An AI world-beater with further to run

But when it comes to equities, what goes up doesn’t always come down. Especially when the company in question is enabling a technology that could disrupt the world in the way steam power, electricity and the internet did before it.  

On 64 times trailing profits, the stock’s valuation appears extreme. But factor in Nvidia’s current momentum, and things look much more reasonable. In the three months to October, revenue rose 206 per cent year on year to $18.1bn. As a capital-light semiconductor designer that outsources manufacturing, Nvidia’s operating margin climbed to a massive 57 per cent, enabling operating profit to rise 1,259 per cent to $10.4bn year on year.

Such rapid growth has some fund managers making a value-based argument for the shares. “Nvidia’s share price growth has been driven by its earnings growth rather than valuation expansion,” argues Tony Wang, T Rowe Price’s co-portfolio manager of its US technology strategy. “At 25 times forward earnings it still looks pretty affordable.”

PEG ratio discount 

One metric that is used to evaluate growth stocks is the price/earnings growth (PEG) ratio. This metric takes the PE ratio and divides it by the growth rate over a certain period. Like the PE multiple, it can be historic or forward-looking. FactSet currently has Nvidia shares on a one-year forward PEG ratio of 0.6, down from 2.4 in May. For comparison, the S&P 500 is trading on 1.6, Apple (US:AAPL) is on 3.4 and rival semiconductor designer AMD (US:AMD) is on 1.23.

The justification for Apple’s valuation is that its strong brand and software ecosystem have created economic moats. It’s hard for customers to switch away from Apple because their data and photos are saved in its cloud. That familiarity and comfort creates high switching costs, meaning customers are happy to cough up for newer, pricier iPhone iterations.

Nvidia has a similar competitive advantage. While most investors know Nvidia designs graphics processing units (GPUs), fewer appreciate it has developed the proprietary software on which AI researchers across the world are trained on.

In 2007, Nvidia launched its programming platform, CUDA, which allows developers to utilise its GPUs for general purpose computing beyond just graphics design. The platform has libraries of computation templates, saving developers time and effort, and making CUDA the default platform in AI computing. Switching away from Nvidia’s GPUs would also mean retraining in a whole new program.

Market dominance drives growth

Nvidia was founded by its chief executive, Jensen Huang. His initial aim for the GPU was to improve computer graphics. By adding many smaller processors than the computer processing units (CPUs) that came before, his innovation would sharpen the rendering of 3D images. However, researchers eventually realised the GPUs could be repurposed for training AI programs. Noticing AI’s potential in the mid-2000s, Huang started to invest in CUDA and build the AI ecosystem around the GPU.

That bet has paid off. Today, every major cloud computer company is scrambling to build out their AI computing capabilities. All want Nvidia's GPUs, which reportedly make up 90 per cent of the enterprise market. Last quarter, Nvidia's AI data centre revenue grew 279 per cent year on year to $14.5bn, while its gaming revenue grew 81 per cent to $2.8bn.

Unsurprisingly, given the huge profits it is generating, others want in. Nvidia’s customers, Amazon (US:AMZN), Microsoft (US:MSFT) and Alphabet (US:GOOGL), have all developed their own parallel computing chips. Last month, AMD launched its MI300X chip, which it says is the “most advanced AI accelerator in the industry” and which should reach sales of $1bn by mid-2024, according to the company’s chief executive – and first cousin once removed of Huang – Lisa Su.

AMD’s claim for the MI300X isn’t exactly correct, however, given it was benchmarked against Nvidia’s H100 chip. Nvidia’s H200 chip, released in November, is up to twice as fast as its older chips on some metrics.

For AMD, the problem isn’t just confined to competing against Nvidia’s chip designs. It’s up against an entire business ecosystem. Nvidia’s 2020 acquisition of Mellanox Technologies, for example, gave it the cabling technology to connect multiple GPUs, which is essential to building supercomputers. Throw in CUDA, and Nvidia has a full-service offering.

To Melius Research analyst Ben Reitzes, this amounts to a “sustainable ecosystem edge with unmatched software capabilities” and a dominant and defendable share of the market for training large language models. As such, he believes AMD can only gain up to 10 per cent of the GPU market in the long run.

However, the concern for all players is the potential size of the addressable market. In recent weeks, the US government has made it clear it does not want its AI chipmakers to supply China. At a recent AI forum, US commerce secretary Gina Raimondo warned industry participants: “I am telling you, if you redesign a chip around a particular cutline that enables them to do AI, I am going to control it the very next day.”

She later said that the US would be happy for Nvidia to sell less powerful chips for commercial purposes. But restrictions are forcing Chinese companies such as Huawei to design chips in-house, potentially removing a valuable source of demand at the lower end of the market. Last year, China made up 21 per cent of Nvidia’s revenue; with the restrictions imminent, Nvidia has said it expects sales to China to “decline significantly” in the three months to 31 January.

At the same time, management is confident that growth elsewhere will more than offset the China headwinds. This is probably true, but for growth to maintain its current levels, the AI enterprise software needs to show clear productivity improvements in the workplace. If this happens – as now seems probable – the scramble for Nvidia's hardware will accelerate in 2024, and its 25 times forward earnings multiple could soon look like a bargain.