You went to replace your router — same tier as last time, nothing exotic, the kind of box you bought three years ago without thinking twice — and the price stopped you. Or your internet provider’s gateway “upgrade” quietly added a couple of dollars to the monthly line. The instinct is to blame the brand, or to assume someone decided to charge more simply because they could.
Branding and margins still matter. But in 2026, one invisible part inside the box is doing more work than most buyers realize. It’s a memory chip you will never see, made on production lines that have been quietly redirected toward an AI buildout you will never touch. That is the real reason routers are getting more expensive in 2026 — and it’s a stranger story than “RAM went up.”
This is the first in a three-part series on what the industry has started calling the memory squeeze. This piece explains why it’s happening. The next looks at why there are fewer genuinely new routers on the shelf this year, and the last one tackles the question you probably came here with: buy now, or wait?
The short version
Here is the whole thing in one breath, and everything below is just this sentence slowed down:
AI servers need a special, expensive kind of memory. To make it, the handful of companies that dominate memory production have redirected their factories toward it. That has drained the supply of the ordinary, older memory that routers actually use — so the cost of building a router has climbed, and you’re seeing the result at the checkout.
The counterintuitive part — the part worth staying for — is that the cheap, old memory inside your router is exactly the memory getting squeezed hardest.
What’s actually inside the box
A router, or an ISP gateway, is a small computer. It has a main processor and a modest amount of memory — often somewhere from hundreds of megabytes to a few gigabytes, depending on the class of the box. That is still nowhere near the memory footprint of a gaming PC or an AI server.
And the boxes don’t all carry the same load. A basic retail router may only need a small pool of DRAM to move traffic around. A modern Wi-Fi 7 gateway — the kind juggling multi-gig Ethernet, built-in security, telemetry, remote management, and sometimes voice or optical functions — needs noticeably more. But the important point isn’t the exact number of megabytes. It’s that nearly all of these boxes are built around memory families that were supposed to be mature, stable, and cheap.
Many mainstream routers and ISP gateways still depend on those mature DRAM families — names like DDR3, DDR4, and LPDDR4. These are the workhorse parts: well understood, more than fast enough for the job, and, for a decade, reliably cheap. The safe assumption was always that this kind of memory only ever got cheaper as it aged. That is the assumption 2026 has broken.
Here’s the paradox. The memory makers are racing to retire these older families so they can build the newer, higher-margin parts that AI servers want. As they pull back, the legacy memory that routers rely on isn’t just getting more expensive — in some cases it’s getting more expensive faster than the new stuff. One Counterpoint analyst has specifically flagged that LPDDR4 supply is shrinking faster than the newer standards meant to replace it. The old memory has stopped behaving like the cheap memory.
How AI drained the supply
The mechanism is simpler than it sounds. The memory that AI accelerators need is enormously demanding to manufacture — every gigabyte of it consumes far more factory wafer area than ordinary memory does. So every wafer a factory commits to AI memory is several wafers’ worth of ordinary memory it never made. Making the scarce thing makes the common thing scarcer.
And the makers have committed hard. The exact capacity math shifts every quarter, but the direction is unmistakable: suppliers are prioritizing server-related memory, especially the high-capacity parts demanded by AI infrastructure, and that leaves far less flexible supply for everyone else. The companies building data centers have signed long-term contracts that lock up that supply years in advance, pushing smaller buyers to the back of the line.
The price record tells the rest. In its June update, TrendForce reported that conventional DRAM contract prices rose roughly 93 to 98 percent in the first quarter of 2026, and expected another 58 to 63 percent increase in the second quarter. Those aren’t year-over-year figures — they’re quarter-over-quarter, stacked one on top of the other. As a spot-market signal of how strange things got, DDR5 contract chip prices roughly quadrupled in a single quarter late in 2025, and at one point DDR4 spot pricing even traded above DDR5 in some configurations: the moment the older, supposedly cheaper memory stopped being the cheaper memory at all.
There’s a quieter signal that matters even more for the people building hardware. Memory quotes, which a manufacturer used to be able to hold for weeks or months while planning a product, have collapsed to validity windows of roughly one to thirty days, with the final price often not set until the parts actually ship. When you can’t know what a component costs until it leaves the factory, you can’t confidently price a product you’re designing for next year. Hold that thought — it’s the hinge of this whole story.
Why routers get hit harder than phones or PCs
You’d expect phones and PCs to be the obvious casualties, and they are. But broadband boxes are getting hit worse, and there’s a clean reason why.
According to Counterpoint Research’s teardown analysis, over roughly nine months the memory cost inside smartphones climbed about threefold — while the “consumer memory” inside broadband products like routers, gateways, and set-top boxes climbed close to sevenfold. The effect on a router’s bill of materials is dramatic: memory has gone from contributing around 3 percent of a low-to-mid-range router’s build cost to more than 20 percent in about a year.
The reason routers absorb the worst of it comes down to leverage. The giants who buy memory by the shipload — the big phone makers — negotiate long-term contracts and sit near the front of the allocation line. A router manufacturer buys in far smaller volumes, which puts it at the back of that line precisely when supply is tight. The less negotiating power a maker has, the harder it gets squeezed.
And this is not just a consumer-router story. Cisco has already discussed price increases and contract-term changes tied to rising memory costs. The same pressure that shows up first in enterprise networking tends to work its way down the hardware stack — toward the gateway your provider hands you and the router on the retail shelf.
The part most shoppers never see
Here’s where the easy version of this story — “memory got expensive, so routers got expensive” — quietly falls short. Because the obvious move, if one memory chip suddenly costs five times more, would be to swap in a cheaper one. From the outside, that sounds like a purchasing problem. In a serious networking box, it quickly becomes an engineering problem.
A memory chip in a gateway is not just bought. It is qualified.
That word carries a lot of weight in this industry. The memory in a serious networking box isn’t a loose part you slot in; it’s tied to the specific processor it runs with, the physical board it sits on, the way the device boots, the heat it has to survive, the firmware that’s been validated against it, the production-line tests it has to pass, and the long-term supply guarantees behind it. Change the memory and you haven’t made a purchasing decision — you’ve triggered an engineering project, with re-testing and re-certification attached, and weeks or months of lead time before a single new unit ships.
So when memory prices spike, the chain doesn’t end at a higher price tag:
- Memory prices go up.
- The router’s bill of materials takes the hit.
- But the manufacturer can’t simply switch to a cheaper memory part.
- Because that part is bound to processor compatibility, board layout, boot stability, thermal behavior, firmware validation, production test, and long-term supply qualification.
- So the response isn’t only higher prices. It’s fewer model variants, slower refresh cycles, and longer reuse of designs that have already been proven.
That is why a memory shortage doesn’t only raise the price of a router. It changes the product map. Fewer variants. Fewer risky redesigns. Longer life for boxes that have already been qualified. That is the part most shoppers never see — and the part we’ll unpack next.
What it means for your wallet — and your options
How it reaches you depends on whether you buy your own gear or rent it from a provider.
In retail, you notice it directly: a higher shelf price, or a model that simply refuses to go on sale the way last year’s did. There are also fewer brand-new models to choose from, and the ones already on the shelf are sticking around longer.
On the carrier side, the pressure may show up less visibly. You rarely see a component price on your bill. Instead it surfaces in equipment economics, upgrade timing, installation offers, or how long an existing gateway stays in the field before a provider swaps it. A box that got meaningfully more expensive to build doesn’t disappear from the equation — it just changes how and when your provider replaces it.
For the people making these boxes, the squeeze comes down to three levers, and most will pull some combination of all three: raise the price, cut the specification, or both. Some version of that cost eventually has to land somewhere — in price, specification, availability, or refresh timing.
None of that means you’re powerless. Two things are worth doing before you spend anything. First, make sure you actually need a new box at all — a surprising share of “my router is bad” problems are really the wrong box for the job, or a coverage issue a new unit won’t fix. (Our guide on what each device in your setup actually does is a good place to start.) Second, on the question of when to buy — whether to move now or wait this out — that deserves its own answer, and it’s the subject of the third piece in this series.
How long does this last?
Honestly: longer than anyone would like. This isn’t a 2021-style shortage where a factory was offline and capacity came back. This is a deliberate, structural reallocation of capacity toward AI, and suppliers have limited incentive to reverse it quickly while AI and server demand keep absorbing the highest-margin capacity. Analysts and at least one major chip executive are pointing at 2027 and even 2028 before meaningful relief arrives, and a growing number have stopped calling it a shortage at all — they’re calling it the new normal. Plan as if elevated prices are the baseline, not a blip.
The bottom line
Other sites review the box. The point of this one is to explain the system that put the box on the shelf — and right now that system runs through a memory factory that decided your router wasn’t the priority.
Which leaves the cleanest way to remember all of this:
AI is not inside your router. But AI may still decide how expensive your next router becomes.
FAQ
Because the older, ordinary memory chips inside routers have become scarce and costly. Memory makers have redirected their factories toward the high-margin memory that AI data centers need, draining the supply of the mature memory that routers and ISP gateways actually use. That pushes up the cost of building a router, and you see it at the register.
No. AI servers use high-bandwidth memory and the newest standards. Routers mostly use mature families like DDR3, DDR4, and LPDDR4. But because making AI memory crowds out production of that older memory, the shortage still lands on routers — and in some cases hits the older memory even harder.
Not really. The earlier shortages were largely about factory capacity being knocked offline. This one is a deliberate reallocation: the memory makers are choosing to build AI memory instead of ordinary memory, because the margins are far better. That makes it more structural and slower to unwind.
Don’t count on it soon. Analysts and chip-industry leaders point to 2027–2028 before any meaningful relief, and some now treat elevated memory pricing as a lasting baseline rather than a passing cycle.
That’s the question the third part of this series is built to answer. The short version: it depends on whether you need one now and which tier you’re buying — but waiting indefinitely for prices to “return to normal” may not be the winning move it once was.
