← Back to list
Colossus (Invest Like the Best / Business Breakdowns)Podcast7 Sep 2022Source: joincolossus.comHost: Colossus

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

In plain words

This piece explains how the chip industry is shifting from general-purpose chips to custom chips for specific tasks, driven by slower performance gains (Moore's Law slowing down). Author Jay Goldberg says AMD's 2008 decision to sell its factories and focus on design was a game-changer—now it uses TSMC's advanced manufacturing to take market share from Intel. He also notes that internet giants like Google and Apple are designing their own chips (e.g., Google's TPU, Apple's M-series) for strategic advantages, not just cost savings. Key holdings: AMD (gaining share via TSMC, margins improving), Intel (stuck on older tech, factory utilization may be only 50%), and Nvidia (its CUDA software locks in AI developers).

AI SummaryAI-generated · may contain errors · verify against the original

AMD (Advanced Micro Devices) is neither the largest nor consistently the best chipmaker globally, but as a quintessential example of the cyclical and structural shifts in the semiconductor industry, its trajectory reveals the core dynamics of the sector. In this episode of Business Breakdowns, semic

~11 min full read · 7 sections
Deep Analysis

This Issue at a Glance

Jay Goldberg (Semiconductor Industry Advisor at D2D Advisory, Partner at Snowcloud Capital) deconstructs the structural shifts in AMD and the semiconductor industry. Core thesis: The chip industry is transitioning from a 40-year era of general-purpose computing to an age of customized, specialized chips. The fundamental driver of this shift is the slowdown of Moore's Law—when CPU performance no longer doubles every 18 months, the economics of designing specialized chips for specific tasks are fundamentally transformed.


Theme 1: The Slowing of Moore’s Law — The Fundamental Driver of Industry Restructuring

Jay Goldberg argues that the deceleration of Moore’s Law is the starting point for understanding all current changes in the semiconductor industry.

  • Historical Context: Moore’s Law (named after Gordon Moore, former CEO of Intel) originally stated that transistor density doubles every 18 months, with performance doubling as well. This law drove a 40-year productivity miracle — the computing power of a smartphone in one’s pocket today exceeds that of all computers on Earth in 1950.
  • Current Shift: The performance doubling cycle has lengthened from 18 months to 3–4 years. More critically, its implicit corollary is that the cost of manufacturing chips grows at nearly the same rate. Currently, building an advanced process wafer fab requires $7 billion (for construction and equipment alone).
  • Industry Divergence: Cost pressures have driven two key splits — in the 1970s, electronic equipment manufacturers separated from chip manufacturers; in the 1990s, chip design (fabless) separated from chip manufacturing (foundry). Today, AMD, NVIDIA, Qualcomm, and Broadcom are all fabless companies, outsourcing manufacturing to foundries such as TSMC.

> Data Chain: The global fabless semiconductor market is approximately $400–500 billion; including foundry, equipment, software, and test & packaging, the total reaches $800–900 billion.


Theme 2: AMD's "From Rags to Riches" — How Asset Divestiture Reversed Its Fortunes

Goldberg argues that AMD's 2008 decision to spin off its wafer fabrication plants was the most critical turning point in its history, serving as a textbook case in business schools of "how specialization and focus create shareholder value."

  • Historical Struggles: Since the 1980s, AMD had been locked in a battle with Intel in the CPU market, but suffered from chronic execution failures—its product roadmap repeatedly faced delays, missing the Christmas selling season and driving customers to Intel. This created a self-reinforcing vicious cycle: delays → loss of market share → reduced profits → insufficient R&D → further delays.
  • Key 2006 Acquisition: AMD acquired GPU company ATI in an attempt to diversify and counter Intel, but instead found itself competing on two fronts against NVIDIA, becoming the second player in both the CPU and GPU markets (roughly 20-30% share in CPUs, and about 30% in GPUs).
  • 2008 Turning Point: AMD realized it could no longer afford the "arms race" of Moore's Law and spun off its wafer fabs into Global Foundries (which went public in 2021). Both sides improved significantly after the split—AMD became a fabless company, partnering with TSMC, and was able to fully capture every performance gain from Moore's Law.
  • Intel's Reversal: In 2015-2016, Intel itself hit a process technology bottleneck, stalling at 10 nanometers, falling two nodes behind TSMC (roughly 3-4 years). AMD, leveraging TSMC's advanced processes, steadily captured Intel's market share in PCs and, most critically, in the data center market.

> Data Comparison: Intel's wafer fab utilization needs to be close to 80% to be profitable, but is rumored to have fallen to around 50%; in contrast, as a fabless company, AMD converts nearly all of its operating profit into free cash flow.


Theme 3: The Rise of Custom Chips – The Disruptive Impact of Internet Giants’ In-House Chip Development

Goldberg notes that the chip industry is swinging back from full abstraction toward partial vertical integration, like a pendulum—but this time, it is internet companies, not electronics manufacturers, that are developing their own chips.

  • Mechanism Breakdown: As Moore’s Law slows, general-purpose CPUs can no longer efficiently solve all computing problems, making it economically viable to design application-specific integrated circuits (ASICs) for particular tasks. Data centers are shifting from 100% CPUs to "heterogeneous computing"—the CPU share drops to 50-60%, with the remainder filled by specialized chips such as GPUs, AI accelerators, and video encoding chips.
  • Key Examples:
  • Google VCU (Video Coding Unit): Designed specifically for YouTube video compression and decompression, saving Google hundreds of millions of dollars annually in operating and capital expenditures.
  • Google TPU: Dedicated to AI mathematical operations. Google states that introducing TPUs reduced data center construction needs by 50% (at the time, one data center was built every nine months, each costing $1 billion).
  • Apple M-series/A-series: Apple’s in-house chips combined with its proprietary operating system enable deep hardware-software integration, delivering superior graphics performance, longer battery life, and fanless designs—translating into consumer purchasing decisions and customer lock-in.
  • John Deere: Currently developing its own chips for autonomous tractors.
  • Industry Impact: The seven largest internet companies (Google, Facebook, Amazon, Microsoft, Baidu, Alibaba, and Tencent) consume approximately 70% of server CPU output. They are transitioning from being the largest customers to becoming competitors. However, Goldberg argues this is not an existential threat—these companies will only develop a few strategic chips in-house and will still require a vast array of other chips to populate their data centers.

> Competitive Dynamics: Major chip companies (AMD, Broadcom, Marvell) are choosing to "embrace rather than fight"—assisting internet companies with ASIC support tasks (handing designs to foundries, debugging, and testing) in exchange for opportunities to embed their own other chips into the overall system design of these clients.


Theme 4: High Barriers in Chip Design – The "Lock-In Effect" of Software Ecosystems

Goldberg explains that the moat of the CPU market lies not in the hardware itself, but in the vast software ecosystem built around the x86 instruction set.

  • Nature of the Barrier: The x86 instruction set is the underlying interface between software and chips. Generations of software engineers have spent decades optimizing code to run on Intel/AMD's x86 chips. Switching to a different instruction set like ARM, while 80% of the work can be automated through cross-compilation, the remaining 20% requires extensive manual optimization — "not fun, no one wants to do it, but enough to prevent switching."
  • Current Shift: ARM is approaching a "tipping point" — more companies are willing to invest in that 20% optimization effort due to better performance and lower power consumption. The Chinese market is particularly active, with numerous startups designing ARM-based CPUs/GPUs.
  • NVIDIA's Software Advantage: The CUDA software layer created by NVIDIA years ago provides a low-barrier interface for AI programming, making it the default choice for AI GPUs. Goldberg believes that if he were AMD's capital allocator, he would prioritize investment in software capabilities.

> Quote: "Chips do not operate in a vacuum. They are vehicles for executing software. And the software ecosystem is massive and complicated."


Mentioned Positions

Position Guest Stance Key Data
AMD Bullish (fabless model + TSMC partnership + market share growth) 2021 revenue $16 billion, gross margin 48%, operating margin 20%
Intel Risk warning (process node lag + capital-intensive difficulties) Stuck at 10nm, 2 nodes behind TSMC; fab utilization may be only 50%
NVIDIA Neutral to positive (CUDA software moat) GPU market share approximately 70%
TSMC Bullish (dominance in leading-edge process nodes) 2022 capital expenditure $44 billion; only TSMC and Samsung can produce advanced nodes below 7nm
Google Not explicitly stated (in-house chip case study) VCU (YouTube chip) saves hundreds of millions of dollars annually; TPU reduces data center demand by 50%
Apple Not explicitly stated (benchmark for in-house chips) M-series/A-series annual operating expenditure approximately $1 billion
Broadcom Neutral (ASIC support business) Gross margin above industry average
Qualcomm Neutral (high-margin fabless company) Gross margin above industry average
Xilinx Neutral to cautious (AMD acquisition price too high) AMD acquired for $49 billion, corresponding to revenue of approximately $800 million
Pensando Neutral (AMD acquisition) Acquisition price approximately $2 billion
Samsung Risk warning (competitiveness in advanced process nodes wavering) As of August 2022, "looks a bit unstable"
Global Foundries Neutral (spin-off from AMD) Listed in 2021

Judgments Worth Remembering

1. Goldberg: The implicit corollary of Moore's Law is more important than the law itself — Chip performance doubles every 18 months, but the cost of manufacturing chips grows at nearly the same rate. Building an advanced fab today costs $7 billion, which is the fundamental reason the industry has split into the fabless + foundry model.

2. Goldberg: AMD's 2008 spin-off of its fabs was a "textbook" case of focus creating value — Before the split, AMD was trapped in a vicious cycle of "delays → lost market share → reduced profits → further delays" due to its inability to afford the Moore's Law arms race; after the split, both companies improved significantly.

3. Goldberg: The moat in the chip industry is software, not hardware — The x86 instruction set has accumulated decades of optimized software ecosystems. The final 20% of manual optimization work required to switch to ARM is "not fun but enough to prevent a switch."

4. Goldberg: Internet companies design their own chips not to save a few dollars in chip costs, but for strategic advantage — Google's TPU reduces data center demand by 50%, and Apple's M series delivers consumer-perceptible performance and battery life advantages, translating into customer lock-in.

5. Goldberg: The best strategy for large chip companies is to "embrace rather than fight" the trend of customer-designed chips — Help Google/Amazon with ASIC support work in exchange for opportunities to embed their own chips in the customer's overall system design.

6. Goldberg: Intel's predicament is a problem of organization and capital allocation, not money — Intel has generated massive cash flows for years, but management missteps led to process lag; the U.S. $52 billion CHIPS Act is even smaller than TSMC's annual $44 billion capital expenditure.

7. Goldberg: NVIDIA's CUDA software layer is the core moat of its AI GPU market — If AMD wants to catch up in the high-growth AI market, it must significantly enhance its software capabilities.

8. Goldberg: The chip industry has consolidated from about 2,000 companies in 2000 to around 200 today — This is the core driver behind the industry's gross margin improvement from 20-30% to over 40-50% over the past 20 years.