The RAM Shortage Is Worse Than You Think, and Nowhere Near Over

The world's memory supply is being vacuumed into a handful of data centers, and the bill is landing in your hands. The scale of the boom driving that is hard to overstate. Back in April of this year, the global market collectively bought $110.5 billion of semiconductors, the first time the industry has ever cleared $100 billion in thirty days, and an astonishing 93.9 percent jump over the same month a year earlier^[1]. The trade body that tracks these things now expects the full year to close above $1.51 trillion, a milestone it did not expect to reach for years^[2]. For the semiconductor manufacturing industry, it's an unprecedented period of profits. For the rest of us, it is just the start.

There seem to be two competing narratives being pushed through the media at the moment. One says the memory manufacturers are gouging and abandoning ordinary buyers to chase the AI gold rush. The other, unsurprisingly argued most directly by the people selling the memory, claims you've misread the situation entirely, that the shortage is simply a brutal limitation of physics. While the manufacturers are technically correct, their latest business decisions and forced lock-in contracts paint a clear picture of their intent to continue brazenly abusing those limitations. The shortage is real, it is by design, and it has no chance of letting up soon. As long as the companies involved are allowed to get away with it, they will continue posting the kind of margins normally reserved for software monopolies. Digging into the numbers as they are today is the only honest way to untangle what is happening.

It started inside the fabs

Three companies, Samsung, SK Hynix, and Micron, control more than 95 percent of the world's DRAM production^[3]. When demand exploded for High Bandwidth Memory, the stacked, interconnected DRAM that is necessary for AI accelerators to function at higher bandwidths, those three made a rational choice. HBM reportedly carries margins near 60 percent against roughly 40 percent for ordinary commodity DRAM^[4], so they aimed their best capacity at it. The problem with this harsh shift is that wafer fabrication is close to a zero sum game. Every line they convert to produce memory for AI accelerators is a line no longer making the memory that goes in laptops, phones, and routers. The vacuum has created a force on the market that is just begging to have its long-term ramifications become very clear.

The rush to fill that vacuum has been fiscally brutal, even to the largest names in the market. Conventional DRAM contract prices rose somewhere between 90 and 98 percent quarter over quarter in the first three months of 2026, then climbed another 58 to 63 percent in the second quarter on top of that already inflated base^[5]. NAND flash, the storage side, actually outpaced DRAM in the second quarter with increases of 70 to 75 percent^[5]. The strangest symptom is at the bottom of the stack. Supplies of older nodes tightened so hard that DDR4 spot prices climbed past newer DDR5 in some configurations, an inversion that has started to turn the normal march of progress on its head^[6], and demand has since reached back even further, slowly dragging up the price of legacy DDR2 and DDR3 parts that were supposed to be effectively obsolete^[7].

The shortage comes for everything with a chip in it

The Verge put the consumer side of things plainly: the price of RAM has tripled, quadrupled, even sextupled depending on the chip, and it is not just a problem for people who build their own PCs^[8]. The damage continues to show up everywhere a computer hides. IDC figures memory is now 15 to 20 percent of the materials cost of a midrange phone, and the same firm that once expected average phone prices to rise about $9 now expects as much as 8 percent, with steeper hikes on cheap phones where the cost gets passed straight to you^[8]. Qualcomm's chief executive told investors a coming dip in its phone business would be "100 percent" down to the memory shortage, and added, with unusual candor for an earnings call, "we just wish there was more memory"^[8]. Game consoles, handhelds, even the next generation of PlayStation and Xbox have been repriced or pushed back for the same reason^[8].

The most overlooked casualties are the cheap boxes nobody thinks about. Counterpoint Research found that memory prices for consumer gear like broadband routers, gateways, and set top boxes jumped more than 600 percent, nearly seven times, over a roughly nine month stretch into early 2026, against a threefold rise for smartphone memory in the same window^[9]. These are devices built on razor-thin margins, and those margins have rapidly disappeared. A year ago, memory was around 3 percent of the bill of materials for a low- to mid-range router. Today it is over 20 percent^[9]. That is the kind of shift that quietly stalls fiber rollouts and keeps a whole generation of affordable consumer hardware from being built at all.

Micron says you have the story wrong

Into the chaos walked Christopher Moore, Micron's vice president of marketing for its mobile and client business, who sat down with Wccftech in January and politely told everyone they had misunderstood the situation^[10]. His argument, while delivered with PR-flavored and lawyer-approved fluff, is worth examining for what it is. Micron, he said, is still serving consumers, just through different channels, supplying memory to the Dells and Asuses of the world rather than selling sticks directly^[10]. The real driver, in his telling, is not malice but scale. The share of the market that data centers demand "used to be 30, 35 percent, and then to 40 percent, and now to 50 and 60 percent," and the entire industry is simply short. "This is not a Micron issue," he said, "it's an industry issue"^[10].

His second point is the one most worth explaining, because the industry issue is one of an industry with very few players. Building modern memory is not a simple endeavor. Every time a fab switches the density of chip it makes, say from a 12 gigabyte chip to a 16 gigabyte chip, total output drops while the lines retool^[10]. So Micron is now asking customers to hold their configurations steady to allow running as few variations as possible, and keep output volume up. And the new capacity everyone keeps pointing to will not arrive on the schedule people imagine. Micron broke ground on its Idaho facility three years ago and pulled its start into mid-2027, but in Moore's own words you will not see "real output, meaningful output" until at least 2028, once qualification and customer acceptance are done^[10]. On the engineering, he is right. The shortage cannot be quickly resolved, and the nearest hope of meeting demand is genuinely years out.

The margins say otherwise

Here is where the numbers start to warp reality, the income statements. In the quarter it reported on June 24, Micron booked $41.5 billion in revenue, up 74 percent in three months, at an operating margin of 81.2 percent^[11]. Its own breakdown shows the gains were almost entirely price, not volume: DRAM bit shipments rose in the low single digits while average selling prices climbed in the low 60s^[11]. That is not a company simply struggling to meet demand. That is a company having the best year of its life, made painfully clear in May when its market value crossed $1 trillion for the first time^[12].

Micron is not alone in this market, and it is not even the biggest winner. Samsung, the volume leader in conventional DRAM, raised memory prices 100 percent year over year in the first quarter and then, unbothered, raised them another 30 percent for the second^[13]. Its semiconductor division alone turned 53.7 trillion won in operating profit for the quarter^[14]. SK Hynix, meanwhile, posted a 72 percent operating margin and a net cash position around 35 trillion won, having paid down its debt to a rounding error along the way^[15]. The three suppliers have each enjoyed the most profitable quarters they have ever seen because of this sudden overwhelming demand, and the continued eye-watering price hikes are unlikely to break their pace for the foreseeable future.

The cavalry is not coming, at least not on time

The obvious fix, you'd rightly think, is to build more fabs at home. This is already underway, and already too slow. The U.S. CHIPS and Science Act put roughly $280 billion behind the effort, including $52.7 billion in direct support for chip manufacturing^[16]. The tier one projects are enormous. TSMC has expanded its Arizona footprint to about $165 billion, and Micron is anchoring a $100 billion campus in New York as well as a $50 billion operation in Idaho^[17]. But these facilities take years to build, and the timelines don't look great. Meaningful memory output won't be on the table until 2028 at the earliest. Subsidies can derisk a fab, they cannot make concrete cure faster.

And subsidies do not work everywhere. A $1.9 billion chip plant proposed by EMP Shield in Coffey County, Kansas, pitched as 1,200 jobs and a slice of the supply chain for rural America, was denied CHIPS Act funding and has gone essentially nowhere since^[18]. It is a painful reminder that you cannot conjure a semiconductor manufacturing industry by simply trying to fund and manufacture the fabs. The money will go first to the places that already have the labor, the logistics, and the incumbents, which is to say the places that have decades of supporting infrastructure and logistics ready to support this type of facility.

The only real exit is rebuilding the machine

If brute force manufacturing cannot close the gap fast enough, the next strategy might be to need less memory in the first place. That work is happening in university cleanrooms rather than megafabs. Through the JUMP 2.0 program, run with the Semiconductor Research Corporation and DARPA, researchers are chasing architectures that merge memory and logic instead of endlessly shuttling data between them^[19]. At the University of Kansas, that means atomically thin memory resistors aimed at neuromorphic computing, hardware that mimics the brain's synapses to fold storage and processing into the same tiny components^[20]. Done right, it would cut the sheer volume of DRAM an AI workload needs. While improving efficiency is far from a promise of satiated demand, the efforts are unlikely to make things worse, at least as far as Jevons Paradox, the tendency for efficiency gains to just spur more total consumption, can be avoided.

While memory used to be a commodity, bought and sold on a volatile spot market that rose and fell with PC demand, that market is effectively dead^[5]. The big buyers now lock in binding, multiyear supply agreements just to guarantee they can keep their machines running, and the suppliers have every reason to keep it that way. While trillion-dollar companies run by billionaires will continue to get taxpayer-funded subsidies to help pay the ever increasing rates, the rest of us will just continue to foot the bill.