A year ago, a 32-gigabyte kit of DDR5 memory cost around $72. Today the cheapest good 32-gigabyte kit sits around $450. The sticks are the same, the speed is the same, yet the price has risen roughly six times.
Graphics cards have followed the same path. The RTX 5090 launched with an MSRP of $2,000. Last month the cheapest one in stock in the U.S. was $9,400, and most units sit above $10,000. The RTX 5080 jumped 2,200 percent in a single month. Anyone still waiting for prices to fall is facing a longer wait than expected.
Memory Makers Are Choosing AI Over Consumers
One memory maker has already stated that the worst of the shortage has not even begun. Micron shut down Crucial, the RAM and SSD brand many buyers had used for 29 years. The company walked away from ordinary consumer sales so it could supply AI data centers instead.
Contract prices for regular DRAM are expected to rise another 10 to 15 percent this quarter, and that figure is being presented as a slowdown. High-bandwidth memory consumes roughly three times the wafer capacity per gigabyte that ordinary DDR5 requires. Every wafer committed to HBM is a wafer that never becomes a consumer stick of RAM.
Samsung has said its 2027 HBM supply is sold out. SK Hynix has indicated that DRAM and HBM are effectively sold out through 2029. Micron has stated it has no clear view of when supply and demand will balance. Team Group, another major memory manufacturer, has warned that the peak of the shortage is still ahead, with supply becoming critically tight into the fourth quarter of 2026 and the second quarter of 2027. The company sees no meaningful relief before 2028.
In September the contract price of 16-gigabyte DDR5 chips rose from $8 to $48, an increase of about 5.5 times. Samsung has more than doubled DDR5 contract prices this year; one cited unit moved from $19.50 to around $72. Mainstream DDR4 chips went from $12 to $47. The same pressure is already visible in finished systems: the next laptop is expected to cost nearly double, approaching $1,000.
Smartphones, PCs, laptops, GPUs, gaming consoles, and SSDs are all expected to carry higher prices as memory tightens. Pre-built systems are absorbing the increases more slowly than individual components, but industry figures warn the shortage could last well into next year or even stretch toward a decade.
Robots and Data Centers Are Locking Up Supply
Micron has said every humanoid robot will need more than 200 gigabytes of memory. Tesla has stated it eventually wants to build one million Optimus robots a year. That volume of memory will never reach ordinary computers. Elon Musk noted on X that Tesla cut the memory of its robot chip in half because that was the only way to secure enough RAM to build Optimus. If Tesla is already constrained, the rest of the market is feeling the same pressure.
Apple and Tesla have both looked to China as an alternative source this week. AMD has told partners to expect roughly a 10 percent price increase on GPUs. Average card prices in the U.S. and Canada have climbed 58 percent above the lowest in-stock figures, and many models sell for 44 percent or more above MSRP. The 5090 itself is now listed 370 percent above its original MSRP after a brief scare that NVIDIA would stop production; NVIDIA later confirmed production continues.
Six-Year-Old GPUs Are Still Useful
A six-year-old RTX 3090 that launched in 2020 still trades around $2,000. That same card continues to run a brand-new 125-billion-parameter Qwen model on a desk machine. Older hardware that was expected to become worthless is holding value because new cards are either unavailable or priced far beyond most budgets.
With 64 gigabytes of system RAM, a single 3090 can run Qwen 3.8 27B or Qwen Flash Next at practical speeds. Reported figures land around 60 to 70 tokens per second in real use and as high as 88 tokens per second in optimized single-slot setups. Newer llama.cpp builds with MTP support have raised speeds roughly 30 to 40 percent on the same card. Cloud-code-style workloads typically deliver 30 to 40 tokens per second, so the local numbers are competitive for daily coding.
Openmonoagent (OMA), an open-source framework built for local operation, runs this exact configuration. The recommended setup is a 3090, 4090, or 5090 paired with 64 gigabytes of RAM. Systems with only 16 or 32 gigabytes of RAM are better suited to the smaller Qwen 3.8 27B variant. The framework is configured by default for four concurrent slots so multiple agents or developers can share one box; that concurrency lowers per-slot speed slightly compared with a single-user setup, but the model remains usable. Users who prefer one-to-one operation can adjust the configuration themselves because the software is open source.
Building Scrappy Local Machines
The practical response is to buy usable hardware now rather than wait. The 3090 and especially the 3090 Ti remain attractive; the Ti version offers roughly 15 percent more tensor cores and overall speed for an extra $100 to $200. Multi-GPU arrangements can increase concurrency, but the accuracy of a strong single model often matters more than raw speed for many workloads.
One example machine was assembled from used parts for under $2,500: an older motherboard and Core i9 processor with 64 gigabytes of RAM purchased on eBay for about $450, plus two V100S cards. Once cooling is resolved, the system is expected to deliver speeds comparable to a single 3090, potentially 100 to 120 tokens per second. The same box can support a development team of at least four people. Even DDR3 or DDR4 RAM is acceptable when the goal is local inference rather than maximum gaming performance; accuracy usually outweighs the speed penalty of older memory.
Why Local Hardware Wins
Cloud pricing is unlikely to stay flat while memory costs rise. Plans that currently cost $200 a month are expected to move toward $2,000 a month within the next 12 months. A local machine has a fixed hardware cost, no per-token fees, no telemetry, and full ownership of the stack. Models can be swapped or upgraded whenever newer open weights appear. Data never leaves the machine.
Openmonoagent.ai is being built as a terminal-native AI coding agent that runs entirely on local LLMs. It carries zero API costs and zero telemetry. The same principle applies more broadly: AI that runs as infrastructure you control rather than a rented service whose price schedule can change overnight.
The people who make memory have already shown their priorities. Supply is being allocated to AI servers and humanoid robots first. Consumer prices for the same components have already multiplied, and manufacturers are warning that the tightest period is still ahead. Waiting for relief that is not expected before 2028 means paying more for less availability. Grabbing usable GPUs and memory now, even used or older-generation parts, keeps local AI practical while the rest of the market is locked out.
