Summary of Key Points
NVIDIA has informed cloud providers such as Microsoft and Google through its contract manufacturers that the prices of AI servers, which will be shipped early next year, will increase by more than 15%. The reason for this increase is not the high cost of the GPUs themselves, but the soaring prices of memory chips (DRAM), which could rise by 250%-280% for the entire year. This price hike will be transmitted along the supply chain (from cloud providers to computing power rental services, and ultimately to end-users). It also reveals a shift in the AI computing power landscape: from a situation where GPUs were the dominant factor to one where memory chips have become the key. Memory chip manufacturers (Samsung, SK Hynix, and Micron) are the biggest beneficiaries, while cloud providers and contract manufacturers are under pressure.
1. The Price Hike Was Not Directly Announced by NVIDIA, but by the Contract Manufacturers
Many people think that NVIDIA directly raised the prices for cloud providers, but in fact, it was the contract manufacturers (such as Foxconn and Pegatron) that passed on the message. The supply chain consists of three layers:
- Upstream: NVIDIA sells AI chips.
- Midstream: Contract manufacturers purchase chips and assemble them with memory, motherboards, etc., to create complete servers.
- Downstream: Cloud providers buy these servers to build data centers.
NVIDIA first raised the prices for memory chips, which results in a gross profit margin of only around 5% for the contract manufacturers (they make a 5% profit on a product sold for $100). Unable to absorb the double increase in the cost of chips and memory, they had no choice but to pass the extra cost on to the cloud providers. For cloud providers, a 15% increase in the price of a server that costs several million dollars means an additional expense of hundreds of thousands of dollars. If cloud providers cannot afford this increase, they will likely pass it on to the providers of computing power rental services and those using large AI models, ultimately leading to higher prices for end-users.
The increase in prices is not uniform; servers with more memory will see a larger increase in cost (for example, versions with full memory configurations can be several times more expensive than basic versions). Additionally, the price hike coincides with NVIDIA's financial report release, which is a somewhat strategic timing.
2. The Root Cause of the Price Hike: Memory Chips Going from a Supporting Role to a Dominant One, with Prices Soaring
The reason for the high prices of memory chips is a severe imbalance between supply and demand:
- Supply Side Monopoly and Capacity Bias: 90% of the global DRAM (general-purpose memory) production is controlled by Samsung, SK Hynix, and Micron. These companies have shifted their production capacity towards HBM (high-speed memory dedicated for AI use), which generates several times the revenue per wafer compared to regular DDR5 memory. As a result, there is a shortage of DRAM, and HBM has been completely sold out by its major customers as early as last October. Moreover, it takes 2-3 years to build new memory wafer production facilities, so capacity cannot be increased in the short term.
- Demand Side Explosion: The demand for memory in AI servers is 8-10 times that in traditional servers. For instance, NVIDIA's new GPUs require 288GB of HBM4 per unit, and the total memory capacity in a single server can exceed 20TB. Data from Morgan Stanley shows that the cost of memory in Vera Rubin servers has increased by 435%, rising from 10% to 26% of the total cost of the server. Memory is no longer a minor component but has become a major factor in the cost structure.
3. Who Is Affected and Who Is Benefiting?
- Cloud Providers Are in the Worst Position: Companies like Microsoft and Google are developing their own chips, but they still rely on NVIDIA for their AI systems. The intense AI competition means that they cannot afford to delay purchasing servers and must accept the price hike. They can either absorb the increased costs (resulting in reduced profits) or pass them on to users (for example, by charging more for services using technologies like ChatGPT).
- Contract Manufacturers’ Profits Are Being Drained: Foxconn, the world's largest AI server contract manufacturer, saw its revenue double in the second quarter of this year, but its gross profit margin dropped from 7% to 5%. Foxconn is now encouraging cloud providers to purchase memory on their own, meaning it will only earn the assembly fee and no longer bear the full cost increase.
- The Biggest Winners Are the Memory Chip Giants: Samsung, SK Hynix, and Micron have a monopoly on memory production capacity and are now in a position to raise prices at will. Even companies like Apple and Qualcomm are affected by the memory shortage and have had to increase their prices. NVIDIA, which used to profit heavily from its GPUs, is now being impacted by the demand for memory that it itself helped create, similar to a boomerang effect.
4. The Changing Landscape of Computing Power: From a Race for GPUs to a Race for Memory, and the End of the “Cheap Era”
For the past two years, everyone has been competing for NVIDIA GPUs, assuming that the bottleneck in computing power lay with the GPUs themselves. However, the reality in 2026 is that the shortage of memory is more severe than that of GPUs, and capacity expansion for memory is even slower (taking 2-3 years), with the shortage expected to continue until at least 2027. This means that the power to set prices for AI computing power has shifted from GPUs to memory.
NVIDIA has been preparing for this by investing $14.5 billion in securing HBM supply for 2026-2027 and has also invested in data center infrastructure (such as Cloverleaf and SB Energy). It has even partnered with institutions like Blackstone to invest $50 billion in building AI infrastructure. NVIDIA's concerns now extend beyond just manufacturing chips; it also has to ensure that its customers have the funds to purchase its products and that data centers have the necessary power supply.
For the entire AI industry, the competitive landscape has changed. No longer does the company with the most GPUs have the upper hand; instead, the one that can afford the high costs of memory will be the dominant player. The “cheap era” of computing power is truly over.