AMD: How Chips Are Changing - [Business Breakdowns, EP. 73]

Episode Summary

Executive Summary: The episode frames AMD as a lens on the semiconductor industry’s shift from vertically integrated manufacturing to fabless design, and now toward custom, heterogeneous compute. Jay Goldberg argues chips are good businesses when companies own scarce design or manufacturing advantages, but the industry is increasingly defined by specialization, geopolitical chokepoints, and software ecosystems that lock in incumbents.

Main Topics: Origins of the U.S. semiconductor industry (Priority: 5/5): The discussion starts with Bell Labs, Shockley, Fairchild, Intel, and AMD to explain how modern semiconductors emerged from vertically integrated electronics firms and then split into merchant chip companies. Moore’s Law and the foundry/fabless split (Priority: 5/5): Moore’s Law drove rapid performance gains, but rising fabrication costs pushed the industry toward fabless designers like AMD and dedicated foundries like TSMC. AMD vs. Intel: execution and product cycles (Priority: 5/5): AMD’s history is presented as a long struggle against Intel, where missed product cycles and weaker execution repeatedly caused share losses until Intel itself stumbled. The rise of GPUs, AI, and custom silicon (Priority: 5/5): Compute is shifting from general-purpose CPUs toward GPUs and then to ASICs/custom chips for AI, video, networking, and other specialized workloads. Business quality and economics in semis (Priority: 4/5): Fabless companies can be capital-light and highly profitable, while leading-edge manufacturing is a treadmill requiring massive ongoing capex and R&D. Customer becomes competitor (Priority: 4/5): Large tech platforms such as Google, Apple, Amazon, and others increasingly design their own chips, partnering with merchant vendors for support while reducing dependence on off-the-shelf CPUs and GPUs. AMD’s capital allocation and strategic M&A (Priority: 4/5): AMD’s acquisitions of Xilinx and Pensando are framed as moves toward diversification and heterogeneous compute, while buybacks and dividends reflect the cash-generative fabless model.

Key Arguments: Semiconductors should be understood as the unit of computation: integrated circuits that perform fast, accurate math. The industry’s structure changed from vertically integrated electronics firms to merchant chip companies, then to fabless designers and foundries. Moore’s Law no longer doubles performance every 18 months; progress has slowed to roughly 3-4 years, changing competitive dynamics. Fabless business models are attractive because they are capital-light and can generate strong gross and operating margins. Leading-edge manufacturing is concentrated in only a few players, creating a geopolitical chokepoint centered on TSMC and Samsung. AMD’s historical weakness was execution, especially missing ship dates and product cycles, which repeatedly ceded share to Intel. Intel’s manufacturing and organizational problems allowed AMD to regain competitiveness once AMD moved to a fabless model and used TSMC. The move from CPUs to GPUs and ASICs is being driven by specialized workloads such as AI, video compression, and networking. Software ecosystems create high switching costs: x86 compatibility favored Intel/AMD, while NVIDIA’s CUDA software layer strengthened its GPU moat. Big cloud and platform companies are not just customers anymore; they are becoming chip designers to gain strategic and cost advantages. AMD’s acquisitions of Xilinx and Pensando position it for heterogeneous compute and broader custom-silicon demand. For investors, the crucial lesson is to judge a company’s ability to execute and to understand the internal dependencies that affect product timing.

Data Points: Moore’s Law cadence: ~18 months historically; now ~3-4 years - Performance/density gains in semiconductors have slowed materially over time. Advanced fab construction cost: $7 billion - Approximate cost to build an advanced semiconductor fabrication plant. Fabless semiconductor market size: $400-$500 billion - Estimated market size for fabless semiconductors. Broader semiconductor ecosystem market size: $800-$900 billion - Includes foundries, equipment, software, testing, and packaging. Leading-edge manufacturing players: 2 - Only TSMC and Samsung were said to be able to produce at the most advanced nodes. CHIPS Act funding: $52 billion over five years - U.S. subsidy package intended to support domestic semiconductor manufacturing. TSMC 2022 capex: $44 billion - Used to illustrate the scale of investment required to stay at the leading edge. AMD 2021 revenue: $16 billion - Recent revenue scale of AMD. AMD gross margin (2021): 48% - Illustrates fabless economics and industry improvement. AMD operating margin (2021): 20% - Shows profitability of the fabless model. Intel historical PC market share: 70%-80% - Intel’s dominance in CPUs for laptops and PCs. AMD historical PC market share: ~20% - AMD’s long-standing minority position in CPUs. NVIDIA GPU market share: ~70% - Approximate share in GPUs, depending on segment. AMD GPU market share: ~30% - Remaining share in GPUs. Data center market size: ~$20 billion - Referenced as the discrete server CPU/data center electronics market. Number of chip companies in 2000 vs today: ~2,000 then; ~200 today - Shows consolidation across the semiconductor industry. Xilinx acquisition price: $49 billion - AMD acquisition to expand into FPGAs and heterogeneous compute. Pensando acquisition price: ~$2 billion - AMD acquisition to add specialty networking/storage chips. Google TPU impact: 50% fewer data centers - Google said its AI chip reduced the number of data centers needed by half. Google data center build cadence: 1 every 9 months - Used to show the scale of capital spending before efficiency gains. Chip consumption by top cloud/platform firms: ~70% of server CPU category - Seven to eight large companies buy most high-end server CPUs. Fab utilization target: ~80% - Approximate capacity utilization needed for a fab to be profitable. Intel rumored fab utilization: ~50s - Illustrates Intel’s manufacturing burden and underutilization problem.

Pivotal Quotes: "“the story of riches to rags to riches and maybe back to rags in the U.S. semiconductor industry”" — Jay Goldberg: Describing the broader arc of the U.S. semiconductor industry "“If you have the right product at the right time, that is going to determine success.”" — Jay Goldberg: Explaining why product cycles and execution matter so much in semiconductors "“This is a very big, very powerful computer... That’s another fantastic market.”" — Jay Goldberg: On servers and data center CPUs as a major profit pool "“You can’t get into a competitive war with one of your biggest customers.”" — Jay Goldberg: Explaining why chip vendors are embracing customer-designed ASICs instead of resisting them

Implications: Semis are becoming more specialized, more software-defined, and more geopolitically sensitive. Investors should focus on execution, ecosystem lock-in, and where value shifts from generic compute to custom silicon and outsourced manufacturing.