AI Ate the RAM: Micron Says the Memory Shortage Gets Worse Through 2028. What It Means for Your Laptops, Phones, Servers and AI Plans (AI Trends, 2 October 2026)

What’s trending in AI on 2 October 2026: the AI boom has found a new casualty, and it sits inside every laptop, phone and server you own. On 30 September, Micron, one of the three companies that make most of the world’s computer memory, reported record results and told investors that memory will be even scarcer in 2027 and 2028 than it is today, with no end in sight. Within 48 hours the bill reached the shelves. Samsung added $100 to most Galaxy S26 phones in the US, Huawei said memory has added about $200 to the cost of every phone it makes, Raspberry Pi raised prices on its cheapest boards, and Nvidia lifted its 128GB DGX Spark AI workstation to $6,950 while launching a version with half the memory. This guide explains why AI is soaking up the world’s RAM, what each forecast says about how long it lasts, the security mistakes a squeezed hardware budget invites, and eight steps to protect your IT plans.

Key takeaways

AI is outbidding everyone for memory. Chipmakers are steering a limited supply of wafers toward high-bandwidth memory (HBM) and server memory for AI data centers, which pay far more than PC and phone makers.
It gets worse before it gets better. Micron expects demand to exceed supply in both 2027 and 2028, with conditions tighter than in 2026, and says its new factories won’t bring quick relief.
Prices have already exploded. A typical 64GB DDR5 desktop memory kit went from about $191 to $1,118 in a year.
Finished products are now going up. Galaxy S26 phones rose $100 overnight (the 1TB Ultra $200), 2GB Raspberry Pi boards rose $12.50, and Nvidia’s 128GB DGX Spark is about 74% dearer than a year ago.
Forecasts disagree on timing. Acer’s CEO expects PC prices to plateau in early 2027 and then fall; SK hynix says the crunch lasts until 2030.
For your business: buy what you must now, stretch healthy devices safely, right-size memory and AI workloads, and never let cost pressure push you onto unsupported or gray-market hardware.

AI data centers are buying up the world’s memory chips, and everyone else is paying for it.

1. Micron’s warning: tighter in 2027 and 2028

Micron’s fiscal 2026 results, released on 30 September, show how much money the AI buildout is pouring into memory. Revenue for the year reached $133.2 billion, more than three times the $37.4 billion of fiscal 2025, and fourth-quarter revenue alone was about $54.2 billion. For the current quarter Micron guided to roughly $61.5 billion in revenue at a gross margin of 86.25%, the kind of margin usually seen in software, not factories. In a year, the gross margin of its core data-center memory unit jumped from 41% to 90%, and that of its cloud memory unit, which sells HBM for AI chips, from 59% to 83%.

The forecast matters more than the profits. CEO Sanjay Mehrotra said Micron has already sold most of what it will make next year, that customers will pay much higher prices than in 2026, and that the industry will be more constrained in 2027 and 2028 than it is this year. He added that Micron has no “line of sight” to when supply and demand will balance again. Industry shipments of DRAM and NAND flash are expected to grow by the low-to-mid 20s percent in each of the next two years, which still won’t be enough. Micron plans about $25 billion of capital spending in the first half of its new fiscal year, but its new factories come online in 2028 and executives warned they won’t immediately ease availability or prices.

Micron’s revenue more than tripled in fiscal 2026, and it expects the shortage that drove it to tighten further.

2. Why AI is eating your RAM

Samsung, SK hynix and Micron make most of the world’s DRAM, with China’s CXMT a growing fourth supplier. Their output is limited by how many silicon wafers their factories can process, and that number grows slowly: a new fab takes years and tens of billions of dollars to build. Every AI accelerator from Nvidia, AMD or Google needs stacks of high-bandwidth memory, and HBM is built by stacking many layers of DRAM on top of each other. So each wafer turned into HBM yields far fewer of the ordinary memory chips that go into laptops, phones and standard servers. AI data centers also buy vast amounts of conventional server memory to sit alongside those accelerators.

When one class of customer will pay almost anything, factories serve it first. Hyperscalers have reportedly locked in almost all of next year’s DRAM production with advance deposits, leaving PC, phone and device makers to compete for what remains. It is the same pattern we described for electricity in AI’s energy crunch and the data center growth surge: the AI buildout is now big enough to set prices for resources everyone else depends on.

A fixed supply of wafers, split by who pays most. Ordinary devices get what’s left.

Retail prices show the result. PCPartPicker data compiled by Tom’s Hardware put the average 2x32GB DDR5-5600 kit at $191 in August 2025 and $1,118 in August 2026. Even older DDR4 memory rose roughly 120% to 180% as buyers retreated to older platforms, and in Europe average RAM prices were up 345% on a year earlier, with hard drives and SSDs up more than 125%. Industry trackers recorded conventional DRAM contract prices rising 90% to 95% in the first quarter of 2026 alone, one of the sharpest quarterly jumps on record.

3. The memory tax reaches the shelves

Phones. On 1 October, Samsung raised US prices for the Galaxy S26, S26+ and S26 Ultra by $100, and for the 1TB Ultra by $200 to $1,999.99, seven months after the phones went on sale. Korean prices rose too, while the S26 FE and the foldables were unchanged. The base S26 now starts at $999.99 for 256GB, $200 more than the base Galaxy S25 cost a year ago. There is an irony here: TNW reports that Samsung’s own semiconductor division generated 94% of the group’s profit while its phone business absorbed the higher memory costs. Huawei’s consumer chief Richard Yu said the same day that memory has added an average of about $200 to the cost of making each of its phones and that “measured” price rises will follow. Budget phones are hit hardest, because memory and storage are reported to make up more than 60% of a cheap phone’s parts cost, compared with over 30% for a premium one. Counterpoint Research estimates that around 230 million sub-$200 phones could disappear from the market by 2030.

Small computers. On 2 October, Raspberry Pi raised the 2GB Pi 4 to $67.50 and the 2GB Pi 5 to $77.50, $12.50 each. The 16GB Pi 5 already sells for about $305 at one US retailer, against a $120 launch price in January 2025. The company says the cheaper memory it stockpiled in 2025 is now largely used up, suggests customers consider older models or check how much RAM their projects really need, and has promised to reverse the increases once memory prices fall.

AI workstations. Nvidia raised its 128GB DGX Spark, a desktop box for running AI models locally, to $6,950, nearly 75% more than a year ago, and introduced a 64GB version expected to sell for about $4,999 through partners such as Dell, HP and Asus. Nvidia’s pitch for the smaller model is telling: it argues that mid-size models of around 26 to 35 billion parameters are now good enough for many private AI agents. Hardware makers are starting to design around memory scarcity rather than wait for it to end.

Finished products are rising by double digits. Raw memory has risen by multiples.

Business PCs and servers. The pressure on corporate hardware started earlier. Dell, Lenovo, HP and HPE began raising server and PC prices late in 2025, and Dell put through a list-price increase of about 17% across its PowerEdge servers, OptiPlex desktops, Latitude laptops and Precision workstations on 30 March 2026. Several vendors have also shortened how long quotes stay valid, and some have reserved the right to reprice an order between quote and shipment. Memory is now the biggest single cost in a typical server, and the crunch has even reached cars, with automakers warning of higher costs for driver-assistance hardware.

4. How long will it last? The forecasts disagree

The memory makers are united and gloomy. Besides Micron’s outlook, SK hynix CEO Kwak Noh-jung has called 2027 the worst year for memory supply in the industry’s history and expects demand to outrun production into 2030. ADATA’s chairman has suggested the DRAM shortage could last a decade.

Their customers are more skeptical. On 21 September, Acer chairman and CEO Jason Chen argued that the big suppliers have every reason to talk up the shortage to protect their margins. He said genuine shortages are now limited to high-end parts, and predicted PC prices will still rise 5% to 20% toward the end of 2026, plateau in the first half of 2027 and then start falling as cheaper memory from Chinese makers comes online. Acer, HP and Asus are already using some CXMT chips. There are also signs that buyers are hitting their limit, which slows how fast retail prices can climb.

Memory makers see a long crunch; PC makers expect relief sooner. Plan for the first and hope for the second.

Both sides have an interest in their story: suppliers enjoy record margins, and PC makers want customers to keep buying. Our planning view is to budget for high prices through at least 2027 and treat any earlier relief as upside. The bigger unknown is AI spending itself. As we noted in AI’s multi-trillion-dollar revenue gap, the buildout depends on AI eventually earning enough to justify it. A sharp slowdown in data center spending would free up memory fast; until then, AI customers set the price.

What you buy	What’s happening	What to do
Laptops and desktops	Vendors raised list prices through 2026; Acer expects another 5% to 20% into year-end	Replace only what must go; lock quotes for the rest of 2026
Phones	Galaxy S26 up $100 to $200; Huawei and others raising prices; budget models hit hardest	Extend handset refresh cycles; standardize on fewer models
On-premises servers	Memory is the largest cost; RAM-heavy configurations rose most; quotes expire faster	Right-size memory per workload; check repricing clauses before you sign
Cloud	Hyperscalers locked in supply early, but their memory costs eventually flow into prices	Review reserved capacity and renewal terms; watch for price changes on memory-heavy instances
Local AI hardware	128GB AI workstations up about 74%; vendors launching lower-memory versions	Use smaller models where they are good enough; rent capacity for peaks
IoT and embedded	Single-board computers and low-end memory chips rising; some older memory types squeezed	Order critical spares early; qualify alternative parts

Where the memory crunch hits a typical business, and the first move for each. Sources: Micron, Samsung, Raspberry Pi, Nvidia, Dell, Acer, Tom’s Hardware.

5. The security catch: what squeezed budgets get wrong

Higher hardware prices are a finance problem. How companies react to them can become a security problem. Five mistakes are worth heading off now.

Stretching devices past their support date. Keeping a laptop an extra year or two is sensible, but only while it still receives security updates. Windows 10 support ended in October 2025, and Google has said it will phase out ChromeOS by mid-2034, cutting short the promised support window on some Chromebooks. Unpatched endpoints are where attackers start, as we covered in AI-powered threats facing small businesses.
Buying gray-market or counterfeit memory. When a memory module costs five times what it did, relabeled, recycled and outright fake parts follow. Buy from authorized distributors and certified refurbishers, test new modules before deployment, and treat any deal far below market price as a warning sign.
Refurbished gear without provenance. Refurbished servers and laptops can be a smart way through the crunch, but insist on a certified data wipe, current firmware and BIOS, and a clear record of where the device came from.
Under-specifying machines so far that security suffers. Endpoint detection, disk encryption, browser isolation and on-device AI features all need memory. Microsoft’s Copilot+ PCs, for example, require at least 16GB of RAM. Starved machines run slowly, and users then disable the tools that slow them down.
Ignoring where the components came from. Chinese-made memory from makers such as CXMT and YMTC is now appearing in mainstream PCs and SSDs. That is not automatically a problem, but YMTC has been on the US Entity List since December 2022, so organizations with government contracts or strict supply-chain rules should ask vendors for component details and check them against their obligations.

One more risk is contractual rather than technical. If a vendor can reprice your order before it ships, your budget is only as firm as the shipping date. Read the terms before you rely on a quote.

Sort every device into one of three buckets before you approve a single purchase order.

8 steps to protect your IT budget (and your security)

Map your fleet by age and support date. List every laptop, desktop, phone and server with its purchase date, memory, condition and the date it stops receiving security updates. You can’t decide what to replace, stretch or retire without it.
Buy what you must now, and lock the price. For devices that are failing, out of support or unable to run your security tools, order now rather than hoping for relief. Ask how long quotes are valid, push back on clauses that allow repricing before shipment, and consider staggered deliveries with fixed prices.
Stretch healthy devices safely. A clean reimage, a new battery and a storage health check can add 12 to 24 months to a business laptop. Only do this for devices that will keep getting security updates over that period, and record the new retirement date.
Right-size memory to the job. Stop ordering one maximum specification for everyone. Most office laptops need around 16GB to run a current operating system, security tools and on-device AI features comfortably; reserve 32GB to 64GB for developers, analysts and designers who will actually use it. Apply the same discipline to server and virtual machine memory.
Right-size your AI. Memory-hungry AI is now expensive to run locally and in the cloud. Use smaller or open-weight models where they are good enough, rent capacity for peaks rather than buying for them, and track inference spend closely, as we explained in why AI pilots overrun their budgets.
Buy only from trusted channels. Use authorized distributors and certified refurbishers, test memory and storage on arrival, and keep proof of origin. For regulated work, ask vendors which memory and storage makers’ parts are inside the devices you buy.
Retire devices securely and capture their value. Used hardware is suddenly worth more. Encrypt, wipe with a verifiable method and document every retired device before you resell or trade it in. Our Green IT guide covers responsible reuse and disposal.
Budget in scenarios and review every quarter. Plan a base case with hardware costs 15% to 30% above last year’s, a downside case for further increases and an upside case in which Acer’s forecast comes true. Review the plan each quarter against what memory makers and PC makers are saying, and against your cloud renewal dates.

Frequently asked questions

Why are RAM prices so high in 2026?

Because AI data centers are absorbing most of the world’s memory production. Memory makers have a limited number of wafers, and they are prioritizing high-bandwidth memory for AI accelerators and server memory for AI data centers, which pay far more than PC and phone makers. HBM stacks many layers of DRAM, so it uses up capacity that would otherwise become ordinary memory chips. The result is that a 64GB DDR5 kit that cost about $191 in August 2025 averaged $1,118 a year later.

How long will the memory shortage last?

Forecasts differ. On 30 September 2026, Micron said demand will exceed supply in both 2027 and 2028, with conditions tighter than in 2026, and that its new factories arriving in 2028 won’t bring immediate relief. SK hynix expects the shortage to last until 2030. Acer’s CEO is more optimistic, predicting PC prices will plateau in the first half of 2027 and then fall as cheaper Chinese memory arrives. A cautious plan assumes high prices through at least 2027.

Why did Samsung raise Galaxy S26 prices?

On 1 October 2026, Samsung raised US prices for the Galaxy S26, S26+ and S26 Ultra by $100 and for the 1TB Ultra by $200. Samsung has blamed memory costs for its earlier price rises on the S26, and the AI-driven shortage is the common thread across the industry. The base Galaxy S26 now costs $999.99. Huawei said on the same day that memory has added about $200 to the cost of each phone it makes and that it will also raise prices.

Should my business buy laptops now or wait?

Replace devices that are failing, out of security support or unable to run your security tools now, because most forecasts point to higher prices into 2027. Keep healthy devices that still receive updates for another 12 to 24 months. Lock quotes where you can, and check whether vendors can reprice an order before it ships.

Does the memory shortage affect cloud prices?

Indirectly. Large cloud providers secured supply early through long-term contracts and advance deposits, which shields them from the worst of the spot market. But they are paying far more for memory too, and those costs tend to show up in pricing, discounts and renewal terms over time. Review reserved-capacity commitments and renewal dates, especially for memory-heavy and AI workloads.

Is it safe to buy refurbished or discounted memory?

Refurbished hardware from certified sellers can be a good option, as long as it comes with a verified data wipe, current firmware and a clear record of origin. Be wary of memory sold far below market price on open marketplaces, where relabeled, recycled and counterfeit parts become more common during shortages. Buy from authorized channels and test modules before deploying them.