The DRAM inversion

What this report concludes

Three findings, in the order they change a purchase decision.

One. The cheap listings are fiction. Australian price comparison engines still index 32GB kits at $110 to $149. We opened those product pages. They are marked End of Life or Out of Stock. Real buyable stock starts around $369 and runs past $1,000. Any plan built on a screenshot of a comparison site is built on inventory that no longer exists.

Two. The China arbitrage has closed. Twelve months ago, importing modules from Shenzhen beat Australian retail comfortably. It no longer does. Alibaba is quoting 32GB DDR5 SODIMM at $930 and Lexar is selling a 32GB kit built on domestic Chinese chips for the equivalent of $592 inside China. Chinese prices converged with global prices because Chinese supply is sold into the same shortage.

Three. One spread survives, and it is not the obvious one. CXMT is moving 32GB DDR4-3200 ECC server modules at roughly $138 against a global average of $300 to $400. That is the only genuine two-to-three-times gap left in memory, and it exists in server ECC parts, not in gaming sticks.

The cause is wafer arithmetic, not demand

This shortage was not caused by people wanting more memory. It was caused by memory being converted into a different product.

High Bandwidth Memory, the stacked DRAM that feeds AI accelerators, consumes roughly three times the wafer area per gigabyte of conventional DDR5. Every fab that shifts a line to HBM therefore removes about three units of commodity supply for each unit of HBM it gains. The margins make that trade obvious for the manufacturer and brutal for everyone downstream.

The supplier table has been rewritten in the process

The scramble for HBM has reshuffled the industry in a single year. SK hynix owns 62 percent of HBM but has paid for it in commodity share, dropping from 39 percent of total DRAM to 26 percent in twelve months. Samsung, which was behind on HBM, took the conventional DRAM demand and the price rises that came with it.

What Australia actually pays, and what it can actually buy

Two different questions, and confusing them is the expensive mistake.

Australian retail moved fast and early. In January 2026 the cheapest 32GB DDR5 kits at Centrecom, PCCG and Scorptec jumped 38 percent in a single month, from roughly $499 to $689. Globally, contract prices have run up 172 percent year on year, and TrendForce still projects a further 13 to 18 percent through Q3 2026.

The subtler problem is that the listings did not keep up with the warehouses. Price comparison engines index whatever a retailer's catalogue exposes, including products the retailer has no intention of ever stocking again. We pulled the live Australian listings for a 32GB DDR5-6000 kit and then opened each product page to check what the retailer says about stock.

The same trap exists one generation back

A Kingston 32GB DDR4-3200 module is still listed at $111.95 by a major Australian retailer. Its product page says Currently Out of Stock. Real DDR4 supply now sits behind the inversion described at the top of this report: it is end of life, rationed, and priced above the newer generation per bit.

Why shipping from China no longer wins

This was the obvious play, it worked for years, and it stopped working during 2026.

The reasoning used to be sound. Chinese module houses assemble at lower cost, sell direct, and skip two layers of distribution. In a normal market that is worth 30 to 40 percent. In this market it is worth nothing, because the input, the DRAM die itself, is priced globally and is short everywhere at once.

Current quotes tell the story without commentary. Alibaba business-to-business listings are asking $930 for a 32GB DDR5 SODIMM and $400 to $445 for a 32GB DDR5 registered server module. KingBank kits on AliExpress, the enthusiast favourite, run $400 to $780 for 32GB. Lexar has listed a 32GB DDR5 kit built on domestic Chinese chips at 3,999 yuan, about $592, inside China.

The clearest single signal is that CXMT's 64GB server DDR5 is now priced above Samsung's. When the state-backed domestic challenger charges more than the global leader, the discount thesis is finished.

There is also a quieter development that removes the last reason to treat Chinese modules as a separate category. Corsair is now shipping Vengeance DDR5 with CXMT dies inside. A Chinese die is no longer a marker of a budget brand. It is simply what is available, in everyone's product.

CXMT is the only variable that can end this early

Everything else in the supply chain is fixed until new fabs finish. One producer is still adding capacity aggressively, and it is the Chinese one.

ChangXin Memory Technologies held 3 percent of the global DRAM market a year ago and holds 7 percent today. Its capacity is projected at roughly 350,000 wafer starts per month by December 2026, against Micron at about 375,000. If those numbers hold, China becomes the world's second largest DRAM producer by raw capacity, and it is adding around 85,000 wafer starts per month every year through 2028.

Two new fabs are under construction in Shanghai and Hefei, targeting 600,000 wafers per month, with a stated goal of 17 percent of global supply by 2028. CXMT is now shipping 16Gb DDR5 at 8,000 MT/s with on-die ECC and JEDEC-compliant timings, and analysts expect it to reach yield parity on DDR5 around the end of 2026.

The catch matters as much as the promise. CXMT's cost per bit on DDR5 remains more than 30 percent above the three incumbents, and a large share of its output is absorbed domestically. That is precisely why its arrival has not translated into cheaper memory yet, and why the relief, when it comes, will arrive gradually rather than as a price crash.

The shortage has produced a counterfeit industry

This is the risk that turns a good margin into a liability, and it is new this cycle.

Fully counterfeit DDR5 modules are now circulating, not merely relabelled or downbinned parts. Documented cases include fiberglass and plastic dummy packages under printed labels, and heatspreaders concealing desktop packages sold as something else. Counterfeit modules carrying G.Skill and V-Color branding have appeared in China and are visually much closer to genuine product than earlier generations of fakes.

Visual inspection is no longer sufficient. The only reliable check is electronic: read the SPD on arrival and confirm module size, rated bandwidth, module manufacturer, DRAM die manufacturer, part number, serial, production week and year, and the supported timing table, per slot, on every unit in a sample.

The exposure nobody prices in

Australian Consumer Law statutory guarantees attach to the seller, not the overseas factory. Whoever sells the module in Australia carries the warranty obligation regardless of what the supplier promised, and cross-border RMA on a $400 part is economically dead on arrival. Budget a 2 to 4 percent failure reserve, and negotiate free spare stock into the purchase order rather than relying on returns that will never physically travel.

Import mechanics, for completeness

Memory modules classify under HS 8473.30. Duty into Australia is generally free and ChAFTA covers Chinese origin regardless, but GST at 10 percent still applies. Any consignment above AUD 1,000 requires a formal N10 import declaration lodged through the Integrated Cargo System, which means an active ABN and, in practice, GST registration so the credits are claimable. Bare memory modules are passive components rather than finished electrical apparatus, so they normally sit outside the EESS and RCM regime, but that should be confirmed in writing by a customs broker before any volume commitment rather than assumed.

Nobody credible expects relief this year

The forecasts disagree, and the disagreement is itself the finding. There is no analyst consensus that this ends soon, only a spread of opinion on how much worse it gets first.

The practical consequence for a buyer is that waiting is not a strategy. There is no near-term point at which this gets cheaper, and the base case from most of the table above is that 2027 is harder than 2026 before it is easier.

The decision, by what is actually being bought

The right answer diverges sharply depending on the part. Treating memory as one category is what produces the wrong call.

If the server DDR4 route proceeds, these are the gates

The same discipline that governs any Synthera sourcing programme applies, tightened for a market where counterfeiting is now industrialised.

1. Name the die in the purchase order. Not the brand on the heatspreader. Specify CXMT, SK hynix, Samsung or Micron, and make a substitution a breach rather than a surprise.
2. Sample before balance payment. Ten units minimum, shipped ahead of the batch. Read the SPD on every one, photograph the labels, and record the production week.
3. Test electrically, not visually. Full SPD dump plus an extended memory test per module. Fiberglass dummy packages pass a visual check and fail the first read.
4. Third-party pre-shipment inspection. An independent inspector at the factory before the container closes. This is the step that catches a mid-run die swap.
5. Pilot batch, then commit. Fifty to one hundred units, sold through, with the real failure rate measured before any bulk order is placed.
6. Reserve for warranty. Two to four percent of units held as spares locally, because cross-border RMA will not happen and the statutory obligation stays here.