AI chip startup boom

The computing power required to do machine learning is more rigorous than when people use computer chips to do other things. That demand has created a thriving market for chip startups for the first time in years and helped them double their funding.

Nvidia estimates that most machine learning or artificial intelligence tasks require 25 times more processing power every two years, while NLP requires 275 times more computing power every two years. In contrast, Moore's Law states that the number of transistors that can be accommodated on an integrated circuit will double in price and double in performance every 18 months. In other words, the computer performance that every dollar can buy will more than double every 18 months. This law reveals the speed of two information technology progress.

The massive increase in demand for computing power has helped Nvidia become the most valuable chip company in the U.S. because of its graphics chips and software stacks that can be used for machine learning. But the boom has also created opportunities for a new breed of chip companies focused on making specialized chips for artificial intelligence.

Nicholas Brathwaite, founding partner at Celesta Capital, said: "As these [AI] workloads start to grow and scale, it provides startups with the opportunity to use specialized semiconductor equipment that can better meet demand than general-purpose equipment." said in an interview.

Global sales of AI chips surged 60% to $35.9 billion last year compared to 2020, with about half of that coming from specialized AI chips in phones, according to PitchBook.

Through 2024, PitchBook expects the market to grow by a little over 20% annually, suggesting it could reach $64.9 billion by 2024. Allied Market Research predicts that figure could rise to $194.9 billion by 2030.

With the advent of artificial intelligence, semiconductor investment has begun to recover.

Before 2015, only a handful of funding rounds saw the opportunity presented by AI, and there was little overall interest in funding chip companies.

First, semiconductor manufacturing is expensive. Costs associated with new factories are measured in the tens of billions of dollars, and chip development costs for companies that outsource manufacturing to companies such as TSMC range from $30 million to $40 million, and chip design talent is scarce and costly.

Now, some of those conditions have changed. Chip development costs are still high and more start-up capital is needed to get off the ground, but that's no longer deterring investors.

Venture capital investment in semiconductor companies more than doubled from 2017 to $1.8 billion last year, according to PitchBook data. This year is expected to rise again, with funding close to $1 billion as of early April. The numbers include funding from chip companies such as Intel, Samsung and Qualcomm, which have their own venture capital arms that keep tabs on the latest technologies and advise on acquisition strategies.

"After years of R&D, many unicorns among data center startups have already offered trial chips to customers," said PitchBook analysts. "The high valuations of Nvidia and AMD show how much data center AI is growing, and greatly improve startups." valuation.”

Micron's investment arm said that Micron is very interested in these startups. "Micron's manufacturing teams also have pain points or gaps, and if those teams can meet the requirements, they can be our helpers."

AI applications often mean finding chips and software developed around parallel processing, and some startups can offer solutions that take on simpler tasks in greater capacity than traditional processors that do a single calculation at a time. In the case of Cerebras’ system, the company didn’t use state-of-the-art manufacturing techniques to make its silicon wafer-sized AI chips, but instead rethought how the chips communicated.

Drive like Moore's Law 

The chip giants themselves have poured billions of dollars into annual R&D spending, including research into parallel processing in GPUs at AMD, Intel and Nvidia. Investors want to fund companies that take a different approach, and that's what incumbent semiconductor companies want to try.

“The market is looking for companies that offer completely different use cases,” say companies using an older chip technology to make a special chip that performs AI processing at a fraction of the power needed to run a chip used in a data center .

"The way they did it was not just by optimizing or tweaking the circuit, but by using a whole new approach to computing," the investment firm said. "This new approach doesn't require some fancy new process or fancy physics concepts. , it reuses existing technologies that have cost tens of billions of dollars and takes them to do other things.”