The Scalper Flair of RAM: What on Earth Is Happening on the Market

The AI model boom has created a hardware shortage that hit RAM before processors or GPUs. The market has descended into scalper behavior reminiscent of the cryptocurrency mining and GPU shortage eras.

Current Prices: 96 GB DDR5 Server RAM

• Legitimate retail (Regard): 410,000 rubles (~$4,100)
• Gray market dealers: 330,000 rubles, fluctuating throughout the day
• Underground dealers (Gorbushka/Savelovskaya): 250,000-280,000 rubles, cash only

For context: In summer 2025, the same module cost 90,000-100,000 rubles.

Buyers describe "quests" rather than transactions: sellers negotiate with dozens of people simultaneously, inventory disappears mid-negotiation, and there are no guarantees until the physical handoff.

The Market Broke Free

The crisis began subtly in autumn. A Russian supplier pressured RUVDS to lock in annual volumes by October. Within four days of negotiations, prices jumped 10%. By December, the situation became absurd.

TrendForce Data

• Year-over-year increase: 171%
• Q1 2026 forecast: Additional 90-95% increase
• Historical significance: The largest quarterly jump ever recorded — exceeding 1990s or dot-com crisis levels

PCPartPicker Retail Trends (May to December 2025)

• DDR4-3200 16 GB: $50 → $120 (+140%)
• DDR5-6000 32 GB: $120 → $410 (+242%)
• DDR5-6000 64 GB: $210 → $750 (+257%)

Global Indicators

• Japan (Akihabara): Tsukumo and Sofmap implemented purchase limits
• USA (Micro Center): Removed DDR5 price tags, directing customers to ask sales staff
• Framework laptops: Stopped selling memory modules separately
• eBay scalpers: 150-250% markups with willing buyers
• CyberPowerPC: Announced system price increases citing 500% RAM cost increase

DDR4 Became More Expensive Than DDR5

A market inversion occurred: previous-generation memory became more expensive than the current generation. ChangXin Memory Technologies shifted production to scarce DDR5, making cheap memory vanish instantly.

Companies with DDR4-based server infrastructure face a dilemma: buying expensive DDR4 upgrades or replacing servers entirely with DDR5 systems costing astronomical amounts.

The secondary market came alive: old servers are being stripped for DDR4 modules. Even used DDR3 saw price increases.

But Why?

1. AI Consumption

OpenAI signed agreements with Samsung and SK Hynix (Project Stargate). 900,000 DRAM wafers monthly — approximately 40% of global DRAM production. Two clients consuming nearly half the worldwide supply.

2. High Bandwidth Memory (HBM) Competition

NVIDIA's H100, H200, and Blackwell GPUs require specialized HBM with a 3-4x larger physical footprint per bit compared to DDR5. The yield rate for HBM chips is only 50-60%. One HBM wafer sacrifices approximately 10 standard memory wafers. Higher margins incentivize producer reallocation. HBM now consumes around 30% of total production capacity.

3. Producer-Coordinated Shortage

Samsung switched to monthly pricing — a signal that "prices will rise tomorrow." Micron exited the consumer business in December 2025, having controlled approximately 20% of the market. This amounts to intentional supply reduction to maintain margins.

4. US Tariff Response

The Trump administration announced 25-100% import tariffs on chips. Server builders (Dell, HP, Supermicro) began panic-buying from warehouses before tariff implementation.

5. Taiwan Risk

Three earthquakes in 18 months (April 2024, January 2025, December 2025). Each causes production line shutdowns. TSMC lost $92 million from a single earthquake. Micron lost 4-6% of quarterly shipments. Quartz tubes fracture, calibrations misalign, and output drops.

Additional Factor

Samsung and SK Hynix coordinated more than 10% NAND production cuts to maintain prices, causing NVMe SSD cost spikes alongside RAM.

I Have Seen This Before with GPUs

2017-2018: The Bitcoin Wave

Bitcoin hit $19,000. AMD RX 580 sold at 3x MSRP or disappeared entirely. Nvidia implemented "two per person" limits.

2020-2021: The Ethereum Wave

A perfect storm: COVID lockdowns, chip shortages, and the Ethereum rise. In April-May 2021:

• RTX 3080: 329% price increase
• RTX 3060 Ti: 230% increase
• Bot armies grabbed inventory in seconds
• Gray market dealers bought stocks for cash resale at 50% markup

Recovery Timeline

The market healed only after Ethereum's shift to Proof of Stake (September 15, 2022). Combined with the FTX and Terra/Luna ecosystem collapses, prices normalized over 8-9 months. Used mining GPUs flooded secondary markets; Chinese dealers sold them by weight.

RUVDS GPU servers purchased at 60,000 rubles dropped to 20,000-30,000 rubles on Avito.

What Will Happen to RAM?

Pessimistic Outlook

Unlike the GPU crisis, which had a clear ending event (the Proof of Stake transition), the RAM crisis lacks any off-ramp.

Why prices will not collapse:

• AI shows no signs of stopping training or inference
• A crash would require an AI bubble burst — possible but uncertain
• Investor skepticism is emerging (questions about ROI vs. flashy demos)
• Even if an "AI winter" arrives, massive inertia remains
• OpenAI, Google, and Microsoft have pre-paid chip manufacturing quotas years ahead
• Long-term contracts cannot be cancelled
• Memory is needed universally (unlike GPUs, which were for miners and gamers only)

Supply-Side Bottleneck

New Samsung P4 and SK Hynix M15X factories will not produce until 2028 at the earliest. No physical capacity increase is expected before then.

Forecasts

• Gartner: 47% additional DRAM price increase in 2026
• Counterpoint: 64 GB DDR5 server modules could double by end of 2026
• SK Hynix statement: All 2026 capacity "essentially sold"

Kingston's Honest Assessment

Representative Cameron Crandall: "If you are planning upgrades, do it now. In 30 days it will cost more. In another 30 days, more again."

What We Are Doing in Russia

RUVDS Decision Framework

We are not rushing purchases; we follow our normal acquisition schedule.

Currency risk scenario: If global prices drop 40% but the ruble weakens 40% against the dollar (plausible with import resumption and trade balance shifts), the benefit cancels out. There is also the risk of a "super-combo": an expensive dollar plus expensive memory.

Strategic Constraints

• Stop buying: Lose market share as the market grows
• Buy everything: Excessive risk exposure
• Milk old servers: A path to bankruptcy — hardware ages, costs rise, clients leave permanently (it is very difficult to win cloud/hosting customers back)

Resolution

A 30% server cost increase means 30-40% longer payback periods. The only solution: pass costs to clients while market demand exceeds supply. Russia faces a chronic datacenter rack deficit since 2022.

Final approach: Raise client prices AND accept lower margins.