AI server prices are set to rise next year. Nvidia, unable to absorb the surging cost of memory semiconductors, has decided to pass the increase on to its customers.
Bloomberg reported on the 22nd (local time) that Nvidia has notified major clients including Microsoft, Google, and Oracle that it will raise the price of systems shipping in early next year by more than 15%. Systems equipped with next-generation chips Vera Rubin and Grace Blackwell are also subject to the increase. The size of the hike will vary depending on the chip generation and memory configuration.
The company that effectively dominates the global AI semiconductor market has cited memory as the reason for the price increase. A single component is rewriting the price tag of the finished product.
The increase is far from minor. Nvidia’s latest AI servers are said to cost tens of billions of won per rack. A 15% increase means several hundred million won more per rack. For major customers buying tens of thousands of units, that could translate into additional costs in the trillions of won.
The reason servers are getting more expensive
Inside an AI server, compute chips and memory are like workers and a supply warehouse. Nvidia’s chips handle the calculations, while high-bandwidth memory (HBM) and other DRAM constantly deliver the data needed for those calculations. If the warehouse is slow, the workers sit idle. That is why competition in AI performance has turned into competition in memory performance.
Memory prices move according to supply contracts signed with major customers on a quarterly or semiannual basis. Each contract renewal reflects the higher market price, a structure that keeps pushing up server costs.
The problem is supply. The global DRAM market is effectively an oligopoly dominated by Samsung Electronics, SK hynix, and Micron. Although the three companies are increasing production, demand growth driven by AI adoption is even steeper. DRAM prices have continued to surge since the second half of last year.
HBM’s characteristics are intensifying the supply crunch. HBM is made by stacking multiple layers of DRAM, so even with the same capacity it uses several times more wafers than ordinary DRAM. As memory makers devote more capacity to HBM production, supply of standard DRAM also tightens. That is why memory prices are rising not only for servers but also for PCs and smartphones.
In the HBM market, SK hynix has maintained its position as Nvidia’s key supplier, while Samsung Electronics and Micron have been pursuing it. With ordinary DRAM prices also rising, all three companies are seeing profits expand.
The power balance has reversed
Bloomberg pointed out that Nvidia’s warning of a price increase shows the strengthening bargaining power of memory makers. For a long time, memory was treated as a volatile subordinate, with prices fluctuating according to the business cycle. Now the strongest finished-product maker is the one sending customers a notice of price increases because of memory costs.
Customers have limited options. Microsoft, Google, Amazon, and Meta are developing their own AI chips. Even so, building large-scale data centers still requires Nvidia chips. That is because Nvidia bundles chips optimized for data centers together with its development ecosystem, making it difficult to switch suppliers.
The price increase is also a way for Nvidia to protect its margins. If the company absorbed the rising memory costs itself, profitability would fall. A company with monopoly-like power can pass that burden on to customers, and it is now clear that Nvidia has used that leverage.
Whether the increase will actually be implemented as announced remains to be seen. Customers are also pushing back with their own chip development and diversification of suppliers. The final increase may still be adjusted during negotiations.
Will my cloud bills also go up?
The higher costs cascade down the chain. As server prices rise, cloud companies face higher costs to build and run data centers. If cloud fees increase, the cost structure of AI services running on top of them also rises. Enterprise pricing would likely move first, and consumers could eventually feel it through services that were once free becoming paid, or through higher usage fees.
The impact is already being felt. Rising ordinary DRAM prices are also pushing up manufacturing costs for PCs and smartphones. The memory supercycle is extending beyond data centers and shaking prices for everyday devices.
For South Korea’s economy, this is initially positive news. Memory is the country’s top export item, and when prices rise, Samsung Electronics and SK hynix see profits expand. In a market where higher prices still lead to sales, the impact is reflected directly in the two companies’ earnings. If investment in capacity expansion continues, orders will also spread to the materials, parts, and equipment industries.
Consumers who have delayed replacing a PC or buying high-capacity memory should keep an eye on price trends. Industry watchers widely expect the supply shortage to persist and the uptrend to continue into next year.
Since semiconductor stocks have led the domestic stock market’s direction, some analysts say news that even finished-product makers are willing to absorb price increases can be read as a sign that demand remains solid and the industry cycle is continuing.
Still, there is also caution about the cycle. The memory market has repeatedly seen capacity expansions during boom periods turn into oversupply two to three years later. A classic example was the sharp price collapse in 2019 after the 2018 boom. If AI demand weakens just as the factories now being built begin to flood the market with supply, prices will move in the opposite direction.
Companies’ responses also differ. Cloud users can soften the impact by locking in rates with long-term contracts or spreading out purchase timing. AI service providers now have no choice but to absorb rising costs through efforts such as caching and other cost-saving technologies.
The memory-driven price increase is a sign that the cost structure of the AI industry is changing. Behind the race for chip performance, the competition to secure memory has emerged as the deciding battleground. How much of the increase is passed on, and when new supply floods the market, are the two key points to watch in next year’s semiconductor market.