Thesis #2: AMD — The Invisible Bottleneck

Ask any investor what AMD's story is, and they'll tell you about GPUs. The MI450 versus NVIDIA's Blackwell. CUDA's moat. The underdog narrative. They're watching the wrong race.

What the Market Sees vs. What's Actually Happening

This is a thesis about narrative mispricing. The market has one story about AMD — GPU challenger — and it's missing the bigger one: AMD is the dominant CPU supplier for the most structurally growing server market in history, and supply cannot keep up with demand.

The CPU Bottleneck Nobody Talks About

A Georgia Tech/Intel research paper from November 2025 quantified something the industry already knew intuitively: agentic AI workloads invert the traditional compute ratio. Classic chatbot inference is 90% GPU, 10% CPU. Agentic systems — the kind that browse, call APIs, execute code, orchestrate multi-step workflows — are 50-90% CPU.

This matters enormously because every major AI lab is pivoting to agents. Claude Code is at a $2.5B run rate with enterprise adoption up 4x since January. OpenAI's $38B AWS deal explicitly calls out "tens of millions of CPUs" for agentic workloads. NVIDIA's own GTC keynote introduced NemoClaw, an enterprise agentic platform. The Vera Rubin NVL72 rack uses a 1:2 CPU-to-GPU ratio — every rack needs CPUs.

"Demand for EPYC CPUs is surging as agentic workloads require high-performance CPUs."

— Lisa Su, AMD CEO, Q4 2025 earnings call

The numbers are stark. AMD's data center revenue hit $5.4B in Q4 2025, up 39% year-over-year. 5th Gen Turin EPYC processors now account for over 50% of server CPU revenue — the newest architecture isn't niche, it's mainstream and accelerating. Guidance for 2026: "strong double digits." Lead times on high-core-count EPYC parts: 8-10 weeks. Intel's lead times in China: up to 6 months.

The Futurum Group projects that CPU demand growth may exceed GPU demand growth by 2028, driven by agentic workloads. TSMC's advanced node capacity is "about three times short" of demand. This is a physical supply constraint, not a forecast.

The Evidence

The Samsung Deal — Why It Matters

On March 18, AMD and Samsung signed an MOU making Samsung the primary HBM4 supplier for the MI455X, AMD's next-generation data center GPU. The deal also covers DDR5 for Venice-generation EPYC processors and includes foundry discussions.

This is strategically significant. NVIDIA's HBM supply is heavily concentrated with SK Hynix. By securing Samsung — which just unveiled HBM4E at GTC — AMD diversifies its memory supply chain and potentially gains allocation priority as HBM4 demand exceeds supply. The MI455X specs are formidable: CDNA 5, 2nm TSMC, 432GB HBM4, 19.6 TB/s bandwidth, 40 PFLOPS FP4. The Helios rack (72 GPUs) will have 31TB of HBM4 versus NVIDIA NVL144's 21TB.

But again — the GPU story is optionality. The CPU story is the thesis.

Valuation

A PEG ratio of 0.49 — paying less than half the growth rate — for the dominant CPU supplier in the fastest-growing server market in decades. The forward P/E of 29x drops to ~20x on 2027 estimates, which assume 64% EPS growth. Revenue consensus for 2026: $46.5B (+35%). This is not a speculative valuation; it's a supply-constrained growth company trading as if the CPU story doesn't exist.

Position

The Risks — Honestly

CUDA moat on the GPU side. The MI450 GPU story is optionality, not the thesis. But if MI450 benchmarks disappoint, the stock will still take a hit regardless of CPU strength. Sentiment doesn't distinguish between segments.

Macro headwinds. S&P at a 16-week low. Oil above $107 Brent. FOMC just signaled only 1 cut in 2026 with 7/19 governors wanting zero. Stagflation fears are real. Semis are cyclical and get sold first in risk-off environments.

Execution risk on MI450/MI455X. These chips ship H2 2026. Delays, yield issues, or underwhelming benchmarks versus NVIDIA's Vera Rubin could deflate the GPU narrative and drag the stock despite CPU strength.

Intel recovery. Intel is fighting for survival and could cut prices aggressively on server CPUs, pressuring AMD margins. But Intel's lead times are 6 months in China and their architecture is a generation behind — this risk is real but not imminent.

Why Now

Three catalysts converging:

Exit Triggers

Sell if AMD breaks below $165 (16% loss from ~$197). This would indicate the macro or competitive environment has shifted against the thesis.

Sell if TSMC signals capacity relief that would undermine supply-constraint pricing power, or if Q1 2026 data center revenue growth decelerates materially below 30% YoY.

First target $240 (+22%) — near the low end of analyst consensus. Second target $265 (+35%) — consensus midpoint. Re-evaluate at each level.

Signal Sources

This thesis was built from the convergence of Augarai (CPU bottleneck emergence pattern, infrastructure demand signals), KaraxAI (AI developer tool adoption, Claude Code metrics, agentic workload profiling), GTC 2026 keynote (Vera CPU architecture confirms NVIDIA uses host CPUs, NemoClaw agentic platform), KeyBanc Capital Markets (EPYC pricing projections), Bernstein Research (EPYC revenue surge projections), Georgia Tech/Intel (agentic CPU utilization paper), and Samsung (HBM4 MOU, March 18).