For companies looking to expand their AI server fleets, the first obstacle they now hit isn't the price of memory—it's whether any allocation is left for them at all. On August 31, Seoul Economic Daily reported that the spot price of 36GB HBM3E, the leading high-bandwidth memory (HBM) product used in AI chips, has reached roughly $2,100 per unit—four to five times the price available through long-term contracts. The reason the same product carries two different price tags lies on the supply side. Samsung is reportedly planning to allocate 60-70% of its memory production capacity to long-term contracts running through 2031, and whether a buyer is included in those contracts is beginning to split procurement terms in two. The ability to secure AI infrastructure is increasingly determined not by how much a buyer can pay, but by whether they have a seat inside the contract framework.
An Actual Gap of $1,600-1,700: The Distance Between Spot and Contract Prices
According to Seoul Economic Daily's August 31 report, 36GB HBM3E is trading on the spot market at around 2.87 million won per unit, or roughly $2,100. The same product purchased through a Long-Term Agreement (LTA) falls in a price range of 500,000-700,000 won, which converts to roughly $370-510 using the same exchange rate applied to the spot price. TrendForce and the newspaper describe this gap as "four to five times." In absolute terms, that works out to a difference of roughly $1,600-1,700 per unit.
However, the contract price is reported as a range of 500,000-700,000 won rather than a single figure. The multiple shifts depending on whether one uses the lower or upper bound, so the "four to five times" figure should be treated as an approximation. The $370-510 figure is also a secondary estimate, derived by applying the effective exchange rate implied by the spot price (2.87 million won ÷ $2,100, or roughly 1,367 won per dollar) to the contract price—it is not a dollar figure directly reported in the original article. Even so, the underlying fact remains: the same 36GB HBM3E carries a price difference of roughly $1,600-1,700 per unit depending on the channel.
A long-term contract locks in volume in advance and caps price upside through annual negotiations over a base five-year term that can be extended. In exchange for being able to plan a return on capital investment, manufacturers give up the potential profit from any spike in market prices. What reaches the spot market is only the production volume left over after contracted quantities are subtracted—and as that remainder shrinks, prices rise not based on production cost but based on the maximum a buyer is willing to pay. The reason the two prices have diverged so sharply is that each market is being priced according to a different logic.
In a July analysis, TrendForce noted that the existence of long-term contracts caps the rise in contract prices. Buyers locked into contracts only see price increases within the bounds set by their agreements, even if the market spikes. Conversely, all of the upside gets concentrated in the spot market. The widening gap between prices inside and outside contracts is less a result of supply and demand than a consequence of how the contracts are structured.
The 16-layer HBM4, which is moving toward mass production, is reportedly trading at around $3,500 per unit on the spot market. However, this figure has not been independently corroborated beyond the Seoul Economic Daily report and should be treated as a reference point for now. If long-term contracts increasingly shift toward next-generation products, the products remaining on the spot market will skew even further toward older generations—and the gap could widen further for next-generation memory.
60-70% Allocated Through 2031, and Prepayments Already Being Collected
Samsung's memory business is reportedly planning to allocate roughly 60-70% of its production capacity to long-term contracts running through 2031. NVIDIA, Microsoft, and Google have been named as major customers, though a late-July report listed a different set of major data center operators among the top five customers—the customer list shifts depending on the report. What is confirmed is that Samsung has already signed contracts with five companies and is negotiating with five more, putting it on track to reach the 60-70% target. What matters more than who exactly is on the list is how far into the future the 60-70% of production capacity will be locked in.
Details of the contracts emerged partly through Seoul Economic Daily's July 30 report and partly through Samsung's August earnings call. The base contract term is five years, using a rolling structure where the term is extended through annual negotiations rather than being renegotiated from scratch upon expiration. At the time, Samsung had already signed with five major data center operators and was negotiating with five more. The contracts include prepayment terms, and on the earnings call, Samsung stated that it has already received one-quarter of the total prepayment amount. The company declined to disclose the amount, citing non-disclosure agreements (NDAs).
Since cash flows in ahead of delivery under contracts that include prepayment, Samsung gains early certainty on its return on investment. In exchange for giving up the upside from price spikes, the company effectively buys insurance that lets it commit to capacity expansion even during periods when future demand is hard to read. Because both sellers and buyers find rationality in this arrangement, the structure is likely to persist.
How much of the 60-70% of production capacity is accounted for by the five already-contracted companies versus the five still in negotiation has not been disclosed. Samsung stated on its earnings call that it intends to preserve sufficient production capacity for customers without multi-year contracts, meaning buyers outside the long-term contract framework are not necessarily forced into spot pricing alone. Still, whether a buyer ends up procuring through standard contracts or the spot market depends on their individual negotiating leverage—and the fact remains that 36GB HBM3E is actually trading on the spot market at around $2,100.
Samsung's semiconductor inventory assets rose 32% in the first half of 2026 compared to the end of the previous year, reaching 38.06 trillion won, with work-in-process inventory up 35% to 29.71 trillion won. SK hynix's total inventory also grew 26% to 17.99 trillion won, with work-in-process up 20% to 11.08 trillion won. Citing industry sources (Dealsite), TrendForce reported that this reflects advance production tied to long-term contracts, distinguishing it in nature from stagnant inventory built up due to insufficient demand. Under this reading, at least part of the inventory buildup represents raw materials and production volume secured in advance for demand anticipated under long-term contracts.
SK hynix is reported to have nearly sold out its 2026 full-year production of HBM, DRAM, and NAND, and Samsung's 2026 HBM production capacity is also reported to be nearly fully booked. Within 2026, only a thin sliver of volume left over from contracts remains available for new buyers to enter.
NVIDIA's Supply Commitment Grows from $119 Billion to $279 Billion
NVIDIA's long-term purchase commitments grew from $119 billion in the first quarter of fiscal year 2027 to $279 billion in the second quarter—a $160 billion increase in three months. The company has explained that this increase is primarily related to memory procurement. This supply commitment is a disclosure of future purchases that pledges to secure memory or manufacturing capacity, and a portion of it may still be subject to cancellation, delay, or renegotiation before formal orders are placed.
Even so, only a limited number of buyers can commit to supply on the scale of $279 billion. Just as Samsung has attached prepayment conditions to some of its long-term contracts, when memory makers are in a position to select their buyers, they tend to favor companies with the scale and creditworthiness to back up such contracts. The competition to secure contracted production capacity is beginning to resemble a contest of procurement power more than one of technology selection.
According to data released by the Korea International Trade Association (KITA) on August 31, export volume of AI-related DRAM (including HBM and low-power LPDDR memory) fell 13.2% from May to July 2026, from 681.79 million units to 591.74 million units. Export value, meanwhile, rose 18.5%, from $11.43 billion to $13.55 billion, and the average unit price climbed 36.6%, from $16.76 to $22.90. The number of units shipped is falling while the amount received is rising—the same shift is visible from the export statistics as well.
If volume falls 13.2% while unit price rises 36.6%, multiplying the two yields an 18.6% increase in export value—consistent with the reported 18.5% figure within rounding. These three figures are mutually consistent, confirming that Korea's memory exports have shifted from a volume-driven business to one driven by unit price. That said, this data aggregates AI-related DRAM as a whole, including LPDDR, rather than HBM alone, so HBM's specific contribution cannot be isolated. The May-to-July period represents the most recent data available, indicating that the shift from volume to unit price has continued since then.
TrendForce forecasts that server DRAM contract prices will rise 13-18% quarter-over-quarter in the third quarter of 2026. Individual company estimates will vary depending on product mix, contract type, and exchange rates, but for server manufacturers and data center operators in Japan without long-term contracts, this growth rate serves as the starting point for their next procurement negotiations. The price increases are not confined to AI-specific products—the scramble for HBM is reaching all the way into the cost calculations of general-purpose servers.
The Next Number to Watch for Buyers Outside the Contract Framework
The price gap hits hardest for those unable to secure five-year contracts with prepayment. Mid-tier server manufacturers and startups trying to build their own AI infrastructure are more likely to end up filling the gaps left by standard contracts with spot-market purchases. A price level of $2,100 per unit means they cannot offer the same service at the same price as competitors procuring at $370-510. In applications where the unit cost of computing resources directly determines profitability, this gap can decide the outcome at the product design stage.
As XenoSpectrum covered on August 30, Korea's DRAM export unit prices had already spiked due to production allocation favoring HBM. What's new here is the fact that this allocation has now been locked in by contract through the year 2031. A skewed production allocation is the kind of thing that normally reverses once demand eases—but a five-year contract does not ease. The thinness of the spot market will not be resolved by the near-term business cycle.
Companies inside the contract framework can calculate their AI infrastructure costs having secured both volume and a cap on price upside. Companies outside it must build the same plans around prices that shift every quarter. Even when launching servers of identical performance at the same time, how far into the future each company can forecast its costs differs entirely. That difference in visibility matters more for business planning than the instantaneous price gap of $1,600-1,700 per unit.
The first thing to watch is whether third-quarter server DRAM contract prices stay within TrendForce's forecast range of 13-18% or exceed it. Next is how many of the five companies Samsung was negotiating with as of July end up signing contracts, further filling out the allocation through 2031. If HBM4 mass production ramps up and production capacity beyond what's committed to long-term contracts becomes available, companies outside the contract framework may also be able to expand their AI servers at realistic prices.