Nvidia closed out the earnings season for big tech with a bang. Revenues grew 56% from a year ago, and net profit by 59%. Those are phenomenal results for a company of any size, let alone the world’s first $4 trillion business. Yet the shares are down since then. When expectations are high, even great results aren’t good enough to produce a positive surprise and drive prices higher. The flip side of that is more worrying: when expectations are high, so is risk.
Many of the technology giants are in a similar boat. Expectations are high, but competitive intensity and capital intensity are heating up. What were once monopolies are now fierce competitors. What were once capital-light businesses are now investing more capital than all the oil majors combined. As intensity rises, profitability usually falls.
Meta is an instructive case in capital intensity. Since the end of 2019, net income has grown by about 25% per annum, but so have capital investment and share-based compensation. Stripping out share-based compensation, Meta’s growth in free cash flow has been zero since 2019. The company is now depreciating AI servers over five years, yet those servers are stuffed with chips that lose their edge in a single year. When depreciation schedules become interesting, you’re talking about a capital-intensive business.
But competition is the scarier of the two intensities. Nvidia faces aspiring competition from AMD, Broadcom, and custom chip designs from its tech giant customers. The giants are in a money-throwing contest to win in offering AI services. OpenAI, Anthropic, xAI, and others are likewise in a money-throwing contest to win in building AI models. All while some of their competitors, including upstarts like DeepSeek, give their models away. Each company believes the battle will be winner-take-most, and each worries that the threat to their golden goose could be existential. Under the circumstances, AI competitors will find it very difficult to stop spending on chips.
This makes Taiwan Semiconductor Manufacturing Company (TSMC) a compelling way to participate in AI adoption. The company makes all of Nvidia’s leading-edge chips, as well as all the leading-edge chips for Nvidia’s competitors and Apple. No matter who wins in AI chip design, TSMC wins. No matter who wins in AI services, TSMC wins. No matter who wins in AI model building, TSMC wins. In our view, TSMC is the only big tech company that deserves to trade at a monopoly multiple. Yet it trades at less than 20 times forward earnings, while Nvidia trades at 40 times. The market undervalues TSMC because Taiwan is in the name, but what happens to Nvidia or Apple if something bad happens to TSMC? Risks to TSMC are also risks to its customers.
TSMC is not alone among chip manufacturers. AI chips run zillions of similar calculations all at once, and to do that well, the processor needs vast amounts of data available instantly. Think of trying to hold thousands of phone numbers in your head at once. As a result, AI chips are much more memory intensive than conventional chips. A single Nvidia Blackwell processor has more gigabytes of short-term memory on the chip than most iPhones have in total storage.
That on-chip memory is also more specialised than traditional memory. To simplify a little, imagine some playing cards (the memory) and a book (the processor) on a table. They used to sit side-by-side, communicating only through the edges. The cards are now stacked 8- or 12-high on top of the book, allowing more surface area for faster communication. For memory makers, this high-bandwidth memory (HBM) is more profitable and less commoditised than their traditional products.
That bodes well for the HBM leader, Korean memory maker SK Hynix, which is seeing rapid growth yet trades at less than 10 times earnings. A fifth of SK Hynix is owned by SK Square, and that stake accounts for most of Square’s value. Though holding company discounts are not always quick to close, Square effectively lets us access Hynix at a lower price. American manufacturer Micron Technology is just behind Hynix in HBM, and is seeing similar fundamental improvement. But the real beauty for the memory makers can be seen in the laggard, Samsung Electronics. Historically the leader in memory, Samsung has fallen behind in selling HBM chips to Nvidia. That hardly spells doom for its fundamentals, however, because HBM takes manufacturing capacity away from traditional memory. Less capacity leads to a tighter supply-demand balance and better pricing for those chips. Spot prices for 16GB of memory are up by more than 50% over the past two years.
Memory remains a capital-intensive business, and competitive intensity in the industry used to be brutal. Yet the difficulty of keeping up with new manufacturing technology has winnowed the field from a dozen makers to three, allowing for a more rational competitive environment.
AI adoption will create growth opportunities for many companies. Investors have great expectations for today’s leaders in AI chip design, AI services, and AI models, even as those companies see rising competitive and capital intensity. With the chip manufacturers, we believe we’ve found an appealing combination of lower expectations and lower competitive intensity. In a money-throwing contest, the winners are the catchers, not the throwers.
Eric Marais is an Investment Specialist at Orbis Investments, a sponsor of Firstlinks. This article contains general information at a point in time and not personal financial or investment advice. It should not be used as a guide to invest or trade and does not take into account the specific investment objectives or financial situation of any particular person. The Orbis Funds may take a different view depending on facts and circumstances.
For more articles and papers from Orbis, please click here.