NVIDIA announced on November 12 that Jensen Huang had resigned as president, chief executive, and director, effective immediately. Huang founded the company in 1993 and ran it for roughly 33 years. The board named Colette Kress, the company’s chief financial officer since 2013, as interim CEO while she keeps the CFO role, opened a search for a permanent successor that will consider internal and external candidates, and appointed Harvey Jones as its first board chair. The filing offered no reason for the departure, and the competitive question does not require one.
Nothing in that announcement erases what NVIDIA is. The company still commands extraordinary technology, an enormous balance sheet, deep engineering talent, and customer relationships that took decades to build. Nobody yet knows what the transition will cost it, if anything. The industry will spend months debating NVIDIA’s future, and that debate is legitimate. The more interesting long-term question sits across the aisle. AMD enters this moment with an executive team that has worked together for more than a decade and a technical strategy that was set long before this week’s headlines. The story may ultimately be less about the leader NVIDIA lost than the engineering organization AMD has spent more than a decade building.
Lisa Su and the Organization Behind AMD’s Revival
Popular accounts of AMD’s revival center on Lisa Su, and with good reason. Su joined AMD in January 2012 and became CEO in October 2014, bringing an MIT doctorate in electrical engineering and semiconductor research leadership from IBM and Freescale. A chip company run by someone who understands chips has an advantage that is easy to underrate. The hero narrative still misses her more durable achievement, which is the team she has kept together around her:
- Mark Papermaster, chief technology officer, joined in October 2011, before Su did. For about 15 years he has owned AMD’s end-to-end technology roadmap, and the company credits him with redesigning its engineering process and leading Zen, its high-performance GPUs, and the modular Infinity Architecture.
- Forrest Norrod joined in 2014 and runs the Data Center Solutions group. His organization delivered multiple generations of EPYC server processors and connects that engineering to the customers who buy it.
- Jack Huynh joined in 1998 as a microprocessor design engineer. Nearly three decades later he runs Computing and Graphics, tying CPU, GPU, and neural-processing work to the platforms partners actually ship.
- Jean Hu (finance, since 2023), David Wang (GPU engineering, returned in 2018), and Vamsi Boppana (AI, by way of Xilinx) add financial, graphics, and AI depth.
What matters is how these responsibilities interlock. Architecture, customer requirements, product investment, and commercialization each have an identifiable owner, and those owners share years of overlapping experience. Institutional memory lives in those relationships: who remembers why a design choice was made six years ago, which supplier interface broke last time, and which customer will test hardest before signing.
This is not a frozen team. Hu arrived in 2023, Wang left and came back, and the 2022 Xilinx acquisition brought an entire engineering culture of adaptive computing and AI engines. AMD combines long tenure with selective renewal, which is healthier than either one alone. Su’s most important accomplishment may be the creation and maintenance of a system capable of producing competitive technology consistently, rather than any single chip she shepherded to market.
Engineering Continuity Is a Competitive Asset
A processor that ships next year reflects decisions made years ago. Chip development runs in overlapping stages, and each generation inherits the work of the one before it:
Workload selection and architecture, several years before launch.
- Intellectual property and interconnect development, spread across overlapping product generations.
- Implementation, verification, and tape-out with foundry and packaging partners.
- Firmware, drivers, compilers, and software enablement after first silicon.
- Customer qualification with PC makers and cloud operators.
- Field lessons carried into the successor product.
Much of what makes this pipeline work never reaches a document. Experienced teams know how their design tools misbehave, where verification has blind spots, and which shortcuts failed before. Some of that knowledge is relational. A CPU architect who understands the GPU team’s memory needs, or a compiler group that anticipates a hardware change, prevents coordination failures that would otherwise cost months. Stable groups reuse proven methods, and reuse shrinks the number of new problems each generation has to solve.
Lag cuts in both directions. Products NVIDIA ships over the next year or two largely reflect decisions made under Huang, so any effect of the transition would appear later, if it appears at all. Nothing suggests NVIDIA has experienced execution problems, and its own bench is deep: Kress since 2013, Jay Puri running worldwide field operations since 2005, Debora Shoquist overseeing manufacturing and suppliers since 2007, and Bill Dally leading research since 2009. NVIDIA also reported overall employee turnover of just 2.5 percent for fiscal 2025, while AMD reported 4.9 percent voluntary and 8.9 percent total turnover for calendar 2025. The definitions differ, but the numbers give AMD no claim to superior retention. The point here is narrower. Replacing an executive can happen in months, while rebuilding years of shared engineering experience cannot.
AMD Is Fighting Two Different Competitive Battles
Reading AMD only through its rivalry with NVIDIA misses half the picture. The company is competing in two distinct markets with different customers, different products, and different rivals.
Enterprise and hyperscale AI
NVIDIA remains dominant in AI acceleration, networking, software integration, and rack-scale systems, where a single GB200 NVL72 rack links 72 Blackwell GPUs in one liquid-cooled machine. AMD competes here with EPYC processors, Instinct accelerators, and its Helios rack-scale systems. Its customer momentum is real and predates the NVIDIA news. OpenAI signed a multigeneration agreement in October 2025 covering up to six gigawatts of AMD GPUs, with a binding initial one-gigawatt commitment, and AMD’s August filing said Meta intends a deployment of similar scale. By November 3, AMD reported volume shipments of its MI455X accelerator and initial Helios racks running with lead customers.
Consumer and prosumer local AI
Apple Silicon established a compelling model for high-memory, integrated personal computing that can run large AI models at a desk. AMD’s Strix Halo family offers a comparable architectural approach inside the native x86 world that most businesses already run. Its natural buyers are developers, researchers, creators, technical enthusiasts, and organizations that want private AI inference without sending data to a cloud provider. AMD does not need to defeat NVIDIA in the data center to become a major competitor to Apple in local AI computing.
Apple Demonstrated the Opportunity. Strix Halo Offers an x86 Alternative.
Apple Silicon places the CPU, the GPU, and a large pool of unified memory in one tightly integrated package. Because the graphics processor can address the same memory as the main processor, a Mac with a lot of memory can hold the weights of a large language model that would otherwise require costly dedicated graphics memory. The Mac Studio with M3 Ultra offers up to 512GB of unified memory and more than 800GB/s of bandwidth, along with strong energy efficiency and Apple’s Metal and MLX software. Developers noticed, and high-memory Macs became the default local AI machine for many of them.
AMD’s response
The Ryzen AI Max+ 395, the chip most people mean by Strix Halo, brings the same idea to x86:
Up to 128GB of shared memory.
- 256GB/s theoretical memory bandwidth.
- 16 Zen 5 CPU cores.
- Integrated Radeon graphics with 40 compute units, plus a neural processing unit.
- Native x86-64 execution on Windows and Linux.
The newer Ryzen AI Max+ PRO 495, code-named Gorgon Halo, pushes capacity further:
Up to 192GB of installed memory.
- Up to 160GB available to graphics in AMD’s specified implementation.
- Approximately 273GB/s theoretical memory bandwidth.
Capacity matters because model weights must sit somewhere. A 70-billion-parameter model compressed to 4 bits per parameter needs about 35GB for its weights alone, more than the 32GB on NVIDIA’s flagship consumer graphics card. A 300-billion-parameter model at the same compression needs roughly 150GB, which squeezes into the 495’s 160GB graphics allocation with little room to spare. A large shared pool lets these models stay resident without buying equivalently large dedicated graphics memory.
Greater capacity does not automatically mean faster inference. The 495 raises maximum memory by 50 percent but bandwidth by only about 7 percent, and generating each token requires reading model weights from memory. Dividing 256GB/s by 35GB of weights yields a rough ceiling of about seven tokens per second for a dense 70-billion-parameter model, before real-world losses. Apple’s high-end bandwidth remains roughly three times greater, and NVIDIA’s DGX Spark desktop matches the 495’s 273GB/s while bringing its CUDA software stack along. Newer mixture-of-experts models, which activate only a fraction of their weights per token, soften the bandwidth problem, and AMD reports 42 tokens per second on one such model in its own testing.
The central competitive proposition
Apple offers a tightly integrated hardware and software ecosystem. AMD offers a high-memory integrated platform that runs the enormous established body of x86 software natively. The opportunity lies with people who want local AI capability without migrating their computing environment. One caution belongs here: x86 compatibility covers the CPU side, not the GPU side. Software written for NVIDIA’s CUDA does not simply run on AMD graphics, and AMD’s ROCm support varies by chip, operating system, and framework. Strix Halo does not need to outperform every Apple Silicon configuration to become a consequential competitor. It needs to offer a compelling combination of memory capacity, application compatibility, performance, and cost for the users it serves.
Why Local AI Changes the Hardware Equation
The next generation of AI computing will not consist only of giant training clusters and cloud inference. Consider Dana, a composite drawn from several small-firm owners I work with. She runs a four-person engineering consultancy whose client files cannot leave the building, and she wants an AI agent that reads project documents, checks a local database, runs the firm’s aging Windows estimating software, and drafts reports overnight. That kind of agent may need to:
- Run a substantial language model.
- Maintain databases and retrieval systems.
- Execute conventional applications.
- Operate browsers and development tools.
- Run containers and background processes.
- Coordinate multiple agents.
- Keep sensitive information under local control.
These requirements make balanced systems more important than benchmark champions. The central engineering distinction is between a computer optimized for an individual inference test and a computer that can complete complex tasks while performing inference. Shared memory makes that distinction concrete, since every gigabyte assigned to the model is a gigabyte unavailable to browsers, containers, and databases. AMD’s combination of large unified memory, native x86 execution, and capable integrated graphics fits mixed workloads like Dana’s well.
AMD is not alone in seeing this. NVIDIA’s DGX Spark explicitly targets local agents with 128GB of unified memory and the full CUDA toolkit, and its new RTX Spark platform brings unified memory to Windows PCs on Arm processors. Dana could also split the job, running her applications on an ordinary x86 workstation that calls a separate inference box. The x86 argument is therefore strongest as a case for simple, consolidated deployment. Local hardware is not free either. A 128GB Framework desktop that launched at $1,999 now lists around $3,449 for the do-it-yourself configuration, a 192GB system starts near $6,799, and NVIDIA’s reference price for DGX Spark is $4,699. The right comparison is cost per successfully completed task, local against cloud.
My own engineering view is simple. The fundamental measure of an autonomous computing system should not be the size of the model it can load or even the number of tokens it can generate per second. It should be the amount of useful work the complete system can perform. Strix Halo’s importance lies in its potential to make that kind of computing practical on familiar consumer and professional hardware.
NVIDIA’s Strength Is Real, and AMD’s Opportunity Is Still Real
Fair analysis requires stating NVIDIA’s advantages plainly:
- CUDA and the developer ecosystem built around it over nearly two decades.
- Leadership in high-performance AI acceleration.
- Sophisticated networking and rack-scale systems integration.
- Vast financial resources, including $96.2 billion in revenue for its most recent reported quarter, against AMD’s $13.4 billion.
- Experienced executives and technical specialists across research, operations, and sales.
- Deep relationships with the largest AI infrastructure customers.
A leadership transition erases none of this. History offers reassuring precedents: Apple under Tim Cook and Microsoft under Satya Nadella both grew enormously after succession. A 2017 meta-analysis of more than 13,000 CEO successions found short-term performance dips but no significant direct long-term effect. NVIDIA may manage this transition well and keep delivering outstanding products. AMD’s opportunities do not require NVIDIA to fail. The relevant question is whether AMD can execute consistently enough to gain customers in the segments where its architecture and relationships offer something different.
The Compounding Effect of a Decade of Engineering
Each successful product generation makes the next one easier to build and sell. AMD’s progression shows the pattern:
- Ryzen restored AMD’s competitiveness in high-performance CPUs.
- EPYC established a significant presence in enterprise and data-center processors.
- Instinct expanded its ambitions in AI acceleration.
- Helios signals a growing commitment to integrated rack-scale infrastructure.
- Strix Halo and Gorgon Halo extend its capabilities into high-memory local computing.
These are not separate bets. They share accumulated expertise in processor design, memory architecture, packaging, system integration, and engineering management, much of it carried by the same people across generations. The Infinity Architecture approach that lets AMD assemble chips from reusable building blocks is a direct product of that continuity, since a modular design only pays off when the teams maintaining the modules stay intact.
None of this progress depends on NVIDIA’s transition. AMD’s agreements with OpenAI and Meta and its first Helios deployments were in motion months before November 12. Those commitments demonstrate an established competitive position, not any benefit from Huang’s departure, and any order announced in the coming weeks will most likely reflect negotiations completed long before. The advantage of continuity is not simply avoiding disruption. It is allowing the knowledge acquired from one product generation to strengthen the next.
What the Next Five Years Will Reveal
The most consequential effects of organizational continuity may not become visible until future product generations reach customers. Through 2031, these are the developments worth watching:
- NVIDIA’s leadership succession and its retention of key technical personnel.
- AMD’s continued execution of its server and AI accelerator roadmaps.
- Successor architectures to Strix Halo, and whether bandwidth grows alongside capacity.
- Improvement in AMD’s GPU software ecosystem and local inference performance.
- Adoption of large-memory x86 platforms for consumer and professional AI.
- Apple’s response through future unified-memory systems.
- Growth of local autonomous computing and changes in its hardware requirements.
- Whether AMD’s continuity produces measurable gains in execution, customer adoption, and competitive position.
AMD faces its own succession question. Su holds both the chair and CEO roles, and the scrutiny now applied to NVIDIA applies to her company eventually. AMD’s voluntary turnover also rose from 4.1 percent in 2023 to 4.9 percent in 2025, a modest drift worth tracking. Leadership continuity is valuable only when it translates into continuing innovation, execution, and customer adoption.
Conclusion: The Power of People Who Know How to Build Together
NVIDIA’s leadership transition is a highly visible event, and its long-term consequences remain uncertain. AMD’s organizational development happened gradually, through years of hard engineering decisions, successful products, commercial disappointments, and accumulated experience. Lisa Su’s achievement is best understood not simply as rebuilding a company but as sustaining an engineering organization capable of competing in several demanding markets at once.
Strix Halo illustrates the value of that capability. AMD has combined substantial unified memory, sophisticated processor architecture, and conventional software compatibility in a platform that offers a credible alternative to Apple Silicon for important segments of local AI computing. Meanwhile, its server processors, AI accelerators, and rack-scale systems keep competing in markets dominated by a much larger rival.
The semiconductor industry has always celebrated its visionary founders and extraordinary individual engineers. Those individuals matter, sometimes enormously. But the most durable competitive advantages are often built by organizations whose people have worked together long enough to transform individual brilliance into institutional capability. NVIDIA must now navigate a change in leadership. AMD has an opportunity to demonstrate what continuity can accomplish. Its greatest advantage may not be any single chip or executive, but the accumulated knowledge of a team that has learned how to build together.