Behind Jensen Huang’s Departure as NVIDIA CEO: Four Pivots and an Organizational Experiment Across 33 Years
When a founder hands over the leadership of a chip company after 33 years, the legacy extends far beyond a multitrillion-dollar valuation. NVIDIA’s technical and business history reveals four decisive moments of strategic retreat and reinvention.
science
Editor’s note (November 16, 2026): The original article retains its qualification about transition details at the time of publication. NVIDIA’s official announcement of November 12 states that Huang resigned as president, CEO, and board member with immediate effect. Colette Kress became interim CEO while continuing as CFO.
In November 2026, a major leadership announcement emerged from Silicon Valley: Jensen Huang, who had served as NVIDIA’s president and chief executive since co-founding the company in April 1993, announced his resignation from the CEO role and other management positions. (At the time of writing, the precise effective date, board arrangements, and succession structure remained subject to NVIDIA’s formal filings with the U.S. Securities and Exchange Commission.) In semiconductors—a fiercely competitive industry where technological paradigms shift roughly every decade and even giants can fall—a founder spending 33 consecutive years at the forefront of product definition and engineering architecture is an unusual business case. Taking this historical moment as its starting point, this article draws on public technical literature and historical records to revisit four pivotal decisions in NVIDIA’s evolution, alongside the distinctive organizational structure Huang designed for the company.
1. Intellectual Honesty at the Brink: From NV1’s Failure to a Six-Month Engineering Cadence
Before NVIDIA became a computing giant, it first had to learn how to survive a mistaken architectural bet. In early 1993, Huang—formerly a microprocessor designer at AMD and a director of a core business unit at LSI Logic—met two Sun Microsystems engineers, Chris Malachowsky and Curtis Priem, at a Denny’s on Berryessa Road in San Jose to work out their startup plans [1]. The PC market was on the verge of a battle over 3D graphics accelerators; at its peak, dozens of startups worldwide would compete in the same arena.
In 1995, NVIDIA introduced its first product, NV1. Under the technical constraints of the time, the founders made an architectural choice that seemed ingenious but departed from the mainstream. Instead of the triangle meshes becoming standard across the industry, they chose quadratic texture mapping (QTM). They also attempted to integrate 2D and 3D graphics, sound, and a Sega Saturn game-controller interface onto a single chip. Only months later, Microsoft’s Direct3D graphics interface arrived in the Windows 95 era, establishing a triangle-based rendering pipeline as the common industry standard. NV1’s incompatible approach quickly became a commercial failure, and mounting returns nearly drained the company’s cash.
The real existential test came in 1996. Sega had signed a $7 million contract with NVIDIA to develop the NV2 chip, based on quadratic-surface technology, for its next-generation Dreamcast console. Partway through development, Huang and his engineers realized that the approach was a dead end in both its mathematics and its hardware scalability. Completing the contract would earn NVIDIA its final payment but produce a chip with no future. Stopping work immediately would breach the agreement, while the company had less than a month of cash left.
Huang made an early decision that exemplified his intellectual honesty. He flew to meet Sega of America president Shoichiro Irimajiri and openly acknowledged that NV2’s technical direction was wrong. He advised Sega to switch suppliers immediately to avoid delaying the console. At the same time, he made an extraordinary request: Sega should still pay the remaining $5 million, later converted into an equity investment in NVIDIA. Without it, NVIDIA would go bankrupt immediately, and Sega’s earlier investment would be lost as well [1]. Irimajiri agreed, giving NVIDIA six months of breathing room.
With this last lifeline, Huang introduced three severe internal reforms:
- Start over and cut the workforce by more than half. NVIDIA abandoned its proprietary graphics-interface efforts, reduced its workforce from more than 100 to roughly 40, and redirected everyone toward a triangle-based pipeline fully compatible with Microsoft Direct3D.
- Invest heavily in hardware emulation. To make the next chip work on its first fabrication run, Huang spent much of the remaining cash on an expensive IKOS hardware-emulation system. Engineers could write and debug drivers before the chip had even been fabricated.
- Break with convention through a six-month cadence. Semiconductor development typically took 18 to 24 months. Huang compressed NVIDIA’s architectural iteration cycle to six months, with three engineering teams advancing in staggered relay.
In August 1997, the third-generation RIVA 128 (NV3) arrived. It sold more than a million units in four months and pulled NVIDIA back from the edge of bankruptcy. Just two years later, in August 1999, NVIDIA launched the GeForce 256, integrating hardware transform and lighting (T&L) on a single chip and formally defining its “graphics processing unit,” or GPU, category [2]. The conviction that the company was always “thirty days from going out of business” became part of NVIDIA’s engineering culture.
2. Paying for a “Zero-Billion-Dollar Market”: The CUDA Bet and General-Purpose Computing
If RIVA 128 saved NVIDIA’s life, CUDA (Compute Unified Device Architecture), introduced in 2006, changed what kind of company it could become.
In the early 2000s, Ian Buck and other doctoral researchers working in the Stanford laboratories of Pat Hanrahan and Bill Dally noticed that the parallel shaders used to calculate pixel colors could be “tricked” into performing general-purpose matrix and vector operations. In 2004, Buck and his colleagues published the foundational paper Brook for GPUs: Stream Computing on Graphics Hardware at SIGGRAPH [3]. It systematically demonstrated how a streaming extension to C could run general-purpose scientific computation on graphics hardware: GPGPU.
Huang quickly recognized the engineering implications. He recruited Buck and reached an agreement with NVIDIA’s then-chief scientist, David Kirk: researchers should no longer have to disguise their computations as texture maps. NVIDIA needed to erase the boundary between graphics and computing at the level of the chip’s hardware architecture.
In November 2006, alongside the G80-based GeForce 8800 GTX, NVIDIA introduced its unified shader architecture and CUDA programming model [4]. To enable gaming graphics cards to execute parallel programs written in C, Huang decided to add dedicated compute-control logic and on-chip shared memory to every GPU shipped.
To Wall Street analysts at the time, the economics looked terrible:
- Every customer would bear the hardware cost. Even a graphics card bought simply to play World of Warcraft contained additional transistors intended for scientific computing. That increased consumer chips’ die area, power consumption, and manufacturing cost.
- Six years with little corresponding revenue. From 2006 to 2012, NVIDIA invested hundreds of millions of dollars annually in maintaining and upgrading the CUDA software stack. Customers for its specialized Tesla compute cards were largely confined to a handful of weather agencies, oil-exploration organizations, and university physics departments. During what Huang later called the “zero-billion-dollar market” phase, the company’s market capitalization spent long periods below $10 billion. Its share price also fell sharply in 2008 amid packaging defects and pressure from mobile chips.
Yet Huang maintained a strategy few professional managers would have dared pursue: put CUDA into every ordinary graphics card sold to gamers. By the eve of 2012, hundreds of millions of CUDA-capable GPUs were quietly sitting in university laboratories, graduate students’ rooms, and personal computers around the world.
The turning point arrived in autumn 2012. At the University of Toronto, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton trained the deep convolutional neural network AlexNet without applying for expensive supercomputer time. Instead, they used two consumer GeForce GTX 580 cards, each with 3 GB of memory, and CUDA programming across the GPUs to train on ImageNet. They reduced the image-classification error rate by more than ten percentage points [5]. CUDA’s already widespread availability on computer-science desktops gave the ensuing deep-learning boom a ready-made software foundation on NVIDIA hardware.
3. Strategic Retreat and a New Unit of Computing: From Leaving Phone Chips to Buying Mellanox
An often overlooked feature of Huang’s decision-making is that he was just as decisive about subtraction and retreat as he was about expansion.
Around 2010, as smartphones swept the world, NVIDIA tried to enter the mainstream handset market with its Tegra mobile processors. In 2011, it even paid $367 million for British baseband-chip company Icera. Within a few years, however, Qualcomm’s advantages in integrated baseband and radio-frequency technology had secured its dominance in Android, while Apple shifted to its own A-series chips. Smartphone system-on-chip competition rapidly became a brutal contest of scale and thin margins.
Between 2014 and 2015, Huang made a deliberate strategic retreat, shutting down and seeking to divest Icera’s mobile baseband business and leaving the mainstream smartphone-processor market. Rather than continue spending engineering resources on a battlefield whose rules others set, NVIDIA redirected Tegra’s low-power embedded-architecture capabilities toward the emerging NVIDIA DRIVE autonomous-driving platform, Jetson robotics modules, and the chip for the Nintendo Switch.
Meanwhile, recognizing the deep-learning scaling trajectory revealed by AlexNet, NVIDIA began a dramatic evolution from chip supplier to full-stack computing-systems company:
- Define computational operations through hardware–software co-design. NVIDIA launched the cuDNN deep-neural-network acceleration library in 2014. Its engineers worked closely with AI laboratories’ code to optimize foundational operations, from convolution and pooling to, later, attention. In 2016, NVIDIA introduced its first integrated deep-learning supercomputer, DGX-1; Huang personally delivered one to the recently established OpenAI.
- Redesign chip area around dedicated matrix cores. The 2017 Volta (V100) architecture introduced Tensor Cores for mixed-precision matrix multiply-and-accumulate operations [6]. Subsequent generations progressed through Ampere (A100), Hopper (H100) with its FP8 Transformer Engine, and then Blackwell and Rubin. Hardware evolution became tightly coupled to the expansion of large-model workloads.
The most consequential acquisition during this transition was NVIDIA’s $6.9 billion purchase of Israeli high-speed networking company Mellanox Technologies, announced in March 2019 and completed in April 2020 [7]. At the time, it was the largest acquisition in NVIDIA’s history.
**Concept The Data Center as the Unit of Computing** In distributed training and inference for large models, a single chip’s peak computational performance no longer determines the system’s ceiling. The basic unit of computing has expanded from a board to an entire data center. High-speed interconnects within and between racks—such as NVLink and InfiniBand—work with scheduling software to determine the effective throughput of clusters containing tens of thousands of accelerators.
Many observers initially struggled to understand why a graphics-card company would spend so much on a business making InfiniBand switches and network adapters. Huang’s reasoning reframed computing infrastructure: once models reached tens or hundreds of billions of parameters, a single GPU—or even an eight-GPU server—could no longer contain the entire workload. The data center itself had become the computer. If communication bandwidth could not keep up, tens of thousands of high-end GPUs costing hundreds of millions of dollars could spend more than half their time waiting for gradient synchronization. Combining Mellanox’s remote direct memory access (RDMA) capabilities with NVIDIA’s NVLink and NVSwitch strengthened the company’s system-level offering. Rack-scale products such as GB200 NVL72, connecting 72 GPUs through high-speed copper interconnects, could operate as a much larger integrated computing system. A competitor could no longer overcome that advantage simply by producing a faster individual chip.
4. An Organizational Experiment Against Corporate Ageing: Sixty Direct Reports and T5T Emails
Beyond the technology roadmap lies another question of interest to management researchers and the technology industry: why have established giants such as Intel and IBM faced divisional barriers and slowing innovation as they grew, while NVIDIA retained so much of a startup’s agility even after reaching a multitrillion-dollar valuation?
Part of the answer lies in the unconventional organizational structure Huang designed and sustained [8].
First, an exceptionally flat reporting structure. In most multinational technology companies, management textbooks suggest that a CEO should have six to ten direct reports, each responsible for a substantial business unit. Huang instead maintained roughly 50 to 60 direct reports. His reasoning was straightforward: fewer layers mean less distortion as information travels. Such a wide span of control depends on senior leaders with substantial professional maturity, who require little hand-holding or performative “managing up.”
Second, the “Top-5 Things” (T5T) email mechanism. In a complex hierarchy, technical blind spots or emerging signals identified by frontline engineers often pass through managers, directors, and vice presidents, accumulating filters and polish until they become a presentation for the top leadership. By then, the signal may be six months old. At NVIDIA, employees from engineers to executives send short, plain-text emails listing the five most important things they are working on or observing. Huang reads hundreds of these messages early in the morning and late at night. The mechanism gives the company’s central decision-making system direct access to its periphery. Early signals—researchers using GeForce cards for deep learning, or robotics kits for simulation—can reach the people allocating resources without layers of mediation.
Third, remove independent business-unit barriers and let “the mission be the boss.” NVIDIA does not divide its product lines into separate fiefdoms, each defending an independent profit-and-loss account. Chip architecture, systems hardware, and foundational software instead function as shared capabilities. When the company selects a new mission—whether autonomous driving, the Omniverse industrial digital-twin platform, or BioNeMo for digital biology, virtual cells, and protein design [9]—resources come together across functions. When the mission ends or its premise fails, people can regroup around the next task.
A technology company ages when departmental walls and the instinct to defend existing profits sever the connections between weak signals at the front line and the people who allocate resources.
Huang also largely dispensed with conventional private one-on-one reporting meetings. Strategic discussions and technical reviews instead took place in meetings attended by dozens of people. The engineering rationale was to give every key participant the same context at the same time, reducing information asymmetry and exposing decisions to collective scrutiny.
5. Open Questions After Huang: Inference Economics and Organizational Succession
Huang’s departure closes a chapter in which a first-generation semiconductor founder personally led the charge. Looking ahead from late 2026, NVIDIA retains formidable hardware and software advantages, but its next leadership team faces three unresolved structural questions.
First, how long can full-stack systems sustain their premium as the industry shifts from pretraining at almost any cost to tightly budgeted inference at scale? During pretraining, rapidly changing model architectures encourage customers to buy highly general-purpose flagship GPU clusters with the fastest interconnects. As inference accounts for a growing share of workloads, hyperscalers including Google, Amazon, Microsoft, and Meta are advancing custom chips—such as TPU and Trainium/Inferentia—to reduce the cost per token. Without its founder at the helm, can NVIDIA’s rack-scale NVLink integration and CUDA/NIM software stack continue to defend the very high margins associated with its hardware business?
Second, can a flat network of 60 senior leaders avoid reverting to a conventional hierarchy once the founder’s authority and technical intuition are gone? Huang’s flat structure, direct T5T channel, and avoidance of independent business-unit boundaries worked in large part because he combined a co-founder’s authority with deep architectural judgment. He could engage with both semiconductor packaging parameters and bottlenecks in attention operations. Just as Apple after Steve Jobs moved toward increasingly refined supply-chain operations, NVIDIA’s post-founder era will test whether its organizational legacy can survive the tendency to rebuild layers of approval and departmental walls in the name of risk management.
Third, can digital biology and physical AI succeed large language models as the next decade’s engines of exceptional growth? In the final years of his tenure, Huang repeatedly bet on two frontiers that had yet to reach their full scale. One was Omniverse and the Cosmos world models, intended to help embodied robots understand the physical world. The other was BioNeMo, a foundation for computational molecular generation, protein folding, and virtual-cell perturbation studies. Neither field can rely solely on text already available on the internet. Their data are expensive, noisy, and constrained by the rhythms of physical experiments and biological processes. When computational predictions must confront wet-lab validation and real mechanical control, whether these two seeds grow into major businesses will be a demanding test of his strategic legacy.
Key Takeaways
-
Four strategic choices shaped NVIDIA’s rise. In 1996, it acknowledged the failure of NV2’s architecture and moved to Direct3D and a six-month development cadence. In 2006, it began embedding CUDA in consumer GPUs despite years of limited immediate returns. From 2014, it withdrew from the mainstream smartphone-chip battle. In 2019, it agreed to acquire Mellanox for $6.9 billion, extending its unit of computing from the individual chip to the data center.
-
Organizational design helped resist corporate ageing. Huang’s roughly 50 to 60 direct reports, the T5T email channel, and mission-based coordination across shared capabilities were intended to reduce information distortion and bureaucratic delay at an enormous company.
-
The post-founder era brings three tests. NVIDIA must respond to hyperscalers’ custom inference chips, preserve its organizational agility without the founder’s technical authority, and determine whether physical AI and digital biology can become major growth engines despite the constraints of real-world validation.
References and Sources
- Witt S. The Thinking Machine: Jensen Huang, Nvidia, and the World’s Most Coveted Microchip. New York: Viking / Penguin Random House, 2025.
- NVIDIA Corporation. Corporate Timeline and Form 10-K Annual Reports (1999–2026). Santa Clara: NVIDIA Investor Relations, 2026.
- Buck I, Foley T, Horn D, et al. Brook for GPUs: stream computing on graphics hardware. ACM Transactions on Graphics (SIGGRAPH), 2004, 23(3): 777–786.
- Lindholm E, Nickolls J, Oberman S, et al. NVIDIA Tesla: A unified graphics and computing architecture. IEEE Micro, 2008, 28(2): 39–55.
- Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS), 2012, 25: 1097–1105.
- Choquette J, Giroux O, Foley D. Volta: Performance and programmability. IEEE Micro, 2018, 38(2): 42–52.
- NVIDIA Corporation. NVIDIA Completes Acquisition of Mellanox Technologies, Creating Major Force Driving Next-Gen Data Centers. NVIDIA Press Release, April 27, 2020.
- Stanford Graduate School of Business. Jensen Huang: View From The Top & Organizational Design Discussions. Stanford GSB Lecture Series, 2024.
- NVIDIA Corporation. NVIDIA BioNeMo Framework for Digital Biology and Generative AI in Drug Discovery. NVIDIA Developer Technical Overview, 2024–2026.