RAM, GPU and SSD prices: why AI is driving them up (and making them swing)

Since late 2025, the price of the components that go into an artificial intelligence workstation or server has been changing almost every week. RAM, graphics cards, SSDs: our suppliers revise their prices constantly, sometimes upwards by several tens of percent in a single quarter. That is why the prices shown on our GPU servers and AI workstations are indicative, and why we prefer to draw up a quote based on your needs. Here is what is happening in the market, backed by figures and sources.

RAM: the first component to soar

DRAM (the RAM in your machines) is at the heart of the increase. Research firm TrendForce, which tracks the prices negotiated between memory manufacturers and large buyers, issued revision after revision in 2026:

• First quarter of 2026: in January, TrendForce forecast a 55–60% rise in contract prices for conventional DRAM, and more than 60% for server memory. The actual increase was even steeper, around 90–95% over the quarter according to figures reported by StorageSwiss.
• Second quarter of 2026: a new forecast of +58–63% for DRAM and +70–75% for NAND flash memory, the kind used in SSDs.
• Fourth quarter of 2026: the pace is slowing but prices are still climbing, with +10–15% expected for DRAM and +15–20% for NAND.

For consumers, the effect is dramatic: a 32 GB DDR5-6000 kit listed at $134.99 in September 2025 had passed $420 by early December, according to Tom's Hardware. A 1 TB SSD has gone from around $45 to nearly $90 since late 2025.

Why: AI is tying up memory factories

Data-centre AI accelerators (NVIDIA H200, B200, B300, AMD Instinct…) use HBM memory, stacked right next to the chip. And this memory is extremely silicon-hungry: according to Micron's CEO, HBM3E requires around three times as much wafer area as DDR5 for the same capacity. Every wafer devoted to HBM therefore means roughly three fewer wafers of conventional memory.

Manufacturers are making a logical economic choice: HBM and server memory sell for much more, to cloud giants who pay upfront. Three groups, Samsung, SK hynix and Micron, produce more than 95% of the world's DRAM, and all of them have steered their new capacity towards AI servers. As early as October 2025, SK hynix announced that its HBM, DRAM and NAND output was sold out for the whole of 2026. Micron even announced it was discontinuing its Crucial consumer brand to better serve its “strategic customers”.

Graphics cards are following suit

A graphics card is a chip plus a lot of memory. On some high-end models, video memory now accounts for more than 80% of the component cost, according to an industry source quoted by Astute Group. When memory prices soar, the card follows:

• the spot price of GDDR6 has risen from around $2.50 to $7.50 per gigabyte since late 2025, according to TrendForce data reported by Tech Times;
• in Germany, the 3D Center website measured in October 2026 an average year-on-year increase of 66% for the GeForce RTX 50 series, 132% for the RTX 5090 and 48% for the Radeon RX 9000 series;
• AMD once again raised the price of the GPU and GDDR6 memory kits it sells to card manufacturers on 1 October 2026, after several increases since the start of the year;
• an NVIDIA H200 card was already trading at between $30,000 and $40,000 in early 2026, with HBM3E stacks themselves up by around 20%.

Manufacturers are also juggling volumes: priority goes to AI accelerators and the most profitable products, making some more affordable models harder to find.

SSDs and storage: AI comes into play here too

Inference (the everyday use of AI models) requires a huge amount of fast storage. According to a Nomura note reported by Tom's Hardware, SanDisk planned to double the price of its NAND memory for enterprise SSDs in the first quarter of 2026. At the end of September, TrendForce said enterprise SSDs were seeing the steepest increase of all memory categories, with new supply already almost entirely reserved.

Cloud giants are booking production in advance

To secure their supply, the major cloud providers are signing multi-year contracts and accepting higher prices. Smaller buyers, integrators and resellers like us, share what is left. DRAM lead times exceed 40 weeks for large orders, and some suppliers advise spreading purchases over several quarters to smooth out price swings.

An effect visible across the entire IT market

Research firm Gartner estimates that the combined price of DRAM and SSDs will have jumped by 130% by the end of 2026 compared with 2025. As a result, the average price of PCs is expected to rise by 17% and that of smartphones by 13% over the same period, while PC shipments are set to fall by 10.4% in 2026, “the sharpest contraction in more than a decade” according to analyst Ranjit Atwal.

What next? Some relief is possible, but not just yet

Building and starting up a new memory fab takes at least 12 to 18 months, and the new capacity from Samsung, SK hynix, Micron and Kioxia will only reach significant volumes between late 2026 and 2027. What is more, it has already been earmarked first and foremost for HBM and server products. Research firm IDC forecasts supply growth of 16% for DRAM and 17% for NAND in 2026, well below the usual 20–30%.

That said, prices will not keep rising forever. In October 2026, memory manufacturers' shares fell sharply, amid the rise of China's CXMT and fears of overproduction by 2029. Counterpoint Research expects DRAM prices to fall by around 23% in 2028 compared with 2027, while remaining around 230% above 2025 levels. In other words, prices will stay high and, above all, volatile, in either direction.

What this means at Happy Mining

A price displayed on Monday may no longer match our purchase cost by Friday, whether up or down. Rather than charging you a safety margin “just in case”, we prefer to work on a quote basis:

• Right-sizing first. Use case, target AI models, required video memory, number of users: your actual needs determine the configuration. Often, a different choice of graphics card or amount of RAM delivers the same result for less.
• Today's price. We check with our suppliers at the time of your request and offer you the best price available, confirmed in writing in the quote.
• Indicative prices on the website. The prices of our AI workstations and servers give you a ballpark figure. They may change, and the quote is what counts.
• A reservation deposit on certain products. It is refunded in full if you decide not to go ahead with the quote.

Our advice if you are planning an AI project

1. Start by defining your use cases and the models you want to run: video memory (VRAM) is what determines the size of the models you can use.
2. Don't buy “big” on principle: every unnecessary gigabyte of RAM or VRAM costs more today than it did a year ago.
3. Once you have received and approved the quote, don't wait too long: in this market, a price can change within a few weeks.
4. Think about scalability: a machine designed to take a second graphics card later saves you from replacing everything.

To compare configurations, see our GPU and AI page, our ready-to-use AI workstations and our GPU servers for AI. For a price tailored to your project, ask us for a quote or call +33 6 32 49 26 41.

Sources

• TrendForce, memory price forecast for Q1 2026 (5 January 2026)
• TrendForce, forecast for Q2 2026 (31 March 2026)
• TrendForce, forecast for Q4 2026 (30 September 2026)
• Gartner, rising memory costs and their impact on PCs and smartphones (26 February 2026)
• Tom's Hardware, why HBM is weighing on PC RAM (19 December 2025)
• Tom's Hardware, SanDisk and NAND pricing for enterprise SSDs (9 January 2026)
• Guru3D, Micron: HBM consumes three times the wafer area
• StorageSwiss, memory and flash prices will not come down before 2027 (6 May 2026)
• Astute Group, the memory shortage and GPU pricing (February 2026)
• Tech Times, graphics card prices rise again (12 October 2026)
• Digital Today, memory stocks fall, Chinese competition and overproduction risk (16 October 2026)