Trillion-Dollar Chipmaker

At the beginning of the last decade, Nvidia was a niche company making computer chips for video games. Riding the AI wave, it’s now one of the most valuable companies in the world. OpenAI, whose technology set that wave in motion, was recently valued at $86 billion. But ChatGPT wouldn’t exist without Nvidia chips to train and run it, which is why, as of this writing, Nvidia’s market value is over 25 times greater than OpenAI’s.

It’s worth putting that into perspective.

Nvidia’s market capitalization—outstanding shares multiplied by price per share—is $2.28 trillion. That’s more than the combined market value of chipmakers TSMC, Broadcom, Samsung, AMD, Qualcomm, and Intel. The company’s share price has nearly doubled in the first three months of the year and is up some 450 percent since OpenAI launched ChatGPT in November 2022. At this point, only Microsoft and Apple are more valuable than Nvidia.

What’s so special about Nvidia’s chips? In short, they’re stellar for AI. 

Unlike central processing units (CPUs)—the type of chip Intel rode to fame—Nvidia’s graphics processing units (GPUs) are tailor-made to run many simple operations in parallel. As it turns out, this is an excellent match for both video game graphics and neural networks. Recognizing the opportunity years ago, Nvidia jumped in with both feet, engineering AI-specialized GPUs, or accelerators, and the software to make them go. Nvidia chips have led the pack ever since. The company owns over 80 percent of the AI chip market, forcing startups and big tech firms alike to do battle for limited supply.

That doesn’t look likely to change appreciably this year.

In a recent keynote at Nvidia’s GTC, founder and CEO Jensen Huang announced the company’s latest chip, Blackwell, has more than double the transistors of its predecessor. Performance gains make it four times faster for training, 30 times faster for inference, and 25 times more energy efficient. Huang told CNBC the chips will cost $30,000 to $40,000 apiece. (He clarified this price range is an estimate—the chips won’t be sold à la carte but rather bundled into larger data-center systems.) Still, if recent trends continue, there’s little doubt Nvidia will sell as many as it can make.

How long the hot streak continues depends on a number of factors.

Like any company, Nvidia’s market value is on paper only and at the mercy of investors. The company is in rarified air by share price, but its revenues and profits, though growing swiftly, are lower than similarly valued companies. That makes it a bet on the future—on generative AI’s staying power and Nvidia’s ability to maintain growth and market share. Not long ago, Tesla was a trillion-dollar company worth more than its biggest peers combined. Now the waters are choppier, and its valuation is lower.

Boom times also breed competition. AMD, Intel, and Samsung are in the AI game too. And although Google, Amazon, Meta, and Microsoft bought $13 billion worth of Nvidia H100 chips last year—a big chunk of the company’s $60 billion in annual revenue—all are developing inhouse chips as alternatives. Meanwhile, OpenAI’s Sam Altman wants to build a mega chip consortium to further boost supply. (It’s worth noting that, for now, Nvidia still has big tech in its corner: The Blackwell press release included testimonials from the CEOs of Alphabet, Amazon, Meta, Microsoft, OpenAI, and Google DeepMind.)

There are stranger offerings in the works too. Cerebras, in which Altman is also an investor, makes AI chips the size of dinner plates. The latest, announced this month, has four trillion transistors. A computer built with the chip can train models over 10 times bigger than GPT-4 and Gemini. Groq, a language processing unit (LPU), is focused on running—not training—the latest chatbots at lightning speed on the cheap. Getting an instant response from a chatbot is a “shocking” experience, Steven Levy wrote for Wired. Low-latency AI could make chatbots more conversational and have implications for robotics.

The argument for startups like Groq is that while last year was about training new AI models; this year is about serving them up to people fast and cheap. But Nvidia is aware of this shift: 40 percent of its data center business in the last year went to inference, and Huang made sure to call out just how much faster and more efficient Blackwell will be at running trained algorithms. Still, even as demand for inference grows, training next-generation models is ongoing. Nvidia’s chips are positioned to satisfy both sides of the coin.

All of which is to say: AI gets the glitz, but for now, there’s no AI without Nvidia.

More News From the Future

Humanoid robotics startup Figure raises over half a billion dollars.

A big number. Robotics startup Figure, founded by Brett Adcock in 2022, announced a $675 million Series B funding round valuing the 80-person company at $2.6 billion. The round included big names like Microsoft, OpenAI, Nvidia, and Amazon, whose billion-dollar investment arm is reportedly focusing heavily on AI in robotics this year. It’s a big number. By comparison, Agility Robotics raised $150 million for its Series B in 2022—seven years after its founding.

GenAI dividend. The deal suggests generative AI is rubbing off on robotics. Traditionally, robots are great at performing one narrowly defined task over and over again. Figure wants to build humanoid robots—that is, two-legged robots that resemble people—that can flexibly handle a range of tasks. The startup is partnering with OpenAI to accomplish some of this. In one impressive example, a video shows a Figure 01 robot having a conversation with a person. The robot is able to identify an apple as edible and hand it over, pick up trash while explaining its reasoning, identify that dishes go in the dish rack, and then put them there. It’s able to do this thanks to the language and image abilities of multimodal AI algorithms like GPT-4 mashed together with control systems that carry out actions physically.

Next wave. Figure is moving fast. In addition to raising ample cash and showing off progress with slick videos (another here), it also recently announced a Figure 01 pilot in BMW warehouses. But it has company. Other humanoid robotics startups and projects include 1X, Apptronik, Tesla’s Optimus, and Agility. The latter, which also took investment from Amazon in 2022, showed off a generative AI interface for its Digit robots last year and is piloting Digit in Amazon warehouses. Even Nvidia is getting in on the action. Still, it’s worth tempering expectations. The robots are slow—Figure 01 moves at 1/6 human speed—and the AI isn’t local, making latency and connectivity potentially problematic. Also, pilots aren’t the same as adoption, and it’s still not clear the humanoid form is always best. There is real progress though, so stay tuned.

A developer just crammed OpenAI’s GPT-2 onto an Excel spreadsheet.

When announced in 2019, OpenAI said GPT-2 was worryingly capable: "OpenAI Built a Text Generator So Good, It’s Considered Too Dangerous to Release," a TechCrunch headline read. Clearly, a lot’s changed since then. Ishan Anand, a software developer, has now laid out the smallest version of GPT-2 in a spreadsheet. The results beautifully illuminate the inner workings of an algorithm not so different from the one behind ChatGPT. "The actual algorithm for GPT-2 is mostly a lot of math operations, which is perfect for a spreadsheet," Anand told Ars Technica. And in a twist worthy of the times, Anand turned to ChatGPT for help when he got stuck on the project.

CRISPRed bacon is near.

Breeding company, Genus, used CRISPR gene editing to make a line of pigs immune to a devastating virus. The pigs, of which there are over 400 already, have a modified gene that prevents the virus from entering their cells. When exposed to high levels of virus in experiments, none of the pigs got sick. Genus is now seeking regulatory approval from the FDA and hopes to receive it by 2025. Other research into CRISPR-edited animals is underway, from fish to cows, and more may gain approval in coming years. 

Scientists slash cholesterol in mice with a gene-silencing injection.

Last year, Verve Therapeutics said that 10 people prone to high levels of cholesterol had received a gene editing infusion to knock out a mutated gene, PSK9, associated with overproduction of the artery-clogging lipid. But what if we didn’t have to directly change the genome to get similar results? A study in mice targeting PSK9 shows we might “silence” genes without editing them. Called “epigenetic editing” the approach builds on the body’s ability to control gene expression with chemical tags. The new study is an early indication epigenetic editing might work in animals and perhaps humans.

 Upcoming Events

EXECUTIVE PROGRAM

Join us for a LIVE Executive Program information session

In this interactive 30-minute webinar hosted by Singularity VP Alix Rübsaam, you'll get to know more about the 5-day program's content, speakers, and logistics. Please bring whatever questions you may have!

Register for April 18 at 11am ET

Ready to apply? Choose one of our 2024 EP dates:

April 7–11, 2024 (sorry, we're at capacity!)

October 6–10, 2024

November 10–14, 2024

THE FUTURE OF AI 

Full speaker list announced

From May 8–10, we're convening the sharpest minds in AI at the Computer History Museum in Mountain View, California. Learn more and secure your seat for this exclusive event.

Thank you for stepping into the future with us! Join our global community of over 200,000 futuremakers - sign-up and unlock early access to the Singularity Monthly newsletter and discover the most impactful technology breakthroughs - get a glimpse of tomorrow, today.

The Singularity Team