SAN FRANCISCO – Ever since generative artificial intelligence began making headlines in 2022, investors poured money into Nvidia, convinced that the company’s leading position in AI hardware would bring them riches.
The gamble paid off in spades. I arrived at the company
Valued at US$5 trillion (S$6.5 trillion) in late October
In 2025, it is expected to report a net profit that exceeds the combined sales of its two main rivals. A flurry of multibillion-dollar data center investments in recent months suggests the AI gold rush is still accelerating.
But the huge rise in stock prices and questions about the nature of some AI deals have some industry insiders wondering if it’s all happening too quickly. There are concerns that when the dust settles, there won’t be enough profitable jobs to justify the huge spending on AI infrastructure.
Nvidia CEO Jensen Huang dismissed concerns that AI is a bubble that will eventually burst. He travels around the world to persuade politicians who remain skeptical of the technology to lift national security restrictions that prevent Nvidia from selling its most powerful and profitable chips in China, the world’s largest semiconductor market.
Currently, NVIDIA’s most profitable product is its Blackwell series of AI accelerators. The product is named after mathematician David Blackwell, the first African American to be inducted into the National Academy of Sciences. As with the previous Hopper series, Blackwell adapted the graphics processing units used in video games. It comes in a variety of formats, from individual “cards” to large computer arrays.
Hopper and Blackwell include technology that turns clusters of computers powered by Nvidia chips into a single unit that can process vast amounts of data and compute at high speeds. This makes it ideal for the power-intensive task of training the neural networks that power the latest generation of AI products. Nvidia has fine-tuned the service to help the AI platform better handle inferences that identify objects through common characteristics (for example, distinguishing between cats and dogs). As more people rely on AI to assist with an ever-increasing number of tasks, demand for this capability is rapidly increasing.
Nvidia offers Blackwell in a variety of options, including as part of its GB200 superchip, which combines two Blackwell GPUs and a Grace central processing unit (the heart of the computer that executes program instructions). (Grace CPU is named after Grace Hopper.)
Founded in 1993, Nvidia was the king of GPUs, the components that generate images displayed on computer screens. The most powerful of them are built with thousands of processing cores running multiple simultaneous computational threads. This makes it possible to produce complex 3D renderings such as shadows and reflections, which are characteristic of fast-paced video games.
Nvidia engineers realized in the early 2000s that these components could be retuned for other applications. Meanwhile, AI researchers have discovered that by using this type of chip, they can finally put their research into practice.
So-called generative AI platforms learn tasks such as translating text, summarizing reports, and compositing images by ingesting vast amounts of existing material. The more you absorb, the better your performance will be. They develop through trial and error, attempting billions of times to achieve proficiency, consuming vast amounts of computing power in the process.
According to Nvidia, Blackwell delivers 2.5 times the performance of Hopper in training AI. For customers rushing to train AI systems to perform new tasks, the performance advantage offered by Hopper and Blackwell chips is critical. The U.S. government has restricted sales of these components to strategic rival China, as they are seen as key to AI development.
Advanced Micro Devices and Intel are working to match the capabilities of Nvidia’s AI products. However, according to market research firm IDC, the company currently controls about 90% of the data center GPU market. The lack of a credible competitive edge is a concern for Nvidia’s cloud computing customers Amazon.com Inc., Alphabet Inc.’s Google and Microsoft Corp., which are seeking to develop chips for their cloud computing businesses in-house.
AMD, Nvidia’s closest rival in graphics chips, has signed a deal to supply ChatGPT maker OpenAI with a huge number of new AI accelerators. This agreement and the deal with Oracle suggest that AMD has gained some credibility as an alternative to Nvidia.
Nvidia has been updating its products, including the software that supports its hardware, at a pace that no other company can match. The company has also devised a cluster system that allows customers to buy chips in bulk and quickly deploy them. Mr. Huang continues to promote new products and partnerships by appearing at technology shows and corporate events around the world at a breakneck pace.
Nvidia is committed to introducing new flagship products annually for years to come, reflecting what Huang says is an unprecedented effort to drive innovation in the industry. Such a pledge serves as a warning that a rival is about to board a running train.
Hwang and his team have repeatedly said the company has more orders than it can handle, even for older models. In late October, Nvidia predicted revenue from its data center division would be about US$500 billion over the next five quarters. This has forced even the most optimistic analysts to raise their expectations. This increased Nvidia’s market value by US$400 billion in one week.
Microsoft, Amazon, Meta Platforms, and Google have announced plans to spend hundreds of billions of dollars on AI and the data centers that support it. OpenAI is continually purchasing computing power that will be deployed in the near future. Building better AI is moving forward at a rapid pace, even in the face of concerns that the technology’s underlying use cases may not yet justify such large investments.
The US and Chinese governments have done much more to help the company’s sales than NVIDIA’s competitors. Nvidia announced in April that it would record a $5.5 billion inventory writedown due to the U.S. ban on supplying H20 chips to Chinese companies. The H20 is a reduced-featured chip designed to overcome previous U.S. restrictions on sales to China. The US then gave Nvidia the green light to resume sales in the second half of the year, but Beijing retaliated by ordering Chinese companies to avoid Nvidia’s products.
Huang traveled to Washington to try to convince President Donald Trump that increasing business with China is good for U.S. national security. If U.S. companies don’t provide the building blocks for AI, he argues, other countries, especially China with Huawei, will step in and threaten U.S. technological leadership.
This perspective has received some attention among politicians in Washington, with the president name-checking the Nvidia product and talking about discussing it with his Chinese counterparts. But so far, no concrete agreement has been reached that would allow Nvidia to sell to China again. bloomberg
