The new gold rush in the cloud is no longer GPUs, but memory. Several industry reports point to a growing mismatch between supply and demand for DRAM and HBM for data centers, and major cloud service providers (CSPs) are reacting very differently. Among the Americans, Google and Microsoft are clearly prioritizing speed and availability over cost, even taking on extra expenses to secure preferred supply.
Memory Shortage Amid the AI Race
The catalyst is well known: the explosion of generative AI and large-scale models has driven up demand for high-performance memory, including HBM (High Bandwidth Memory) for GPUs and DDR5 for servers. Additionally, the main manufacturers—Samsung, SK Hynix, and Micron—are shifting factories and production lines toward these premium products, reducing the availability of cheaper, traditional memories.
The result is a tense market where:
- Shipments to servers are prioritized over PCs and consumer devices.
- Some players have been engaging in of modules for months to ensure future capacity.
- Prices for DDR5 and HBM have experienced double-digit increases throughout 2025, particularly for high-capacity server modules.
Google and Microsoft: Paying More to Be Faster
According to memory channel sources and industry analysts, among the US hyperscalers there are two clear strategies. On one side, Google and Microsoft are accepting higher prices to guarantee fast deliveries and assured volumes for their AI clusters.
This makes sense:
- Both are engaged in a public race to deploy AI superclusters, such as Copilot services, Gemini, proprietary models, and enterprise AI solutions.
- A memory bottleneck would leave their top-tier GPUs underutilized, wasting billions invested in hardware.
- For their business accounts, the extra cost of memory is marginal compared to the strategic value of continuing to launch AI products rapidly and securing major enterprise contracts.
In other words: for Google and Microsoft, the “real risk” isn’t paying 20–30% more per module, but running out of RAM or HBM when clients want to train models, fine-tune LLMs, or deploy AI agents at a global scale.
Other CSPs: More Price-Sensitive
Other major players—such as some second-tier cloud providers and parts of the Asian ecosystem—are being more conservative: delaying orders, diversifying suppliers, or combining modules from different generations to contain costs.
This approach reduces immediate pressure on their CAPEX but has an hidden cost:
- Less margin to aggressively scale their AI platforms.
- Increased exposure to future price spikes if shortages worsen.
- Risk of falling behind in performance compared to hyperscalers that have heavily invested in high-speed HBM and DDR5.
The Role of Samsung, SK Hynix, and Micron
The other piece of the puzzle is the manufacturers. Samsung and SK Hynix, in particular, are accelerating plans for new production phases and more advanced nodes, aiming to strengthen their leadership in HBM3E and upcoming HBM4 dedicated to AI data centers.
This strategy involves:
- Increased investment in factories and DRAM capacity almost exclusively for AI.
- Less focus on conventional memories for PCs and mobile devices, which pushes prices down across the entire chain.
- Stricter negotiations with customers: the highest-paying and most demand-visible clients get the best production slots.
Hence, the hyperscalers with stronger financial muscle, like Google and Microsoft, are in a better position to secure long-term contracts and guaranteed volumes.
What Does This Mean for the Cloud and AI Markets?
In the short term, the message is clear:
- Clients will see higher prices for GPU instances and high-memory nodes.
- Startups and smaller companies may find it harder to access cutting-edge AI resources at competitive prices.
- The gap between major hyperscalers and the rest widens, not only in software and models but also in basic physical infrastructure.
Mid-term, this tension could accelerate:
- The development of more memory-efficient alternatives, such as architectures with better DDR utilization, new cache hierarchies, or more compact models.
- Interest in on-premise and specialized bare-metal solutions, where some companies prefer to invest in their own hardware to avoid dependence on the cloud amid soaring prices.
- Greater regulatory and geopolitical movements concerning chips, fabs, and export controls, with the US, Europe, Korea, and Taiwan as key players.
Speed Today, Bill Pay Tomorrow
That Google and Microsoft prioritize speed over cost doesn’t mean that pricing isn’t a concern. They have simply decided that, at this stage of the AI race, the critical factor is to arrive first (or not fall behind) and capture market share.
The big question is how long this dynamic can be sustained:
- If memory prices stay high for years and AI demand doesn’t translate into sufficient margins, even giants will need to reconsider their strategies.
- Conversely, if corporate AI and business agents become the new foundational infrastructure, today’s expensive contracts could be seen in a few years as a winning bet to dominate the next decade of cloud computing.
Meanwhile, the clear message from the market is: in the new AI economy, memory is power. And those hyperscalers willing to pay for it will set the pace for the next wave of innovation.