The latest plot twist in the investment trajectory of hyperscalers isn’t just about more data centers, more GPUs, or expanded networks: according to UBS, memory is beginning to “shift” CAPEX visibly. The bank estimates that roughly one-third of this year’s CAPEX increase can be attributed to rising memory costs, a component that previously often went unnoticed in the overall spending picture.
After the earnings season, UBS upgraded its CAPEX forecasts for leading hyperscalers: +43% for 2026 and +28% for 2027, reaching $827 billion and $915 billion, respectively. This would imply year-over-year growth of 61% in 2026 and 11% in 2027.
UBS’s interpretation is clear: memory is becoming a much more dominant cost line. Its estimates indicate that memory expenditure for hyperscalers will rise from $53 billion in 2025 to $155 billion in 2026 and $252 billion in 2027, adding about $100 billion in CAPEX annually. Regarding the “culprits” behind this growth, UBS calculates that this increase accounts for 32% of the CAPEX growth this year.
Memory already has a bigger weight in the BOM (and in rack economics)
UBS also breaks down the impact at the product level. In general-purpose servers, the firm estimates that memory could add around $10,000 per server, raising its share in the BOM (bill of materials) from 4–6% to 6–7%. It may seem like a small percentage change, but it’s significant across fleets deployed in hundreds of thousands of units.
The more “dramatic” shift occurs in large-scale AI platforms. UBS notes that, in a NVL72 rack, memory cost share could jump from 6% to 16% as the platform evolves from GB200 to VR200, a move it describes as “significant”.
Summary table (UBS estimates)
| Concept | 2025 | 2026 | 2027 |
|---|---|---|---|
| Hyperscale CAPEX (forecast) | — | $827 billion | $915 billion |
| YoY CAPEX growth | — | +61% | +11% |
| Hyperscale memory costs | $53 billion | $155 billion | $252 billion |
| Additional annual CAPEX attributable to memory | — | ~$100 billion | ~$100 billion |
| Proportion of CAPEX increase due to memory | — | 32% | — |
| Impact per general-purpose server (memory) | — | +$10,000/server | — |
| Memory share in BOM (general-purpose server) | 4–6% | 6–7% | — |
| Memory share in NVL72 (platform transition) | 6% | — | 16% |
Implications for the market
If this scenario becomes consolidated, the message for investors and operators is twofold:
- The bottleneck is no longer just “compute”: memory availability and pricing (and its integration into AI systems) are now conditioning deployment rates and budgets.
- The infrastructure mix is changing: as memory gains weight in the BOM and in rack economics, small price variations translate into large numbers at hyperscale levels.
- Efficiency becomes the hidden KPI again: with a larger portion of CAPEX absorbed by memory, optimizing configurations, platform refreshes, and purchasing cycles can have a more direct financial impact.
via: Jukan X