AMD trades at roughly 13x our base intrinsic value with a negative 10-year ROIC and an OpenAI warrant that can dilute shareholders by ~10%. The qualitative story is exciting; the price tag is not.

Plain English

AMD designs computer chips — the brains in laptops, servers, game consoles, and AI machines. It does not make them; Taiwan does. Lisa Su rebuilt the company from near-bankruptcy and now competes with Intel and NVIDIA. The chips are good. The price of the stock is not. You are paying about thirteen dollars today for one dollar of what the business is worth. Over the last ten years, AMD has actually destroyed money on average for every dollar it invested. Hope is doing the work, not cash. A wonderful business at a terrible price is still a terrible investment. Wait for cheaper.

Thesis

Advanced Micro Devices designs x86 CPUs (Ryzen, EPYC), GPUs (Radeon, Instinct MI-series), adaptive computing (the Xilinx FPGA franchise) and custom semi-custom silicon for game consoles. It is fabless: TSMC manufactures the wafers, AMD owns the architecture, the software stack (ROCm), and the customer relationship. The bull case is straightforward and seductive — AMD is the only credible second source to NVIDIA in AI accelerators, just won a 6-gigawatt OpenAI commitment for the MI450, and rides Intel's ongoing manufacturing stumbles in server CPUs.

The problem is not the story. The problem is the price. Owner earnings TTM are roughly $2.37B against a market capitalization that, at $360.54, runs well above half a trillion dollars. The deterministic scorecard puts trailing P/E at 680x, EV/FCF at 516x, and a 10-year average P/E of 251x — the latter telling you this stock is structurally a hope-discounting machine, not a cash machine. Ten-year average ROIC is negative 2.71%; AMD has destroyed capital on average across a full cycle. FCF conversion of 95% is a bright spot, but FCF is small.

The scorer's IV range is $24.67 (low) to $36.38 (high), with a base of $28.62. At today's $360.54, the price-to-IV ratio is 12.6x — even on the bullish IV-high case, the stock would need to fall ~90% to reach a price where margin of safety becomes meaningful. A Buffett-Munger compounder requires buying durable cash flows at a discount; AMD offers volatile cash flows at a 12-fold premium. Owning it here is paying for a perfect AI-accelerator decade with no room for error and a built-in OpenAI warrant that dilutes shareholders if the bull case actually plays out. Verdict: Avoid at this price; revisit only below ~$45.

Moat

AMD's moat is real but narrow, and far narrower than the ~$580B market cap implies. Run the five moat lenses against Damodaran's framework [2][4]:

Pricing power (intangibles / brand). AMD has earned genuine architectural respect: Zen, EPYC, and the Xilinx FPGA platform are credible engineering. But pricing power in semis is governed not by brand affection but by performance-per-watt versus the latest competing silicon. ASPs in client CPUs collapsed twice in the last decade when Intel re-tooled. In AI accelerators, NVIDIA's CUDA-driven gross margin runs in the 70s; AMD's company-wide gross margin runs in the 40s. That gap is the market telling you who has pricing power and who is the price-taker. No durable brand premium.

Switching costs. This is AMD's most plausible moat — and it is structurally weaker than NVIDIA's. Damodaran's Microsoft case [1][3] explains why switching costs are decisive: file-format lock-in, ecosystem integration, retraining cost. NVIDIA owns the analogous lock-in in AI via CUDA — fifteen years of accumulated kernels, libraries, frameworks (PyTorch, TensorRT), and developer mindshare. AMD's ROCm is years behind, and the very fact that hyperscalers are forcing AMD into deals (OpenAI's 6 GW commitment came with a 160M-share warrant struck at $0.01 — a giant subsidy demanded by the customer) reveals that switching to AMD is costly to the customer, not the other way around. In x86 server CPUs, switching costs are modest: socket-compatible motherboards and Linux mean migrating a fleet from Intel Xeon to AMD EPYC is annoying but routine. Switching costs: present in CUDA's favor, weak in AMD's.

Network effects. Effectively none for AMD. Network effects in compute live in the software ecosystem — CUDA, the Apple/iOS app store, the Windows/x86 developer base. AMD benefits from x86's network effect but does not own it; Intel co-owns it and ARM is eating into it (Apple Silicon, Graviton, Ampere). In GPUs, the network effect runs against AMD: every paper, every model checkpoint, every tutorial is written for CUDA first.

Intangibles (patents, IP, R&D productivity). Damodaran [4] is explicit: it is not who spends most on R&D, it is who converts R&D into commercial value. AMD spends roughly $6B+/year on R&D against $2.4B of owner earnings — the conversion ratio is poor. Patents are mutually neutralized via cross-licensing among Intel, NVIDIA, AMD, ARM. The Xilinx acquisition added FPGA IP but at a price ($49B in stock) that already presupposed a successful integration.

Cost advantages. AMD has no cost advantage. It is fabless, so TSMC controls the cost of bleeding-edge wafers and charges every customer (Apple, NVIDIA, Qualcomm, AMD) the same. There is no scale economy in design that AMD has and NVIDIA does not — both buy from the same fab. The only cost lever AMD has is chiplet architecture (cheaper to manufacture than monolithic dies), and NVIDIA is following them there.

Stress test: Give a competitor $10B and five years. NVIDIA already spends ~$12B/year on R&D and earns 70%+ gross margins to fund it; Intel still has ~$50B of revenue and is restructuring its foundry; ARM-based hyperscaler silicon (Graviton, Maia, Trainium) is being designed by customers who used to buy from AMD. Each of these is a credible $10B/5-year threat. AMD survives the stress test in client CPUs (a duopoly with Intel) but fails it in AI accelerators, where the moat is largely NVIDIA's, not AMD's.

Erosion risk. The Damodaran disruption framework [5] applies directly: hyperscaler in-house silicon is the classic disruptive entrant — initially inferior, targeted at narrow workloads (inference), but improving. Within five years, Trainium/Maia/TPU-class silicon could absorb 30-40% of the merchant accelerator TAM that AMD is being valued on.

Moat verdict: NARROW.

Management

Lisa Su has been an excellent operator. Taking over in 2014 with the company on the brink of insolvency, she executed the Zen architecture turnaround, navigated the foundry split with GlobalFoundries, and re-established AMD as a credible competitor in CPUs and a credible challenger in GPUs. As a technologist-CEO she is among the best in the industry. The harder question is capital allocation through the Buffett-Munger lens, where the record is mixed.

Reinvest in the business. R&D spending has scaled from ~$1B to $6B+ annually. That is appropriate for a fabless designer, and the Zen and EPYC results validate much of it. But ROIC over the trailing 10 years has averaged negative 2.71% — meaning the cumulative reinvestment cycle has failed to clear the cost of capital. ROIIC over the last 5 years is not even meaningful (NOPAT declined). Engineering wins do not automatically equal capital-allocation wins.

Acquire. The Xilinx deal (closed 2022, ~$49B all-stock at AMD's share-price peak) was paid for with the most expensive currency AMD has ever had. Strategically defensible — adaptive computing complements the data-center stack — but the price-paid relative to Xilinx's standalone earning power means the synergies have to be enormous to justify the dilution. The Pensando acquisition ($1.9B) added DPU capability. The pending ZT Systems acquisition extends AMD into rack-scale systems integration. Pattern: AMD acquires capability, not cash flow. That is justifiable in semis but punishing if the cycle turns.

Debt. Net debt to EBITDA is -8.7 (i.e., net cash). Interest coverage is 3.78x — adequate but unremarkable for a net-cash company, indicating EBITDA is depressed relative to historical interest expense. Balance-sheet posture is conservative, which is correct for a cyclical fabless designer dependent on a single foundry.

Buybacks. AMD has executed buybacks, but share count is up 8.68% over 10 years — meaning issuance for stock-based compensation and the Xilinx deal has more than offset repurchases. There is no evidence management has bought back stock at attractive prices relative to intrinsic value; with the stock now trading at ~12.6x our IV base of $28.62, any buyback at current levels would be active value destruction. The recent OpenAI warrant — 160 million shares at $0.01 strike — is the single most consequential capital-allocation decision of 2025: AMD is paying its largest customer in equity to take its product. That is a tell about pricing power, and a structural ~10% future dilution to existing holders if the bull case actually triggers vesting.

Dividends. None. Appropriate for the lifecycle.

Communication. Lisa Su's investor communication is unusually clear and technically grounded for a CEO. Guidance has been honest, including downward revisions in 2022-2023. No accounting drama, no governance scandals.

Net assessment. A great engineering CEO who has run the company well operationally but whose major capital-allocation acts (Xilinx at the peak, OpenAI warrant) implicitly priced AMD's currency at top-of-cycle multiples and gave it away cheaply. The negative 10-year ROIC is an indictment of the whole allocation cycle, not of Su personally — but it is what the scorecard records.

Capital allocator: B-.

Industry

Apply Porter's Five Forces to merchant logic semiconductors:

Rivalry (HIGH). AMD competes simultaneously against (a) NVIDIA in discrete GPUs and AI accelerators, where NVIDIA owns ~80%+ share and 70%+ gross margins; (b) Intel in x86 CPUs, a re-awakened competitor with new fabs and the Gaudi accelerator line; (c) ARM-architected designs from Qualcomm, Apple, Ampere, and the hyperscalers themselves; (d) Broadcom and Marvell in custom silicon. The product cycle is 18-24 months — every generation, the moat is renegotiated. Rivalry intensity is structurally high and rising as AI capex draws every actor into the same TAM.

Threat of new entrants (HIGH for the merchant model). This is unusual and matters. The capital cost of designing leading-edge silicon used to be a moat; now hyperscalers (AWS Graviton/Trainium, Google TPU, Microsoft Maia, Meta MTIA) routinely fund their own silicon teams. Their motivation is direct: AMD/NVIDIA gross margins are a cost line on the hyperscaler P&L. The same buyers AMD is being valued on are the most credible new entrants against AMD. OpenAI's warrant is a related signal — buyers extracting equity from the merchant supplier.

Supplier power (HIGH — concentrated in TSMC). AMD is fabless. TSMC is effectively the only credible supplier of leading-edge (3nm and below) capacity. TSMC's pricing power has visibly increased: wafer ASPs at the leading node have roughly tripled in five years. AMD competes for TSMC capacity against Apple (priority customer), NVIDIA (highest-margin customer), Qualcomm, and others. AMD is a price-taker on its largest cost input. Geopolitical concentration in Taiwan adds tail risk.

Buyer power (HIGH and rising). Three customer concentrations matter: hyperscalers in data center (Microsoft, Meta, Amazon, Google, Oracle, plus OpenAI), Sony/Microsoft in semi-custom (consoles), and OEMs in client (Dell, HP, Lenovo). The hyperscalers buy in volumes large enough to dictate roadmap, demand custom SKUs, and — as the OpenAI warrant proves — extract equity. The semi-custom console business is a known low-margin annuity controlled by two buyers.

Substitutes (HIGH). GPUs substitute for CPUs in some workloads; FPGAs substitute for GPUs in others; TPUs and custom ASICs substitute for GPUs in inference and increasingly training; ARM CPUs substitute for x86 in cloud servers. Every dollar of AMD revenue has multiple credible substitutes from credible suppliers.

Value-pool location. Most of the economic profit in the merchant logic stack today accrues to (1) TSMC at the manufacturing chokepoint, (2) ASML upstream of TSMC at the lithography chokepoint, and (3) NVIDIA at the AI software/ecosystem chokepoint. AMD sits in none of these. The Damodaran semiconductor industry data [1] shows long-run sector ROIC around 14-17% — and AMD has earned less than zero. The value pool is real; AMD's share of it is small.

Industry Verdict: Average. The semiconductor industry has produced extraordinary winners (TSMC, NVIDIA, ASML), but the average fabless designer fights cycles, foundry pricing, hyperscaler in-housing, and 18-month product clocks. AMD is a competent participant in a structurally difficult industry, not a structural beneficiary of it.

Inversion

I am now playing short-seller against AMD at $360.54.

1. The single event that kills this. The kill shot is a credible inference benchmark from a hyperscaler-designed ASIC (Google TPU v6, AWS Trainium2/3, Microsoft Maia 200) showing 1.5-2x better performance-per-dollar than MI450 on a frontier-model workload, accompanied by that hyperscaler publicly cutting its merchant-GPU order. Within one quarter, AMD's data-center segment guidance gets cut, the AI-accelerator multiple resets, and the 6 GW OpenAI commitment is exposed as the only major non-hyperscaler customer of scale. Secondary kill shot: a Taiwan supply disruption that exposes AMD's single-foundry dependence and forces it to compete with Apple and NVIDIA for residual TSMC capacity — AMD loses that contest because Apple pays in cash and NVIDIA pays in highest gross margin per wafer.

2. Why the moat is narrower than bulls think. Bulls treat AMD as the duopoly partner to NVIDIA in AI accelerators. The reality is more like Damodaran's search-engine analogy [3]: 'there is little cost to an end-user from switching from one engine to another, and no barriers to new search engines being developed.' For an AI accelerator buyer the switching cost is into CUDA, not into ROCm. The fact that the OpenAI deal required a 160M-share warrant struck at $0.01 — a billion-dollar-scale subsidy demanded by the customer — is a confession that AMD's product alone could not justify the purchase at NVIDIA-equivalent economics. In x86, the moat is real (a duopoly) but Intel's manufacturing recovery, even partial, removes AMD's last cycle of share gains. In semi-custom (consoles), the moat is contractual and known to be a low-margin annuity. In FPGA, Xilinx is a stable franchise but acquired at a price that already capitalized the moat.

3. Why management is worse than it appears. Lisa Su deserves enormous credit for the operational turnaround, but capital allocation tells a different story. Ten-year ROIC of -2.71% is the scoreboard. The Xilinx acquisition was paid in $49B of stock at peak multiples — the strategic logic is debatable, the price was indefensible. The OpenAI warrant gives away 160M shares (~10% of float) for purchase commitments AMD will recognize as revenue and fund the buyer's purchase of with dilution. Buybacks have been outpaced by issuance: share count is up 8.68% over a decade in which the stock 30x'd. A great manager of a great business buys back stock when it is cheap; AMD has done the opposite, issuing into strength.

4. What bulls are extrapolating that won't hold. Three extrapolations are doing the work in the current price: (a) AI accelerator TAM grows to $400-500B by 2030 and AMD captures 20%+ of it and at NVIDIA-class margins; (b) the merchant-GPU model survives hyperscaler in-housing rather than being displaced by it; (c) the current cycle does not produce a glut. All three are needed simultaneously. The base rate from prior compute cycles (mainframe, mini, PC, mobile, x86 server) is that the merchant arms-supplier captures 5-15% of the eventual TAM, not 25%, because the largest customers vertically integrate. The Damodaran disruption template [5] applies: hyperscaler ASICs are 'initially targeted at small and less profitable markets and thus not viewed as a threat.' That is exactly inference today.

5. Valuation trap (multiple compression / regime change). Trailing P/E is 680. EV/FCF is 516. Owner earnings are $2.37B against a market cap north of $580B. The 10-year average P/E is already 251 — the stock has been priced for a future that has not arrived for a decade. Mean-reversion in the multiple alone, even without an earnings disappointment, takes the stock down 80%+. With a single bad accelerator quarter, the multiple compresses and the earnings denominator falls — the classic value-trap double-whammy. Owner earnings would need to grow ~40x to justify today's price at a normal 25x P/E; that requires AI accelerator revenue to scale to $80-100B at industry-leading margins, with no share loss to NVIDIA, no share loss to ASICs, no foundry-cost squeeze, and no warrant-driven dilution. The probability-weighted expected outcome is not in the bulls' favor.

Bear-case fair value. Normalize owner earnings on a through-cycle basis at $4-6B (generously crediting some AI ramp), apply a 25-30x multiple appropriate for a cyclical fabless designer with a narrow moat, divide by ~1.7B shares pre-warrant or ~1.85B post-warrant: $50-100B equity value, or $30-55 per share.

If I am right, the stock could be worth $40 within 3 years.

Lollapalooza Bias Check

Working through Munger's bias checklist for biases active in me as the analyst right now:

Anchoring (active, strong). The deterministic IV range is $24.67-$36.38. The current price is $360.54. That ten-fold gap is so large it threatens to anchor me toward extreme bearishness — once the price/IV ratio crosses some absurd threshold, my brain stops doing fresh thinking and just defaults to 'Avoid.' I tried to mitigate this by genuinely steel-manning the AI accelerator bull case in the moat and inversion sections, but I am aware the mathematics of 12.6x IV makes any other recommendation feel intellectually unserious. I chose to trust the math.

Authority (active, mixed). The scorer is deterministic Python and the brief explicitly tells me not to redo its math. That is correct epistemics for this exercise — but it also means I am deferring to a model I did not build. I would normally want to inspect the maintenance-capex assumption and the discount rate; the scorer's note that 'maintenance capex uncertain (>50% spread); widen IV range' tells me even the model is hedging. I priced this in by widening the conviction interval rather than overruling the IV.

Recency / availability (active against the bull case). AMD's stock has run hard on the AI narrative; the 'NVIDIA is the only AI winner' counter-narrative has been everywhere. Both are recency-biased in opposite directions. I tried to ground the analysis in the through-cycle ROIC (-2.71% over a full decade including the AI boom) rather than the trailing 12 months.

Social proof (active against the bull case). Every value-investing newsletter and every crypto-adjacent trader I know thinks AMD is overvalued. That alone is suspicious — when a contrarian view becomes consensus among contrarians, it is no longer contrarian. I asked: what would change my mind? Answer: a credible MI450 deployment that expands gross margins toward 60% within 18 months. I do not see that in the filings.

Confirmation bias (active). Once I saw the trailing P/E of 680 and the OpenAI warrant terms, I went looking for reasons to support 'Avoid.' I tried to write the bull thesis (real engineering, second source to NVIDIA, net cash balance sheet, strong FCF conversion at 95%) before writing the bear thesis. The math still came out where it did.

Deprival super-reaction (active for the bull case in the market, not in me). This bias is driving the price right now: investors are terrified of missing AI. That is exactly the social-proof + deprival lollapalooza Munger warned about. I am trying not to be subject to the inverse — terror of missing the next leg up.

Commitment / consistency (low risk). I have no prior public position on AMD. I am free to recommend Avoid without ego cost.

Net. The strongest active bias is anchoring to the IV math, which I judge appropriate given the magnitude of the gap. I would only override the math if I saw evidence of a structural margin step-up that the IV model missed; I do not.

10-Year Outlook

Same fundamental business model in 2036? Probably yes in shape — AMD will still be a fabless designer of CPUs, GPUs, and adaptive silicon — but the mix and the competitive position are highly uncertain. The fabless model itself is stable; AMD's share within it is not.

Customer base larger? Probably modestly larger in absolute terms (AI compute TAM grows), but the concentration of customers is increasing. Five hyperscalers plus OpenAI plus Sony/Microsoft will likely represent the majority of AMD's data-center and semi-custom revenue. Concentration that high erodes negotiating power and increases the probability of in-house substitution.

Profit per customer higher? Unlikely on a sustained basis. The OpenAI warrant is a clear signal: large customers are now extracting equity, not paying premium prices. As hyperscaler ASICs mature (Trainium 3+, TPU v7+, Maia 2+), the merchant-GPU per-customer ASP and gross margin face downward pressure. The most likely path is lower gross margin per customer offset by higher unit volume — a recipe for revenue growth without ROIC growth.

Moat wider? Probably narrower. The CUDA ecosystem keeps deepening; ROCm gains ground but does not catch up. Hyperscaler in-housing erodes the merchant TAM. Intel's foundry recovery (whether or not fully successful) limits AMD's share-gain runway in x86. The only direction the moat could widen is if AMD becomes the dominant open-stack alternative to CUDA — possible but not the base case.

Single biggest threat? Hyperscaler vertical integration in AI silicon. This is the same disruption pattern Damodaran [5] describes: initially inferior, narrow workload, ignored by incumbents, then good enough. Trainium and TPU are already past the 'inferior toy' phase.

Reasoning. A 10-year forecast for a fabless designer in a 24-month product cycle, dependent on a single foreign foundry, with three credible substitution vectors (NVIDIA, hyperscaler ASIC, ARM), is structurally low-confidence. The historical 10-year ROIC of -2.71% is the empirical answer to 'how predictable have AMD's returns been over a decade.' The answer is: not very, and not good.

CONFIDENCE: low

Position Guidance

• Recommendation: Avoid
• Conviction: high
• Target buy price: $45 (roughly 1.25x IV-high of $36.38, leaving a meaningful margin of safety on the bullish IV case; only at this level does owning AMD make sense as a Buffett-Munger position)
• Target trim price: N/A — current price already exceeds even bull-case IV by ~10x; do not initiate; if held, trim aggressively above $50
• Position sizing: 0% at current price. Watchlist position only. If price reaches $45 and the through-cycle owner-earnings trajectory has improved (sustained $5B+ owner earnings, gross margin >55%, no further dilutive warrants), consider 1-2% starter position. Maximum position even in a deep drawdown: 3-4% — the industry is structurally cyclical and the moat is narrow, so this is never a core compounder holding.
• Triggers to revisit: (a) MI450 ramps with sustained 55%+ gross margin; (b) ROCm achieves credible CUDA parity for major frameworks; (c) hyperscaler ASIC programs visibly stall; (d) stock falls below $60 for non-fundamental reasons.