Mark Papermaster retweeted
$AMD CPU vs $ARM based CPU w/ CTO & Chief Architect 👂
Advanced Insights, @AMD CTO and EVP @mpaper2000 Mark Papermaster sits down with Mike Clark, AMD SVP, Corporate Fellow and Chief Architect, to explore the architecture, engineering decisions and culture behind the Zen CPU family.
Agentic AI changes the CPU:GPU balance. Prefill/decode stays GPU heavy, but orchestration, spawning sub-agents, tool calls, context management, sandbox execution, and feeding accelerators run on the CPU. AMD projects the ratio moving toward roughly 1-2:1 (versus ~1:4 in earlier generative-AI setups) and says agents create highly parallel, throughput oriented work that plays to high core/thread counts plus strong per-core performance.
1. Scaling large numbers of agents
AI Labs, Enterprises and Hyperscalers must choose AMD x86 CPU for mmaximum threads per socket, compatibility with existing x86 tools/libraries/enterprise software, mixed workloads, or strong per-core performance alongside density. SMT gives AMD more logical threads per core (typically 2 vs 1 on most ARM server cores), which is a real advantage for some agentic and RL workloads.
2. Training agents with RL
Reinforcement learning for agents usually has two phases that stress different parts of the system:
~Data collection / environment interaction: many parallel rollouts or simulated environments. This is often CPU-bound (or mixed) and benefits from high thread counts and good single-thread performance so you can scale the number of concurrent agents cheaply.
~Policy/value network updates: typically GPU-bound matrix math.
x86/Zen advantages here include SMT for more parallel environments, mature vector extensions, and the fact that many existing simulators, tools, and enterprise applications that agents call are already x86 optimized.
Thread density is a strong AMD selling point for scaling agents and RL rollouts.
Full Video here: youtube.com/watch?v=7sBzosci…
$AMD at $630/share is dirt cheap on 0.1-0.6 PEG🧵
Not Financial Advice! DYOR! I do own AMD shares!
Projected FY2026 Revenue: $55-$60B ($57.5B Mid)
Net margin keeps rising with better operating leverage
PEG = (today’s price ÷ that year’s EPS) ÷ EPS CAGR from 2026.
2027 will have explosive YoY growth "well over 100%", which we can think of 120-140%. Fwd PE is in the 26-29x with 3 digits growth, that is cheap. And yes everyone is trying to project 2028-2029 growth RIGHT NOW. I'm sticking to my $800 Personal PT by end of 2026. AMD this week broke my old $620 PT.
This thread is more of average revenue growth projections for the next 3-5 years to judge its PEG Ratio. Less than 1 is mad cheap for @AMD growth and potential.
I believe 40-50% CAGR is very reasonable base case for AMD in the next 3-5 years. 60%+ is the bull case where AMD gets more TSMC allocation as #1 TSMC customer in term of WPM.
Dr. Su has been explicit that the old 1:4 / 1:8 CPU to GPU host ratio is moving toward 1–2:1, and that lots of agents could mean much more CPUs than GPUs. In recent discussion, people are talking abt 40 CPU : 1 GPU in extreme Agentic AI case. In a $META @Muse Consumer Agentic world, the most logical Ratio would be 8-12 CPU : 1 GPU IMO. Consumers are actually find AI Agents useful, and Muse is WINNING!
Projected GPUs market share FY2027 15-25%
CPUs market share will be in the 50%+ medium-long term.
Net Margin Path assumption, the advantage of Fabless:
40% case: 24% → 26% → 28% → 30% → 32% → 33%
50% case: 24% → 27% → 30% → 32% → 34% → 35%
60% case: 24% → 28% → 31% → 33% → 35% → 36%
Revenue > Net Income > EPS > PE > PEG
1. 40% revenue CAGR
2026: $57.5B, $13.8B NI, $8.31 EPS, 76x
2027: $80.5B, $20.9B NI, $12.61 EPS, 50x , PEG 0.97
2028: $113B, $31.6B NI, $19.01 EPS, 33x, PEG 0.65
2029: $158B, $47.3B NI, $28.51 EPS, 22x, PEG 0.44
2030: $221B, $70.7B NI, $42.58 EPS, 15x, PEG 0.29
2031: $309B, $102B NI, $61.48 EPS, 10x, PEG 0.21
2. 50% revenue CAGR
2026: $57.5B, $13.8B NI, $8.31 EPS, 76x
2027: $86.2B, $23.3B NI, $14.03 EPS, 45x, PEG 0.65
2028: $129B, $38.8B NI, $23.38 EPS, 27x, PEG 0.40
2029: $194B, $62.1B NI, $37.41 EPS, 17x, PEG 0.26
2030: $291B, $99.0B NI, $59.62 EPS, 11x, PEG 0.17
2031: $437B, $153B NI, $92.06 EPS, 7x, PEG 0.11
3. 60% revenue CAGR
2026: $57.5B, $13.8B NI, $8.31 EPS, 76x
2027: $92.0B, $25.8B NI, $15.52 EPS, 41x, PEG 0.47
2028: $147B, $45.6B NI, $27.49 EPS, 23x, PEG 0.28
2029: $236B, $77.7B NI, $46.82 EPS, 13x, PEG 0.17
2030: $377B, $132B NI, $79.45 EPS, 8x, PEG 0.10
2031: $603B, $217B NI, $130.76 EPS, 5x, PEG 0.07
=> 3 year PEG (2028)
40% 0.65 / 50% 0.40 / 60% 0.28
=>5 year PEG (2030)
40% 0.29 / 50% 0.17 / 60% 0.10
AMD may look like at a premium with TTM, but we are only more than a month away from Q3 Nov ER, where Q4, the first biggest J-curve quarter guidance will be issued.
Analysts, Investors and Institutions are fighting on where Q4 guidance is going to land. Otherwise if "they" just strictly judging AMD on TTM, AMD would just trade in the $300-$400 range at most.
I do not believe in outdated/old(TTM) data, market is forward looking. You may be able to buy stocks at a discount in businesses that have declining revenue/profitability or slower growth, but most will not get to buy high quality businesses when growth is ACCELERATING! Get it?
Not Financial Advice! DYOR! I do own AMD shares!