As 2025 drew to a close, it became increasingly clear that something big was going to go down in the coming year. Few predicted that mere days into said new year, a president would have themself snatched in the night, protests would once again burst onto the streets of Iran, and the eyes of the world would anxiously await the fate of a frozen Arctic territory.

But one certainty that no one seemed to notice until it was far too late was that of the memory shortage. How did nobody spot this? Isn’t somebody paid exorbitantly to monitor and anticipate such events as these? Are we so blinded by shiny GPUs that we find ourselves sleepwalking into yet another semiconductor shortage?

Someone at OpenAI must have spotted something when they convinced Samsung Electronics and SK Hynix to commit to supplying advanced memory chips for their Stargate project.

The ChatGPT developer somehow struck a deal with the two companies that equates to 900,000 dynamic random access memory (DRAM) wafers per month. That’s roughly 40% of the world's total DRAM output. Did neither of the South Korean giants think to check OpenAI’s dubious (and rather well-documented) financial position?

The result? Scramble. Everyone and their dog raced to pick up memory chips before the year’s end, when prices would inevitably soar.

How was this ever allowed to happen? Or rather, how did nobody see this coming?

To paraphrase Dr. Ian Malcolm: Were we so preoccupied with the newest GPUs and the need for bigger AI models that we didn't stop to think if we had the RAM to keep the damn thing in its cage?

Systemic stack issues

Now, there are always folks who can sit on the side when everything turns bad and say, "I told you so."

But instead of trying to work out who our memory Nostradamus or Captain Hindsight is, let’s look back, as over the past few years, a patterning of hints here, there, and everywhere showed this was coming, but which no one really seemed to piece together.

First up, the "Moore's Law is Dead" talk. We’ve all heard it repeatedly throughout the past few years, that Intel co-founder Gordon Moore’s long-held empirical guideline for increasing processing power just isn’t feasible with the current-day generational leaps we’re seeing from Ampere to Ada Lovelace, Hopper to Blackwell.

And with that, there were a handful of industry voices warning of a RAMageddon, with each generational leap requiring more and more high-bandwidth memory for them to work as they were designed to do.

Sure, James Powell’s 2008 prediction that Moore's Law would be obsolete by 2036 appears quaint, given the sheer computing power we’re able to muster today. But surely firms were not working on this assumption as a means to not think about the wider stack.

The memory predicament we find ourselves in then is indicative of folks still thinking of AI as a software story, all the while ignoring the physical infrastructure needs like memory bandwidth and capacity until those constraints show up as higher prices or worse specs on everyday devices.

Case in point is the annual conversation between those responsible for a business' technology stack and their superiors. Every year, they struggle to explain that costs are increasing, and that just because the lights are on and blinking, it doesn't mean the technology is operating optimally, and that everything is fine. Meanwhile, those controlling the purse strings furrow their brow in a vain attempt to understand what they’re saying.

Trying to explain to a CFO or even a CEO that you need a ton more capital to build out your inference stack to support their precious AI chatbot is like trying to organize a stampede. All they want to hear is AI, AI, AI, Nvidia, Nvidia, Nvidia.

It’s super telling when a new generation of GPU comes out that every business under the sun crawls out from the depths to say, "Hey, we got ourselves Rubin, check it out." We get it, you got your hands on a shiny new chip. Now, do you have the storage capabilities in your stack to get the best out of that bad boy?

Imagine having secured yourselves some Blackwell, only to then tell your C-suite report that you need a ton of interconnects, cooling, and, in this instance, memory hardware, just to get it to perform a task that, in the time they had all this up and got it shipped, they’ve made marginal efficiency gains.

To bring this section of the ramble to a head, one of the reasons RAMageddon came about was likely because folks didn’t think about AI holistically from an infrastructure perspective.

OpenAI, for all its shenanigans, got in first. Jensen Huang’s infamous "chimaek summit," in which he, along with two of Asia's most powerful men, walked into a fried chicken restaurant, chugged beer, and ultimately secured Nvidia's supply of memory hardware, shows the game being played in this period of dearth was "you snooze, you lose."

The supply chain’s fault … again?

More broadly, then, the systemic failure to identify the importance of the underlying tech infrastructure is one thing. But the overall preciousness of the semiconductor supply chain further compounds that lack of solid thinking.

Now, you might say, "Did we learn nothing from the pandemic?" after the world found itself chipless and wanting. The answer to that question is, of course we bloody did. Every country and its dog moved to try to shore up their own supply chain. It is far from sustainable to have production concentrated in one corner of the globe. Or having one company in the Netherlands be solely responsible for the extreme ultraviolet photolithography (EUV) machines required to manufacture advanced chips.

But efforts like the U.S. CHIPS Act (which is still clinging onto life) and the EU Chips Act don’t magically solve the problem overnight. Fabs take years to build, let alone get up and running. The sheer precision required to get one of these up and running is astronomical.

Reshoring will help solve some of the problems. But with tensions high and the requisite materials just happening to find themselves behind the borders of frenemies, future disruption is inevitable.

The semiconductor industry is no stranger to peaks and troughs. In fact, it’s woven into the fabric of the market. From dearth to glut, scarcity to oversupply, it is by nature cyclical. But with the sheer volume of chips here, there, and everywhere, in everything from your car to your air fryer, demand has already strained supply. Meeting future demand with increased supply is only going to intensify the market's dynamics (also, why does your air fryer need voice assistance?).

To avoid this piece turning into yet another case study on the ridiculous precariousness of the semiconductor industry as a whole, on the memory front, the memory segment is arguably the most volatile.

For starters, geographical concentration is a major concern. Samsung and SK Hynix control roughly two-thirds of the global DRAM market. When you factor in third-placed Micron, and a near oligopoly concentrated largely in South Korea, alarm bells should start ringing.

Then there’s the fact that each wafer used to create high-bandwidth memory (HBM) lines, in fact, takes away from standard RAM and SSD storage, with the former using up to three-times more wafer capacity per unit than standard DRAM, and those alarm bells should be sounding a little louder.

And then comes the inventory crash, with those key manufacturers seeing stock levels that were previously at around 17-week inventory buffers as late as 2024 to just a couple of weeks in late 2025.

Ultimately, the failure of the industry to identify full-stack costs and considerations, and the fragility of the broader supply chain, come together and congeal into a mess of our own making.

You will pay the price for your lack of vision

But now, it’s already too late. Consumer DRAM prices are through the roof, which is why this writer can’t afford to upgrade his rig and is stuck with a GTX 1660 Super in the Year of Our Lord 2026. Heck, even Micron has packed up its bags in the consumer space to follow the money.

Fingers might point at Sam Altman and OpenAI’s dirty DRAM deal, or the time-honored tradition of the chip supply chain being held together with twine and scotch tape. But in essence, this was a failure of imagination, an industry so drunk on GPU benchmarks and having the latest shiny toy that it forgot you need memory to actually use the damn things.

Ultimately, the memory shortage we find ourselves in is what happens when an entire industry mistakes software hype for hardware strategy. It’s building cathedrals to process power, only to forget to install the plumbing.

The real winners then will be the memory makers themselves. Samsung Electronics is projecting operating profits for its memory division to soar by 422%. Micron’s stock has jumped approximately 230% compared to this time last year. And SK Hynix is looking at raising prices by as much as 70% compared to the final quarter of 2025.

But the losers in all this may well be the AI developers unlucky enough to have spotted this inevitability. The PC builders will just have to sadly suck it up and wait. But ultimately, the consumer will have to pay for AI’s greed.

We’re already seeing laptop makers offer baseline units with just 8GB of system RAM, going back to computer memory levels not seen since the early 2010s. This new normal of just accepting surging RAM costs and manufacturing shortages is unfair. But at least ​almost everyone got caught out, so we’re all in this together. And that’s something.