In October 2014, when
Lisa Su took the CEO title at Advanced Micro Devices, the company's market capitalization was $3.2 billion — roughly the valuation of a midsize regional bank, less than 2% of Intel's, and a rounding error against the semiconductor industry's titans. AMD's stock traded under $4. The company had lost money in four of the previous five years. Its share of the server processor market, once a credible 26%, had cratered to single digits. Employees joked, bleakly, about the stock price converging with the price of a sandwich. Wall Street analysts did not so much cover AMD as eulogize it.
A decade later, AMD's market capitalization would exceed $200 billion. Its stock had multiplied more than fifty times from the Su-era trough. It had surpassed Intel in market value for the first time in the two companies' intertwined half-century — a sentence that, had you written it in 2014, would have gotten you uninvited from serious semiconductor conferences. And it had positioned itself as the only merchant silicon company on earth with credible products across CPUs, GPUs, FPGAs, and AI accelerators — the full stack of what the data center of 2025 actually requires. The question animating the next chapter is whether this extraordinary resurrection is the setup for a durable franchise or a cyclical peak dressed up in AI-era clothing.
By the Numbers
AMD at a Glance (FY2024)
$25.8BAnnual revenue
$12.6BData center segment revenue
$5B+AI GPU (Instinct) revenue in 2024
~$180BMarket capitalization (mid-2025)
26,000+Employees worldwide
50xStock price appreciation, 2015–2024
$49BXilinx acquisition value (2022)
69%YoY data center revenue growth, Q4 2024
The Showman and the Second Source
AMD's founding mythology is inseparable from the personality of W. Jerry Sanders III — a man who descended staircases at company parties in matching fur coats with his wife, drove Ferraris before his company turned a profit, and minted the axiom that defined an entire era of the semiconductor industry:
"Real men have fabs." Sanders was not an engineer. He was a salesman, a marketing executive from Fairchild Semiconductor — the ur-company of Silicon Valley — who cofounded AMD in 1969 with seven colleagues, a year after the so-called "traitorous eight" had already splintered Fairchild into the fragments that became Intel, National Semiconductor, and the rest. Where Intel had
Robert Noyce's patrician authority and
Gordon Moore's scientific gravitas, AMD had Sanders's bravado and a willingness to play whatever role the market would give it.
For much of its first two decades, that role was second source. In the early microprocessor industry, customers — particularly the U.S. military and large enterprise buyers — insisted on having at least two suppliers for critical chips. IBM, designing the original PC in 1981, required Intel to license its x86 architecture to a second manufacturer. AMD became that second manufacturer. The arrangement was elegant in its simplicity and corrosive in its implications: AMD existed, in a fundamental sense, because Intel's customers demanded an insurance policy against Intel's monopoly. The company built its identity, its revenue base, and its engineering culture around the x86 instruction set — someone else's intellectual property, licensed under legal agreements that would generate decades of litigation, a 1995 settlement, and the permanent structural condition of AMD's competitive life.
The relationship between Intel and AMD is one of the strangest dyads in corporate history. They were not partners and not quite competitors — more like a dominant species and its obligate parasite, each needing the other for reasons neither would fully admit. Intel needed AMD to maintain the legal fiction of a competitive x86 market, deflecting antitrust scrutiny. AMD needed Intel's architecture to have any market at all. For long stretches of the 1980s and 1990s, AMD reverse-engineered Intel's chips to produce compatible alternatives at lower price points, selling to the customers Intel didn't want — the budget PC buyers, the second-tier OEMs, the markets where price mattered more than prestige.
It was almost a joke, right? Because for decades they had these incredible performance problems. And that's changed.
For a company biography that captures the Sanders-era swagger in full, Jeffrey Rodengen's
The Spirit of AMD: The Legend of Advanced Micro Devices remains the most complete account of the founding generation's culture — part hagiography, part time capsule of an industry that barely exists anymore.
The x86 Trap
The x86 instruction set architecture is simultaneously AMD's greatest asset and its most persistent constraint — a paradox that has defined every strategic era of the company's existence. Understanding AMD requires understanding what x86 is and what it costs.
An instruction set architecture (ISA) is the language that software speaks to hardware. When a developer writes code that runs on a PC or server, it ultimately compiles down to x86 instructions — a set originally designed by Intel in 1978 for the 8086 processor. Over four decades, the x86 ecosystem accumulated an almost inconceivable depth of software compatibility: operating systems, compilers, enterprise applications, databases, games, drivers, firmware. Trillions of dollars of global software infrastructure assumes x86. This installed base is the moat around the moat — the reason why ARM, MIPS, RISC-V, and every other alternative architecture have found it so brutally difficult to penetrate the PC and server markets despite periodic technical superiority.
AMD is one of only two companies on earth licensed to design and sell x86-compatible processors. (Intel is the other. VIA Technologies holds a vestigial license but has been commercially irrelevant for years.) This duopoly status gives AMD something approaching a legal monopoly on being Intel's alternative — a structural advantage worth tens of billions of dollars in server and PC revenue. But the x86 license also chains AMD to Intel's strategic decisions, Intel's roadmap cadence, and Intel's definition of what a processor should be. Every AMD chip must maintain binary compatibility with software written for Intel's architecture. Every AMD innovation must fit within the x86 framework or risk breaking the compatibility that justifies the company's existence.
The trap has a second jaw. Because x86 is a complex instruction set (CISC), the chips that implement it are inherently larger, more power-hungry, and more expensive to design than chips built on reduced instruction sets (RISC) like ARM. As computing shifts toward workloads where power efficiency matters — mobile devices, edge AI, custom silicon for hyperscalers — the x86 tax becomes heavier. Amazon's Graviton processors, Apple's M-series chips, and the explosion of ARM-based data center silicon all represent a world learning to live without x86. AMD's x86 license is a castle with rising waters.
The ATI Bet and the Accidental GPU Company
In 2006, AMD made what many analysts at the time considered a catastrophic acquisition: the $5.4 billion purchase of ATI Technologies, a Canadian chipmaker whose primary business was graphics processing units for video games and personal computers. The deal was expensive — AMD paid a significant premium — and the strategic logic seemed dubious. AMD was a CPU company. ATI was a GPU company. The synergies were theoretical at best, and the debt burden was crushing.
The acquisition nearly destroyed AMD. Combined with an aggressive and ultimately failed attempt to compete with Intel in quad-core server processors, the ATI deal left AMD hemorrhaging cash. By 2008–2009, the company was in existential crisis. It had tried to maintain its own semiconductor fabrication facilities — Sanders's "real men have fabs" philosophy — while simultaneously integrating a major acquisition and fighting Intel's process technology lead. Something had to give.
What gave was the fab. In 2009, AMD spun off its manufacturing operations into a joint venture with the Abu Dhabi sovereign wealth fund, creating GlobalFoundries. The decision was heretical by the standards of the Sanders era but rational by every financial metric that mattered. AMD became a fabless chip designer, outsourcing production to GlobalFoundries and, increasingly, to Taiwan Semiconductor Manufacturing Company (TSMC). The move freed AMD from billions in capital expenditure obligations and, more importantly, gave it access to TSMC's relentlessly advancing process technology — a advantage that would prove decisive a decade later.
The ATI acquisition looked like a disaster for years. Then the world changed. The rise of GPU computing — first for cryptocurrency mining, then for deep learning, then for the generative AI explosion — transformed graphics processors from a niche gaming component into the most strategically important semiconductor category on the planet. Nvidia's
Jensen Huang saw this first and captured the vast majority of the value. But AMD, thanks to the ATI deal, was the only other company with both the GPU intellectual property and the x86 CPU franchise to compete across the full data center stack. The worst acquisition of 2006 became the essential acquisition of 2023.
The Fabless [Pivot](/mental-models/pivot) and the TSMC Lifeline
The decision to go fabless was, in retrospect, the single most important structural decision in AMD's history — more consequential than any product launch, any CEO appointment, any individual chip design. It reframed AMD from a vertically integrated manufacturer competing against Intel's colossal capital expenditure machine into a design house competing on architectural innovation, with TSMC's process technology as an equalizer.
The significance of this cannot be overstated. Intel's competitive advantage for decades rested on the integration of design and manufacturing — the ability to co-optimize chip architecture with fabrication process, achieving performance and density that competitors couldn't match. Intel's process technology lead was real and substantial, often a full generation ahead of merchant foundries. This advantage began eroding in the mid-2010s, as Intel stumbled on its transition from 14-nanometer to 10-nanometer manufacturing — a delay that stretched from months to years to what increasingly looked like a structural inability to keep pace with TSMC's cadence.
As Chris Miller documents in
Chip War: The Fight for the World's Most Critical Technology, the global semiconductor supply chain's convergence on TSMC as the sole leading-edge manufacturer was one of the most consequential geopolitical developments of the early 21st century. For AMD specifically, TSMC's ascendance meant that a fabless chip designer could now access better process technology than Intel's own internal fabs — a sentence that would have been nonsensical in 2010 but was empirically demonstrable by 2019. AMD's Zen 2 processors, manufactured on TSMC's 7-nanometer process, outperformed Intel's competing products still stuck on a node Intel misleadingly labeled "14nm++." The student had access to better tools than the master's own workshop.
GlobalFoundries, AMD's former manufacturing arm, acknowledged the reality in 2018 when it abandoned leading-edge process development entirely, focusing instead on mature-node specialty chips. AMD shifted its most advanced designs exclusively to TSMC. The move was the final severing of Sanders's manufacturing philosophy and the completion of AMD's transformation into a pure design company — asset-light, TSMC-dependent, and suddenly dangerous.
Lisa Su and the Engineer's Turnaround
Lisa Su arrived at AMD in 2012 as a senior vice president, became CEO in October 2014, and has since engineered one of the most remarkable corporate turnarounds in technology history. Her biography is a study in compressed competence. Born in Tainan, Taiwan — the same city that produced Nvidia's Jensen Huang, to whom she is a distant relative — Su emigrated to the United States at age three, earned a PhD in electrical engineering from MIT at 24, and spent over a decade at IBM, where she ran the emerging products division and worked on silicon-on-insulator technology and strained silicon. A stint as CTO of Freescale Semiconductor followed. She is, in the most literal sense, a chip engineer who happens to run a chip company — a rarity in an industry where CEOs are increasingly drawn from finance, operations, or sales.
When Su took over, AMD was a company with mediocre products, crushing debt, declining market share, and a demoralized workforce. Her strategic bet was deceptively simple in concept and ferociously difficult in execution: design a completely new CPU microarchitecture from scratch — what became the Zen architecture — and bet the company's survival on its success. The project had been initiated under her predecessor, Rory Read, with the hiring of legendary chip architect Jim Keller in 2012. Keller, a peripatetic genius who had previously designed AMD's K8 architecture (the basis of the successful Athlon 64) before stints at Apple and elsewhere, laid the architectural foundations of Zen before departing AMD in 2015. Su's contribution was the relentless, grinding execution required to bring that architecture to market, iterate on it across multiple generations, and deploy it across every segment AMD served.
There are so few companies in the world that have access to the intellectual property that we have and the customer set that we have, and the opportunity frankly to really shape how AI is adopted across the world. I feel like we have that opportunity.
Su's management style is often described as deliberate, plainspoken, and people-oriented — a temperamental opposite of Sanders's flamboyance and a contrast to Huang's operatic intensity. Where Huang manages by overwhelming velocity and personal involvement in every technical decision, Su operates through disciplined prioritization and a willingness to say no. She killed unpromising product lines, focused R&D spending on the segments where AMD could differentiate, and — critically — maintained long-term roadmap commitments that rebuilt trust with OEM partners who had been burned by years of AMD's broken promises and missed schedules.
The results were staggering. AMD's revenue grew from $4.3 billion in 2016 to $25.8 billion in 2024. Its stock price went from under $2 in early 2016 to over $200 at its peak. It surpassed Intel in market capitalization in 2022 — a milestone that would have seemed hallucinatory to anyone who had watched AMD struggle through the 2010s. Time Magazine named Su "CEO of the Year." Fortune named her Businessperson of the Year in 2020, noting that the stock price had risen from under $2 to over $85 during her tenure, on revenue approaching $9 billion.
Key milestones in AMD's transformation under Lisa Su
2012Lisa Su joins AMD as SVP and GM of global business units; Jim Keller hired to design new CPU architecture.
2014Su becomes CEO. Market cap: $3.2 billion. Stock price: ~$4.
2017Zen architecture launches as Ryzen (consumer) and EPYC (server), AMD's first competitive products in years.
2019Zen 2 on TSMC 7nm leapfrogs Intel's stalled 14nm process. AMD begins taking meaningful server share.
2020AMD announces $35 billion acquisition of Xilinx. Revenue approaches $9 billion.
2022Xilinx acquisition closes at $49 billion (stock appreciation). AMD surpasses Intel in market cap for the first time.
2023Launches MI300X AI accelerator, AMD's first serious competitor to Nvidia in data center AI.
Zen and the Art of Architectural Warfare
The Zen microarchitecture, which first shipped as the Ryzen consumer processor and EPYC server processor in 2017, is the technical foundation of everything AMD has accomplished under Su. Understanding why Zen mattered requires understanding what it replaced.
AMD's previous CPU architecture, codenamed Bulldozer and its derivatives, was a design catastrophe. Launched in 2011, Bulldozer bet on a "cluster-based multithreading" approach that prioritized throughput over single-threaded performance at a time when most software still depended heavily on single-thread speed. The result was a chip that consumed more power and delivered less per-core performance than Intel's competing Sandy Bridge and subsequent architectures. OEMs stopped recommending AMD. Enterprise customers stopped buying AMD servers. The architecture was so uncompetitive that AMD's server market share, which had peaked near 26% in the mid-2000s during the Opteron glory days, collapsed to low single digits.
Zen was a clean-sheet design that abandoned Bulldozer's failed approach and instead prioritized the fundamentals: high single-thread performance, power efficiency, and a modular "chiplet" design that would prove prescient. The chiplet approach — building a processor from multiple smaller silicon dies interconnected on a package, rather than a single monolithic die — was both an engineering innovation and a manufacturing strategy. Smaller dies have dramatically higher yields than large ones (the probability of a defect killing the chip rises geometrically with die area), and chiplets allow AMD to mix and match die configurations to serve different market segments from a common set of building blocks. A consumer Ryzen and a 64-core EPYC server chip could share the same core chiplets, with different I/O dies and interconnects.
Intel, committed to monolithic die designs and its own manufacturing process, could not easily replicate this approach. The chiplet strategy effectively allowed AMD to extract more usable chips per wafer from TSMC's leading-edge processes while offering a product lineup that scaled from 6-core desktop parts to 128-core server monsters. It was capital efficiency through architecture — a fabless company's answer to the manufacturing advantages it no longer possessed.
Each successive Zen generation — Zen 2 (2019), Zen 3 (2020), Zen 4 (2022), Zen 5 (2024) — delivered meaningful improvements in instructions per clock (IPC), power efficiency, and feature integration. By Zen 3, AMD's EPYC processors had achieved clear performance leadership over Intel's Xeon in most data center workloads. The server market share numbers told the story: from approximately 1–2% in 2017 to roughly 23% by 2024, with the trajectory still climbing. For the first time since the Opteron era, IT buyers were choosing AMD not as a budget alternative but as the performance leader.
The GPU Question and Nvidia's Shadow
If Zen is the story of AMD's resurrection against Intel, the GPU business is the story of AMD's aspiration against Nvidia — a fundamentally different competitive dynamic with a fundamentally different set of odds.
AMD's GPU heritage, inherited from the ATI acquisition, produced the Radeon line of consumer graphics cards — a perennial competitor to Nvidia's GeForce in the gaming market, usually competitive on raw hardware specifications but consistently behind on software, drivers, and the kind of developer ecosystem support that determines which GPU a game studio optimizes for first. In consumer gaming, AMD maintained a respectable but minority position, frequently winning on price-performance at lower tiers while ceding the high-end enthusiast market to Nvidia.
The data center is where the stakes became existential. Nvidia's dominance in AI accelerators is not primarily a hardware story — it is a software story. CUDA, Nvidia's proprietary parallel computing platform launched in 2006, accumulated nearly two decades of developer investment, libraries, frameworks, and tooling before the generative AI explosion of 2023. When researchers at OpenAI, Google DeepMind, or Anthropic write code to train large language models, they write it in frameworks that are optimized for — and in many cases, require — CUDA. This software moat is wider and deeper than any hardware specification gap, and it represents the central strategic challenge of AMD's AI business.
AMD's answer is ROCm (Radeon Open Compute), an open-source software stack intended as a CUDA alternative. Su has been candid about the software gap and aggressive about closing it: by early 2025, AMD claimed that more than one million models on Hugging Face could run "out of the box" on AMD hardware. The company has invested heavily in ROCm optimizations, hired aggressively in software engineering, and pursued partnerships with major AI labs. In October 2025, OpenAI announced a partnership with AMD — a symbolic and commercially significant endorsement that AMD's hardware could serve the most demanding AI workloads.
Our strategy is to establish AMD ROCm as the industry's leading open software stack for AI, providing developers with greater choice and accelerating the pace of industry innovation.
Yet the numbers remain sobering. Nvidia controls an estimated 80–90% of the AI accelerator market. AMD's Instinct GPU revenue, while growing rapidly — exceeding $5 billion in 2024 — represents a small fraction of Nvidia's data center revenue, which surpassed $100 billion in the same period. The MI300X, AMD's current-generation AI accelerator, has won meaningful deployments at Meta, Microsoft, and other hyperscalers, but often in supplementary roles rather than as the primary training platform. Wall Street's verdict has been mixed: BofA Securities lowered its AMD price target in early 2025, and Bank of America analyst Vivek Arya observed that the company had yet to "articulate how it can carve an important niche" relative to Nvidia.
The competitive dynamic here is subtle and worth unpacking. AMD does not need to beat Nvidia to succeed — it needs to be good enough to capture the second source position in a market expanding so rapidly that even 10–15% share translates to tens of billions in revenue. The AI accelerator market, which Su has projected will reach $150 billion by 2027, is large enough that a credible number two can build an enormous business. The question is whether "credible number two" is a stable equilibrium or a transitional state before hyperscalers complete the shift to custom silicon.
The Xilinx Gambit: Buying the Missing Piece
In October 2020, AMD announced the acquisition of Xilinx for $35 billion in an all-stock transaction — a deal that would close in February 2022 at a final value of approximately $49 billion, inflated by AMD's rising stock price during the intervening period. The acquisition was the largest in semiconductor history at the time and represented Su's most ambitious strategic bet: the thesis that the data center of the future would require not just CPUs and GPUs but also field-programmable gate arrays (FPGAs) and adaptive computing platforms, and that the company controlling all four would hold a decisive advantage.
Xilinx, founded in 1984, had invented the FPGA — a chip that can be reprogrammed after manufacturing to perform different functions, offering a middle ground between the rigidity of application-specific integrated circuits (ASICs) and the generality of CPUs. FPGAs found homes in telecommunications infrastructure, aerospace and defense, industrial automation, and increasingly in data center workloads where the computational task was specialized enough to benefit from hardware customization but not large enough to justify a full custom chip design.
The strategic logic was the portfolio argument: hyperscale data center customers — Microsoft, Google, Meta, Amazon — run enormously diverse workloads. Some are best served by general-purpose CPUs. Some by GPU-accelerated parallel computing. Some by FPGA-based acceleration for specific networking or inference tasks. Some by combinations of all three. A vendor offering a coherent platform across all compute types, with unified software tools and co-designed silicon, could capture a larger share of total data center spend than any single-product competitor. AMD, post-Xilinx, was the only merchant silicon company with leading products in all three categories.
The integration has been quieter than the acquisition itself. Xilinx's revenue — reported as AMD's "Embedded" segment — generated $3.3 billion in FY2024, down from a post-acquisition peak as the inventory correction in embedded and industrial markets worked through the supply chain. The more strategic value lies in the IP portfolio, the customer relationships in defense and telecom, and the eventual integration of FPGA adaptive computing capabilities into AMD's data center platforms. Whether this portfolio vision justifies the $49 billion price tag is a question the market is still adjudicating.
The Custom Silicon Threat
The most existential challenge to AMD's long-term franchise may not come from Intel or Nvidia but from its own customers. The rise of custom silicon — chips designed by hyperscale cloud providers for their own specific workloads — represents a structural shift that could erode the addressable market for merchant chip companies across the board.
Amazon's Graviton ARM-based server processors have already captured a significant share of AWS's internal compute. Google's TPUs (Tensor Processing Units) handle a substantial portion of the company's AI training and inference workloads. Microsoft is developing custom silicon under its Maia AI accelerator program. Meta has invested in both custom training chips and inference accelerators. Each of these efforts represents a hyperscaler deciding that the premium for general-purpose merchant silicon is not justified for its largest, most predictable workloads — that it is cheaper and more efficient to design the exact chip the workload requires rather than buy the general-purpose chip that the market offers.
This trend strikes at the heart of AMD's value proposition. AMD sells general-purpose processors — CPUs and GPUs that can run any workload — and charges a margin for that generality. Custom silicon eliminates the generality premium. A hyperscaler that designs its own ARM-based CPU doesn't need an AMD EPYC. One that designs its own AI accelerator doesn't need an AMD Instinct GPU. The x86 installed base provides insulation — the switching costs of recompiling or rewriting software for a new ISA are enormous — but the insulation is finite, and the hyperscalers are precisely the customers with the engineering talent and financial motivation to bear those costs.
AMD's defense against custom silicon is three-pronged: continuing to advance its general-purpose products so aggressively that the performance gap versus custom silicon remains narrow; deepening the software ecosystem (ROCm, development tools, middleware) so that the switching costs of leaving AMD's platform remain high; and using the Xilinx FPGA portfolio to offer the customization benefits of bespoke silicon within AMD's existing platform. Whether this defense holds against customers who collectively spend over $200 billion annually on capital expenditures is, to put it mildly, uncertain.
The Intel [Inversion](/mental-models/inversion)
The competitive relationship between AMD and Intel has undergone a reversal so complete that it reads less like a business story than a parable about institutional decay. For forty years, from the founding of both companies through approximately 2017, the relationship was asymmetric: Intel was the standard, AMD was the alternative. Intel set prices. Intel got the best OEM placements. Intel's process technology led the industry. Intel's margins — regularly above 60% gross — reflected a monopolist's pricing power. AMD survived by being cheaper, occasionally being faster, and always being available as the bargaining chip that enterprise buyers used to extract better terms from Intel.
Then the inversion began. Intel's 10-nanometer process delays, which started in 2015 and extended through the end of the decade, broke the cadence of
Moore's Law execution that had been Intel's core competitive advantage. AMD, now manufacturing on TSMC's processes, shipped 7nm products while Intel was still iterating on 14nm. The performance gap flipped. The market share gap began to close. By 2022, AMD surpassed Intel in market capitalization — a crossing that, for semiconductor industry observers, carried the symbolic weight of a changing of the guard.
Intel's struggles were not merely technical. The company cycled through strategic pivots — mobile, IoT, autonomous driving, foundry services — with a frequency that suggested confusion rather than agility. Its foundry ambitions under CEO Pat Gelsinger, while strategically rational in the abstract, required tens of billions of dollars in investment and years of execution before yielding competitive results. Intel recently acknowledged that its AI accelerator offerings had not gained meaningful traction and canceled an upcoming AI chip. For AMD, Intel's stumbles created a decade-long tailwind: the vacuum Intel left in the server market, AMD filled with EPYC. The vacuum in performance-competitive desktop processors, AMD filled with Ryzen.
But Intel's weakness also removed the urgency — and in a strange way, the clarity — that AMD had derived from being the underdog. When Intel was dominant, AMD's strategy was obvious: build something better, price it lower, take share. Now that AMD is competitive or leading across most x86 product categories, the strategic question becomes more diffuse. Where does growth come from when you've already taken the share you were chasing?
For a detailed account of the Intel-AMD rivalry from the Intel perspective, Michael Malone's [
The Intel Trinity: How Robert Noyce, Gordon Moore, and Andy Grove Built the World's Most Important Company](https://www.amazon.com/dp/B00G2A7WL2) provides essential context on the institutional culture that AMD spent decades fighting and that ultimately cracked from within. And for AMD's own account of the antitrust battle, Hector Ruiz's [
Slingshot: AMD's Fight to Free an Industry from the Ruthless Grip of Intel](https://www.amazon.com/dp/B00CHCOKZ0) tells the story from the inside, complete with the acrimony and litigation that defined the relationship through the 2000s.
The Software Deficit
In semiconductors, the axiom has always been that software eats margins. The company that controls the software ecosystem captures the bulk of the economic value; the company that merely manufactures the hardware commoditizes into a price-taker. Nvidia understood this earlier and more deeply than any other chip company of its generation. CUDA, launched in 2006, was a bet that GPU computing would become general-purpose — that the parallel processing architecture designed for rendering video game graphics could be repurposed for scientific computing, machine learning, and eventually artificial intelligence. By the time the AI training boom arrived in 2023, CUDA had been accumulating developer mindshare for seventeen years.
AMD's software story is the inverse of its hardware story. Where the hardware has gone from joke to juggernaut, the software has remained a persistent weakness. ROCm, while improving rapidly, is years behind CUDA in ecosystem maturity. Developers who write GPU-accelerated code overwhelmingly target CUDA first and ROCm second, if at all. Major machine learning frameworks — PyTorch, TensorFlow, JAX — support AMD GPUs, but the optimization depth, debugging tools, and library completeness lag behind the Nvidia equivalents.
Su has been forthright about this challenge and has made software investment a central pillar of AMD's strategy. The company has hired thousands of software engineers, contributed aggressively to open-source AI frameworks, and pursued an "open" strategy — positioning ROCm as an open-source alternative to Nvidia's proprietary CUDA lock-in. The OpenAI partnership announced in late 2025 suggests that this strategy is gaining traction at the highest-profile AI companies.
Yet the software deficit is structural, not merely a function of investment. CUDA's advantage is a network effect: more developers means more libraries means more tools means more developers. Breaking this cycle requires not just technical parity but a compelling reason for developers to switch — which usually means either significantly better hardware (difficult against Nvidia's pace of innovation) or significantly lower cost (difficult when TSMC manufacturing costs are similar for both companies' chips). The most likely path for AMD's software ecosystem is not to replace CUDA but to thrive in the gaps — inference workloads, cost-sensitive deployments, cloud instances where AMD's price-performance ratio justifies the software migration cost, and the growing universe of open-source AI models that are increasingly hardware-agnostic.
The $1 Billion Supercomputer and the Government Bet
In October 2025, the U.S. government announced a $1 billion partnership with AMD to construct two supercomputers designed to tackle problems ranging from nuclear power to cancer treatments to national security. The deal was emblematic of a broader strategic dynamic: AMD's positioning as the credible alternative in high-performance computing at a moment when governments worldwide are reassessing their dependence on any single semiconductor supplier.
The geopolitics of semiconductors have become inseparable from corporate strategy. U.S. export controls on advanced AI chips to China, enacted in October 2022 and tightened in subsequent revisions, directly affect AMD's addressable market — the company's AI accelerators and advanced server CPUs fall under the restricted categories. China represented a meaningful and growing customer base for AMD's data center products; the export controls effectively closed that market for the most advanced chips. At the same time, the CHIPS Act, signed in August 2022, committed $52 billion in U.S. government subsidies to domestic semiconductor manufacturing and R&D — a windfall that primarily benefits Intel and TSMC (both building fabs on U.S. soil) but also creates funding streams for AMD's design and research operations.
The government supercomputer deal serves multiple purposes for AMD. It validates the company's high-performance computing credentials at the highest level of technical rigor. It diversifies revenue away from the hyperscaler concentration that has increasingly characterized the AI chip market. And it strengthens AMD's relationship with a customer — the U.S. government — that operates on procurement timelines measured in decades, not quarters, offering a base of durable revenue that private-sector AI spending cycles cannot.
The Paradox of the Complete Portfolio
AMD in 2025 is a company that has achieved something no other merchant semiconductor firm can claim: competitive products across CPUs, GPUs, FPGAs, and AI accelerators, all designed under a unified architecture strategy and manufactured by the world's leading foundry. The portfolio is broader than Nvidia's (which lacks CPUs), deeper than Intel's (whose GPU and AI accelerator efforts have faltered), and more general-purpose than any hyperscaler's custom silicon. It is, on paper, the most complete compute platform in the industry.
The paradox is that breadth creates its own vulnerabilities. AMD must simultaneously invest in CPU architectures competing against Intel and ARM-based alternatives, GPU architectures competing against Nvidia, FPGA platforms competing against Lattice and Intel's remaining FPGA business, and AI accelerator roadmaps competing against Nvidia, Google's TPUs, and a proliferating set of AI ASIC startups. Each of these competitions demands world-class engineering talent, billions in annual R&D (AMD spent approximately $5.9 billion on R&D in FY2024), and the kind of focused obsession that is difficult to maintain across four simultaneous fronts.
Nvidia, by contrast, is a company built around a single unifying abstraction: the GPU as universal accelerator. Every dollar of Nvidia's R&D investment compounds within a single architectural paradigm. AMD's R&D must be distributed across multiple paradigms. In a business where the quality of each transistor's contribution to competitive performance is what matters, concentration has an inherent advantage over diversification. The complete portfolio may be a strategic asset in selling to CIOs who prefer a single vendor — but it is a strategic liability in the R&D lab, where focus wins.
We're still in the very early stages of the deployment of AI, and we do know that the AIs are not always right. And so part of what we have to do as a set of leaders is figure out how to use the technology for good and also protect the downsides.
On a Tuesday in February 2025, AMD reported fourth-quarter results. Revenue of $7.1 billion beat analyst estimates. Data center revenue of $3.86 billion grew 69% year-over-year. But Wall Street had expected $4.14 billion. The stock dropped 9% in after-hours trading, closing the next day at $112.01 — its lowest level since November 2023. The gap between the actual numbers and the expected numbers was $280 million, less than 4% of data center revenue. The market's reaction — erasing roughly $16 billion in market capitalization — was less a judgment on the quarter than a question about the trajectory: whether AMD's AI business was accelerating fast enough to justify the valuation premium the market had assigned to the AI narrative.
In the server rooms of Meta's data centers, AMD MI300X chips were running the company's largest Llama model through Meta AI. At Microsoft, the same chips powered OpenAI-based Copilot services. The technology worked. The customers were real. The revenue was growing at rates that any company outside the semiconductor industry would celebrate. The question — the one that the stock price encodes daily, that analysts debate in their target-price revisions, that Su herself must weigh against every R&D allocation decision — is whether "second" is a position you can build a great business from, or just the most profitable way station on the road to displacement. AMD has been here before. The last time, it almost died. This time, it is armed with better chips, better process technology, a broader product portfolio, and the advantage of being the one company left standing with licenses to build x86, build GPUs, build FPGAs, and compete in AI — all at once, on the same roadmap, led by an engineer who speaks the language of the silicon herself.
