The Memory Line Item: How HBM Rewired the $725B Hyperscaler CapEx Stack
The Memory Line Item: How HBM Rewired the $725B Hyperscaler CapEx Stack
Somewhere in the last two years, memory stopped being an afterthought in the AI infrastructure story and became its central constraint. Amazon, Alphabet, Meta, and Microsoft are on track to spend roughly $725 billion combined on capital expenditure in 2026 -- up 77% from $410 billion in 2025 -- and by some industry estimates, memory alone now accounts for as much as 30% of that spend, up from single digits just two to three years ago. High-bandwidth memory (HBM) is sold out at every major supplier through 2027. Having watched several capital cycles compress and release since the 1990s, I think the more interesting question isn't whether hyperscalers keep spending -- they clearly will -- but which companies actually control the pricing power inside that spend. Increasingly, the answer sits in Korea.
- $725B -- combined 2026E capex across Amazon, Alphabet, Meta, and Microsoft, up 77% from $410B in 2025 (Source: CreditSights, via CNBC and Seeking Alpha, Aug 2026)
- $54.6B -- Bank of America's 2026 global HBM market forecast, a 58% increase year-over-year (Source: SK Hynix Newsroom, 2026 Market Outlook)
- Sold out through 2027 -- Micron's HBM capacity commitment status, with BofA pushing its supercycle-end estimate as far as 2030 (Source: IG International, Jul 2026)
All figures cross-verified against at least two independent sources unless noted otherwise.
1. The Capex Number Keeps Climbing
The headline figure is now familiar to anyone following Big Tech earnings: the four largest hyperscalers plan to spend approximately $725 billion on capital expenditure in 2026, up from roughly $410 billion in 2025 and $226 billion in 2024. That is a 77% year-over-year increase on top of an already-elevated base. Goldman Sachs has pushed its longer-run estimate even further, now projecting $5.3 trillion in combined capex for Meta, Microsoft, Amazon, and Alphabet between fiscal 2025 and fiscal 2030 -- up from $4.5 trillion before the latest round of earnings.
What is less understood outside semiconductor circles is where the incremental dollar is actually going. Two to three years ago, memory was a relatively minor line item in AI server bills of materials -- by some estimates around 8% of total capex. Today, some analysts put that figure as high as 30%, a roughly 3.75x increase in relative weight even as the absolute capex base has also grown enormously.
2. From a Rounding Error to the Central Constraint
The mechanism behind this shift is straightforward once you look at how AI accelerators are actually built. GPUs like Nvidia's H200 and B200 sit next to stacks of high-bandwidth memory that feed them data fast enough to keep the compute busy. As models and context windows have grown, the HBM content per accelerator has risen sharply, and HBM itself is more expensive and harder to manufacture than conventional DRAM -- Bank of America estimates it requires three to four times the production capacity per unit. That combination -- more memory per chip, at a higher cost per gigabyte -- is what has pushed memory's share of the AI infrastructure bill so much higher, so quickly.
Single-source caution: the 8%-to-30% capex-share figure originates from a single industry analysis and has not been independently corroborated by a second source at the time of writing. Treat it as directionally indicative of the trend rather than a precise measurement.
Table 1 -- Hyperscaler CapEx Trajectory, 2024-2026E
| Year | Combined CapEx | YoY Change |
|---|---|---|
| 2024 | ~$226B | -- |
| 2025 | ~$410B | +81% |
| 2026E | ~$725B | +77% |
Source: CreditSights estimates, company guidance compiled via CNBC and Seeking Alpha (Aug 2026)
3. The Three-Way Race for HBM Supply
With memory now the binding constraint, the question of who actually supplies it becomes a capital-allocation question in its own right. SK Hynix remains the HBM revenue-share leader, with estimates in the 50-56% range depending on the source and quarter. Samsung is the fastest-gaining challenger, with revenue share estimated between 28% and 41% depending on whether the metric is revenue-based or bit-capacity-based -- UBS's bit-capacity projection actually has Samsung edging ahead of SK Hynix in 2027, at 41% versus 39%. Micron trails at roughly 21-22% but has the advantage of being effectively sold out of HBM capacity through 2027, giving it pricing visibility regardless of market-share rank.
Both Korean suppliers will present next-generation HBM work at Hot Chips 2026, running August 23-25 at Stanford -- Samsung on the base die, SK Hynix on advanced packaging. That conference is not a product launch venue, but the industry reads these technical disclosures as roadmap signals for who leads the coming HBM4 generation, which several suppliers are targeting for volume production later this year.
4. Why the HBM Market Itself Is Compounding
Underneath the market-share competition, the total addressable market is expanding fast enough that share battles matter less than raw capacity. Bank of America estimates the global HBM market will reach $54.6 billion in 2026, a 58% increase from the prior year, with average selling prices for both DRAM and NAND rising sharply as demand outstrips available capacity. Some forecasts suggest the 2028 HBM market alone could exceed the entire DRAM market's 2024 size -- a reflection of how central this specific memory category has become to AI infrastructure economics.
Table 2 -- HBM Market Share by Supplier (2026 Estimates)
| Supplier | Revenue Share | Positioning |
|---|---|---|
| SK Hynix | ~50-56% | Incumbent leader; first to qualify each HBM generation with Nvidia |
| Samsung | ~28-41% | Fastest-gaining; may lead on bit-capacity basis by 2027 (UBS) |
| Micron | ~21-22% | Smaller share but sold out through 2027 |
Source: TrendForce, UBS, The Motley Fool (ranges reflect differing revenue vs. bit-capacity methodologies, Q1-Q2 2026)
Both Samsung and SK Hynix are responding with historic capacity commitments. Samsung is reportedly expanding HBM production capacity by roughly 50% in 2026, while SK Hynix has more than quadrupled its previously announced infrastructure investment. South Korea's government has separately announced a ten-year plan that includes roughly 800 trillion won (about $530 billion) of combined investment from the two companies to build new fabrication sites in the country's southwest -- Samsung's P5 facility in Pyeongtaek is targeted for 2028, and SK Hynix's M15X facility for mid-2027.
5. What This Means for Capital Allocation Discipline
For institutional readers, the practical implication is that "AI capex" is no longer a single undifferentiated line to underwrite. A hyperscaler's capex-to-operating-cash-flow ratio matters, but so does the composition of that spend -- specifically, how much of it is locked into multi-year memory supply agreements at fixed or semi-fixed pricing versus exposed to spot-market volatility. SK Hynix confirmed in earlier reporting that it had completed long-term supply agreements for HBM4 with more than ten customers, the kind of contracted demand that, in my experience, tends to matter more for earnings durability than any single quarter's headline market-share print.
The counterpoint worth naming directly: at close to 25x EV/EBITDA for US AI-exposed equities according to Allianz Research, valuations are near historical extremes not far from telecom multiples that preceded the 2000 peak. When four companies are all making the same structural bet simultaneously, they also share the same downside if enterprise AI monetization disappoints. Memory suppliers are, in that sense, both the primary beneficiary of the current build-out and one of the more exposed links if the capex cycle ever reverses.
6. Market Outlook: Two Paths
I want to avoid a deterministic call here. What follows is the most probable direction current data supports, alongside the counter-risks that would invalidate it.
The base case rests on memory remaining the physical bottleneck of this cycle: HBM sold out through 2027 caps how quickly any single hyperscaler can scale regardless of capital committed, which in turn preserves pricing power and margin durability for the two Korean suppliers specifically. The counter-scenario is that the widening gap between AI infrastructure spending and realized AI revenue eventually forces a correction -- a single soft cloud-growth quarter from any of the Big Four could re-rate the entire capex-linked stack at once, given how synchronized the spending commitments already are. New capacity coming online from Samsung's P5 and SK Hynix's M15X facilities in 2027-2028 would also gradually ease the physical bottleneck, though on a slower timeline than near-term demand growth.
7. Portfolio Considerations for Institutional Readers
- Distinguish capex growth from capex composition -- a hyperscaler's memory-supply contract structure is now a material input to earnings-durability analysis, not a footnote.
- Korea-specific semiconductor exposure captures a structural piece of the AI capex story that a US-only tech allocation typically does not.
- Treat single-source capex-composition estimates (such as memory's rising capex share) as directional signals to monitor for corroboration, not settled figures to model precisely.
- Diversification and a long time horizon remain the most reliable tools through a cycle this synchronized and this richly valued.
8. Related Reading
9. Conclusion
The AI infrastructure story has quietly split into two distinct narratives: how much hyperscalers are spending, and what they are actually able to buy with it. On the second question, the physical constraint increasingly runs through a small number of Korean and US memory fabs rather than through capital availability itself. Having watched capacity-driven cycles play out before, I'd treat sold-out HBM through 2027 as a genuine structural signal, not just a supply-chain headline -- while staying honest that a cycle this synchronized, and this richly valued, carries real correlated downside if the underlying AI revenue thesis disappoints.
10. Investment Disclaimer
This article is published for informational and analytical purposes only. Nothing contained herein constitutes financial, investment, legal, or tax advice. All investment decisions involve risk, including the possible loss of principal. Past performance is not indicative of future results. Readers should conduct their own research and consult a qualified financial advisor before making any investment decisions. The views expressed are solely those of the author based on publicly available information and do not represent the official position of SEJONG DnC.
- CreditSights estimates, via CNBC, "Tech AI spending approaches $700 billion in 2026," 2026.2.6
- Seeking Alpha, "4 Hyperscalers, One Message: The AI Trade Isn't Over," 2026.8.2
- Goldman Sachs research, via Yahoo Finance, "Meta, Microsoft, Amazon, and Alphabet capex to 2030," 2026.6.3
- SK Hynix Newsroom, "2026 Market Outlook -- Focus on the HBM-Led Memory Supercycle"
- IG International, "Memory chip supercycle 2026: what is driving the DRAM, HBM rally," 2026.7.8
- Tenbrief, "Hot Chips 2026: Samsung brings the base die, SK hynix brings the packaging," 2026.8.11
- DataCenterDynamics, "Samsung and SK Hynix to scale up memory production capacity in 2026," 2026 (3 weeks prior)
- Allianz Research, EV/EBITDA valuation commentary, via Forbes, 2026.6.2
Comments
Post a Comment