At a briefing in early November, Gerry Chen of Team Group described a table he looks at daily: server memory, two prices.
Customers with long-term agreements take delivery at the contract price. Those who signed but cannot get their full allocation buy spot and pay more.
Between the two: 30 to 50 per cent.
Two price lists
Contract prices are negotiated in advance, once a quarter — bigger volumes, lower unit prices, no guarantee on quantity. Spot is what clears on the day: available now, more expensive, not always there.
One product, two lists, a gap of 30 to 50 per cent.
Team Group is a memory module maker. It does not fabricate chips. It buys DRAM from Samsung, SK Hynix and Micron, mounts it on modules and sells to channels and enterprises. The position makes it price-sensitive: when chips rise a cent, its costs rise a cent; when it cannot get chips, it has nothing to sell.
Behind the two lists sits a reallocation. The three big manufacturers have shifted lines toward high-bandwidth memory and server-grade DDR5, thinning what is left for general-purpose DRAM. Server customers sign long contracts, lock quantities and care little about unit price. Consumer buyers watch the unit price and stop buying at this level. Both bid for the same wafers.
Chen expects spot supply to tighten over three quarters, and contract prices to keep rising into 2027. His estimate for restored balance: the second half of 2028 at the earliest.
From 8.70 to 48 dollars
Down at the chip level: 16Gb DDR5 carried a contract price of USD 8.70 in October 2025. In November 2026 it is above USD 48, before packaging, thermal design, module assembly, branding and distribution.
48 divided by 8.70 is 5.52.
One year, and the same part trades at 5.5 times its old contract price.
Chen adds a second observation. Past cycles turned every few months — up fast, down fast. This one has run a year and is still rising. The lengthening matters more than the size of the move.
Why 550 per cent does not add up
Coverage of the briefing has generally reported an increase of "more than 550 per cent."
The arithmetic does not close. From 8.70 to 48 is a gain of 452 per cent — a multiple of 5.5. A price that is 5.5 times the old one and a price that has risen by 5.5 times are different claims. The second requires USD 56.55. The gap between the readings is over 8 dollars.
It sounds like bookkeeping, but the direction misleads buyers: feed "up 4.5 times" into a cost model and "up 5.5 times" into the same one, and two quotes come out.
The chip bill inside one module
A 16GB DDR5 module carries eight 16Gb dies.
At a contract price of USD 48, eight times 48 is USD 384 for chips alone. A year ago, at 8.70, it was USD 69.60.
Same module: chip cost from about 70 dollars to 384, excluding board, packaging, testing, branding and margin.
It takes two inputs: dies per module, and contract price per die. Both appear in coverage, rarely multiplied.
Long contracts do not guarantee volume
What hurts buyers is not the price. It is the quantity.
Chen is direct: a customer can sign in advance and still receive less than the agreed volume, then buy the shortfall at spot. A contract fixes price, not tonnage.
So the cost of a module splits in two: the covered portion at contract price, the shortfall at spot. A procurement team holds one contract and runs the arithmetic twice.
That is why a 30 to 50 per cent spot premium travels into finished-goods quotes. The missing portion is precisely the portion bought at the worst price.
A module maker that changed customers
Team Group's own numbers tell it better than any forecast. In the first nine months of 2026, enterprise business was 64.6 per cent of revenue. A year earlier, 24.6 per cent.
Twelve months, and the customer base turned over. Third-quarter gross margin: 49.27 per cent.
A company assembling modules at 49 per cent is unusual here. The same module sells at different prices to a server customer and to someone building a PC, and whoever can get supply sets the price.
That margin breaks into three layers: the customer spread, inventory timing (cheap stock sold at today's price), and allocation. All three recede when the shortage eases.
Notebooks fell first
Downstream, the cost lands on finished machines.
TrendForce data: notebook shipments fell 15.8 per cent in the third quarter of 2026, with a further single-digit decline expected in the fourth. Consumer DDR4 and entry-level DDR5 take the first hit — the DIY market's volume products, thin margin, first pushed back when capacity is tight.
PC makers moved before consumers. Some have trimmed memory purchasing for late 2026 and early 2027 — partly because they sell less, partly because they will not carry inventory at these prices. Memory is one of the few components still rising, and every repricing of the bill of materials loses orders. The first casualty is not the person buying a stick; it is the company assembling the machine.
What would signal relief
Three conditions must arrive together: new capacity actually comes online, demand growth for high-bandwidth memory slows, and channel inventory returns to normal.
Miss one, and the price stays.
Chen's read is that Samsung, SK Hynix and Micron are in no hurry to expand. While prices climb, nobody lets go first.
The signal to watch is the spot premium over contract. When it narrows from 30 to 50 per cent down to single digits, supply has genuinely loosened. Until then, anyone building a PC or quoting a machine works from two price lists.
One fixes the price and not the volume.
One is available today, at whatever today costs.
Disclaimer: This article is based on public information and is for reference only; it is not investment advice.
