Memory Got This Expensive For Two Separate Reasons

What HBM Actually Is And Why It Is A Bottleneck

High Bandwidth Memory, or HBM, is not ordinary DRAM sold as a stick or built into a budget laptop. It is a stack of DRAM dies bonded vertically and placed directly beside an AI accelerator die, connected through TSMC's CoWoS 2.5D advanced packaging process. That physical proximity is the entire point: HBM gives a GPU the bandwidth to move data in and out of memory fast enough to keep its compute cores fed, and bandwidth, not raw computing power, is now the limiting factor in AI training and inference.

Every NVIDIA H100, H200 and Blackwell-class GPU sold for AI workloads needs HBM stacked beside it to hit the performance level buyers expect. That single requirement links two industries that used to move independently: DRAM chipmaking and advanced semiconductor packaging. When either one runs short, GPU supply runs short with it, regardless of how many logic dies a foundry can print in a given month.

Who Makes HBM And What They Just Reported

Three companies build nearly all of the world's HBM, and each one reported a version of the same story in mid-2026. SK Hynix, Micron and Samsung all confirmed that HBM capacity for this year is effectively spoken for, with allocations locked in months or quarters ahead of delivery.

SK Hynix began HBM4 mass shipments in the second quarter of 2026 and is ramping further in the second half of the year, backed by long-term supply agreements with roughly ten customers. Micron finished allocating its entire calendar-2026 HBM3E and HBM4 capacity by the fourth quarter of 2025 and has been shipping HBM4 for NVIDIA's Vera Rubin platform since March 2026. Samsung, for its part, is warning that shortages will extend across its wider memory lineup, not just HBM, through at least 2027.

Two Separate Bottlenecks, Not One

Most coverage treats the memory shortage as a single problem, but it is actually two separate constraints compounding each other. The first is DRAM fab capacity: TrendForce reports that the three major DRAM makers have shifted more than 80% of their advanced fabrication capacity toward high-margin HBM and server DRAM, leaving far less room for the ordinary consumer DRAM that goes into laptops, phones and budget PCs. Data centers now consume an estimated 70% of all memory chips produced worldwide, and AI workloads alone account for roughly 20% of total 2026 DRAM production.

That reallocation explains why DRAM shortages spread well beyond AI accelerators into everyday consumer hardware. A fab line making high-margin HBM for an AI customer is a fab line not making the DDR5 modules that go into a mainstream laptop, and chipmakers have a clear financial reason to favor the former. The squeeze started inside data centers, but it did not stay there.

Why More Memory Alone Would Not Fix It

TSMC's CoWoS advanced packaging is the second constraint, and it has nothing to do with how much DRAM exists. Every AI GPU that uses HBM needs a CoWoS packaging slot to bond the memory stack onto the logic die, and that step is a completely separate manufacturing process from growing DRAM wafers. TSMC has been scaling CoWoS capacity from roughly 35,000 wafers a month in late 2024 toward a targeted 120,000 to 140,000 wafers a month in 2026, with outside assembly and test partners such as Amkor adding another 50,000 to 60,000 wafers a month on top of that.

Even so, the gap between CoWoS supply and demand ran around 20% earlier in this cycle and is only narrowing toward roughly 10% by the end of 2026. That means even if every DRAM fab in the world doubled its HBM output overnight, finished AI GPUs would still queue for a limited number of CoWoS packaging slots. The two bottlenecks compound each other instead of substituting for one another, and that is the detail most consumer coverage of the shortage leaves out.

What This Already Costs EU And UK Buyers

This shortage has already reached ordinary consumer prices in Europe and the UK, not just AI infrastructure budgets. Apple raised Mac prices on 25 June 2026, citing memory and storage costs directly: the MacBook Air rose 200 EUR to 1,399 EUR, the 14-inch MacBook Pro rose 300 EUR to 2,199 EUR, and the 16-inch MacBook Pro rose 400 EUR to 3,399 EUR. Samsung's DDR5 32GB module pricing jumped from $149 to $239 in September 2026, a 60% increase, while broader DDR5 contract pricing has more than doubled, from roughly $7 to $19.50 per unit.

JPMorgan research estimates that DRAM prices rose more than 400% from the start of 2024 to the end of 2026, and that the memory shock alone is adding roughly 0.2 to 0.4 percentage points to core inflation, since every 10% rise in hardware costs adds about 0.1 percentage point to core CPI. Server DRAM contract prices have reportedly risen 90% to 95% in a single quarter, and TrendForce forecasts conventional DRAM contract prices rising a further 13% to 18% quarter on quarter in the third quarter of 2026.

How Long This Shortage Is Expected To Last

SK Hynix's own chief executive gave the clearest timeline estimate available in July 2026, warning that the shortage will probably persist well beyond 2030. That estimate comes from the company best positioned to know: SK Hynix posted 79.3 trillion won in Q2 2026 revenue, up 51% quarter on quarter and 257% year on year, with a 76% operating margin, numbers that reflect just how much demand is chasing limited supply. Micron and Samsung have made similarly long-dated commitments and warnings rather than promising a quick return to normal pricing.

A filed European Parliament question asks the European Commission whether the memory shortage should trigger crisis powers under the EU Chips Act, citing risk to the bloc's AI-infrastructure buildout and its broader digital strategic autonomy. Whether or not Brussels acts, the shortage is not behaving like a temporary supply hiccup that resolves once one company adds capacity. It is being driven by two separate, compounding constraints, and only one of the three major suppliers has offered a specific end date, and that date is a decade away.

Read next: Three Firms Now Decide If Your Phone Stays Cheap  |  A 14 Percent Crash Will Not Cut Your Memory Bill