Memory, The Most Popular Commodity

When stuff costs more, that’s not necessarily good. It’s a cost. Memory prices are famously volatile, and the technology is generally considered a commodity. That means that the current revenue boost and profit margins should be considered temporary, as history has shown.

Rising prices are also not good for building infrastructure. Memory is to infrastructure what gasoline is to car travel. When the price goes up, it adds to the cost of everything in the technology supply chain, from your mobile device to the servers driving AI consumption. The inflation is feeding through to every piece of equipment in the system.

Memory pricing has been popping up in many technology earnings calls, as executives decide how much of the cost they can pass on by rising prices. In fact, it was a topic of discussion in the earnings calls from network infrastructure providers Fastly and Akamai’s calls last week. Akamai said it was passing on costs, and Fastly said it wasn’t. Fastly’s stock fell 30% and Akamai’s stock went up. Case closed.

This demand boom has been great for memory companies such as Samsung, SK Hynix, and Micron, which are estimated to control more than 90% of the DRAM market worldwide, but it’s not good for the rest of the world. It’s like celebrating the performance of Exxon just as you are about to go on a summer vacation road trip.

Creative Workarounds

Here’s the big question: How much of the increase in tech revenues, capex, and electronics spending is attributable to price inflation?

The stock market might be over-optimistic about the sustainability of memory pricing, but this also causes other problems. Volatile pricing causes difficulty in the supply chain, where some companies might start hoarding or double-ordering memory. That sets the stage for more volatility.

The memory shortage is driving companies to get creative every day. New software approaches and innovation are being deployed to emphasize techniques to optimize memory, such as offloading data to less expensive storage. For example, VAST AI recently launched a program to reclaim Flash memory. VAST’s architecture emphasizes the use of underutilized legacy SSD memory and porting it to areas of AI in demand for memory infrastructure.

Cloud companies are also actively seeking solutions. The rise of the KV cache, which is used to store data closer to AI inference demand, is a contributor to the need for memory. Google has developed a custom memory compression system called TurboQuant that targets the KV cache in language models, compressing data at the hardware level to free up processing and memory space.

But these are just small fixes in the scheme of things. Everybody is still buying very expensive memory.

The Cloud Capex Inflation Question

The memory price surge is contributing to the outsized cost of the capex spending by cloud companies, which also makes future paths of spending difficult to predict.

Capex spending keeps increasing because capex spending drives memory demand, which makes memory more expensive, which requires more capex.

In other words, as the spending is increasing, it’s fueling yet more inflation, which in turn adds to the cost. Economics 101. More dollars chasing fewer goods increases the price of everything.

Here’s why this adds to the risk: It’s very hard to measure what happens when the inflation subsides. If memory prices plunge by 50%, how much does capex slow down?

As in past bubble cycles, this could fuel a spiral that feeds on itself. Memory prices fall. Companies that hoarded memory have too much supply and unload it. Prices fall further. The stocks of memory companies crash. Capex numbers start to come down because pricing are falling. And so on.

Following the most recent earnings cycle, Raymond James analyst Simon Leopold wrote that total web-scale capex is up 80% year-over-year, driven by AI investment. But Leopold also pointed out that some of this is driven by inflation of the goods themselves.

Many of the tech investors are just buying inflation.

“Inflation, primarily from memory, drove the increases,” wrote Leopold. Leopold has written in his research that investors might be wary of the compounding effects when the inflation subsides, potentially in the second half of the year. He calls this a potential “air pocket.”

Compounding Risks Are Growing

At the same time that capital expenditures have risen sharply over the past three years, driven primarily by AI infrastructure buildouts, operating cash flow has also increased, but not as quickly as capex. This has been compressing free cash flow and raising capex intensity across the sector.

In the chart below, the capital intensity of the top five hyperscalers, which is the share of their cash flows allocated to capex, has reached nearly 100%. This means the industry is increasingly turning to debt, in addition to cash flow, to fund the buildout.

If memory pricing comes down, it will be good for the hyperscalers because they don’t have to spend as much. But it likely won’t be good for the stock market—especially memory stocks—which has been fueled by enthusiasm about capex spending.

Just as inflation and capex are working together to compound the rise in prices and stocks, the same system can work in reverse. If demand subsides, inflation subsides, and the inflation component of the entire capex supply chain comes down. We don’t yet know how much capex will fall when memory prices and inflation come down, as they are a significant part of the spending, but we do know that it will have a compounding effect.

In the long term, this isn’t something to celebrate, just like you don’t celebrate higher gas prices. This is creating much higher risks and unpredictability for the entire technology supply chain.

Futuriom provides paid research and marketing services to technology companies, with the goal of providing accurate insight into how cloud and AI infrastructure markets are evolving. These services include subscription research, custom research, and report sponsorships. In the past twelve months, Futuriom has not had a research relationship with any of the companies mentioned in this article, although individuals may subscribe to our premium services. The author holds no positions in individual technology stocks mentioned in this article.