Something unusual happened on April 29, 2026 that the headline financial coverage did not fully process. Meta Platforms raised its 2026 capital expenditure guidance from USD 115 to 135 Billion to USD 125 to 145 Billion, a USD 10 Billion increase at both ends of the range. The reason Meta gave in its 8-K filing matters more than the number: "This reflects our expectations for higher component pricing this year and, to a lesser extent, additional data center costs to support future year capacity."
Read that sentence carefully. Meta is not buying meaningfully more units. It is paying meaningfully more per unit. And it is telling the SEC, in plain language, that the year-over-year capex increase is driven primarily by price, not by additional capacity. This is the single most important framing change in the AI infrastructure investment cycle since ChatGPT launched, and the rest of the supply chain is now arranging itself around the same logic.
The unit-to-price shift
Through 2024 and most of 2025, hyperscaler capex growth was a story about deploying more accelerator silicon. Microsoft, Google, Amazon, and Meta were buying more H100s, more H200s, more Blackwell B200s, and more custom ASICs. The unit count went up. The dollar count went up faster than the unit count, but not catastrophically faster.
In 2026 that relationship inverted. Samsung Electronics has signalled high-teens to low-20% HBM contract pricing increases for 2026 deliveries. CoWoS advanced packaging at TSMC is oversubscribed through 2027. UBS estimates NVIDIA alone could demand 678,000 CoWoS wafers in 2026 — approximately 40% above 2025 — and TSMC will reach 120,000 to 130,000 CoWoS wafers per month by year-end, up from 13,000 to 16,000 per month as recently as 2023. That capacity ramp is unprecedented in scale. It is still not enough.
The result is a market where every component in the AI accelerator bill of materials is pricing toward its supply ceiling. Memory is now projected to consume approximately 30% of hyperscaler capex in 2026, four times the 2023 share. Micron has publicly used the word "unprecedented" to describe the current memory shortage. Smartphone OEMs are absorbing the second-order effect: IDC now predicts smartphone shipments will decline nearly 13% in 2026 as elevated DRAM and NAND component costs compress handset margins.
The HBM allocation tax
The DRAM market is now bifurcated into two products that share a wafer line. Standard DDR5 and LPDDR5X — the memory that goes into servers, PCs, and smartphones — entered 2026 with spot prices that fell 18% in Q1 as Samsung continued to produce at volume into a soft consumer market. HBM3E and the early HBM4 generation, manufactured from the same base DRAM wafers but with an additional stacking and through-silicon via process, is fully allocated through 2026 with contract pricing rising.
SK Hynix has structurally chosen this allocation pattern. Every wafer start the company directs to HBM processing is a wafer not making DDR5 modules. The company's Q3 2025 operating profit of KRW 7.03 Trillion — its highest in six years — came almost entirely from this allocation choice. The trade-off is rational at SK Hynix's individual scale. The aggregate result is a memory market where the part of the segment that serves AI infrastructure is sold out at premium pricing and the part that serves the consumer device market is in oversupply.
What CEOs are actually saying
The framing change is visible in earnings call language. TSMC Chairman C.C. Wei was unusually direct on the Q4 2025 call when pressed on whether AI demand is real: "I'm also very nervous about it. We have to invest about USD 52 to 56 Billion for the capex. If we did not do it carefully, that will be a big disaster to TSMC for sure. So of course I spent a lot of time in the last three-four months talking to my customers and then customers' customers. I want to make sure that my customers' demands are real."
BE Semiconductor's Q1 2026 call captured a different dimension. CEO Richard Blickman confirmed orders being booked in 2026 for volume production starting in 2028 — a two-year forward visibility window that backend equipment makers have not had in any prior cycle. ON Semiconductor's Hassane El-Khoury closed his Q1 2026 call describing the period as "one of the most challenging cycles our industry has seen" while simultaneously delivering 5% year-over-year growth, the gap between difficulty and reported numbers being a function of how concentrated the AI-driven demand has become at a small number of supply chokepoints.
The dispersion of fortune across the semiconductor industry in 2026 is wider than the headline "USD 1 Trillion year" suggests. TSMC, SK Hynix, ASML, BESI, and the HBM-allocated portions of Micron and Samsung are operating in a structural shortage that supports premium pricing. Mainstream DRAM, the smartphone supply chain, consumer NAND, and large parts of the analogue and mixed-signal segment are in cyclical correction with extended duration.
For procurement teams, this means the 2026 budget is not a single budget. It is two budgets: one for AI-adjacent component categories that are operating in a sellers' market with multi-year forward commitments required, and one for everything else where buyer leverage has returned for the first time since 2021.
The smartphone collateral damage
The most overlooked consequence of the AI-driven memory allocation shift is what it has done to the smartphone industry. Mid-range Android handset OEMs cannot pass elevated DRAM and NAND costs through to consumers without losing volume to lower-priced Chinese competition. Apple has the brand premium to absorb the cost increase, and the company's Q1 2026 iPhone revenue held up. The mid-range Samsung Galaxy A series, Xiaomi, Vivo, Oppo, and Motorola portfolios are facing a different problem: their pricing power is constrained from above by used-iPhone competition and from below by domestic Chinese price points.
The result is the IDC projection of nearly 13% smartphone shipment decline in 2026. Memory and storage cost compression is the largest single contributor to this decline, exceeding the impact of replacement cycle extension or consumer macro weakness. The same DRAM allocation logic that delivers SK Hynix its record operating profit is removing approximately 150 million units from the global smartphone shipment forecast versus what would have happened at 2024 memory contract pricing.
What investors should price differently
The investment thesis that dominated 2024 and 2025 — "AI infrastructure is real, NVIDIA is the principal beneficiary, buy NVIDIA" — produced 240% NVIDIA stock returns from January 2024 to June 2026 and was the correct trade for that period. It is incomplete for the 2026 to 2028 horizon for one specific reason: NVIDIA's revenue growth from this point requires either substantially more CoWoS capacity, substantially more HBM allocation, or NVIDIA capturing more of the value within the same hardware envelope. None of those three options is fully within NVIDIA's control.
The investors who repositioned in late 2025 and early 2026 toward the supply chain bottleneck names — ASML, BE Semiconductor, Kulicke & Soffa for advanced packaging equipment; SK Hynix for HBM; Lasertec for actinic inspection — captured a structurally different set of returns. UBS estimated that the bottleneck-exposed equipment makers traded at price-to-earnings multiples 30 to 40% below NVIDIA on 2026 estimates as of Q1 2026, while their forward revenue visibility extended further into 2027 and 2028 than NVIDIA's did.
What this means for 2027
The trillion-dollar 2026 will not be the peak. SIA projects the trajectory continues through 2027 and 2028, with DRAM revenue alone potentially approaching USD 600 Billion by 2028 if HBM allocation continues to capture an increasing share of total memory revenue. The structural constraint is not capital — hyperscalers have proven willing to commit any reasonable amount — but execution velocity at the bottleneck nodes. ASML can sell only as many EUV systems as it can manufacture. TSMC can install CoWoS lines only as fast as it can qualify the process at each capacity increment. SK Hynix can produce HBM4 only as fast as its through-silicon via stacking yield will support.
The 2027 question is whether HBM4 12-layer yield at SK Hynix reaches the level required to support the Rubin GPU NVL576 cluster configurations on NVIDIA's published timeline, and whether Samsung closes its HBM yield gap fast enough to provide meaningful second-source supply. Both variables are real, neither is fully observable from outside the supply chain, and procurement teams making 2027 commitments now are essentially expressing a probability estimate on each one without being able to verify their estimate before the commitment is locked.
That uncertainty is the framing every CFO managing 2027 budget allocation should hold in mind. The semiconductor industry crossed USD 1 Trillion in 2026 not because more chips were sold, but because the chips that mattered most became scarce enough to command premium pricing through a complete budget cycle. The structural conditions that produced this outcome are not resolving in 2027.