Memory Is the Capex Constraint

Hyperscaler capital spending is guided toward roughly $635 billion in 2026, and a rising share of that number is memory price, which buys no added capacity. SemiAnalysis puts memory at about 30 percent of hyperscaler capex this year, up from roughly 8 percent in 2023 and 2024, and high-bandwidth memory is supplied by just three companies, two of them Korean, which is why the June repricing came through Seoul.

Read the companion deck (PDF)→

01 · THE NUMBER

A third of hyperscaler capex is now memory

Memory was about 8 percent of hyperscaler capital spending in 2023 and 2024, and SemiAnalysis estimates it reaches roughly 30 percent in 2026. CLSA takes the same category to 35 percent in 2026 and 48 percent in 2027. These are analyst estimates for memory as a whole, wider than high-bandwidth memory alone, and we flag them as estimates. The reason the share climbs is price as much as volume. In the third quarter of 2025 DRAM contract prices rose 171.8 percent from a year earlier, per TrendForce, and TrendForce estimates a further 93 to 98 percent quarter-on-quarter increase in the first quarter of 2026 with another 58 to 63 percent projected for the second. When the price of a large input triples inside a year, a capital-spending line that holds its physical shape still grows sharply in dollars, and that dollar growth reads on the tape as more buildout.

Clients quote the dollar figure as if it were capacity, and Microsoft's 2026 guidance shows why that reading misleads. On the April 29 fiscal-Q3 call the company guided 2026 capital spending to about $190 billion, roughly 23 percent above the 150 billion the sell-side expected, and chief financial officer Amy Hood attributed about 25 billion of the raise to higher prices for memory and components. For context, that 25 billion is roughly 13 percent of Microsoft's entire 2026 budget, and it pays for the same equipment the company would have bought anyway at last year's prices. Meta cited memory pricing when it lifted its own 2026 range from $115 to $135 billion to $125 to $145 billion. The mechanism is the one the reference desk has described for compute, where prices signal scarcity and ration a constrained input, applied now to the memory that sits next to the processor.

02 · THE OLIGOPOLY

Three firms make the memory, two of them Korean

Three companies make the high-bandwidth memory that AI accelerators require, and Counterpoint Research put SK Hynix at 62 percent of HBM bit shipments in the second quarter of 2025, Micron at 21 percent, and Samsung at 17 percent. SK Hynix and Samsung are South Korean and Micron is American, so about four-fifths of the bits ship from Korea. The Q3 2025 prints move some share back toward Samsung, near 22 percent, and the three-firm structure holds. The barrier that keeps the count at three is physical. Gavin Baker of Atreides Management, whose 2025 Micron position is up about fourteenfold, argues that high-bandwidth memory for 2027-class accelerators requires stacking 12 to 16 DRAM dies in a single package, and that only Micron, SK Hynix, and Samsung can do it. We record that as a single investor's view and flag it, though the concentration it describes matches the Counterpoint shares and the reported sold-out state of 2026 supply.

An oligopoly of three producers holding a scarce, hard-to-build input sets price on different terms than a commodity market. SK Hynix said in October 2025 that its DRAM, NAND, and HBM capacity was sold out for 2026, and Micron chief executive Sanjay Mehrotra told Bloomberg in May that meaningful new industry supply does not arrive until 2028. SK Hynix chairman Chey Tae-won said at Computex in June that he expects the shortage to run into 2030. High-bandwidth memory also consumes wafer capacity out of proportion to the bits it yields, since IO Fund reports that current HBM3E needs about three times the wafer area per gigabyte of standard DDR5, so every wafer moved to HBM cuts commodity DRAM supply by more than one-for-one and lifts the price of the memory in ordinary servers and phones as well. Baker's read is that Micron should trade at a premium to equipment makers such as ASML and Lam Research, above the commodity discount memory usually carries. That premium only makes sense if the pricing power is durable, and the same assumption sits under the economics of the entire buildout.

03 · THE READ-THROUGH TO SEOUL

The AI trade reprices through Korea

The marginal supplier of memory for US AI capex sits in Korea, and the equity tape has repriced to match. IO Fund's June 26 chart shows US memory, meaning Micron, up more than 1,200 percent from the August 2025 low, and South Korean memory up more than 700 percent, both far ahead of a semiconductor ETF and the Nasdaq-100 over the same window. Micron alone added more than $1 trillion of market value over ten months to about 1.35 trillion. In Korea the concentration is now a macro fact. Samsung and SK Hynix together were a record 42.2 percent of the KOSPI in May 2026, the KOSPI is up more than 80 percent on the year, and on June 22 SK Hynix passed Samsung to become the country's most valuable listed company for the first time in more than 25 years, briefly worth about $1.3 trillion. SK Hynix has secured roughly two-thirds of Nvidia's initial HBM4 orders.

The concentration explains the shape of the June drawdown better than a story centered on the US megacaps. When the memory-share estimates circulated in late June and the market reassessed hyperscaler returns on invested capital, the assets that moved most were the memory makers and the Korean index, because that is where the scarce input and its pricing power are held. A buyer of the AI-capex complex is long a memory cycle priced by three firms, and short the risk that those three raise the cost of the buildout faster than the buildout produces revenue. That is a different risk than owning the compute layer, and it concentrates in fewer names and one exchange.

SECOND-ORDER EFFECT · WHY REPORTED GROWTH OVERSTATES THE BUILDOUT

A capex number rising 55 percent year-on-year reads as more capacity, since more dollars spent on infrastructure should mean more equipment installed. Part of the increase is price, because memory is both a large share of the bill and in a supercycle, so a chunk of reported capex growth is the same physical order at a higher unit cost. Microsoft's own $25 billion attribution for 2026 is a price effect a hyperscaler has quantified, and if memory is a third of the bill and its price has more than doubled, the gap between dollars spent and capacity added widens across the cohort. For allocators, this means that some portion of AI's measured contribution to GDP growth is a transfer of rents from the hyperscalers to three memory suppliers, and only the remainder represents new compute in the ground.

WHAT WE'D DO · FOR OPERATORS AND ALLOCATORS

Treat the hyperscaler capex headline as a gross figure and net out memory price before reading it as capacity, because the physical buildout and the dollar total have separated and only the physical quantity can be planned against. An allocator long the AI-capex complex is implicitly long a memory cycle set by three firms and short their pricing power, so size that exposure deliberately and watch Korea, where two companies now carry 42 percent of the index. An operator buying compute should write memory-price escalators into multi-year infrastructure plans, since suppliers have guided the shortage into 2028 and beyond and are signing five-year contracts, and should assume no relief in unit cost before new fabs reach output in 2028. And where a workload can run on less memory-intensive inference, moving it there is one of the few direct hedges available against memory-price inflation.

WATCHLIST · WHAT WOULD CHANGE OUR MIND

The thesis rests on memory staying scarce and its price staying high, and two developments together would break it. The first is new supply arriving earlier than 2028, whether from a faster Micron or SK Hynix ramp or a Samsung yield recovery that adds a fourth effective HBM source and softens the three-firm floor. The second is a demand step-down, either a hyperscaler capex cut or an efficiency gain such as Google's TurboQuant that reduces memory per unit of inference enough to loosen the bottleneck. A clean way to be proven wrong is a 2027 in which memory's capex share falls back toward 2024 levels while the dollar capex number keeps climbing, which would say the constraint had moved elsewhere. We are watching the DRAM contract series for a quarter that breaks the run of increases, the memory makers' sold-out language for any easing, and the share of hyperscaler capex that independent trackers attribute to memory as 2026 actuals replace estimates.

SCORECARD · CONSENSUS VS THE TAPE

Three consensus sentences read differently against the 2026 tape. The loudest, that record hyperscaler capex is a clean signal of AI capacity, ignores the price term: Microsoft alone attributes about $25 billion of its 2026 number to higher memory and component prices. The claim that memory is a commoditized, cyclical business is harder to square with a three-firm supply structure that is sold out into 2028 and DRAM prices that rose 171.8 percent year-on-year in the third quarter of 2025. The claim that the AI trade is a US megacap trade understates Korea, where SK Hynix and Samsung are 42 percent of the KOSPI and SK Hynix just became the most valuable company in the country. We flagged the memory estimates as analyst figures and the die-count claim as one investor's, and we will mark the call when 2026 capex actuals and the memory-share breakdowns are published.

NEXT ISSUE · THE POWER BILL

Next week the series follows the buildout past memory to electricity, and to whether the grid can deliver the megawatts the 2026 capex plans assume. We will separate the data-center capacity that is contracted from the capacity that is energized, and trace the same scarcity-and-price logic through interconnection queues and power prices.

SELECTED SOURCES AND CAVEATS

SemiAnalysis on memory as a share of hyperscaler capex (CY23-CY24 ~8%, CY26 ~30%), posted April 2026 and reported by Tom's Hardware; CLSA's 35% (2026) and 48% (2027) via Crypto Briefing, June 23, 2026; Microsoft's April 29 fiscal-Q3 call and the ~$25B memory attribution (CNBC, The Register, Motley Fool); Meta's 2026 capex range (DCD, Fortune); combined 2026 guidance summed by Second Order from company figures, with 2022-2025 levels from reported aggregates (Visual Capitalist, Tom's Hardware); Counterpoint Research HBM bit-share via Mark LaPedus / Semiecosystem (August 5, 2025); TrendForce DRAM contract-price series via Tom's Hardware and IO Fund; supply and shortage statements from SK Hynix, Micron (CEO Sanjay Mehrotra, Bloomberg), and SK Hynix chairman Chey Tae-won; the equity read-through and HBM-market figures from IO Fund / Beth Kindig (June 26, 2026); Korea concentration and the SK Hynix / Samsung crossover from CNBC and KED Global; and Gavin Baker's HBM-share and die-count claims from a Podcast Alpha clip, June 27, 2026.

The memory-share figures are analyst estimates for a category wider than HBM alone, and the 48% for 2027 is a forecast. The 2022-2025 combined capex levels are assembled from reported aggregates and are approximate; the 2026 total is summed company guidance, and third-party aggregates run as high as $725B. HBM bit-share is a third-party estimate. Baker's 30 to 40% HBM-of-capex claim and the 12-to-16-die claim are one investor's view, attributed and flagged. The equity read-through is an investor-published chart; the memory returns are its stated figures and the benchmark values are approximate reads. Second Order is an independent research briefing, for discussion only and not investment advice.