PDF Summary:The Thinking Machine, by Stephen Witt

Most people know Nvidia as the company behind expensive graphics cards for gamers, but you may not know that it also helped create the technological foundation for the AI revolution. Stephen Witt’s The Thinking Machine reveals how Nvidia CEO Jensen Huang’s contrarian bet on parallel computing made this happen. While competitors like Intel focused on making traditional processors faster, Huang spent over a decade investing in academic computing tools that seemed commercially worthless—but that positioned Nvidia perfectly for the moment when AI systems needed massive parallel processing power.

Witt, the journalist behind How Music Got Free, provides an inside look at Huang’s unconventional leadership methods and the technical breakthroughs that transformed gaming chips into the essential infrastructure powering modern AI. Along the way, Witt illustrates how technological revolutions actually happen: through patient work on ideas that others dismiss. In this guide, we’ll explore the principles behind Nvidia’s success, examine competitive threats that emerged after the book’s publication, and dig into the gaming systems that inadvertently enabled the AI revolution.

(continued)...

(Shortform note: AlexNet uses “convolution,” a mathematical operation that slides small pattern-detecting filters across an image—like moving a small window—to look for edges, shapes, or textures. Since this requires thousands of identical calculations performed simultaneously across the image, it perfectly matches what GPUs are designed to excel at. More computational power means having more filters running in parallel, which enables the neural network to detect increasingly complex features, from simple edges to complete objects like faces. This is why scaling up processing power makes AI more capable: More computing power literally means recognizing more sophisticated patterns.)

While the broader AI research community was initially slow to grasp the full implications of AlexNet, Witt explains that Huang immediately recognized it as a huge opportunity for Nvidia’s parallel computing architecture. He swiftly redirected the entire company toward “deep learning,” the type of neural network training demonstrated by AlexNet, and declared Nvidia an “AI company” almost overnight. This rapid, decisive pivot proved crucial to seizing the emerging opportunity that would transform both Nvidia and the technology industry.

From Cat Videos to Whale Songs

While a neural network that can identify cat videos might seem to have limited utility, the same method has yielded breakthroughs with real scientific value. Researchers have trained neural networks to identify humpback whale songs using 187,000 hours of underwater recordings collected over 14 years in the North Pacific. The neural network converts audio recordings into visual spectrograms, turning sound waves into images that show how frequency changes over time. Then, it uses the same pattern recognition techniques that AlexNet used to identify cats to spot the distinctive shapes of whale songs.

This approach has enabled discoveries like finding whales at remote locations far from any known breeding areas, which would have been impossible with manual analysis, and it lets scientists detect the “upcalls” of critically endangered North Atlantic right whales in real time, which helps ships to avoid fatal collisions. Tasks that once required teams of human analysts listening to audio for months can now be accomplished in hours, allowing scientists to discover previously-invisible patterns in whales’ behavior.

How Transformers Created the Language AI Revolution

The next major breakthrough came in 2017, when Google researchers developed transformer architecture. According to Witt, transformers are a type of neural network designed to process language by analyzing the relationships among all of the words in a text simultaneously, rather than processing them one at a time. Instead of reading a sentence from beginning to end like humans do, transformers can examine every word in relation to every other word at the same time. This parallel processing capability made transformers perfectly suited for GPU acceleration—meaning they could take full advantage of Nvidia’s chips’ ability to perform thousands of calculations simultaneously—creating another massive opportunity for Nvidia.

(Shortform note: While transformers represented a major breakthrough in AI language processing, they still face a significant “barrier of meaning”: They don’t understand language the way humans do. In Artificial Intelligence, Melanie Mitchell explains that transformers rely on statistical pattern matching. In other words, they process text without genuine comprehension of its meaning. They also lack intuitive knowledge about the world that we take for granted, which leaves them vulnerable to surprising errors. This suggests the alignment between transformers and Nvidia’s parallel processing capabilities, while commercially successful, may represent sophisticated pattern matching rather than the true intelligence that AI critics worry about.)

Transformers enabled the development of large language models (LLMs), AI systems that use transformer architecture to process and generate language after being trained on vast amounts of text. Models like OpenAI’s GPT series required unprecedented computational power to train. But researchers discovered that as these models grew larger and consumed more computational resources, they became dramatically more capable. Models meant to predict the next word in a sentence could translate languages, write code, or explain scientific concepts. Witt explains that this created a feedback loop: More computing power led to more capable AI, which justified greater investments in infrastructure, all of which benefited Nvidia.

Are LLMs’ Capabilities Real or Exaggerated?

The unexpected leaps forward that Witt describes are what computer scientists call “emergent abilities”—capabilities that appear spontaneously as AI models grow larger. However, Stanford researchers argue that many of these abilities are “mirages” created by how experts measure AI performance. The debate initially focused on tasks where LLMs seemed to develop skills they weren’t trained for, like solving math problems. When researchers switched from “all-or-nothing” metrics, like requiring a model to get every digit of a math problem right, to more gradual metrics, like giving partial credit for getting some digits correct, the sudden jumps in performance disappeared and revealed smooth, predictable improvements instead.

This suggests that what looked like dramatic leaps in ability may have been illusions: The models were getting better at these tasks, but the original evaluation methods made small improvements invisible until they crossed a threshold. If true, this means that the scaling law discovery that drove Nvidia’s success—the observation that bigger models dramatically outperformed smaller ones—was based partly on flawed measurements that exaggerated the differences between model sizes. But the debate continues, as some researchers maintain that even with better metrics, certain abilities still appear suddenly and unpredictably.

How AI Factories Became Essential Infrastructure

The computational demands of transformer-based AI models created a new challenge. Witt notes that traditional computing tasks could be handled by a single powerful machine. But training the most advanced LLMs required breaking the work into millions of pieces and coordinating calculations across thousands of chips in real time. Huang’s solution was what he envisioned as “AI factories,” specialized data centers designed specifically for training and running AI systems. These represented a new form of industrial infrastructure that would consume raw data and produce intelligence, in much the same way that traditional factories consume raw materials and produce goods.

From Factory Floor to AI Factory: A New Form of Alienation

Huang’s AI factory reveals a new manifestation of the observations Karl Marx made in his theory of worker alienation—the disconnect between people and the products they produce. Traditional factories separated workers from the results of their physical labor, but AI factories create a new form of separation between human creators and their intellectual output. LLMs are trained on enormous datasets of human writing, art, and other creative work. This process transforms human creativity into raw material for algorithmic production: Billions of people’s thoughts, experiences, and creative expressions are subsumed into datasets used to reduce human meaning-making to statistical relationships.

The process relies on multiple layers of abstraction and “alienation” that move progressively farther from human understanding. AI systems first convert text into lists of numbers that represent each word’s position in mathematical space. The system then identifies patterns in how those numbers relate to each other, rather than engaging with the actual meaning of the language itself. Therefore, when AI systems generate text or art, they’re doing advanced mathematics: They recombine numerical patterns to produce outputs that mimic human expression. The text or art an AI system creates is a mathematical echo of human thought patterns, but stripped of those thoughts’ original meaning and significance.

The economic implications of this infrastructure shift were unprecedented. Training the most advanced AI models required computational resources measured in thousands of GPU-years (work that would take a single chip thousands of years to complete), but could be accomplished in weeks or months when distributed across massive parallel systems. Witt notes this made the barrier to entry for developing state-of-the-art AI systems so high that only the largest technology companies could compete: Businesses like Microsoft, Google, and Amazon became locked in an arms race for computational capacity, spending billions on Nvidia hardware to build ever-larger AI factories and maintain their competitive advantage.

(Shortform note: While parallel computing allows individual GPUs to handle thousands of calculations simultaneously, the massive scaling of computing power Witt describes became possible only with distributed computing: coordinating thousands of GPUs to train a single AI model. Training an LLM requires processing billions of text examples and adjusting millions of internal settings, called parameters, that determine how the model responds: a task so massive it would take a single computer thousands of years but requires only weeks when distributed across thousands of GPUs. Such coordination creates major challenges, which Nvidia solved with specialized networking hardware and software to optimize how GPUs share information.)

How Huang’s Unconventional Leadership Enabled Nvidia’s Success

By the time AI breakthroughs like AlexNet and transformers created an explosive demand for parallel processing, Nvidia was uniquely positioned to capitalize on it, but not just because of its technology. According to Witt, the difference between Nvidia and its competitors often came down to leadership execution: how quickly it pivoted when opportunities emerged, how it maintained focus during years of losses, and how it scaled operations when the time came.

Huang’s ability to make these crucial decisions stems from his leadership approach, forged during Nvidia’s early struggles for survival. The company nearly went bankrupt in 1996 when their first product flopped, forcing Huang to lay off half the workforce and bet the company’s remaining funds on untested chips. These near-death experiences, Witt explains, shaped Huang’s demanding leadership philosophy, built around three core principles that enabled Nvidia to execute when the AI revolution created unprecedented opportunities.

(Shortform note: The competitive mindset Huang developed in Nvidia’s early struggles is reflected in the company’s aggressive naming conventions. “Nvidia” comes from “invidia,” the Latin word for envy, as the founders hoped to create chips so powerful they’d make competitors “green with envy.” In Rome, invidia was associated with the “evil eye,” a hostile gaze that could bewitch or harm—perhaps a fitting symbol for a graphics company geared to captivate users with visual technology. The theme extends throughout Nvidia’s lineup: Its high-end GPUs are “Titans,” a reference to pre-Olympian gods of Greek mythology, while its consumer graphics cards have names that project power, like the “GeForce” series. )

Principle 1: Huang Uses Public Confrontation to Drive Performance

First, Huang deliberately confronts underperforming employees’ in front of large groups rather than handling problems privately, turning individual mistakes into organization-wide learning experiences. According to Witt, in one notorious incident from 2008, Huang spent over an hour publicly berating the architect responsible for a flawed graphics chip, with more than a hundred executives watching in the company cafeteria. Witt notes that some employees describe Huang’s approach as “verbal abuse,” but Huang employs this method because he believes public accountability creates stronger motivation than private feedback.

(Shortform note: Huang’s use of public confrontation may reflect what Kara Swisher (Burn Book) characterizes as toxic masculinity in Silicon Valley leadership. Some trace the origin of this behavior back to the 1960s, when programming went from being a field full of women to one dominated by men, who brought a new combativeness to the work. Columnist Carolina Miranda describes this new pattern as “mean-boy aggression,” where tech executives treat workplaces like competitive arenas to display their dominance. Tech CEOs’ use of fear as a management tool has become increasingly normalized, but research suggests it undermines the psychological safety that teams need for innovation and risk-taking.)

Despite the intense public criticism he delivers, Huang combines his demanding standards with genuine care for his employees. Witt says that Huang remembers personal details about employees’ lives, provides support during family crises, and rarely fires people for performance issues—partly so employees won’t be afraid to take risks. This combination of demanding standards with personal loyalty creates what Huang’s employees describe as an overwhelming desire not to disappoint him, generating the intense commitment necessary for Nvidia’s ambitious technical goals.

(Shortform note: The effectiveness of Huang’s demanding leadership style may be reinforced by what experts call “golden handcuffs,” financial incentives that make leaving difficult despite a stressful work environment. Nvidia’s stock has gone up over 3,700% since 2019, creating thousands of new millionaires among its employees. In 2025, Huang said 78% of Nvidia’s 42,000 employees are millionaires, but their stock grants typically vest over four years, requiring workers to stay to earn their full compensation. This has created a dynamic where employees work grueling hours in what they describe as a “pressure cooker” environment, yet Nvidia’s turnover rate dropped from 5.3% to just 2.7% after its valuation exceeded $1 trillion.)

Principle 2: Huang Eliminates Information Barriers Through Flat Structure

Second, to avoid missing critical information, Huang maintains more than 60 direct reports, far exceeding the eight to 12 that business experts recommend, because he doesn’t want information to be filtered through layers of management before reaching him. According to Witt, Huang refuses to designate a second-in-command or create hierarchical structures that might slow decision-making or allow bureaucratic power centers to form within the company.

(Shortform note: While Huang’s flat structure eliminates traditional management layers, it differs fundamentally from the truly non-hierarchical organizations described by Frédéric Laloux in Reinventing Organizations. Laloux studied companies where power flows through fluid, interconnected networks rather than radiating from a single central figure like Huang. He explains that these organizations distribute decision-making power throughout the company, where anyone is able to make decisions after seeking appropriate advice from colleagues. They also replace traditional job titles with flexible roles that emerge organically from employees’ talents and company needs, enabling workers to take on responsibilities across different areas.)

To make Nvidia’s flat structure work, Huang requires every employee to send him a weekly email summarizing their five most important activities. With more than 30,000 employees, this generates thousands of emails each week that Huang can’t possibly read. But Witt explains that Huang samples broadly enough to remain aware of what’s happening throughout Nvidia, allowing him to spot problems and opportunities that might be missed in a traditional corporate hierarchy. This approach serves Huang’s goal of staying directly connected to technical and operational details across the company, so that when critical decisions need to be made, he already has the information and relationships needed to act quickly.

(Shortform note: Though Huang reads hundreds of emails a day, he could enhance his information gathering with sentiment analysis, a natural language processing technology that detects emotions in text by analyzing word choice and tone. This would support Huang’s goal of catching “weak signals” of new developments before they’re too far along. Researchers used sentiment analysis to retroactively analyze Enron’s executive emails and pinpoint the moment when communications shifted around the unethical practices that eventually destroyed the company, despite no formal complaints being raised. Similarly, sentiment analysis might achieve Huang’s goal of spotting problems that traditional hierarchies might obscure.)

Principle 3: Huang Cultivates Urgency to Enable Bold Decisions

Third, Huang cultivates what we’ll call a “thirty days to live” mentality throughout Nvidia, reminding employees that the company is always on the verge of failure. While this philosophy originated from Nvidia’s near-bankruptcy in the 1990s, Witt explains that Huang has preserved this mindset even during periods of success because it serves a specific purpose: It justifies taking risks that more comfortable organizations might avoid. When employees believe the alternative to aggressive action is certain death, they’re willing to make dramatic changes quickly.

According to Witt, this philosophy also gives Huang cover to resist short-term pressures from investors. When the company operates in permanent survival mode, it can justify investments that don’t pay off immediately as necessary for long-term survival, like the decade-long commitment to parallel computing. This approach enabled Nvidia to maintain strategic focus during years of losses, positioning the company to seize the moment when AI demand exploded.

(Shortform note: While Huang believes that a constant sense of existential threat improves decision-making, research suggests the opposite may be true. A scarcity mindset (the state of feeling you have insufficient resources) can impair cognitive function. When people experience scarcity and are in survival mode, they exhibit “cognitive tunneling,” where they focus intensely on the immediate threat while neglecting to consider other important information. This creates a mental burden that interferes with executive functions like planning and self-control. Huang’s approach may work for Nvidia because it isn’t actually facing scarcity, but organizations operating under genuine resource constraints might struggle with the strategic thinking Huang prizes.)

What Threatens Nvidia’s Dominance?

Nvidia’s technological advantages and Huang’s leadership have created what appears to be an unassailable market position, but Witt identifies several significant vulnerabilities that could threaten the company’s dominance. These challenges range from geopolitical risks and manufacturing dependencies to Huang’s refusal to engage with AI safety concerns, plus practical questions about energy consumption and corporate succession planning.

Nvidia’s Dangerous Dependence on Taiwan

The first vulnerability Witt identifies is Nvidia’s dependence on Taiwan Semiconductor Manufacturing Company (TSMC) to produce its most advanced chips. TSMC has technical expertise and manufacturing precision that would take competitors years or decades to replicate.

(Shortform note: TSMC was founded in 1987 through a Taiwanese government initiative at the perfect moment to capitalize on the global shift toward outsourced chip manufacturing. While Intel controlled 65% of advanced chip production at the time, Intel’s focus on designing and manufacturing its own chips left an opening for a company willing to manufacture chips designed by others. TSMC filled this “foundry” niche, developing manufacturing processes so precise that they now produce chips measured in nanometers. Experts say that for other countries to replicate TSMC’s capabilities, they’d require not just massive investment in manufacturing technology but ecosystems of suppliers and engineering talent.)

TSMC’s primary operations are in Taiwan, a self-governing island that China claims as its territory and has threatened to take by force. Witt explains that China’s increasingly aggressive stance toward Taiwan creates a direct threat to Nvidia’s business model and the broader AI industry that depends on these chips. Any military conflict or economic disruption in Taiwan could immediately halt production of the processors that power the global AI revolution.

Despite his Taiwanese heritage and cultural connections that helped build Nvidia’s relationship with TSMC, Huang publicly downplays these risks. But Witt suggests that this dependence represents one of the most significant long-term challenges facing both Nvidia and the AI ecosystem as a whole. He also notes that it has led to discussion of a “silicon shield”: the idea that the world’s reliance on Taiwanese semiconductor manufacturing might deter Chinese aggression by making the costs of conflict too high for China to bear.

Testing the Silicon Shield: Why Economic Deterrence May Not Work

The “silicon shield” theory assumes that rational calculation will prevent conflict, but some of China’s actions suggest that strategic and ideological goals may outweigh purely economic considerations. Since Witt’s book was completed, China has escalated military preparations despite the enormous economic risks—analysis shows a Taiwan conflict would shrink China’s GDP by nearly 9% and cost the global economy $10 trillion. Yet US officials report that Chinese military exercises around Taiwan serve as “rehearsals” for invasion, with China targeting 2027 as the year it could take Taiwan by force. More tellingly, China has multiple strategies that could bypass the silicon shield’s protection entirely.

Rather than invasion, China could use cyberattacks, economic blockades, or a “psychological war” convincing Taiwan to lose confidence in Western military protection while appealing to the cultural heritage Chinese and Taiwanese people share. Polling suggests this might work: 43% of Taiwanese view China as more dependable than America, compared to 49% who favor the US. China seems to believe that centuries of cultural commonality will override their political separation, with Taiwan ultimately choosing reunification over alliance with what Beijing portrays as a declining West. This suggests China may accept severe economic costs to restore what it sees as natural, cultural, and territorial unity.

Huang’s Refusal to Engage With AI Safety Concerns

Another critical vulnerability Witt identifies is Huang’s complete dismissal of concerns about the potential risks of artificial intelligence. While prominent researchers have expressed serious concerns about the potential for AI systems to become uncontrollable or even pose existential risks to humanity, Huang’s position is that AI systems are simply processing data and aren’t a threat to human welfare or survival. According to Witt, when he pressed Huang about the potential dangers of the AI systems that Nvidia’s chips enable, Huang refused to engage with the substance of these concerns. Instead, Huang became angry, yelling at Witt and calling his questions ridiculous.

(Shortform note: AI safety concerns fall into two categories, and experts disagree about which poses the greater threat. Those concerned with existential risk worry that AI systems might develop their own goals or pursue tasks too single-mindedly, as in a thought experiment where an AI designed to maximize paperclip production prioritizes that goal over human survival. Such concerns are taken seriously by researchers like Geoffrey Hinton, who won a Nobel Prize for his work on neural networks. But Gary Marcus and Ernest Davis argue in Rebooting AI that these concerns distract from a more credible and immediate danger: the likelihood that we’ll cede decision-making to AI in situations where its lack of understanding proves catastrophic.)

Witt reveals that other Nvidia executives show a similar lack of concern or are reluctant to contradict their CEO. One source told Witt that the executives seem more afraid of Huang yelling at them than they are of possibly contributing to human extinction. Witt presents this as a fundamental tension in the current AI landscape: While some of the most respected researchers in artificial intelligence are warning about potentially catastrophic risks, the CEO of the company enabling these developments refuses to seriously consider such concerns. This suggests that legitimate safety considerations may not receive adequate attention within the company that controls the infrastructure powering AI development.

(Shortform note: Nvidia executives’ reluctance to challenge Huang may reflect a deeper irony: We often mistake validation, from other humans or from AI, for sound reasoning. Mike Caulfield (Verified) explains that AI systems act as “justification machines,” trained to provide responses that validate our views and biases rather than challenge them. Huang’s position on AI safety might exemplify this same dynamic. Research shows that professionals across fields are more likely to trust information that confirms their beliefs, and higher expertise increases this bias. When Nvidia executives don’t openly challenge Huang’s dismissal of AI safety concerns, he may interpret this as rational consensus instead of the suppression of critical thinking.)

Sustainability Questions About Nvidia’s Future

Finally, Witt highlights two critical challenges that could limit the continued growth of AI systems powered by Nvidia’s technology: environmental impact and organizational continuity.

Energy Consumption

The first challenge is environmental. Data centers filled with thousands of GPUs consume massive amounts of electricity, contributing to growing concerns about AI’s environmental impact. Companies like Google and Microsoft have seen their carbon emissions increase dramatically due to their expanding AI infrastructure. Witt notes that early AI systems consumed power comparable to household appliances, but current systems require enough electricity to power entire neighborhoods. As AI models continue to grow in size and capability, their energy requirements are expanding exponentially, raising questions about whether current development trajectories are environmentally sound.

The Real Environmental Leverage Points in AI

The debate over AI’s environmental impact often presents an implicit choice between halting AI development altogether and accepting unlimited energy consumption to let it continue, but this framing obscures where the real leverage lies. While concerns about AI’s energy use are valid and urgent—data centers now rank as the 11th largest electricity consumer globally—the impact emerges primarily from corporate decisions about AI development and deployment rather than personal use. Research suggests that watching Netflix for an hour consumes roughly the same energy as 26 ChatGPT queries, illustrating that an individual’s AI usage has a relatively modest environmental impact compared to other digital activities.

The meaningful decisions happen at the corporate level: how frequently companies like OpenAI release new models, whether tech giants like Microsoft and Google invest in renewable energy sources for their data centers, and whether companies disclose their actual energy consumption and carbon footprints. This creates an interesting position for Nvidia: While the company benefits from increased AI development regardless of its environmental impact, its corporate customers might increasingly demand more energy-efficient chips if environmental costs—or social pressure—become too acute. This dynamic also creates pressure on Nvidia’s customers to fix inefficiencies in their own AI development processes.

As one analyst notes, many tasks don’t need AI models to be “better” at generating creative, polished outputs, but to be “right” in providing accurate answers, a distinction that matters for justifying energy consumption. When companies release new models every few weeks, they often invest vast amounts of energy for marginal gains that users barely notice. For example, for GPT-4, which dramatically outperformed GPT-3 on benchmarks, real-world differences were subtle. Plus, the current approach of retraining models, rather than updating them with new information, represents a massive inefficiency that might be solved through “model editing” techniques that could reduce energy consumption while keeping AI systems current.

Succession at Nvidia

The second challenge is organizational. Witt reveals that Nvidia has no clear succession plan for Jensen Huang. His eventual departure could create significant challenges for a company so closely identified with its leader’s vision. The flat organizational structure that has served Nvidia well under Huang’s leadership may become a liability in transition scenarios, since there is no clear second-in-command, and more than 60 people report directly to the CEO.

This concentration of decision-making authority in a single individual creates vulnerabilities for a company whose market value depends heavily on continued strategic vision. Given Nvidia’s central role in AI development, any disruption to its leadership or strategic direction could have far-reaching implications for the pace and direction of AI progress worldwide.

(Shortform note: Nvidia’s succession challenges spotlight a broader problem with Silicon Valley’s “great man” narratives: founder-focused stories that reinforce the idea that individual genius drives technological progress. This obscures the collective effort that made AI possible: decades of research by computer scientists, an open-source software movement that enabled experimentation, the engineering that built the foundations of neural networks, and social movements demanding more capable technology, which all created the conditions for AI breakthroughs. It also suggests we wouldn’t have modern AI without Huang, when had Nvidia not existed, the same forces might have driven someone else to play a similar role.)