disclosure-quality flag, not an accounting violation, but it lets management frame the AI ramp without granular accountability.
• GAAP vs. non-GAAP gap is wide. FY2025: non-GAAP dil. EPS $4.17 vs. GAAP $2.65 — a ~37% wedge. The bridge is dominated by amortization of Xilinx/Pensando intangibles + stock-based compensation. SBC materially flatters non-GAAP; the comps table's 58x forward P/E is on the non-GAAP number — on GAAP the multiple is far higher. Every bull EPS figure in this dossier is non-GAAP and must be read with the SBC + amortization caveat.
• Dilution overhang. ~320M warrant shares (OpenAI ~160M + Meta) against ~1.63B shares out = ~16–18% potential dilution layered on top of ordinary SBC. Per-share metrics will be structurally pressured as these vest 2026→. The OpenAI tranches vest against purchase milestones and stock-price targets escalating to $600 — partially self-funding, but real share creation.
• Goodwill / intangibles from Xilinx dominate the balance sheet; an Embedded-segment downturn raises (low but non-zero) impairment-test attention. ROE ~8% reflects this denominator.
• Inventory / receivables vs. revenue: not independently verifiable without the 10-QForm 10-QThe quarterly version of the annual report. Lighter, and not audited. (not ingested). Flag as n/a; to verify on next refresh.
• Cash flow quality looks solid: FY25 FCF $5.5B, Q1-26 FCF $2.6B (~25% margin) — FCF is converting, which mitigates the non-GAAP-flattery concern at the cash level even as per-share GAAP earnings lag.

Regulatory findings ``:

• SEC Litigation Releases: none naming AMD (2021-06-17→2026-06-17), per EDGAR EFTS LR search.
• AAER: none found (EFTS AAER search returned HTTP 500 on this run; the LR search completed clean — treat AAER as not-confirmed-clean rather than verified, and re-check next refresh).
• Non-SEC enforcement : no material new FTC/DOJ/EU consent decree, fine, or penalty against AMD surfaced in this run. (AMD's most relevant regulatory exposure is **US export controls on China AI-GPU SKUs** — a policy/revenue risk, not an enforcement action .)
• 10-K Item 3 (Legal Proceedings): not retrieved (filing not ingested per wave constraint) — n/a; verify on next refresh.
• Net: No material accounting/enforcement findings via SEC LR + web as of 2026-06-17. AAER channel and 10-K Item 3 unverified this pass — explicitly flagged, not asserted clean.

---

Phase D — Project & stress-test

Forward Projection

Build from FY2025 actuals (non-GAAP) into FY2026E / FY2027E / FY2028E. All inputs labeled; outputs `` with arithmetic. Per wave constraint, our model create was NOT run (unattended breadth loop).

Anchors ``: FY25 rev $34.6B, FY25 non-GAAP dil. EPS $4.17; Q1-26 rev $10.3B, Q2-26 guide ~$11.2B; ~1.63B shares out (pre-warrant); consensus FY26E EPS ~$6.64, FY27E ~$10.74; Visible Alpha DC-revenue FY27 $33.9B.

The number that matters: does AMD hit "tens of billions of DC AI revenue in 2027"? Visible Alpha says $33.9B total DC FY27 (CPU+GPU) — implying GPU somewhere ~$20B+. That is the load-bearing assumption under the ~58x multiple. If FY27 DC lands ~$34B and non-GAAP EPS ~$10, the forward P/E on FY27 is ~52x today — still rich, but the PEG holds only if growth persists into FY28.

Tracked-forecast line (for a future our model create, not logged this run): AMD FY26 non-GAAP EPS >= $6.50 — p ≈ 0.55 — resolves 2027-02-28 (FY26 results) — tags amd,deep-dive. A second worth logging: AMD FY27 Data Center revenue >= $30B — p ≈ 0.55.

---

Bull vs Bear

Bull case (narrative). AMD is the only merchant-silicon alternative to NVIDIA at the exact moment every hyperscaler and sovereign is desperate to avoid single-vendor dependence — and it now has the signed backlog (OpenAI 6 GW, Meta up-to-6 GW, Oracle zettascale) to prove the demand is real, not theoretical. The MI400/Helios generation closes the system-level gap (rack-scale, UALink, 432 GB HBM4) where AMD has historically lost, and on inference economics (tokens-per-dollar, memory capacity) it can win, not just tie — which matters because inference is the larger long-run market and is memory-bound, exactly AMD's strength . CoWoS capacity *quadrupling* by late-2026 disproportionately frees the number-two vendor . Layer in the still-compounding EPYC server-share story (a second, independent growth engine the market under-credits) and ROCm 7 reaching ~parity, and the path to "tens of billions" of AI revenue + $10+ EPS by 2027 is credible. A best-in-class operator (Su) is executing. AMD aims for 20% of a $433B AI-chip TAM by 2031.

Bear case (2–3 permanent-impairment risks).

1. The software moat never closes enough. "10–30% behind CUDA" `` is fine for inference bake-offs but loses the high-margin training franchise; if ROCm plateaus, AMD is permanently capped as a low-share inference-and-CPU vendor, not a co-leader — and the multiple implodes.
2. It's priced for perfection at ~58x forward. Richer than NVIDIA (24x) and Broadcom (25x) for the follower. Semiconductor history offers ~no examples of sustaining 50–60% EPS growth for five straight years. Any single DC quarter that merely meets (rather than beats) re-rates the stock hard — the Q1 +18% pop shows how much is hope.
3. The growth is bought with equity. ~320M warrant shares (OpenAI + Meta) dilute ~16–18% before any per-share number "means anything"; SBC + Xilinx amortization already push GAAP EPS 37% below non-GAAP; ROE is only ~8%. The reported AI win partly transfers value from existing holders to the customers.

Pre-mortem (18 months out, thesis broke): It's late-2027. MI400/Helios shipped late or with rack-integration/ROCm bugs; OpenAI's vesting milestones slipped and the 1 GW underdelivered; NVIDIA's Vera Rubin + CUDA held training share and quietly cut inference pricing, compressing AMD's tokens-per-dollar edge; HBM4/CoWoS tightness re-tightened and AMD (third at TSMC) got squeezed on allocation again. DC revenue came in at the low end, EPS missed the $10 bar, and a 58x→30x de-rate halved the stock even as revenue still "grew." The diversified base (EPYC, gaming) couldn't offset a GPU-narrative break.

Are multiples too high? On absolute and relative terms — yes, the multiple already prices the bull case. The upside from here is "AMD compounds 50%+ for years and the 58x is justified ex-post"; the asymmetry is unfavorable at spot given the analyst average PT ($486) sits below the price (~$521).

Contrarian view (what the market refuses to see): The market is fixated on the GPU fight and treats EPYC as boring — but the server-CPU share gain is the more durable, higher-confidence value driver, and it's being given away free inside the AI-GPU euphoria. Inversely, the consensus underrates how much of the "AI win" is equity-funded — the warrants mean AMD is buying revenue, and the bull EPS numbers are gross of a dilution that's coming.

---

Devil's Advocate (short-seller)

Dismantling the bull case:

• Revenue concentration is becoming dangerous. The 2027 story rests on a handful of mega-deals (OpenAI, Meta, Oracle). If even one slips or renegotiates, the "tens of billions" headline breaks — and these are sophisticated counterparties who extracted equity warrants as the price of their commitment. That's not a position of strength; that's AMD paying to be designed in.
• The moat is weaker than bulls think — on the part that matters. NVIDIA's ~86% revenue share isn't an accident of timing; it's CUDA + system integration + multi-gen cadence + ~60% of CoWoS allocation. AMD wins inference bake-offs on selected models at selected moments; it does not have a durable, defensible lock anywhere except x86 CPUs. ROCm "within 10–30%" is a trailing metric on a target that keeps moving.
• Most dangerous competitor bulls underestimate: not NVIDIA — hyperscaler custom ASICs (Google TPU, Amazon Trainium, Meta MTIA) + Broadcom. The same hyperscalers signing AMD deals are also building captive silicon with no margin stack ``. AMD could be a transitional second source that hyperscalers use to discipline NVIDIA pricing while they internalize volume to ASICs — a structurally capped role.
• Worst capital-allocation / incentive flags: the warrant structure (equity-for-demand), the wide non-GAAP/GAAP gap (SBC-flattered), 8% ROE on a Xilinx-bloated balance sheet, no dividend.
• Assumptions that must hold for ~$521: ~50–60% EPS growth sustained multiple years; MI400/Helios ships on time and competitive; ROCm reaches functional parity; the mega-deals all convert on schedule; HBM/CoWoS tightness helps not hurts AMD; NVIDIA does not cut inference pricing.
• If growth disappoints 20–30%: FY27 EPS ~$7 instead of ~$10 + a multiple de-rate to ~30–35x ≈ a stock in the ~$230–$280 range — roughly a halving. That's the downside the 58x is ignoring.
• Single permanent-impairment scenario, plausibility: NVIDIA cuts inference pricing materially and hyperscalers accelerate ASIC internalization → AMD's GPU franchise stalls at single-digit share with eroding margins, leaving a (good but cyclical) CPU company valued like an AI co-leader. Plausibility: moderate (~25–35%) over 24 months — not a tail.

---

Management Questions (ordered by information value)

1. Disclose Instinct-GPU revenue separately from EPYC within Data Center — what was it in FY25 and Q1-26, and will you break it out going forward? (Highest-value: it's the one number the whole thesis hinges on and you don't report it.)
2. What is the dollar value and unit/GW schedule of the OpenAI 6 GW and Meta commitments by year, and what are the take-or-pay vs. best-efforts terms — what happens to revenue if they under-purchase?
3. Net of the ~320M warrant shares (OpenAI + Meta) and ordinary SBC, what is your expected fully-diluted share count through 2028, and what FY27 GAAP EPS does the "tens of billions" map to?
4. On TSMC CoWoS + HBM4 allocation for 2026–27 — where do you rank for capacity vs. NVIDIA, and is supply (not demand) the binding constraint on MI400 volume?
5. Quantify the ROCm-vs-CUDA gap today on training (not just inference) for frontier models — what % of MI400 design wins are training vs. inference?
6. What is the realistic gross-margin trajectory for the Instinct franchise vs. corporate average, given HBM cost inflation and rack-scale (Helios) system content?
7. How do you defend against hyperscaler custom ASICs — are your largest GPU customers also your most likely long-run substitutes?
8. MI400/Helios on-time-and-at-spec risk: what is the schedule confidence for first 1 GW (OpenAI) in H2-26, and what are the integration/UALink dependencies?
9. What is the plan for Embedded (Xilinx) to re-accelerate, and how do you answer the per-share value-creation question on the $49B deal given current ROE?
10. Capital allocation priorities 2026–28 — buybacks vs. M&A vs. de-lever — and will you initiate a dividend?
11. China export-control exposure: what share of Instinct TAM is at risk, and what's the contingency?
12. How do you sustain EPYC server-share gains as Intel's roadmap (Panther Lake / foundry) and Arm-server entrants pressure the x86 duopoly?
13. What single competitive development would most change your strategy over the next 24 months?
14. What is your inventory and purchase-commitment posture given the lumpy GPU ramp — what's the write-down risk if a generation transitions faster than expected?
15. At ~58x forward, what return do you believe a shareholder buying today should expect over 3 years, and what has to be true for it?

---