How much has it actually gone up?

Around 485% over twelve months. A 128GB DDR5-6400 kit reached $3,399 in the US — about ten times the lowest price ever tracked for that capacity. 64GB kits that sold for under $200 last summer are now over $1,000.

Why is this happening?

AI datacentre buildouts are absorbing DRAM production. Hyperscalers are outbidding consumer and prosumer buyers for the same manufacturing capacity, and the shortage has begun to spread into DDR4, SSDs and hard drives.

When does it get better?

Not soon. Meaningful new manufacturing capacity is not expected before late 2027 or 2028, and prices are widely expected to stay elevated through 2027.

Does this change whether I should run models locally?

It changes the arithmetic, which is the honest way to put it. Local inference was often justified on a payback period against API spend. A 5x increase in the hardware side of that calculation pushes the break-even out considerably, especially for intermittent workloads.

InfrastructureAugust 18, 2026·2 min read·By the Automation Squad Research

DDR5 Is Up 485% in a Year

This is not a spike you wait out over a quarter. If your plan involved buying memory to run models locally, the plan needs re-costing.

By Robert MacKelfresh

Founder, Automation Squad · August 18, 2026

The short answer

DDR5 memory prices have risen roughly 485% in twelve months. A 128GB DDR5-6400 kit reached $3,399 in the United States, around ten times the lowest tracked price, and mainstream 64GB kits now exceed $1,000 against under $200 a year ago. Hyperscale AI datacentre buildouts are outbidding other buyers for DRAM capacity, and analysts do not expect significant new supply before late 2027 or 2028.

Decision table

Buy, wait, or rent — priced against what memory actually costs now

The instinct with any price spike is to wait for it to pass. The supply timeline is the argument against that instinct here, so the decision is worth making deliberately rather than by deferral.

Price the API alternative before you price the RAM
At $3,399 for 128GB, the break-even against hosted inference is much further out than it was a year ago. Do that division before buying, not after. For a lot of intermittent workloads the honest answer is now the API.
Check whether memory is actually your constraint
Plenty of local-model frustration is bandwidth or VRAM, not system RAM. Buying the wrong component at a 5x premium is an expensive way to not fix a problem.
Separate committed needs from speculative ones
Memory for work you are doing this quarter is a cost. Memory for a project you might start is an option you are buying at the worst price in two decades.
If you buy, buy the configuration you will still want in 2028
Given the supply outlook, the upgrade you defer may be more expensive than the one you make now. That is an unusual thing to be able to say about computer hardware, and it is worth taking seriously.

Your situation	The call
You need it for work now	Buy. Waiting has a known cost and an unknown end date.
Speccing a build for 'someday local models'	Wait. You are paying a 5x premium for optionality.
Inference you could run on an API instead	Rent. Run the monthly API cost against $3,399 honestly.
Existing machine, considering an upgrade	Check whether you are actually memory-bound first.
Buying for a team	Buy what is committed, defer what is speculative.
Hoping for relief in Q4	Do not plan on it. Capacity is a 2027–2028 story.

Memory prices have climbed roughly 485% in twelve months. A 128GB DDR5-6400 kit now sells for $3,399 in the United States, and 64GB kits that went for under $200 last summer are now over $1,000. TrendForce reported on August 17 that German DDR5 prices are approaching five times year-on-year, with China seeing double-digit week-on-week increases.

The facts: the driver is AI datacentre demand absorbing DRAM manufacturing capacity, with hyperscalers outbidding everyone else for the same supply. 128GB kits are around ten times the lowest price ever tracked for that capacity. The shortage has started to spread beyond DDR5 into DDR4, SSDs and hard drives. Analysts do not expect significant new capacity before late 2027 or 2028, and expect prices to stay elevated through 2027.

Automation Squad's take: this quietly rewrites one of the standard arguments for running models locally. That argument was usually a payback calculation — buy the hardware once, stop paying per token, break even in months. Multiply the hardware side by five and the break-even moves a long way out, far enough that for intermittent or bursty workloads the hosted API is now the cheaper answer rather than the lazy one. The second-order effect is more interesting than the first: AI demand has made the hardware that would let you opt out of AI services more expensive, which is a fairly efficient moat that nobody had to design.

Run this now: if you have a build or an upgrade in a plan somewhere, re-cost it today rather than discovering the new numbers at checkout. Do the division against hosted inference honestly, including the months you would not have used the machine. And check that memory is genuinely your bottleneck before spending at these prices — a lot of local-model pain is VRAM or bandwidth, and buying the wrong component at a 5x premium fixes nothing.

Questions people are asking

How much has it actually gone up?: Around 485% over twelve months. A 128GB DDR5-6400 kit reached $3,399 in the US — about ten times the lowest price ever tracked for that capacity. 64GB kits that sold for under $200 last summer are now over $1,000.
Why is this happening?: AI datacentre buildouts are absorbing DRAM production. Hyperscalers are outbidding consumer and prosumer buyers for the same manufacturing capacity, and the shortage has begun to spread into DDR4, SSDs and hard drives.
When does it get better?: Not soon. Meaningful new manufacturing capacity is not expected before late 2027 or 2028, and prices are widely expected to stay elevated through 2027.
Does this change whether I should run models locally?: It changes the arithmetic, which is the honest way to put it. Local inference was often justified on a payback period against API spend. A 5x increase in the hardware side of that calculation pushes the break-even out considerably, especially for intermittent workloads.