Jensen Huang: How Nvidia Survived Near-Bankruptcy, Built CUDA, Backed OpenAI and Became the Engine of the AI Boom

In 1997, Nvidia was running out of money. Its original graphics technology had taken the wrong direction, Microsoft had established an incompatible industry standard, and approximately 30 competitors were fighting for the emerging 3D graphics market.

Jensen Huang had about six months to save the company.

Instead of following the conventional process of manufacturing a chip, testing it and correcting its problems, Nvidia developed a way to test its next processor before it existed physically. It then committed its remaining resources directly to production.

The gamble worked. But Huang does not describe it as blind risk-taking. He believes the important decision was to move as much uncertainty as possible out of the future and resolve it before committing the company.

That approach would become central to Nvidia's development of CUDA, scientific computing, data-center infrastructure and artificial intelligence.

In October 2023, Huang discussed these decisions with Ben Gilbert and David Rosenthal on the Acquired podcast. At the time, Nvidia was valued at approximately $1.1 trillion and ranked among the world's six most valuable companies.

The conversation went beyond the technology. Huang explained why Nvidia pursued markets that barely existed, why he acquired networking company Mellanox, how he organizes employees without a conventional corporate hierarchy, and why he would not willingly repeat the experience of founding Nvidia.

Developments through October 2026 provide an opportunity to examine which parts of that strategy have translated into measurable commercial results, and which questions remain unresolved.

1. The 1997 Decision That Saved Nvidia

Nvidia's original graphics architecture was wrong

Huang founded Nvidia in 1993 with the ambition of transforming personal computing through hardware-accelerated graphics.

At the time, Windows depended heavily on software rendering. Nvidia believed specialized processors could take over computationally demanding graphics operations and make personal computers dramatically more powerful.

The company initially pursued an approach based on curves, forward texture mapping and tessellation. Its NV1-era architecture differed substantially from the triangle-based rendering systems that eventually became standard.

Nvidia believed its approach could eliminate or reduce the need for conventional depth buffers, known as Z-buffers, which help computers determine which objects should appear in front of others.

The problem was not simply that Nvidia's engineers had made technical mistakes. The broader software industry began adopting a different architecture.

Microsoft introduced DirectX, giving developers a common way to use graphics hardware. Nvidia's existing approach was fundamentally incompatible with the direction Microsoft was establishing.

Nvidia had also pursued its own programming interface, DirectNV, hoping developers would build applications specifically for its processors.

But Microsoft already had extensive relationships with software developers and enormous influence over PC computing. Nvidia could not realistically force the industry to follow its proprietary approach.

By 1997, despite having entered consumer 3D graphics early, Nvidia faced approximately 30 competitors.

It was running out of money, time and confidence.

Huang regards this period as one of the most consequential in Nvidia's history. The company had made several serious mistakes, but the decisions taken under pressure established engineering principles it would continue using decades later.

RIVA 128 was designed to outperform everything

Nvidia abandoned its attempt to resist DirectX and decided to build the fastest possible processor for Microsoft's standard.

The resulting product was the RIVA 128.

Huang describes it as a fully hardware-accelerated 3D graphics pipeline, covering transformations, projection, rendering and the framebuffer.

Engineers pushed the hardware close to the practical limits of the period. They incorporated texture caching, fast memory, unusually extensive bus and framebuffer capabilities, and an exceptionally large graphics chip.

The design was deliberately expensive.

Rather than competing primarily on manufacturing cost, Nvidia selected a cost structure that Huang believed competing companies would be unwilling to match.

This strategy depended on the existence of customers prepared to pay substantially more for better performance.

Huang believed the developing PC market had exactly that kind of customer: the enthusiast who continually wanted the fastest computer available.

Graphics also presented a long-term opportunity because performance requirements would keep increasing. Better graphics would encourage developers to create richer games, which would create demand for still more powerful hardware.

Huang considered this a sustainable technology opportunity.

The RIVA 128 nevertheless contained a significant flaw. It supported only eight of DirectX's 32 blending modes, which control how visual elements combine on screen.

Huang acknowledged that Nvidia should have implemented all 32. He discovered the omission too late and had to proceed with the processor already developed.

Nvidia tested the chip before manufacturing it

The conventional semiconductor development process involved designing a chip, manufacturing a prototype, testing it, correcting defects and manufacturing an improved version.

Nvidia could not afford those cycles.

Huang approached Ikos, an emulation technology company he recalled was shutting down because it had no customers. Nvidia purchased its equipment to create a virtual testing environment for the RIVA 128.

The emulator allowed engineers to run software against a simulated version of the processor before manufacturing physical silicon.

Dwight and Nvidia's software team developed the software stack, tested drivers, ran games and checked VGA applications.

The technology was extraordinarily slow. Huang remembered watching Windows graphics appear at speeds somewhere between approximately one minute and one hour per frame.

But slow simulation was preferable to an expensive manufacturing mistake.

By the time Nvidia committed to production, the company had already performed much of the testing traditionally carried out after fabrication.

Huang described the urgency in simple terms:

“we get one shot”

Nvidia proceeded directly from the final design to manufacturing, simultaneously committing resources to marketing and developer relationships.

The company had no financial room for another major redesign.

Why Huang rejects the idea of reckless gambling

The Acquired hosts asked whether founders should interpret Nvidia's experience as encouragement to bet their entire companies on ambitious ideas.

Huang rejected that interpretation.

His argument was that a consequential commitment should be preceded by unusually thorough preparation.

Nvidia did not know the RIVA 128 would work because Huang was optimistic. It had tested the architecture, software and applications before manufacturing.

The practical principle was to identify future risks and resolve as many as possible in advance.

That thinking remained relevant to Nvidia's later work.

Huang believes avoiding unnecessary design cycles improves engineering efficiency, production costs and competitiveness. A superior chip that reaches customers too late can lose much of its value.

As he put it:

“speed is performance”

The 1997 experience taught Nvidia to treat development speed as a technical advantage, not merely a scheduling concern.

John Carmack and Tim Sweeney helped create the market

Nvidia's engineering decisions were necessary, but Huang does not believe they alone explain the company's survival.

The market had to develop in Nvidia's favor.

John Carmack demonstrated extraordinary software-rendering performance with Doom. The game used conventional processors rather than relying on dedicated 3D graphics hardware.

Huang admired Carmack's achievement. The limitation was that obtaining such performance required an exceptional programmer.

Carmack's later move toward OpenGL and hardware acceleration with Quake helped validate Nvidia's belief in specialized graphics processors.

Tim Sweeney's Unreal provided another important consumer application.

Huang credited both developers with helping create demand for accelerated 3D graphics.

Without applications that made dedicated hardware valuable, even an excellent processor might have failed commercially.

This distinction matters because Huang does not believe business success can be explained by collecting a list of good decisions.

Competent founders can fail, and founders who make serious mistakes can succeed when technology, timing, customers and market conditions align.

Nvidia experienced both sides of that reality.

2. CUDA: The Investment That Prepared Nvidia for AI

A graphics processor became a general-purpose computer

Nvidia's original business involved accelerating graphics, but the underlying processors had another characteristic: they could perform enormous numbers of calculations in parallel.

Traditional CPUs are designed to handle varied instructions and sequential tasks efficiently. GPUs can process large quantities of similar calculations simultaneously.

As graphics processors became more programmable, Nvidia began exploring applications beyond rendering.

Its earlier Cg programming technology helped engineers investigate whether higher-level programming languages could express computations performed on GPUs.

Experiments included CT image reconstruction, image processing and other scientific workloads.

These activities produced early evidence that GPUs could serve as general-purpose computing devices.

Nvidia eventually developed CUDA, a parallel computing platform that gave programmers a practical way to use GPUs for non-graphics calculations.

According to Nvidia's technical history, the first CUDA release arrived in November 2006, with a C-based programming environment. CUDA Toolkit 1.0 followed in June 2007.

The investment was enormous relative to the market that initially existed.

During the interview, the hosts estimated that CUDA had consumed approximately 10,000 person-years of engineering work.

Huang did not provide an independently audited calculation for that figure, but the scale illustrates the concern: Nvidia had invested heavily in a computing platform long before machine learning became a major commercial market.

Scientists adopted CUDA before AI companies did

CUDA's early customers were not primarily businesses developing chatbots.

They were scientists.

Researchers used GPU acceleration for molecular dynamics, medical imaging, seismic processing, weather simulation, quantum chemistry and other computationally intensive problems.

CUDA made powerful parallel computing available to researchers who could not necessarily access traditional supercomputers.

Huang described this as democratizing supercomputing.

The research community also gave Nvidia something commercially valuable: direct relationships with people exploring problems that existing computers struggled to solve.

By the time modern deep learning became prominent, Nvidia already had the hardware, programming tools and scientific relationships needed to support it.

Its AI position was therefore not created entirely in response to ChatGPT or the commercial excitement surrounding large language models.

The foundations had been developed over many years.

AlexNet changed Huang's understanding of machine learning

The breakthrough came in computer vision.

In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton developed AlexNet, a deep neural network that achieved a major improvement in image classification.

Their published research reported a top-five error rate of 15.3% in the ImageNet 2012 competition, compared with 26.2% for the second-best entry.

The model used GPUs to accelerate training and contained approximately 60 million parameters.

Those figures are documented in the researchers' published paper, ImageNet Classification with Deep Convolutional Neural Networks.

For Huang, the important discovery was not simply that one algorithm outperformed others.

It appeared to achieve a dramatic advance in a field that researchers had worked on for decades.

Nvidia had already been studying computer vision, but many conventional algorithms were poorly suited to CUDA's parallel architecture.

AlexNet suggested that neural networks and GPUs could be a remarkably effective combination.

Huang began asking why the method worked and whether its effectiveness would improve with larger models, more data and greater computing power.

He interpreted neural networks as potentially universal function approximators: systems capable of learning complicated relationships from examples.

If that interpretation was correct, machine learning could extend far beyond recognizing images.

Prediction opened a much larger commercial market

Huang distinguished traditional scientific programming from machine learning.

Conventional approaches frequently begin by understanding causation. Engineers identify the principles governing a system and develop equations or instructions to simulate its behavior.

Machine learning can approach some problems differently.

A system may produce useful predictions without completely understanding the causes behind them.

Huang illustrated this with consumer behavior.

A retailer does not necessarily need to understand the full psychology behind someone's toothpaste preference. It may be commercially sufficient to predict which product that customer will buy.

The same principle applies to predicting whether someone will purchase mustard with a hot dog, which movie they will enjoy, or which social-media post they will engage with.

Accurate recommendations can produce enormous economic value.

Huang extended the argument to scientific applications such as weather forecasting, where data-driven models can complement calculations based on physical principles.

His broader conclusion was that machine learning might change how much of the world's software was developed.

Instead of programmers explicitly specifying every rule, increasingly capable systems could learn useful relationships from examples.

That would change the architecture of computing itself.

Processors optimized for conventional software would not necessarily be the most efficient machines for training and operating these models.

Nvidia backed scaling before its commercial value was obvious

Huang believed machine-learning systems could improve as their datasets and models became larger.

He saw encouraging evidence even in primitive research.

When he encountered Ian Goodfellow's early generative adversarial networks, their output included small, blurry images around 32 by 32 pixels.

The immediate results were hardly commercial products.

But Huang believed the underlying methods could improve substantially.

Nvidia consequently continued investing in systems capable of training larger networks.

Research papers became another source of feedback. Huang followed developments closely and recalled that significant advances initially appeared every few months.

By 2023, new papers were appearing almost daily.

The pace reinforced his view that deep learning had considerable room for improvement.

Language models extended the same idea to human knowledge

Large language models demonstrated that scaling could produce valuable capabilities in a different domain.

The Acquired hosts discussed substantial improvements as models crossed roughly 10 billion and then 100 billion parameters.

These were observations about particular generations of models, not universal thresholds at which intelligence suddenly appears.

Huang was especially impressed by self-supervised learning.

He admired BERT's use of masked-word prediction, which created training exercises from ordinary text without requiring humans to label every example.

His interpretation was that language models compress information contained in written material.

Human language includes facts, explanations, relationships and examples of reasoning. A sufficiently capable model could learn some of these patterns through training.

That made emerging reasoning abilities plausible to Huang, although he still considered their appearance remarkable.

This development strengthened the logic behind Nvidia's long investment in accelerated computing: larger models required increasingly powerful systems, and useful applications extended across many industries.

3. How Nvidia Became Connected to OpenAI

Research relationships preceded commercial demand

CUDA placed Nvidia in contact with leading researchers before modern AI became a mainstream technology business.

Huang described relationships with Yann LeCun, Andrew Ng, Geoffrey Hinton and Ilya Sutskever.

He attended conferences, followed publications and asked researchers how Nvidia could help advance their work.

These relationships were not simply marketing exercises.

They gave Nvidia feedback about which calculations were difficult, what hardware improvements mattered and which software tools researchers needed.

The company could then design systems in anticipation of future workloads.

This created a connection between academic research and Nvidia's commercial engineering decisions.

Huang knew OpenAI's early figures but was not a founder

OpenAI was publicly announced on December 11, 2015.

Its initial structure was nonprofit, with a stated mission of advancing digital intelligence for humanity rather than maximizing financial returns.

The organization's original announcement identified a research team including Ilya Sutskever, while supporters included Elon Musk and Peter Thiel.

Huang told Acquired that he was not involved in founding OpenAI.

He nevertheless knew several associated figures through his work with the technology and research communities.

Some people who participated in OpenAI's early development later joined Nvidia.

Huang understood that organizations pursuing increasingly large neural networks would need computing systems beyond those available to most researchers.

Nvidia was developing such a machine.

The first DGX-1 went to OpenAI in 2016

On April 5, 2016, Nvidia announced the DGX-1, an integrated supercomputer designed specifically for deep learning.

It combined GPUs, software and development tools in a system intended to simplify demanding AI research.

Nvidia claimed performance equivalent to hundreds of conventional servers for selected deep-learning workloads.

In August 2016, Huang personally delivered the first DGX-1 to OpenAI in San Francisco.

Contemporary reporting by Fortune documented the handover and Elon Musk's acknowledgment that the system had been donated.

The DGX-1 had been announced with a starting price of approximately $129,000, illustrating the substantial value of the equipment.

It was one of the earliest visible examples of Nvidia supplying integrated infrastructure to an organization pursuing advanced AI research.

The commercial importance of language models was far from established in 2016.

Huang acknowledged that early systems had limited practical usefulness. His confidence came from the underlying progress and the belief that successive improvements could eventually produce remarkable capabilities.

The OpenAI relationship also illustrated a broader element of Nvidia's strategy.

Providing infrastructure to researchers helped advance the field while giving Nvidia early insight into the requirements of future AI systems.

4. Nvidia's Advantage Was More Than Its Chips

Nvidia wanted to be a platform company from the beginning

Huang disputes the idea that Nvidia began merely as a graphics-chip manufacturer and became a platform company only after CUDA.

He explained that the original vision was already broader.

Nvidia wanted to build something resembling a game console inside the PC.

A console succeeds partly because developers can create applications for a stable hardware and software environment.

Nvidia hoped to establish the same relationship with PC developers.

Its early Unified Device Architecture was conceived around a family of accelerators rather than isolated graphics products.

Developer relations therefore became an organizational priority almost immediately.

One of Nvidia's first employees focused on working with software developers.

The company understood that hardware would become more valuable when other businesses created applications specifically for it.

DirectX defeated Nvidia's first platform approach

Nvidia initially hoped its DirectNV architecture would attract developers to its own programming interface.

Microsoft's DirectX changed that calculation.

Microsoft already controlled important parts of the PC software environment and maintained strong developer relationships.

Huang acknowledged that Microsoft made good decisions during the transition while Nvidia made serious mistakes.

He also rejected the suggestion that Nvidia had been an equal competitor to Microsoft in the 1990s.

Their resources and market positions were fundamentally different.

Nvidia eventually built its graphics business around standards including DirectX and OpenGL.

Those applications generated demand for GPUs and helped finance the hardware installed in customers' computers.

CUDA later extended that installed base into general-purpose computing.

In effect, graphics applications paid for the development and distribution of processors that could also run CUDA applications.

Compatibility became nonnegotiable

Huang believes computing platforms require dependable compatibility between hardware generations.

Developers cannot make long-term commitments if each new processor requires them to redesign their software.

He estimated that Nvidia had approximately 250 million to 300 million active CUDA-capable GPUs worldwide at the time of the interview.

That figure was his estimate, rather than a separately verified installed-base count.

The essential principle was that Nvidia's accelerator generations should maintain architectural compatibility.

Huang described the rule emphatically:

“everything else is negotiable”

Compatibility, in his account, was the exception.

The platform's value came not only from processor performance but also from the software, expertise and applications accumulated across many hardware generations.

Why ecosystems are more durable than isolated products

Huang believes companies entering new markets can develop structural advantages by helping entire ecosystems succeed.

Developers build applications. Customers adopt them. Partners create complementary products and services.

Over time, those investments reinforce one another.

The resulting network becomes difficult for competitors to reproduce quickly.

Huang does not particularly like describing this as building a moat.

A moat encourages a company to think about protecting its own position. He prefers creating networks in which multiple businesses benefit from the market's growth.

This does not eliminate competition. It changes what competitors must overcome.

A rival processor may match or exceed one technical specification but still face the challenge of supporting an established software environment and its developers.

The importance of CUDA to Nvidia's AI business is therefore greater than the performance of any individual GPU.

5. Why Nvidia Entered Data Centers Before the AI Boom

The original limitation was the desktop computer

Huang traced Nvidia's move into data centers to a problem he recognized roughly 15–17 years before the interview.

At the time, most Nvidia GPUs were installed in personal computers connected to monitors.

That created an obvious limit.

There were only so many desktop computers, and only so many users prepared to buy a dedicated graphics processor.

Huang began questioning whether computation and display needed to happen in the same location.

An Nvidia engineer demonstrated a method of capturing a framebuffer, encoding its output as video and streaming it to another device.

The concept allowed a remote computer to perform graphics calculations while a different device displayed the result.

This became an important foundation for Nvidia's cloud-gaming efforts.

GeForce NOW led to enterprise computing and supercomputers

Huang identified GeForce NOW as Nvidia's first data-center product.

Cloud gaming allowed users to access remotely processed graphics without needing all the computing power locally.

The company recognized that latency would present serious difficulties.

Network connections introduce delays, and physical distance creates unavoidable constraints imposed by the speed of light.

Huang discussed approximate delays in the range of 70–120 milliseconds, while acknowledging uncertainty about the precise figures.

His argument was that reducing unnecessary delays could make remotely delivered computing practical even when physical latency could not be eliminated.

Nvidia's next data-center opportunity involved remote graphics for enterprise customers.

Companies could install GPUs centrally and allow employees to access demanding applications remotely.

The third major step combined GPUs with CUDA for scientific supercomputing.

These stages developed Nvidia's capabilities in centralized computing before AI training became the principal source of demand.

The distinction is important.

Nvidia did not suddenly discover data centers because language models needed GPUs.

Cloud gaming, enterprise graphics and scientific computing had already introduced the company to the technical requirements of remote and centralized processing.

Centralizing computation expanded Nvidia's market

Separating computation from individual computers changed how Nvidia could sell its technology.

A GPU no longer had to be associated with one person's machine.

Multiple processors could be installed in centralized infrastructure, shared among customers and combined into systems far more powerful than ordinary PCs.

Organizations could expand computing capacity according to their workloads.

The market was no longer limited by the number of consumers purchasing graphics cards.

This transformation also exposed a new challenge.

A data-center computer is not defined solely by its processors.

It depends on networking, infrastructure, security and communication between machines.

That realization led Nvidia toward Mellanox.

6. The Mellanox Acquisition and the Economics of AI Infrastructure

Why networking became essential

Huang identified a fundamental difference between conventional cloud computing and large AI training workloads.

In ordinary hyperscale cloud computing, many users can share the resources of individual machines.

A large AI training job may work in the opposite direction.

One workload can require coordination across thousands or potentially far more processors.

Different machines must exchange data continuously as the model is trained.

If communication becomes a bottleneck, adding processors will not deliver the expected performance.

Huang described this as an inversion of traditional hyperscale computing.

Commodity Ethernet networks worked well for many established applications, including search systems and Hadoop workloads.

But tightly coordinated AI training could benefit from networking technologies designed for high-performance computing.

InfiniBand was especially relevant because it had been developed for systems requiring fast communication and low latency.

Mellanox supplied technology Nvidia lacked

Nvidia had already worked extensively with Mellanox in high-performance computing.

Huang knew its technology and respected its engineering organization, particularly the Israeli team.

He estimated Nvidia employed approximately 3,200 people in Israel at the time of the interview.

Nvidia announced its agreement to acquire Mellanox on March 11, 2019, for approximately $6.9 billion.

The transaction closed on April 27, 2020, at a reported value of approximately $7 billion, according to Nvidia's filing with the US Securities and Exchange Commission.

The acquisition combined Nvidia's processors and software with Mellanox's high-performance networking expertise.

For Huang, the objective was to become a more complete data-center computing company.

Instead of supplying only processors, Nvidia could provide important components of the systems connecting those processors.

He considers the acquisition one of his best strategic decisions.

The Acquired hosts reported hearing similarly favorable assessments from people interviewed during their research.

Later financial results show the value of networking

The importance of networking can be measured through Nvidia's later financial disclosures.

According to the company's fiscal 2026 annual report, data-center networking revenue reached approximately $31.4 billion for the year ended January 25, 2026.

That compared with approximately $13.0 billion in fiscal 2025.

The increase was approximately 142%.

Nvidia attributed the growth to the expansion of AI infrastructure, including NVLink computing fabrics, Ethernet and InfiniBand products.

This provides a measurable commercial result behind Huang's explanation.

Mellanox did not merely give Nvidia another product category. It helped establish capabilities in a part of data-center infrastructure that became increasingly valuable as AI systems grew larger.

Nevertheless, Nvidia's current networking business includes technologies and development undertaken after the acquisition. Its entire revenue cannot reasonably be credited to Mellanox alone.

The acquisition also illustrates strategic timing

Huang prefers positioning Nvidia near potential opportunities rather than trying to predict every market development precisely.

He compared the approach to standing near a tree before its fruit falls.

The company does not need perfect timing if it has already developed the capabilities required to capture an opportunity.

The Mellanox acquisition fits that principle.

It provided Nvidia with networking capabilities before the scale of the generative AI infrastructure boom became apparent.

But preparation has a cost.

Companies must invest in capabilities that may take years to become commercially valuable. They also need intermediate businesses capable of supporting those investments.

Huang regards the ability to earn revenue along the way as essential to sustaining long-term technology development.

7. Nvidia's Zero-Billion-Dollar Market Strategy

Entering markets that do not yet exist

Huang describes many of Nvidia's opportunities as zero-billion-dollar markets.

These are markets with little or no established commercial demand but substantial potential if the underlying technology becomes practical.

He explained the idea:

“there's no market yet but we believe there will be one”

The appeal is that established competitors often ignore such opportunities.

A company willing to invest early can develop the technical knowledge, products and relationships needed to participate when the market finally emerges.

Huang identified gaming, professional workstations, scientific computing and AI as examples of Nvidia entering computing categories before their full commercial potential became obvious.

His preferred development horizon is often approximately a decade.

This approach requires more than patience.

A company must have enough financial strength to sustain development, enough technical conviction to continue, and enough flexibility to adapt as the market takes shape.

Software-defined vehicles were an early example

Nvidia entered automotive computing because Huang believed cars would become increasingly dependent on software.

That would require much more powerful onboard computers.

Some automotive executives initially questioned the idea.

Huang remembered one chief technology officer warning that vehicles could not tolerate the equivalent of a Windows blue screen of death.

The objection concerned reliability.

Cars are safety-critical systems, and failures can have consequences far beyond those associated with an ordinary personal computer.

Huang did not believe those difficulties invalidated the broader technological transition.

Looking back approximately 15 years after Nvidia entered automotive computing, he believed the industry's increasing reliance on software supported the original thesis.

The distinction between a promising technology direction and a commercially dominant business remains important.

In fiscal 2026, Nvidia reported approximately $2.35 billion in automotive revenue, compared with $193.74 billion in data-center revenue.

Automotive had developed into a meaningful business, but it remained much smaller than Nvidia's AI infrastructure operations.

Omniverse represented another early-stage opportunity

Huang applied the same strategy to Omniverse, Nvidia's platform for simulation, digital twins and connected 3D environments.

At the time of the interview, he estimated that Omniverse had approximately 40 customers.

Amazon and BMW were mentioned during the discussion.

Despite the small customer base, Nvidia was investing heavily in the technology.

Huang believed simulation and digital representations of physical environments could eventually become important computing markets.

His commitment reflected the same reasoning behind CUDA and data centers: develop the necessary technology before the full commercial opportunity becomes clear.

He also claimed Nvidia's technology had become nearly universal in professional design workstations, presenting that market as an example of how an early technical position can mature into widespread adoption.

The claim about workstation penetration was Huang's characterization rather than an independently documented market-share figure in the interview.

Market creation depends on other businesses

Huang's strategy connects early investment with ecosystem development.

It is not enough to invent technology for a future market.

Developers must create useful applications, customers must adopt them, and complementary businesses must have reasons to participate.

A technology platform becomes more valuable when these relationships grow together.

This explains Nvidia's continuing emphasis on developer relations, compatibility and supporting researchers.

They are not separate from the company's financial strategy.

They help determine whether a technically promising market becomes commercially viable.

8. Nvidia's Management Structure: More Than 40 Direct Reports

Huang organizes the company like a computing system

At the time of the interview, the hosts described Huang as having more than 40 direct reports.

That is unusual for the chief executive of a large technology company.

Huang explained that Nvidia deliberately avoids a conventional military-style hierarchy based on layers of command.

Instead, he thinks about the organization as a computing stack.

A computing system contains architecture, chips, software and higher-level functions.

Different people and teams are responsible for different components.

The person best qualified to lead a function should manage it, regardless of where that individual might appear in a conventional corporate hierarchy.

Huang questions why businesses producing fundamentally different products often use remarkably similar organizational charts.

In his view, a company's structure should reflect the work required to build its products.

A semiconductor company, restaurant and software business do not operate through identical processes. Their organizations should not automatically be designed the same way.

“Mission is the boss”

Nvidia uses a principle Huang described as:

“mission is the boss”

Rather than relying exclusively on departments, employees are organized around specific objectives.

Those missions are concrete engineering and product outcomes.

Examples include building the Hopper processor, developing systems around Hopper, preparing CUDA for the architecture, improving cuDNN and delivering cloud products.

Teams may need expertise from multiple parts of the company.

Huang compares this collaboration to a neural network, with connections formed according to the requirements of a particular problem.

The structure is intended to encourage direct cooperation across organizational boundaries.

It also reflects Nvidia's emphasis on product architecture.

Information does not have to flow through management layers

Huang believes traditional organizational hierarchies often give senior managers an advantage because they control information.

They know about important decisions before the people reporting to them.

Nvidia deliberately distributes information more broadly.

Huang described a robotics meeting attended by recent college graduates, three vice presidents and two executive staff members.

They discussed a technical problem and reached decisions together.

Everyone present learned the result simultaneously.

The newest graduate did not need to wait for information to pass through several management layers.

The approach places greater responsibility on leaders.

Without privileged information as a source of authority, executives must demonstrate their value through knowledge, reasoning, judgment and helping colleagues succeed.

Huang acknowledged that this creates substantial pressure.

The company's product cadence reflects its operating philosophy

The hosts highlighted Nvidia's rapid product-development cycle, contrasting it with the difficulty of imagining some other major technology companies launching flagship products at the same pace.

Huang connected engineering efficiency with organizational design.

The right structure should make it easier for people to solve problems quickly and avoid unnecessary work.

He also studies other organizations without attempting to copy them.

Huang reads business books, observes competitors and follows developments in unrelated industries.

His principle is to learn from their experiences and then determine what those lessons mean for Nvidia's particular circumstances.

He believes imitation without adaptation misses the purpose of learning.

9. The Investors Who Helped Nvidia Survive

A recommendation from LSI Logic helped secure funding

Huang's relationships from his previous employment played an important role in Nvidia's formation.

Before founding Nvidia, he worked at AMD and then LSI Logic.

At LSI Logic, chief executive Wilfred Corrigan was impressed by Huang's work.

When Huang moved toward starting Nvidia, Corrigan contacted Don Valentine and recommended him as an investment prospect.

Valentine became one of Nvidia's important early investors through Sequoia Capital.

Sutter Hill Ventures also backed the company during its founding period.

These relationships provided capital when Nvidia had little operating history and its proposed market was still developing.

Don Valentine combined pressure with personal encouragement

Huang remembered Valentine as demanding, grumpy and personally caring.

He recalled Valentine jokingly warning that he would kill Huang if he lost the investment.

The remark reflected Valentine's blunt manner rather than a literal threat.

Their relationship extended beyond financing.

When newspapers such as the Mercury News published favorable articles about Nvidia, Valentine sometimes wrote short congratulatory notes directly on the clippings and mailed them to Huang.

Huang regarded these gestures as evidence of genuine concern for the company.

Some early relationships lasted decades

Tench Coxe and Mark Stevens continued serving on Nvidia's board long after the company's founding.

Huang associated their involvement with the enduring relationships established through Sutter Hill and Sequoia.

He also credited Jim Gaither, Harvey Jones and Bill Miller among the early supporters who remained committed to Nvidia.

Their contribution extended beyond financing.

Huang described advice, encouragement, experience and emotional support during periods of uncertainty.

He regarded Nvidia as an exceptional investment for its early venture backers, potentially among Sequoia's greatest successes.

But his emphasis in the interview was on continuity.

The relationships survived repeated technological setbacks and severe financial volatility.

10. Jensen Huang's Background and the Importance of Abstraction

AMD and LSI Logic shaped his technical thinking

Huang began his semiconductor career at AMD before moving to LSI Logic.

At LSI Logic, he encountered methods that changed how he thought about computer engineering.

Traditional chip design involved working closely with transistors, logic gates and individual hardware functions.

Higher-level design methods allowed engineers to describe increasingly complex systems through programming languages and abstractions.

Software tools could then translate those descriptions into hardware implementations.

Huang became fascinated with the approach at approximately 21 years old.

He recognized that abstraction could make engineers dramatically more productive.

Instead of manually reproducing every detail, they could describe what a system should accomplish and use tools to implement much of the underlying complexity.

Harvey Jones and the development of chip-design tools

Huang credited Harvey Jones with important ideas surrounding high-level chip design.

He connected LSI Logic's work to developments at Synopsys and the wider industry of electronic design automation.

These tools transformed how semiconductor engineers could design and verify increasingly complicated processors.

Huang also developed relationships with people including Chris, Curtis, Andy Bechtolsheim and Jon Rubinstein during his early technology career.

He remained attached to LSI Logic.

Huang said he did not naturally move between jobs and believed he might have remained there indefinitely if Chris and Curtis had not persuaded him to start Nvidia.

The same principle may eventually apply to biology

Huang sees high-level abstraction as a recurring pattern in technological development.

Complex systems become easier to work with when engineers can describe their intended behavior through increasingly powerful interfaces.

He connected this progression to software development, semiconductor design, machine learning and digital biology.

His longer-term ambition is for biological systems to become more programmable through higher-level representations.

Large language models could potentially contribute to that process.

This remains a technological ambition rather than an established capability comparable to modern electronic design automation.

Nevertheless, the connection explains Huang's interest in technologies that allow people to work with complex systems without manually controlling every underlying component.

11. AI Safety, Copyright and Human Supervision

Physical safety and information safety involve different risks

Huang believes AI safety is necessary for widespread adoption.

He distinguished between systems capable of affecting the physical world and those generating information.

Robotics and autonomous vehicles must address functional safety, active safety and the consequences of incorrect physical actions.

Information-generating systems raise different concerns.

Huang identified bias, misinformation and creators' rights as important issues.

These problems require different technical safeguards, but both categories involve determining when humans must supervise automated systems.

Nvidia pursued licensed creative content

Huang pointed to Nvidia's relationships with Getty Images and Shutterstock as examples of developing commercially appropriate generative AI products.

The agreements are documented independently.

In March 2023, Nvidia announced a collaboration with Getty Images to develop text-to-image and text-to-video models using licensed content.

Nvidia's announcement stated that Getty Images would provide royalties to artists on revenues generated from the models.

During the same month, Shutterstock announced a partnership using Nvidia Picasso to develop generative 3D models trained with Shutterstock assets.

The relationships supported Huang's argument that AI development could involve licensed creative material rather than relying exclusively on unrestricted internet scraping.

They do not, by themselves, establish that every Nvidia AI model uses licensed training material or that all disputes over generative AI rights have been resolved.

Huang opposes uncontrolled self-modification

Huang expressed concern about increasingly autonomous AI systems retraining and changing themselves freely after deployment.

He favored a controlled process.

Developers should collect and curate information, train models, test their behavior and validate updates before releasing them.

Human supervision should remain part of the process, particularly as systems become capable of taking consequential actions.

Huang expects such supervision to remain important for a substantial period.

He compared AI safety to aviation.

Aircraft can use autopilot, but their operation also involves pilots, air-traffic control, redundancy and multiple safety mechanisms.

No single mechanism is expected to guarantee safety in every circumstance.

Huang believes similar principles can help govern autonomous vehicles, robotics and other advanced AI systems.

12. Will Artificial Intelligence Create or Destroy Jobs?

Huang expects productivity to support employment growth

Huang predicted that AI would probably create more jobs overall than it eliminates, although he acknowledged uncertainty about the timing.

His reasoning begins with productivity.

If a company can produce more with the same resources, its profitability may improve.

A more profitable business can invest in new products, pursue additional customers and enter new markets.

That expansion can generate demand for more employees.

Huang argued that productivity improvements would lead directly to workforce reductions only if companies had no worthwhile opportunities for expansion.

He identified drug discovery, transportation, retail, entertainment and technology as industries containing substantial unmet needs.

In his view, human ambition creates a continuing demand for productive work.

The prediction does not protect individual workers

Huang acknowledged an important qualification.

Even if AI creates more employment across an economy, particular workers can still lose their jobs.

Companies may change which skills they require.

Some tasks may disappear while new activities emerge.

Huang believed workers were particularly vulnerable to competition from people who learned to use AI more effectively.

His advice was straightforward:

“I'd encourage everybody just to learn AI”

He similarly encouraged companies to adopt AI to increase productivity, profitability and their ability to expand.

He also expected AI to help industries facing labor shortages.

Automation could allow organizations to increase production even when they struggled to recruit enough employees.

Independent research supports transformation, not a guaranteed jobs boom

Subsequent research provides a more cautious framework than Huang's overall employment prediction.

A May 2025 study by the International Labour Organization and Poland's National Research Institute found that approximately one in four jobs worldwide had some degree of exposure to generative AI.

The ILO's findings suggested that transformation of existing work was more likely than complete replacement for many occupations.

That does not establish whether total employment will ultimately rise or fall.

Exposure to automation is not the same as actual job loss.

Outcomes depend on how companies adopt technology, whether productivity gains create new demand and how quickly workers can transition between roles.

Huang's prediction remains an economic argument rather than a confirmed outcome.

Why cheaper computing can expand entire markets

The Acquired hosts connected Huang's argument with an idea associated with Sequoia investor Michael Moritz.

They called it the Moritz corollary to Moore's law.

The principle is that cheaper computing can make previously uneconomic products and services commercially viable.

The hosts recounted Moritz studying Sequoia's investment returns and wondering how future funds could outperform an earlier fund that benefited from Cisco.

The answer, in their account, was that declining computing costs would continue expanding the range of economically valuable technologies.

Huang agreed with the broader principle.

Greater productivity does not necessarily mean humanity will accomplish the same work with fewer resources and then stop.

It may encourage businesses and individuals to attempt things that previously seemed too expensive or difficult.

That possibility is central to his long-term economic case for AI.

13. The Human Cost of Building Nvidia

Huang's greatest fear is letting employees down

Asked what frightened him professionally, Huang identified the possibility of disappointing Nvidia's employees.

He explained that people join ambitious companies because they believe in their founders' aspirations.

They often make those ambitions part of their own lives.

Huang described his fear as:

“letting the employees down”

He feels responsible for employees' careers, families and opportunities to build successful lives.

He wants the people who contributed to Nvidia's development to benefit professionally and financially from its achievements.

That sense of responsibility helps explain why Nvidia's long periods of uncertainty were personally difficult for him.

Nvidia survived devastating market declines

The Acquired hosts estimated that Nvidia had experienced two or possibly three public-market drawdowns exceeding 80%.

Huang recalled a period when the company's market capitalization had fallen to approximately $2 billion to $3 billion.

He associated some of Nvidia's difficult financial years with its expensive investment in CUDA and related technologies.

The company was spending heavily on a future computing platform while commercial demand remained uncertain.

Huang described the embarrassment of watching Nvidia's valuation collapse.

Employees had questions. Investors questioned the strategy. Outsiders criticized decisions that might not produce commercial results for years.

He distinguished employees asking legitimate questions from employees abandoning the company.

Huang also recognized that many influences on Nvidia's share price were beyond his control.

Nevertheless, market declines made the consequences of long-term commitments painfully visible.

Sustaining those decisions required conviction that the company was pursuing an opportunity large enough to justify the investment.

He would not willingly found Nvidia again

One of the interview's most revealing exchanges came when the hosts asked Huang what company he would start if he were 30 years old again in 2023.

His answer was immediate.

“I wouldn't do it”

He explained that building Nvidia had been:

“a million times harder than I expected it to be”

Huang described the vulnerability, embarrassment, shame, suffering and repeated setbacks involved in creating the company.

He believed that if aspiring founders understood the full difficulty beforehand, many would never begin.

Entrepreneurial inexperience can therefore have an unexpected benefit.

Founders may attempt extraordinary things partly because they do not fully appreciate how difficult those things will become.

Huang continues using a version of that attitude himself.

He asks how hard a new challenge can really be, even after decades of experience have taught him that the answer is usually much harder than expected.

He applied this thinking to newer projects such as Omniverse.

His unwillingness to repeat the entire founding journey did not mean he wanted to retire.

Huang said he continued to enjoy his work and believed he was still contributing value.

Long-serving colleagues became his support system

Huang credited family, friends, colleagues and investors with helping him endure Nvidia's difficult periods.

He described employees who had remained with him for decades.

Chris, Jeff Fisher, Dwight and Joe R were among those associated with roughly 30 years at Nvidia. Jonah and Brian had also remained for approximately 25 years or longer.

These colleagues had witnessed failures, uncertainty and market collapses without abandoning the company.

Huang emphasized their loyalty.

He also valued returning home to a family that remained supportive regardless of Nvidia's immediate commercial performance.

The same principle applied to investors and board members who continued supporting the company through difficult periods.

For Huang, these relationships were not incidental to Nvidia's development.

They made the long-term commitments psychologically sustainable.

14. Huang's Personal Habits, Books and Priorities

He believes there is enough time for important work

Huang said his younger self would have disagreed with his present belief that there is enough time to accomplish what matters.

His reasoning is based on prioritization.

Leaders cannot do everything, but they can choose which activities deserve their attention.

That requires deliberately sacrificing less important commitments.

Huang warned against allowing Outlook or a calendar to control how someone spends the day.

He even referred to arriving late for the interview because he had become absorbed in another activity.

The incident illustrated his preference for focusing on important work rather than automatically treating every scheduled commitment as the highest priority.

Clayton Christensen and Andy Grove influenced him

Huang identified Clayton Christensen's business books among his favorites.

He appreciated their accessible explanations of technological disruption and the forces affecting established businesses.

He also praised Andy Grove's books for their practical management lessons.

Huang reads widely but believes executives should interpret ideas according to their circumstances.

He does not regard another company's successful management structure as a universal template.

Star Trek, a Toyota Supra and a Mercedes-Benz

Despite spending his career developing futuristic technology, Huang said he had never read a science-fiction book.

His favorite television series was Star Trek.

The program influenced the naming of conference rooms at Nvidia.

Huang also recalled his affection for a Toyota Supra.

He and Lori became engaged around Christmas and were driving home in his new Supra when the car was totaled in an accident.

He emphasized that the accident was neither his fault nor the car's.

At the time of the interview, Huang was being driven in a Mercedes-Benz EQS.

He cited security and practical reasons for no longer driving himself regularly and praised the vehicle.

These personal details provide a different perspective on a founder often discussed almost exclusively through Nvidia's technology and valuation.

15. What Huang Believes Founders Should Actually Plan

Elaborate financial forecasts are not the foundation

Huang admitted that he did not know how to write a conventional business plan when Nvidia was founded.

He later concluded that detailed early financial projections are frequently less reliable than founders assume.

An entirely new market contains too much uncertainty for precise forecasts to provide a dependable foundation.

Huang considers more fundamental questions important.

What need will emerge? Why will customers care? How can the company meet that need? And why will the solution be difficult for competitors to reproduce?

A business may identify an enormous opportunity and still fail if another company can copy its product easily.

Strategic insight is harder to learn than business mechanics

Huang does not dismiss pricing, positioning, sales or go-to-market execution.

Those activities are essential to building a functioning company.

But he regards many of them as learnable skills.

Identifying a major future market and building genuinely difficult-to-reproduce technology is a different challenge.

That requires technical understanding, judgment, willingness to experiment and sustained commitment.

It also requires acknowledging the role of circumstances beyond management's control.

Huang believes founders can make intelligent decisions and still fail.

Conversely, Nvidia's history demonstrates that serious early mistakes do not necessarily prevent eventual success.

The relationship between strategy and outcomes is not mechanical.

Preparation must be balanced with the ability to survive

Nvidia's decisions involved large commitments, but they were rarely made in isolation.

Graphics revenue supported early hardware development.

The GPU installed base supported CUDA.

Scientific applications supported Nvidia's investment in accelerated computing.

Cloud gaming and enterprise graphics helped establish its data-center capabilities.

Mellanox added networking expertise before large-scale AI demand became obvious.

Each stage created technologies and relationships useful for the next.

The process did not require Huang to predict every development correctly.

It required Nvidia to keep building relevant capabilities while preserving enough commercial strength to continue.

16. Nvidia's Financial Transformation After the 2023 Interview

Annual revenue increased approximately eightfold

The financial results following the interview demonstrate the scale of the AI infrastructure expansion.

Nvidia's fiscal 2023 results, covering the year ended January 29, 2023, reported annual revenue of approximately $27.0 billion.

Data-center revenue was approximately $15.0 billion.

By fiscal 2026, covering the year ended January 25, 2026, the figures were substantially larger.

Sources: Nvidia fiscal 2023 results and fiscal 2026 SEC filing.

Total annual revenue increased approximately eightfold.

Data-center revenue grew nearly thirteenfold.

The difference reveals how dramatically Nvidia's business shifted toward AI infrastructure.

In fiscal 2026, data centers generated approximately 90% of Nvidia's total revenue.

Gaming remained commercially significant, but it was no longer the dominant source of sales.

This was consistent with Huang's argument that the company's ultimate market was far larger than PC graphics hardware.

Growth continued into the 2026 reporting period

On August 26, 2026, Nvidia reported revenue of approximately $96.2 billion for its second quarter of fiscal 2027, covering the three months ended July 26, 2026.

That represented an increase of approximately 106% from the corresponding quarter a year earlier.

Data-center revenue reached approximately $89.0 billion, up 117% year over year.

These figures come from Nvidia's August 2026 financial announcement.

A single quarter's revenue had become more than three times Nvidia's entire annual revenue in fiscal 2023.

These results establish that AI infrastructure demand translated into extraordinary sales growth through the reporting period.

They do not establish that growth will continue indefinitely or that Nvidia will maintain its competitive position permanently.

The remaining question is whether Nvidia can sustain its advantage

The interview opened with two broad questions.

Would artificial intelligence become an enormous economic opportunity?

And could Nvidia maintain its dominance as the industry developed?

Subsequent financial results provide strong evidence that AI infrastructure became a very large commercial market.

The second question is more complicated.

Nvidia's advantage includes processors, CUDA, networking technology, software libraries, developer relationships and integrated computing systems.

But competitors are investing heavily in alternative hardware and software.

Large cloud customers also have incentives to reduce their dependence on individual suppliers.

Nvidia's own 2026 regulatory filings warn that competition, supply constraints, changes in customer demand and the availability of data-center capacity and energy could affect future results.

These are not hypothetical considerations unrelated to its business. They are risks the company identifies in its public disclosures.

Export restrictions have already affected Nvidia's opportunities

An additional complication comes from restrictions on advanced computing exports.

Nvidia's quarterly SEC filing for the period ended July 26, 2026 described the impact of US export controls on its ability to compete in China's data-center computing market.

The company said it was effectively prevented from competing in substantial portions of that market under the prevailing regulatory environment.

It also warned that losing access could help competitors build larger developer and customer ecosystems.

That risk is especially relevant to Huang's own platform strategy.

A computing platform gains strength as developers and customers build around it.

Restricted market access can therefore have consequences beyond immediate hardware sales. It may influence which alternative technologies gain adoption.

Nvidia's exceptional growth and its exposure to geopolitical restrictions can both be true.

The company's position in 2026 was financially much stronger than in 2023, but the long-term competitive landscape remained unsettled.

17. What Connects Nvidia's Decisions Across Three Decades

Nvidia's history contains several apparently unrelated developments.

The RIVA 128 was a graphics processor intended to save a struggling startup.

CUDA was a programming platform developed before the commercial AI market existed.

GeForce NOW attempted to move graphics processing away from individual computers.

Mellanox brought high-performance networking into Nvidia.

DGX connected processors and software into systems designed for advanced research.

Omniverse and autonomous driving reflected further attempts to develop emerging computing markets.

These decisions share several characteristics.

Nvidia repeatedly looked for computational problems that existing systems could not solve efficiently. It developed technology before demand became fully established, encouraged outside developers to build on its platforms, and used existing businesses to finance new capabilities.

The company also made mistakes.

Its original graphics architecture failed. Its first platform strategy could not withstand Microsoft's influence. Some long-term investments were commercially uncertain for years.

The 1997 crisis showed that Nvidia could survive a strategic failure by changing direction and preparing more thoroughly.

CUDA showed that software compatibility and developer relationships could create value beyond individual processors.

Mellanox demonstrated the importance of understanding the complete computing system rather than focusing exclusively on chips.

OpenAI's early use of DGX illustrated how supporting researchers could connect Nvidia to technologies years before they became major commercial products.

By 2026, the financial results had confirmed much of the commercial potential Huang described in 2023.

They had not resolved every question about competition, sustainability, AI employment or the future economic value of machine intelligence.

Huang's account also resists a simple story of inevitable success.

He credits developers who helped create early demand, colleagues who endured difficult years, investors who maintained their support and market conditions that arrived at critical moments.

Most revealingly, the man who built one of the world's most valuable technology companies said that, knowing the full cost of the journey, he would not willingly begin it again.

Nvidia's development was neither a sequence of perfect predictions nor a series of fortunate accidents.

It was a prolonged effort to build useful technology before its ultimate markets were fully understood, while surviving long enough for some of those markets to arrive.