Memory Is Sold Out for 2026. Your Inference Floor Just Moved

What changed, and what actually matters

• The headline number is behind us. Conventional DRAM contract prices rose roughly 90-95% quarter on quarter in Q1 2026 on TrendForce data reported via Tom's Hardware. Read it as a near-doubling rather than a precise figure.
• The rate is moderating, not accelerating. TrendForce said on 9 July 2026 that server DRAM contract prices are expected to rise 13-18% quarter on quarter in 3Q26. That follows a forecast 58-63% rise for conventional DRAM in Q2. The trajectory is downward.
• But the composition changed. TrendForce says the primary source of server DRAM price increases shifts from Q3 2026 towards customers without long-term agreements, and towards incremental supply sold outside those agreements.
• Capacity, not sentiment, is the constraint. SK Hynix told its earnings call that its HBM, DRAM and NAND capacity is essentially sold out for 2026, and TrendForce says a server DRAM shortage is already anticipated for 2027.
• The practical consequence is a raised floor on owning hardware. If memory is the fastest-rising line in a server bill of materials, the fixed cost of owning inference capacity rises relative to renting it — which moves the self-host break-even in the opposite direction to what most 2025-era spreadsheets assumed.

The quarterly arc: read the right-hand column

Almost every piece of coverage this year has led with the size of the increase. That was the right lead in February. It is the wrong lead now, because the interesting variable has stopped being how much and started being whose bill.

Here is the arc as the primary industry source describes it. Note carefully which entries are reported outcomes and which are projections — the distinction has been collapsed in a lot of secondary reporting, and it matters for anything you plan against.

Two things follow from that table. The first is that anyone still describing this as an accelerating crisis is reading a chart that stopped being current in the spring. A move from a near-doubling to a 13-18% quarterly rise is a sharp deceleration by any reasonable measure, and the honest framing of the aggregate market right now is cooling, not panic.

The second is that the aggregate is no longer a useful proxy for what an individual buyer pays. That is the part worth your attention.

The two-tier memory market nobody put in a headline

TrendForce's own wording is the clearest statement of the mechanism, and it is worth quoting exactly rather than paraphrasing: several U.S.-based CSPs have entered into multi-year long-term agreements (LTAs), which restrict suppliers from raising prices for these clients.

Unpack that sentence and a market structure falls out of it. The largest buyers of server memory on the planet have contractually removed themselves from the price mechanism for the duration of their agreements. Suppliers still want the revenue that a shortage entitles them to. They cannot take it from the LTA cohort. So it comes from everywhere else — which is precisely what TrendForce describes when it says the primary source of increases from Q3 2026 shifts towards customers without LTAs and towards incremental supply sold outside those agreements.

This is a two-tier market. Tier one is price-protected by contract. Tier two carries the residual. And the composition of tier two is exactly the population this publication is written for: the independent AI team, the GPU reseller without hyperscaler volume, the university lab, the systems integrator building on-premise inference boxes, and the self-hosting startup — whether it sits in Bengaluru or Manchester.

There is a specific and uncomfortable consequence for smaller regional providers. A neocloud or hosting company without hyperscale purchasing power is itself in tier two. It buys memory in the unprotected market and either absorbs the increase into thinner margins or passes it through to you. That is the same margin structure we picked apart in our look at how Nscale's $51bn backlog compared with delivered revenue: contracted volume and delivered economics are separate things, and public-market discipline tends to resolve that tension in favour of pass-through rather than absorption.

The supply-side numbers say this is structural

A spot-price spike gets arbitraged away. A capacity constraint does not, and the evidence here points firmly at the second.

Start with the vendor statements. SK Hynix told its earnings call that its HBM, DRAM and NAND capacity is essentially sold out for 2026. That is not a forecast or an analyst inference; it is a supplier describing its own order book. Samsung raised the price of 32GB DDR5 modules to $239 from $149 — a 60% increase — in September 2025, which places the beginning of this repricing well before the 2026 headlines that most people noticed. Over the same period, DDR5 contract pricing surged about 179%, reaching about $19.50 per unit against around $7 earlier in 2025.

Gartner's number is a forecast and should be read as one: the firm has forecast DRAM prices to increase 47% across 2026 on significant undersupply. Set alongside a supplier saying it is sold out, the two are consistent — a forecast of sustained price strength is what an analyst writes when the sell side has no spare capacity to compete with itself.

Then there is where the wafers are going. HBM production for AI accelerators consumes approximately three times the wafer capacity of standard DRAM per gigabyte. That single ratio explains a great deal. Every gigabyte of HBM stacked next to an accelerator displaces roughly three gigabytes' worth of conventional DRAM manufacturing capacity, which means the AI build-out is not merely competing with the conventional memory market for supply — it is consuming that supply at a punitive exchange rate. The economics of an inference chip that reduces HBM dependency get considerably more interesting in that light, which is one reason the $312m raised by OLIX for an inference chip that skips HBM entirely reads less like a curiosity and more like a bet on exactly this constraint persisting.

The forward-looking supply figures are the ones that should shape your 2027 planning. TrendForce says a server DRAM shortage is already anticipated for 2027, and that total RDIMM bit supply is initially projected to grow only 15-20% year on year, significantly lagging projected growth in server CPU shipments. Read those two together. If the modules that populate server memory channels grow 15-20% while the processors that need populating grow faster, the shortfall is arithmetic rather than sentiment. New fabrication capacity takes years rather than quarters to come online, so the gap cannot be closed by anyone deciding to build their way out of it this year.

Demand-side, an AI server carries substantially more memory than a conventional one, which is the multiplier that turns accelerator demand into memory demand.

What this does to your inference cost floor

Here is the mechanism, stated as plainly as it can be. A substantial share of what you are paying for in an inference server is memory: the HBM on the accelerator package, and the conventional DRAM in the host system. When memory reprices upward and stays there, the capital cost of owning a unit of inference capacity rises. Rental prices also move, but they move differently: a provider amortises hardware bought at an earlier price over a multi-year life, and if it holds price-protected supply it may not need to reprice at all for the term of that protection.

So the two curves diverge, and the direction of the divergence is the thing to internalise. Owning gets more expensive first and more sharply. Renting gets more expensive later and more gradually, mediated by whatever cover your provider holds.

The build-versus-buy spreadsheet points the other way now

Most self-host break-even models written in 2024 and 2025 shared an unexamined assumption: hardware gets cheaper while you own it. Under that assumption, buying a box is a bet that improves with time — you lock in a cost, utilisation climbs, and the alternative you were comparing against keeps falling in price more slowly than your amortisation falls. That model is why so many teams concluded that self-hosting pays above some fairly modest utilisation threshold.

Invert the hardware assumption and the conclusion inverts with it. A rising memory price raises the numerator in your cost-of-ownership calculation, which raises the utilisation you need to sustain before ownership beats rental. It also raises the penalty for getting your capacity forecast wrong in the low direction, because an underused owned box is now an underused more expensive owned box. If you have a break-even model in a spreadsheet somewhere, the honest thing to do this quarter is open it and check whether the hardware cost input is a 2025 number. Our walkthrough of the self-host versus API decision sets out the structure of that calculation, and forecasting your LLM bill before launch covers the demand side of the same model.

The regional texture matters here and cuts in different directions. A team in Bengaluru building on-premise inference for a data-residency requirement is buying hardware in the unprotected tier and paying import duties on top, so the ownership premium is felt most sharply there. A team in Manchester or London weighing the same decision against AWS London region pricing is comparing an immediate capital outlay at today's memory prices against a rental rate set by a provider that may well be an LTA holder — which, counter-intuitively, makes the rental side of that comparison look better than it did a year ago, not worse. The same logic applies to an Indian team comparing on-premise against AWS Mumbai. In both markets the direction of travel is the same: the case for renting has strengthened relative to owning, for reasons that have nothing to do with the software.

None of this argues that self-hosting has stopped making sense. It argues that the threshold moved, and that the threshold is now sensitive to an input most teams do not track. It is also worth noting that the biggest reported inference cost reductions of the past year came from changing silicon rather than from owning it — the pattern we examined in the TPU migrations at Anthropic and Meta. Architecture choices still dominate procurement choices by a wide margin.

What to check in a GPU-host contract now

If you rent, the exposure you care about is not the memory price. It is whether your provider can hand you the memory price. Most hosting contracts written in 2024 and 2025 simply did not contemplate a component input rising this fast, which means the pass-through question is frequently unaddressed rather than deliberately answered — and an unaddressed question resolves in the provider's favour by default.

The corollary for anyone signing a longer commitment is that term has become more valuable to you, not just to the provider. In a market where your supplier's input cost is rising and its ability to reprice you is the open question, a locked rate for a defined period is a hedge with real value. The trade is the usual one — you give up flexibility on capacity and accelerator generation to gain certainty on price — but the exchange rate on that trade has shifted in favour of taking the certainty.

What to actually do this quarter

Five concrete actions, in the order we would take them.

1. Re-run your self-host break-even with a current memory price. Not a 2025 quote, not a vendor list price from last year. Get a real quote this month and put it in the model. If the answer flips, better to learn that before you buy.
2. Ask your host the pass-through question in writing. Whether component cost increases can be passed through, capped or uncapped, and with what notice. Get it in email if it is not in the contract.
3. Find out whether your provider is protected upstream. Not every provider will answer, and the refusal is itself informative. A provider with contractual price protection has an incentive to tell you so, because it is a differentiator.
4. Bring forward hardware purchases you were already committed to, and only those. If a box is definitely being bought in the next two quarters and the supply picture for 2027 is a projected shortage, the timing argument favours earlier. This is not a reason to buy hardware you were not going to buy.
5. Optimise the memory footprint you actually need. Quantisation, KV-cache management, batch sizing and model selection all change how much memory a given throughput requires. Every gigabyte you do not need is a gigabyte you are not buying in a two-tier market. This is the cheapest lever available and the one most teams under-use.

The meta-point is about how to read this story at all. The aggregate memory market is cooling from an extraordinary spike, and reporting that says so is accurate. Your own cost exposure may be worsening at the same time, because you sit in the tier that TrendForce explicitly identifies as absorbing the residual increases. Holding both of those in mind simultaneously is the whole discipline here. The number in the headline was never about you.

Primary source for the quarterly figures and the LTA composition shift is TrendForce, whose press releases are published at trendforce.com. The Q1 2026 conventional DRAM figure reached wider circulation through Tom's Hardware, citing TrendForce. The 2026 price forecast is Gartner's, published through the Gartner newsroom.