On August 25, 2026, TrendForce released a forecast showing that DRAM and NAND Flash will rise from 47% of major cloud service providers' (CSPs) total capital expenditure in 2026 to 68% in 2027. The implication is that the costs driving AI infrastructure spending aren't limited to the number of GPUs deployed—memory pricing and capacity per system are becoming equally significant factors.
That 68% figure isn't a locked-in budget. It's a TrendForce projection, and the August 25 announcement didn't specify which companies were included or how the calculation was derived. Still, when combined with separate capex totals TrendForce published for CSPs, the scale of the shift facing major cloud operators comes into focus. Will they absorb higher prices, cap costs through long-term contracts, or cut back on memory per server? Memory is shifting from a procurement line item to a core design constraint for AI systems.
Translating 68% into roughly $884 billion
In a separate report dated August 3, TrendForce forecast that the world's nine largest CSPs would spend more than $886.7 billion on capex in 2026, rising to approximately $1.3 trillion in 2027. The nine companies covered are listed below; the five North American firms account for roughly 90% of the 2026 total.
| Region | Companies |
|---|---|
| North America (5) | Google, Amazon, Meta, Microsoft, Oracle |
| China (4) | ByteDance, Tencent, Alibaba, Baidu |
Assuming these are the same nine companies referenced as "major CSPs" in the August 25 announcement, applying the published percentages to the capex figures yields: for 2026, 47% of over $886.7 billion equals more than $416.75 billion; for 2027, 68% of roughly $1.3 trillion equals approximately $884 billion. That's roughly a 2.12x increase, or about 112% growth. Even if total capex growth is only around 50%, the 21-percentage-point rise in memory's share means memory-related spending expands even faster.
This roughly $884 billion figure isn't a purchasing budget disclosed by CSPs. It's also unclear whether the memory valuation embedded in capex aligns with manufacturers' reported DRAM and NAND revenue. TrendForce hasn't disclosed how memory built into servers and storage equipment was valued, or whether distribution-channel pricing was included. This estimate translates the scale of 68% into a dollar figure—it doesn't confirm actual order amounts.
The underlying forecasts themselves are also shifting. The August 3 report described the nine companies' 2026 capex growth as "approximately 90% year-over-year," while the August 25 announcement cited "98% growth" for major CSPs. TrendForce hasn't explained whether this reflects an upward revision or a change in which companies were counted. As a result, some assumptions remain in connecting the $886.7 billion figure to the 68% share.
Rising prices meet a 27% supply increase
The rise to 68% stems from a forecast in which unit prices and volumes increase simultaneously. Contract prices for Server DRAM—which CSPs purchase heavily—rose a cumulative 64% in the second half of 2025 and are projected to rise a further cumulative 270% in 2026. Enterprise SSD prices are forecast to rise about 35% in the second half of 2025 and roughly 235% in 2026.
Some long-term agreements (LTAs) are curbing these increases. Contracts signed from Q2 2026 onward include price caps, making it harder for suppliers to raise prices freely on the volumes covered by those agreements. Even so, TrendForce projects that 2027 HBM contract prices could rise 70–140%.
On the volume side, concentration in AI servers is intensifying. HBM is a stacked DRAM product, while RDIMM refers to the DRAM modules used by server CPUs. Together, these are expected to account for 51% of DRAM supply bits in 2026. In 2027, combined Server DRAM and HBM supply bits are forecast to grow 27%, driven by process migration and new fabs coming online.
A 27% increase doesn't necessarily mean lower prices. If AI server shipments and per-unit memory configurations grow even faster, that supply increase gets absorbed by demand. Moreover, since HBM typically consumes more wafer capacity than standard DRAM, ramping up HBM production doesn't proportionally increase overall DRAM bit supply. The 68% figure isn't simply a measure of price spikes—it's the product of high unit prices multiplied by large bit volumes.
Cutting memory configurations outside the reach of long-term contracts
Even CSPs with LTAs aren't fully insulated from market pricing. TrendForce forecasts Server DRAM contract prices will rise 13–18% quarter-over-quarter in Q3 2026, with the price increases concentrated among customers without LTAs and among contracted customers making additional purchases beyond their allocated volumes. SK hynix has stated it signed LTAs with roughly 10 customers by Q2 2026, but hasn't disclosed the specific products, volumes, prices, or contract terms involved.
Efforts to reduce per-unit memory capacity are already underway. Some CSPs and server OEMs have been switching from 96GB/128GB RDIMMs to 32GB/64GB modules since the first half of 2026. This reflects adjustments aligning memory configurations with CPU supply conditions and procurement costs—it doesn't indicate growth in total procurement volume or completed server units. TrendForce projects that 2027 RDIMM supply bits will grow only 15–20% year-over-year, failing to keep pace with server CPU shipment growth.
However, reducing capacity carries computational trade-offs. Cutting CPU-side DRAM shrinks the memory available for host-side preprocessing and reduces headroom for concurrent processing. Reducing HBM shrinks the volume of models and intermediate data accelerators can access at high speed. Because the impact varies by model and software configuration, there's no uniform conversion rate—such as "dropping to 32GB reduces performance by X%."
TrendForce also mentioned options such as adjusting HBM capacity in future AI chips or developing custom ASICs with model architectures fixed directly into silicon. This isn't an announcement that CSPs have committed to these approaches. They represent a trade-off—sacrificing general-purpose flexibility for efficiency in specific workloads—which could reduce memory costs but also make it harder to adapt to model changes or repurpose hardware for multiple uses.
DRAM tightness and NAND easing fall under the same 68% figure
The combined 68% for DRAM and NAND doesn't mean both markets are moving in the same direction. TrendForce estimates 2026 DRAM supply-demand balance at negative 1–2%, with the shortfall widening further in 2027. Even if new fabs come online in the second half of 2027, their meaningful contribution to supply won't materialize until 2028. Time is needed after construction completes to install equipment and ramp up yields.
NAND is following a different trajectory. Higher layer counts are increasing bits per wafer, and new fabs are gradually starting operations. Kioxia plans to invest approximately ¥470 billion annually to grow supply at an average rate of 22% per year from 2025 to 2028. The company says it can meet demand through 2029 using its existing Yokkaichi Y7 and Kitakami K2 facilities, without adding new fabs.
If consumer device demand remains weak, TrendForce expects NAND supply constraints to ease in the second half of 2027. That said, QLC products are expected to account for the largest share of bit demand growth in enterprise SSDs. Even if NAND market conditions loosen in the latter half of the year, average annual CSP spending could still be pushed up by first-half pricing, enterprise SSD volumes, and elevated DRAM and HBM prices. The 68% forecast and NAND easing aren't contradictory.
Whether actual spending approaches the forecast depends on the capex totals CSPs ultimately disclose, and on whether they need to purchase beyond the volumes locked in through LTAs. It also hinges on when Server DRAM and HBM supply bits catch up with demand, and whether NAND supply-demand balance turns positive in the second half of 2027. Confirmation of the 68% figure will show up in how many units cloud giants cut memory capacity on—and how much they end up paying outside their contracts.