UBS has substantially raised its artificial intelligence capital expenditure forecast, and the revised numbers are large enough to reframe the entire semiconductor complex.
The bank now expects AI capex to reach $998 billion in 2026 — almost double the $506 billion recorded in 2025 — before climbing to roughly $1.447 trillion in 2027. Its previous framing was “nearly $1 trillion this year”; the new detail is where that money goes.
The punchline is memory
| Segment | 2025 | 2026 | 2027 |
|---|---|---|---|
| Memory | $71bn | $367bn | $923bn |
| Other AI components | — | $631bn | $525bn |
Memory spending is forecast to climb from $71bn to $367bn this year and to $923 billion by 2027. That means higher memory costs account for roughly 60% of the increase in AI capex this year, and — because other component spending actually declines to $525bn — more than the entire net increase in 2027.
This is not a story about more chips. It is a story about the cost of the memory those chips need.
The counter-current
The same news cycle produced the opposing case, which is what makes this a real debate rather than a press release:
- OpenAI projects burning through $278 billion by 2030, per the FT — a liability number that only makes sense if capex keeps compounding.
- Nvidia’s CEO publicly dismissed the doomsday framing, saying there is a “0% chance” the world ends in 2030.
- Anthropic pushed its IPO to November, explicitly to present Q3 results after the safety debate reshaped the landscape — the CEO had publicly called for an industrywide slowdown in development pace.
A capex forecast rising while the loudest labs are openly discussing slowing down is not a contradiction. It is a maturity story: the buildout is now financed by balance sheets rather than narrative, and memory is where the bill lands.
Assets to watch
Memory-levered exposure — SKHY / DRAM / MU
Catalyst: SK hynix sits in the daily most-active list with a $1.3T+ market cap and +2.46% on the day; the Roundhill Memory ETF (DRAM) is now among the top ten ETFs by volume. Memory is no longer a niche cyclical — it is the largest single line in the AI buildout.
Surge case: if memory takes 60% of this year's capex increase, pricing power in DRAM and HBM is the scarcest input in the entire AI supply chain. UBS's own 2027 line implies the tightness persists for two more years.
Crash case: memory has historically been the most brutally cyclical link in semiconductors. A capex air pocket — one hyperscaler delaying a build — would show up in memory pricing before it showed up anywhere else.
Leveraged memory and semi pairs — MUU / SOXL / SOXS
Catalyst: the Direxion Daily MU Bull 2X ETF and the SOXL/SOXS pair are appearing in top-ETF volume repeatedly, which is the clearest sign retail is expressing the memory thesis with leverage rather than with shares.
Surge case: a single hyperscaler capex raise turns the memory trade into a momentum event, and 2x/3x vehicles amplify the arithmetic.
Crash case: leveraged pairs decay in chop, and this trade now depends on a two-year forecast from one bank. The 52-week range on the underlying names is huge — semiconductor equipment and memory names are showing ranges measured in multiples. Volatility cuts both ways when it is levered.
The AI bellwether — NVDA / INTC
Catalyst: NVDA still tops the most-active list at $222 with a $5.37T market cap; Intel is up triple digits over 52 weeks as its foundry comeback play holds.
Surge case: a $1.4tn capex year needs accelerators, and the incumbent capture of that spend is why NVDA's market cap exceeds most national GDPs.
Crash case: the same concentration that makes NVDA the bellwether makes it the single point of failure. A moderation in rollout pace — the very thing Anthropic and OpenAI are publicly debating — hits the largest position in the complex first.
The read
UBS just told you the AI trade’s bottleneck is not compute, it is memory supply. Meanwhile the labs are publicly debating whether to slow down. Both can be true: the infrastructure bill is contracted years out, and the narrative risk sits on top of a spending schedule that no longer depends on sentiment.
Watch next: Nvidia’s next data point on HBM supply, and whether DRAM’s rise in ETF volume is retail momentum or institutional positioning.