Oleksandr

$AMD's $1,200 stock w/ Accelerated CapEx toward Inference IMO 🧵 Not Financial Advice! DYOR! AI Bears are trying really really hard to convince people to sell AI Leaders, or how AI CapEx gonna collapse in some months while this is actually a 20-30 years supercycle. Dr. Su was right on Inference roadmap and Agentic AI, she will be right again on this supercycle. The AI CapEx supercycle is in its earliest stages and has strong structural tailwinds for 20+ years. We are still pre-robotic manufacturing at scale, pre-widespread physical/embodied AI, and pre-AGI. The bulk of long-term AI Total Addressable Market and associated infrastructure spend will shift to inference continuous, usage-driven compute that scales with adoption rather than one-time training runs. 1. Revenue & EPS to reach $1,200 by end of 2027? 45 analysts are projecting FY2027 Revenue $58-$106B EPS: $13-$19.40 I believe this is too conservative, and the current P/S is ~21-22x. I will link the thread where I discussed my projection for FY2027 and various threads on supply chain & Superior TCO system to support this price target. The TLDR: 4GW Helios Rack from $META and @OpenAI + $21B(Other segments with Flat YoY to be conservative) =$90+$21B= $111B. This is excluding many deals that AMD signed up recently and existing deals + Agentic AI Rack Revenue. $ORCL , $MSFT , $GOOGL , $AMZN , $SPCX ,Softbank, Anthropic, TensorWave(1.5GW), LumaAI(0.5-1GW), Softbank, 5C(1.5-2GW 2026-28), Dell, HPE, SMCI, European Countries, SEA, East Asia, and so many more. Morgan Stanley projected 6.75m Venice to be sold in 2027 or $60-$100B from just CPU, excluding other mixed EPYC offerings. We will probably hear more about Agentic AI Rack at Advancing AI event. IMO Sales & Non-GAAP EPS Bear case: $120-$140B( EPS $23-$25) Base case: $150-$160B (EPS $29-$33) Bull case: $170-$200B (EPS $34-$38) P/S at similar range 21-22x would require 40-50% growth for FY2028 or how the market feels about $AMD growth for 2028. But 40-50% growth in 2028 is possible as demand for Inference and TAM for Inference will continue to expand for years to come. While we don't really know the exact P/S be at by end of 2027, but at bear case, that would already be enough to reach >$1,200 at current 21-22x P/S. I believe Institutions will keep Semi stocks in this 15-25x P/S range or Fwd P/E in the 15-30x for years, as they could not stop talking about AI CapEx slowdown for 3-4 years now. All of this will depend on $TSM(supply chain thread linked below) 2. Current State: Early Innings of a Multi-Year (and Multi-Decade) Buildout Hyperscalers (Microsoft, Amazon, Google, Meta, SpaceX & others) are already committing hundreds of billions annually, with projections showing sustained or accelerating spend through at least 2030–2031. Goldman Sachs estimates combined CapEx from the four major hyperscalers at roughly $5.3 trillion cumulative from 2025–2030, with broader industry AI infrastructure (compute, data centers, power) reaching ~$7.6 trillion over a similar window. Annual figures are projected to rise from hundreds of billions toward $1.6 trillion by 2031 in baseline models. Global data center CapEx/infrastructure spending heading toward $2-3 trillion cumulative or annual levels by 2030. AI-driven accelerated servers potentially comprising up to two-thirds of data center infrastructure spend by 2030 This is not a short hype cycle. Analysts describe it as a "multi-year investment cycle" with private markets, debt, and equity stepping in alongside internal cash flows. Upward revisions to forecasts have been consistent as demand outpaces initial expectations. Power infrastructure is both a constraint and a parallel investment theme. Global data center electricity demand is projected to roughly double or more by 2030 (IEA baselines around 945–1,000+ TWh), with AI as the key accelerator. U.S. projections vary but often show data centers rising to several percent of national electricity use. 3. Inference Will Dominate Long-Term AI Compute and TAM Training frontier models is compute intensive but episodic (periodic upgrades or new architectures). Inference running models for real-world queries, agents, generation, and decisions is continuous and scales directly with users, tasks, and complexity. And $AMD is the biggest winner on Inference (EPYC roadmap), market share is TBD. Inference forecasted at 79% CAGR through 2030 vs. ~25% for training; potentially 80% of AI critical IT load by 2030. Shift from ~50/50 or training heavy today toward 80%+ inference by late 2030 (Deloitte, Futurum, McKinsey, Brookfield, Lenovo forecasts of 80/20 reversal). AMD EPYC CPUs as the leading inference chip Among inference-optimized chips, AMD EPYC processors (5th Gen Turin 9005 series and next-gen Venice/Zen 6) stand out as leader for long term, large-scale inference deployments and are positioned as a #1 option especially for agentic AI and production workloads. AMD’s EPYC lineup delivers leadership rack-scale throughput and performance-per-watt in agentic AI inference, with modeled claims of 2.37x advantage over competing solutions like NVIDIA Vera CPUs (and up to 3.3x with Venice) in relevant 100 kW rack scenarios. They offer exceptional core density (up to 192+ cores today, scaling beyond 36,000 cores per rack in future generations), superior single-threaded performance critical for agentic workflows, and strong cost/TCO advantages. As inference demand grows to dominate the TAM, AMD EPYC’s x86 compatibility, power efficiency, and ability to handle both standalone inference racks and GPU-hosting roles make it a foundational choice for scalable, cost-effective production deployments. This complements GPU(MI355X, MI455X + MI500) accelerators for the heaviest workloads while capturing a growing share of the inference ecosystem particularly as agentic AI and diversified workloads proliferate. AMD is explicitly building inference optimized variants and optimizations to further solidify this position. 4. Physical AI and Robotics: The Next Major Leg We are not even at robotic manufacturing or scaled physical AI today. Current AI is overwhelmingly digital/cloud-based. Embodied AI (robots, autonomous systems interacting with the physical world) introduces entirely new compute demands that will sustain and amplify CapEx for decades. Morgan Stanley’s robotics analysis highlights explosive edge computing needs: By 2050, potentially 1.4 billion robots sold globally, driving edge AI compute demand equivalent to millions of high-end GPUs. Tesla has discussed robots as distributed compute nodes; aggregate from large fleets could rival or exceed centralized clusters. Humanoid and service robot shipments are ramping (tens to hundreds of thousands in coming years, with faster growth projected). Each unit needs ongoing inference; fleets create recurring, distributed demand. Training world models and simulators for physical interaction requires frontier-scale (or larger) compute clusters. Robotic manufacturing itself will be AI-driven, creating a self-reinforcing loop for more hardware and infrastructure CapEx. Epoch AI analyses support continued scaling feasibility through 2030 and beyond under plausible assumptions about power, chips, and efficiency. We will see: ~Productivity gains → Higher GDP/economic output → More capital available for investment in AI and supporting infrastructure. ~AI could materially boost overall energy demand via growth ~ Electricity grids, highways, internet/telecom buildouts took decades with sustained CapEx, upgrades, and expansions. ~Longer-term views, SoftBank’s Masayoshi Son envision trillions annually in AI-related investment by 2040, with AI revenue potentially reaching 20% of global GDP, making the spend a “rounding error” economically. Power, chips, cooling, networking, and data centers will see waves of investment: initial mega-clusters for training, then distributed/edge for inference, efficiency upgrades, and new capacity as adoption grows. Custom silicon, better algorithms, and synthetic data help on the supply side but historically have been outpaced by demand elasticity. Conclusion: Dr. Lisa Su and the AMD team have been preparing for precisely this moment for years. While headlines focused on GPU training clusters, Dr. Su consistently highlighted the coming inference J-curve, the point where serving intelligent models at massive scale would drive sustained, and ultimately larger, infrastructure demand than training alone. Recent developments with agentic AI workloads, long-context models like OpenAI and Anthropic, and the shift toward CPU-optimized or hybrid inference architectures have validated this foresight. AMD EPYC processors, with their rack-scale leadership in agentic inference (modeled advantages of 2.37x–3.3x in key comparisons), high core density, and strong TCO for production serving, are ideally positioned to capture a significant share of this expanding TAM. Inference, especially agentic inference, will be the dominant force for the next 20–30+ years. Training runs remain important but periodic; inference is perpetual, usage driven, and compounds with every new application, user, agent, robot, and interaction. As AI moves from digital copilots to embodied physical systems, humanoids in factories and homes, autonomous fleets, robotic manufacturing lines, and world model driven simulation; the compute requirements diversify into continuous, real-time, distributed workloads that favor efficient, scalable solutions like EPYC Racks alongside accelerators. This is not a one-time buildout but a foundational infrastructure layer for intelligence itself, akin to electricity grids or the internet, only with faster iteration and broader economic leverage. The supercycle endures because each advance in capability unlocks more demand: better models enable more ambitious agents and robots; widespread physical AI generates fresh data that improves models; productivity gains create the economic surplus to fund further infrastructure. Projections already point to trillions in cumulative CapEx through the 2030s, with longer-term outlooks (such as calls for annual multi-trillion-dollar AI investment by 2040) reflecting a world where AI revenue could represent a substantial share of global GDP. Power, chips, data centers, and edge systems will see ongoing waves of investment. Initial mega-clusters, then distributed inference networks, efficiency upgrades, and entirely new domains enabled by AGI-level systems. Dr. Su saw the inference supercycle coming, and the evidence increasingly shows she was right. For investors, companies, and economies aligned with this shift, the coming decades represent one of the largest and most durable capital formation opportunities in history. The AI era is not a boom and bust cycle like AI bears called for 200 times in the last 3-4 years; it is the beginning of a multi-decade transformation where intelligence becomes the ultimate infrastructure and inference its primary engine. Not Financial Advice! DYOR!