Advanced Micro Devices
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat Score
September 2026 review: Q2 printed and Helios launched, but the moat did not deepen a grade. AMD's edge is still execution, not structural lock-in — the primary x86 alternative (EPYC Turin/Venice share gains) and a credible #2 AI accelerator, with OpenAI (6 GW), Meta (6 GW), Anthropic (up to 2 GW), and Azure Helios commitments deepening second-source embedding. ROCm.ai and 3 million day-zero models are a capability print, not a CUDA-class developer flywheel. The only status change is systemOfRecord, corrected from weakened to na — the old note treated 'not the default GPU' as a system of record, which the pillar forbids. The score moves only because that weak box drops out, not because of a new advantage.
AMD's advantage lies in Architectural Efficiency:
- Chiplet Innovation: AMD led the transition to chiplets, allowing for higher yields and more flexible SKU creation compared to monolithic designs. Helios extends that into a rack — 72× MI455X, Venice CPUs, and Pensando networking as one scale-up domain — but the architectural lead is still a cost/performance edge, not a lock-in.
- x86 Market Share Capture: Q2 was the fifth consecutive record server-CPU quarter: cloud and enterprise EPYC each grew more than 70% year over year, and Venice is in production with every major OEM. That is share capture against Intel, not a new switching cost.
- Open Ecosystem: ROCm.ai and a claimed 3 million models running out of the box make the port easier. OpenAI, Anthropic, Meta, and Azure are co-optimising on Instinct. The stack is a viable second source, not a CUDA replacement — hyperscalers still buy it for vendor flexibility.
Moat Verdict
AMD is a net AI beneficiary in demand and a laggard in AI-resilient structural moats versus NVIDIA. The Q2 print ($11.5B, Data Center +107%) and the Helios launch with Anthropic 2 GW plus Azure do not change that: they deepen transaction embedding that was already intact, and they leave learned interfaces, bundling, and network effects weakened against CUDA. Proprietary CDNA/ROCm IP and scarce chip talent are the durable sources; the software flywheel is still NVIDIA's. Execution is world-class. The upside is still a CUDA challenger, not a CUDA replacer — and the ~$1T cap prices the challenger path as the base case.
48.8 resilient · 41.5 vulnerable · 80/20 = 47.3 · = 47
Open a moat to read its note.
Held weakened. ROCm.ai and a claimed 3 million models running out of the box make the port faster; they do not make Instinct the interface ML PhDs learn. CUDA remains the default learned surface. A developer-experience launch is not a mindshare shift.
Chip design is not a software-logic business. ROCm is a portability layer, not vendor logic customers encode production pipelines against — that lock-in still sits with CUDA.
AMD does not control access to a public dataset; this pillar is not part of a semiconductor design model.
Chip-design engineers remain scarce; Lisa Su's executive team is hard to replicate. That is not NVIDIA's CUDA-kernel talent flywheel — hold intact, do not mark strong because the other GPU company is strong.
Held weakened. Helios is now a launched rack — 72× MI455X, Venice, Pensando, ROCm — with named deployers (Anthropic, Azure, OpenAI, Meta, Oracle, HUMAIN). That is a product, not NVIDIA's CUDA-anchored bundle economics. A shipping SKU does not create the switching cost.
CDNA architecture IP, ROCm optimisation data, foundry-integration know-how, and the MEXT predictive-memory acquisition remain proprietary. Scale of Instinct deployments is not a unique franchise dataset.
Held weakened. MI430X for HPC and sovereign AI, plus embedded aerospace-and-defense design wins inside an $18B 2026 win-book, widen the government footprint. That is still not NVIDIA-scale export-licence or defence lock-in.
Held weakened. Open-source ROCm contributions up more than tenfold, 3 million models day-zero, and lab co-optimisation with OpenAI, Anthropic, and Meta seed a network. They are still a fraction of CUDA's 4M+ developer flywheel. Lead-customer logos are not a self-reinforcing community.
Deepening, not upgraded. OpenAI (6 GW), Meta (6 GW), Anthropic (up to 2 GW of MI450 in Helios, first GW in H1 2027), and Microsoft Azure Helios for frontier inference make AMD a contracted second source in the largest AI build-outs. The OpenAI commitment was partly bought with equity — a warrant for up to 160M AMD shares that vests as its purchases scale — so it is weaker proof of pull than an unsubsidised order. Alternatives remain. Strong is a rail, not a second GPU.
Corrected from weakened. Semiconductor design is not a system of record — AMD does not own an authoritative store downstream systems must defer to. The prior note used this pillar as 'not the default GPU standard,' which the framework forbids.
Fabless or equipment franchise whose edge is IP and process know-how, rated elsewhere; manufacturing volume does not give it a unit-cost lead rivals cannot buy.
Buyers are enterprises choosing on switching cost, integration and performance, which the other pillars rate. The name carries reputation, not a price premium it could hold on brand alone.
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat Score
September 2026 review: Q2 printed and Helios launched, but the moat did not deepen a grade. AMD's edge is still execution, not structural lock-in — the primary x86 alternative (EPYC Turin/Venice share gains) and a credible #2 AI accelerator, with OpenAI (6 GW), Meta (6 GW), Anthropic (up to 2 GW), and Azure Helios commitments deepening second-source embedding. ROCm.ai and 3 million day-zero models are a capability print, not a CUDA-class developer flywheel. The only status change is systemOfRecord, corrected from weakened to na — the old note treated 'not the default GPU' as a system of record, which the pillar forbids. The score moves only because that weak box drops out, not because of a new advantage.
Growth Score
Q2 2026 (quarter ended June 27, reported August 4) printed revenue $11.536B, +50% from $7.685B a year ago and +13% from Q1's $10.253B, beating the $11.2B guide. Data Center was $6.718B, +107% YoY and 58% of sales (from 42% a year ago). Non-GAAP gross margin 56%, operating margin 27%, diluted EPS $1.66; GAAP EPS $1.38. Q3 is guided at $13.0B ±$300M (~+41% YoY, ~+13% QoQ) with non-GAAP GM ~56%. Helios is in production — first shipments late Q3, a step-up in Q4, then through 2027. Management now puts the data-center AI accelerator TAM at more than 45% annually to ~$1.4T by 2030 and the server-CPU TAM at more than 50% to ~$220B, and says company revenue grows above a >40% high-performance/AI compute market. They expect to significantly exceed the $20 Financial Analyst Day EPS target inside the strategic window. H1 free cash flow was $4.124B; cash and short-term investments $13.111B.
Valuation Score
At ~$624 (fresh ATH, ~$1.0T), AMD trades above the revised base ($570). The Q2 print and Helios/Anthropic/Azure commitments re-rated FY2027 consensus EPS from ~$13 to ~$15.5, and the ~$620–628 average analyst target sits on spot — but that is per current share. OpenAI holds a warrant for up to 160M AMD shares at $0.01 (~9.6% of the ~1.66B diluted count), vesting as its GPU purchases scale from 1 GW to 6 GW; the share-price hurdles escalate to $600 for the final tranche, which spot has already cleared. The base case assumes OpenAI deploys, so it assumes the warrant vests: ~$15.5 becomes ~$14.1 per fully diluted share. The AI thesis is priced as the base case, not as optionality.
The Chiplet Edge
AMD's advantage lies in Architectural Efficiency:
- Chiplet Innovation: AMD led the transition to chiplets, allowing for higher yields and more flexible SKU creation compared to monolithic designs. Helios extends that into a rack — 72× MI455X, Venice CPUs, and Pensando networking as one scale-up domain — but the architectural lead is still a cost/performance edge, not a lock-in.
- x86 Market Share Capture: Q2 was the fifth consecutive record server-CPU quarter: cloud and enterprise EPYC each grew more than 70% year over year, and Venice is in production with every major OEM. That is share capture against Intel, not a new switching cost.
- Open Ecosystem: ROCm.ai and a claimed 3 million models running out of the box make the port easier. OpenAI, Anthropic, Meta, and Azure are co-optimising on Instinct. The stack is a viable second source, not a CUDA replacement — hyperscalers still buy it for vendor flexibility.
Moat Verdict
AMD is a net AI beneficiary in demand and a laggard in AI-resilient structural moats versus NVIDIA. The Q2 print ($11.5B, Data Center +107%) and the Helios launch with Anthropic 2 GW plus Azure do not change that: they deepen transaction embedding that was already intact, and they leave learned interfaces, bundling, and network effects weakened against CUDA. Proprietary CDNA/ROCm IP and scarce chip talent are the durable sources; the software flywheel is still NVIDIA's. Execution is world-class. The upside is still a CUDA challenger, not a CUDA replacer — and the ~$1T cap prices the challenger path as the base case.
48.8 resilient · 41.5 vulnerable · 80/20 = 47.3 · = 47
Open a moat to read its note.
Held weakened. ROCm.ai and a claimed 3 million models running out of the box make the port faster; they do not make Instinct the interface ML PhDs learn. CUDA remains the default learned surface. A developer-experience launch is not a mindshare shift.
Chip design is not a software-logic business. ROCm is a portability layer, not vendor logic customers encode production pipelines against — that lock-in still sits with CUDA.
AMD does not control access to a public dataset; this pillar is not part of a semiconductor design model.
Chip-design engineers remain scarce; Lisa Su's executive team is hard to replicate. That is not NVIDIA's CUDA-kernel talent flywheel — hold intact, do not mark strong because the other GPU company is strong.
Held weakened. Helios is now a launched rack — 72× MI455X, Venice, Pensando, ROCm — with named deployers (Anthropic, Azure, OpenAI, Meta, Oracle, HUMAIN). That is a product, not NVIDIA's CUDA-anchored bundle economics. A shipping SKU does not create the switching cost.
CDNA architecture IP, ROCm optimisation data, foundry-integration know-how, and the MEXT predictive-memory acquisition remain proprietary. Scale of Instinct deployments is not a unique franchise dataset.
Held weakened. MI430X for HPC and sovereign AI, plus embedded aerospace-and-defense design wins inside an $18B 2026 win-book, widen the government footprint. That is still not NVIDIA-scale export-licence or defence lock-in.
Held weakened. Open-source ROCm contributions up more than tenfold, 3 million models day-zero, and lab co-optimisation with OpenAI, Anthropic, and Meta seed a network. They are still a fraction of CUDA's 4M+ developer flywheel. Lead-customer logos are not a self-reinforcing community.
Deepening, not upgraded. OpenAI (6 GW), Meta (6 GW), Anthropic (up to 2 GW of MI450 in Helios, first GW in H1 2027), and Microsoft Azure Helios for frontier inference make AMD a contracted second source in the largest AI build-outs. The OpenAI commitment was partly bought with equity — a warrant for up to 160M AMD shares that vests as its purchases scale — so it is weaker proof of pull than an unsubsidised order. Alternatives remain. Strong is a rail, not a second GPU.
Corrected from weakened. Semiconductor design is not a system of record — AMD does not own an authoritative store downstream systems must defer to. The prior note used this pillar as 'not the default GPU standard,' which the framework forbids.
Fabless or equipment franchise whose edge is IP and process know-how, rated elsewhere; manufacturing volume does not give it a unit-cost lead rivals cannot buy.
Buyers are enterprises choosing on switching cost, integration and performance, which the other pillars rate. The name carries reputation, not a price premium it could hold on brand alone.
Growth Analysis
Growth Drivers
Key Risk
If NVIDIA's Rubin Ultra widens the performance gap on flagship training workloads through 2027 and ROCm fails to convert the 3-million-model / ROCm.ai print into production CUDA-equivalent share, AMD's AI GPU share stalls below 15% of TAM and the 32–42% CAGR collapses toward 15–20% — invalidating the OpenAI/Anthropic/Azure upgrade cycle priced into the ~$1T re-rating. This is a cyclical hardware print: the +50% sits near the top of a hyperscaler capex cycle, and a digestion year would cut the second-source vendor's orders before the incumbent's.
Score Derivation
91.8 base + 2.7 trajectory + 4 margin − 10 risk = 88
Base 91.8 (32–42% CAGR, midpoint 37%, baseFromCagr) + 2.7 trajectory (2 of 3 drivers accelerating) + 4 expanding margins (non-GAAP GM 56%; op margin 27%, +2 ppts QoQ) − 10 high keyRisk (NVIDIA Rubin Ultra still the unmaterialised overhang; Helios just entering production) = 88. Do not bump because they beat. primaryType does not score.
Price Scenarios (12–24 Months)
Valuation Multiples
| Trailing P/E (GAAP) | ~160× |
| Forward P/E (NTM) | ~54× |
| PEG Ratio | ~1.5× |
| Price / Sales (NTM) | ~15× |
| Price / FCF | ~125× |
| OpenAI warrant | ~9.6% |
The ~54× forward P/E is expensive versus the semis peer group and the broad market (~21×), but it is no longer the July 65× on an $8.50 NTM that assumed a flawless MI450 ramp still ahead. A PEG of ~1.5× is premium-but-defensible only if the 32–42% CAGR holds, and every multiple here is on the current share count — full warrant vesting lifts each by ~10%. The gap between ~160× trailing GAAP and ~54× forward still prices a steep 2026–27 earnings ramp — now consensus, not a beat.
Approximate figures as of September 2026.
Where We Are vs Targets
Loading live price…
AI capex digestion in 2027, NVIDIA reasserts dominance with Rubin Ultra, and the OpenAI/Anthropic/Azure deployments slip. A cyclical bear cuts earnings and the multiple together: FY2027 EPS stalls near ~$11 instead of ~$15.5, at ~25×. Slipped deployments also leave most of the OpenAI warrant unvested, so dilution is small here.
- Hyperscaler AI capex growth decelerates from 40%+ to mid-teens as training-cluster overbuild concerns surface, and the second-source vendor's orders are cut first
- Rubin Ultra performance leadership widens vs. MI450/Helios, capping AMD's share of new AI GPU spend at <15%
- Intel 18A plus a softer H2 PC market slow EPYC and Ryzen compounding together
Helios ramps from late-Q3 shipments through 2027, the OpenAI, Anthropic 2 GW, Meta, and Azure deployments land as guided, and Data Center more than doubles in 2027 — ~40× FY2027 EPS of ~$14.1 per fully diluted share (consensus ~$15.5 with the OpenAI warrant's 160M shares issued, since the base case is the one where OpenAI's purchases vest it).
- Helios volume in Q4 2026 and through 2027 converts the named GW-scale commitments into Data Center growth that more than doubles the segment
- EPYC server revenue grows more than 70% in FY2027 off the raised 2026 base, tracking the $220B 2030 server-CPU TAM
- Non-GAAP operating margin holds in the high-20s as AI mix and ROCm adoption offset Helios ramp yields
AMD becomes a credible second AI standard, capturing 25%+ of AI GPU spend as ROCm reaches functional parity with CUDA for inference and key training workloads — FY2027 EPS beats toward the $20 FAD target (~$18.2 per fully diluted share once OpenAI's 6 GW vests the full warrant) at a sustained ~48× multiple.
- MI500 (2027) achieves performance parity with NVIDIA on flagship training workloads and the annual rack-scale cadence holds
- Anthropic's 2 GW plus OpenAI/Meta scale toward a second multi-GW hyperscaler, and ROCm.ai converts the 3-million-model print into production share
- Company revenue stays above the >40% TAM print long enough for FY2027 EPS to clear $20 inside the strategic window