DRAM Is Up 401 Percent and AWS Has Raised GPU Prices Twice. Not One Token Price on Our Tracker Has Moved.

The Korea Trade Statistics Promotion Institute publishes an unglamorous number every month: the average export value of Korean DRAM, excluding modules, per kilogram. For the first twenty days of August 2026 that number was $92,183, up 401.0 percent year over year, with export value up 504.8 percent to $9.807 billion. TrendForce ran the longer arithmetic last week. Against January 2023, when ChatGPT was three months old and nobody had built a data center around it, Korean DRAM is up roughly 12.5 times.

Several outlets converted it into the line that traveled: at that valuation, a kilogram of Korean DRAM is worth about as much as 620 grams of 24 karat gold. It is a good line. It is also a distraction, because the interesting question is not what memory costs. It is where that cost has arrived and where it has not.

I spent this morning walking the chain from the fab to the API price sheet. There are four links. Three of them have moved. The last one, the only one most people reading this actually pay, has not moved at all.

Where the Memory Bill Has Landed

That fourth row is our own data, not a report I am citing. As of the August 30 snapshot, the tracker has Claude Sonnet 5 at $2 in and $10 out, GPT-5.6 Sol at $4 and $20, Gemini 3.7 Flash at $0.75 and $3.75, DeepSeek V4 Pro at $0.435 and $0.87, and Qwen3.8-Flash at $0.16 and $0.47. Not one of those is higher than it was in the spring. Several are lower.

The Part Nobody Priced: Context Windows Are a Memory Commitment

Here is what has been bothering me. Look at the context column on the tracker. Claude Opus 5, Claude Sonnet 5, Claude Fable 5, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, Gemini 3.7 Flash, Grok 4.3, DeepSeek V4 Pro, Qwen3.8-Max, GLM-5.3, Kimi K3, Muse Spark 1.2, Laguna S 2.1. Every one of them ships a context window at or above one million tokens.

A context window is not a marketing number. During decode, the key and value tensors for every token in the window live in high bandwidth memory attached to the accelerator, and the model reads them back on every single generated token. Published work on long context serving puts the KV cache for a one million token window on an 8B class dense model in the neighborhood of 125 GB, which is more memory than most single accelerators have. Frontier models use architectural tricks to compress that, but the direction of the dependency does not change. Longer context means more bytes resident in HBM, and more bytes read per token emitted.

So the entire industry ran a million-token land grab through the same eighteen months in which memory became the scarcest input in the stack, HBM demand is projected to consume something like a quarter of total DRAM wafer output, and a bit of HBM burns roughly three times the wafer area of a bit of DDR5. Everyone shipped the memory-hungriest feature in the product category into the tightest memory market in twenty years, and nobody raised a price.

The Surcharge Is Already There. You Are Paying It in the Output Column.

The pass-through is not entirely absent. It is hiding in the ratio between input and output pricing. Prefill is compute bound and parallel. Decode is memory bandwidth bound and serial. That is why output tokens cost multiples of input tokens everywhere, and the size of the multiple is a rough read on how each provider prices its memory constraint.

The split is clean enough that I checked it twice. The US frontier labs sit at 5x or 6x. The Chinese labs sit at 2x to 3x. DeepSeek prices decode at twice prefill, which is the lowest memory surcharge of any serious model on the board.

There are two readings and I do not think anyone outside those companies can settle it. The charitable one is architecture: multi-head latent attention and aggressive KV compression genuinely shrink the cache, so DeepSeek is charging 2x because 2x is what it costs them. The uncharitable one is that a 2x ratio is a competitive weapon aimed squarely at agent workloads, which are output-heavy by construction, and the memory cost is being absorbed rather than reflected. Both can be partly true. Either way, if memory is the binding constraint of 2027, that column is where the repricing will surface first, and it will surface as the ratio widening rather than as a headline price hike.

Three Reasons the Token Price Has Not Moved

The memory was bought years ago. HBM does not trade on the spot market that the Korean export figure reflects. Hyperscalers and frontier labs buy on multi-year allocation contracts, which is exactly why 2027 capacity could be sold out in the middle of 2026. The fleet serving your requests today was provisioned at prices agreed before the spike, and it will be depreciated over three to five years. The shock is real and it lands on the next purchase order, not this one.

The floor is set by people who do not need to make money on it. Our tracker shows the cheap tier anchored by DeepSeek V4 Flash at $0.14, Qwen3.8-Flash at $0.16, GLM-5.3 Flash at $0.15, and Nemotron 3.5 Lightning at $0.08. Several of those are loss leaders for ecosystems that monetize elsewhere, and at least one is a chip vendor giving away inference to sell accelerators. A Western lab raising the price of its mid tier while that floor holds is handing over volume in a market where volume is the metric investors are underwriting.

Nobody wants to be first. Anthropic is in registration. OpenAI is financing an enormous buildout on the strength of a revenue curve. Google prices inference as a strategic input to everything else it sells. The first published price increase in this category will be read as a margin confession, and it will be read that way whether or not it is true. That is a strong incentive to take the hit on gross margin for another few quarters and let the cost show up in capex commentary instead.

Three Counterarguments

Memory is not the whole rack. The accelerator package, the networking, the power, and the cooling are all substantial lines, and a 401 percent move in one input does not multiply the whole bill by five. This is the strongest objection and it is correct. But HBM is a rising share of accelerator cost per generation, and the argument cuts both ways: if memory is a modest share of the rack, then absorbing the increase is cheap and prices should hold, which is exactly what we observe, and the story becomes about 2027 supply volume rather than 2026 unit cost.

The export figure is the wrong instrument. A per-kilogram export average is a blunt proxy. It moves when the product mix shifts toward high value HBM stacks even if no individual contract reprices, and it is denominated in weight, which is not a unit anyone buys memory in. Fair. It is a directional indicator, not a purchase price, and I am using it as one. The sold-out 2027 capacity reporting is the harder evidence.

Algorithms may eat the shock entirely. KV cache quantization to two bits or lower, latent attention, sliding window and sparse attention, cache offload to CPU memory and to SSD, and processing near memory research all attack the same constraint from the software side, and the gains published in the last eighteen months are not small. It is entirely possible that effective bytes per token of context fall faster than the price per byte rises. If that happens, this article ages badly and the industry gets away with it. I would put it at better than even that the algorithms absorb most of it and worse than even that they absorb all of it.

Our Take

The memory shortage has been covered as a consumer story, because a 485 percent move on a 64GB DDR5 kit is legible and infuriating and easy to write. That framing has buried the part that matters for anyone building on these APIs: memory is now the binding constraint on inference capacity, the 2027 supply is reportedly already claimed, and the price signal that would tell developers about it has been held flat by three separate forces that all expire.

The allocation contracts roll. The open-weight floor is set by labs that are themselves raising capital at real valuations and will eventually be asked about unit economics. And somebody will go first on pricing, at which point going second costs nothing.

What I would not do is read flat token prices as evidence that the memory crunch is overstated. Published prices are the slowest-moving instrument in this entire stack. They are downstream of contracts signed in 2024, and they are also the most competitively sensitive number any of these companies publishes. A price sheet is a statement about market position first and cost structure second. Right now those two things are pointing in opposite directions, and the sheet is telling you about the first one.

Three things I am watching:

One, whether any provider widens its output-to-input ratio without touching its headline input price. That is the quiet way to reprice memory, it does not generate a news cycle, and it would hit agent workloads hardest. We would catch it on the models tracker before anyone announced it.

Two, whether context window pricing gets unbundled. Charging a premium for requests above some token threshold, or metering cache residency separately, is the honest way to price a memory constraint, and the first lab to do it will frame it as a feature.

Three, whether anyone reduces a context window. Every announcement for two years has gone one direction. A frontier model shipping with a smaller advertised window than its predecessor would be the loudest possible signal that the memory bill finally arrived, and it would be announced as a focus on quality over length.