ROB PERRONE: Winning the AI race: the battle between logic and memory

We don’t know who is going to win the artificial intelligence (AI) race, which is why we are betting on the components. Everyone is going to need semiconductors in the production of AI. 

Semiconductors are split into logic and memory. Logic chips, the brains of a computer, are specialised. Some are good for phones, some for servers, some for AI, and much of the value is in the design — Nvidia, which designs chips made by others, recently touched $1-trillion in market value.

Memory chips, the short-term and long-term memory of a computer, are far more commoditised. This makes the memory industry more cyclical than its logical cousin. Memory is cyclical for two reasons: the sector sells a commoditised product, and it has huge, fixed costs. 

Making chips requires enormous upfront investments — a single fabrication plant, or fab, can cost $15bn. With such huge, fixed costs, chipmakers try to run fabs flat-out, as prices need to fall to get below cash costs and make it worthwhile to halt production. The result is that memory makers sometimes flood the market with supply, crushing the price of chips. 

Those cycles used to be brutal. The industry would print money at the top only to bleed it at the bottom. A dozen contenders fought for market share, leading to frequent bouts of oversupply.

When the industry was stitching itself up after down cycles, governments from Japan to Korea to Taiwan prolonged the pain by subsidising home-grown players to enter the ring and increase production. 

But the industry has changed, thanks in part to another difference between memory and logic. In both cases, cutting-edge manufacturing processes produce chips that are faster, smaller and more energy efficient. For logic, leading-edge chips are essential for some uses, like phones or AI, but many applications work just fine with chips made on older, trailing-edge processes.

That is not the case with memory. Substantially all memory is produced on leading-edge processes, which produce the most cost-efficient chips. This makes memory an “up or out” industry. If you’re late to the next process, you start counting your months to bankruptcy and you must either catch up or close down. 

As the industry has pushed the boundaries of physics and balance sheets further, fewer competitors have been able to keep up. In the 1990s there were 14 big memory producers. By 2014 that number had dwindled to just three. Over the long term this consolidation should lead to structurally higher and more stable profits for the survivors — Samsung Electronics, Micron Technology and SK Hynix. Samsung and Micron are held in several of the Orbis funds. 

Though memory is now more consolidated, it still has cycles, and the current one rivals the global financial crisis in its severity. In the worst quarter of this cycle we estimate that memory makers will see year-on-year revenue declines of more than 50%, and late 2022 Micron’s valuation approached one times book value — near the lows of typical prior cycles. 

Yet we are already seeing the benefits of the industry consolidation. Micron and SK Hynix, the smaller two players, announced 15%-25% reductions in production, as well as 40%-50% reductions in capital investment, which drives future production. Samsung, the industry heavyweight, delayed announcing cuts, in part to force rational behaviour from its competitors. Yet even Samsung has announced plans to cut memory production and investment. Both Samsung and Hynix expect that inventories have peaked. These are signs that we believe indicate we are near the trough of the cycle, and tighter supply should lead to sharply higher memory prices and profits. 

As the industry emerges from its cycle it will greet several long-term tailwinds. Memory content in phones and computers continues to increase about 10%-15% per annum. Solid state drives using memory chips have become more cost-competitive with old-school hard disks, which should drive increased purchases among giant cloud services providers. And AI is memory intensive — an AI server needs five times as much memory as a conventional one.  

Rather than trying to guess whether Microsoft or Google will come up with the best AI tools, or whether Nvidia’s processors will win out against self-developed chips from the tech giants, we can invest in the picks and shovels enabling AI adoption. In logic, that is Taiwan Semiconductor Manufacturing, which is also held in the Orbis funds. In memory, that is Samsung and Micron. 

Memory is not without risk. While cycles are largely driven by supply, demand can be hit by economic slowdowns in the short term. Chips are at the centre of tensions between the US and China, which has banned Micron from selling chips to infrastructure operators. Korea could yet be caught in the middle.

The valuations of Samsung and Micron, while discounted versus our assessments of intrinsic value, have bounced from their lows of late 2022. But in a consolidated, global commodity market, competitive forces should drive profitability and valuations over the long term. 

Searching for fundamental value sometimes means buying decent businesses at great prices. In the memory makers, we believe we’ve got good prices for businesses that will become great. 

• Perrone is with Orbis, Allan Gray’s offshore partner.