15 September 2026
Heard In AI

DeepSeek's memory diet challenges what a data center needs to buy

On Moonshots #288, a 4 a.m. chart about DeepSeek's new V4.1-Flash model sent the panel from cache statistics to the shopping list for an AI data center. DeepSeek says the model's lookup memory needs a quarter of the expensive high-bandwidth memory and an eighth of the SSD cache storage of its previous generation. The panel's argument was about what that does to a buildout in which, by one panelist's estimate, 40% of American capital spending goes to that one component.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

Peter Diamandis opened the segment by admitting he had been awake at 4 a.m. texting with Emad Mostaque, who had sent him two charts. The first was about a model released the day before — DeepSeek V4.1-Flash — and about memory. The second plotted quality against cost. "Every day is a freaking new model now, right?" Mostaque said, before explaining what this one had done differently.

What followed was not really a model review. It was an argument about what an AI data center should be built out of.

The expensive memory, and the trick for avoiding it

When a model answers you, it does not re-read your entire prompt for every new word it produces. It keeps a running store of what it has already processed, called the KV cache — Mostaque described it as "the lookup memory, kind of prompt to prompt." That cache has to be reachable fast, which normally means it lives in high-bandwidth memory, or HBM: stacks of memory chips mounted right against the processor, and one of the costliest parts of an AI server.

In its announcement on 10 September, DeepSeek says V4.1-Flash's KV cache requires one-quarter the HBM and one-eighth the SSD cache storage of the previous generation. Those are cache-resource comparisons — the memory used to hold context while the model works, not a claim about the size of the model itself. They matter most for workloads that chew through the same long context again and again, which is what agent-style tools do all day.

On air, Mostaque traced the drop in cruder terms, quoting a fall from "48,000" in the DeepSeek V3 generation "that kind of freaked out everyone" to 35,000 and then "890," without units. The mechanism he described was displacement: the model is "basically routing around the HBM memory that is so ridiculously expensive," pushing work onto SSDs as a lookup table and onto ordinary DDR memory, the cheaper kind used as main system memory. His conclusion: the market and DeepSeek are "finding a way to make memory not an issue to scaling this up." The host's summary was a rule: "Every constraint gets worked around."

Two ways to solve the same problem

The question then went to Alex, another voice on the panel, who framed the release as one side of a contest between two engineering cultures.

One he called "the Western HBM force." Because frontier models are large and densely packed with weights — the numbers a model consults to produce each output — those weights want to sit as close as physically possible to the circuits doing the multiplication. That pressure, he said, pushes designers toward folding memory in three dimensions directly on top of the transistors doing the work, away from the classical computer layout in which memory and processor are separate parts joined by a wire.

The other, "the Eastern School," does not have access to that 3D physical architecture, he said, because of sanctions and other reasons. So its labs attack the same bottleneck in software: making models sparser, tying weights together, cutting the overall weight burden and parameter count "such that they don't need 3D HBM."

He did not expect the split to last. "I don't think this, call it an HBM or a post-Van-Noyman overhang, is sustainable," he said, predicting that Western frontier labs will "adopt every single innovation that's worth adopting from the Chinese labs that are deprived of 3D architectures." For the moment, though, he described a real tension between buying your way past the memory wall and engineering around it.

The part that alarmed the panel

The reaction was about money. "This completely changes what a data center should be built out of," one panelist said, adding that it also changes which fabs and which chip generations are worth investing in, and even what mass you would launch into space. The whole value chain, he said, deserved its own segment.

The figure behind the alarm was Mostaque's: on his estimate, 40% of the current American capital expenditure buildout is HBM memory, against a trillion-dollar total, and the new architecture implies a fourfold cut in the requirement for it. That is his arithmetic applied to DeepSeek's cache comparison, not a number DeepSeek published.

The second point was where that money actually goes. "When you buy an Nvidia rack and put it in a data center, you think you're buying Nvidia, you're not. You're mostly buying Hynix, HBM," one panelist said — SK Hynix being a supplier of the stacked memory. He argued the picture is worse than the sticker price suggests, because the NVIDIA chips are underutilized in ordinary use: even when half the cost feels like GPU, "it's more like 5 to 10 percent of the bottleneck is actual Nvidia."

From that came the entrepreneur's lesson: "look for the restriction, look for the scarcity." Constraints imposed on China, in that telling, force its engineers to invent around them — and the useful instinct is not thinking outside the box but forcing yourself into a very small one.

Another panelist drew a narrower technical lesson. The multilayer neural networks underneath all of this go back thirty or forty years and are fairly settled, he said, but the transformer's attention mechanism dates only from 2017 and remains "new and very raw technology." A tenfold or hundredfold improvement in that part of the stack should surprise nobody: "a lot of people in San Francisco treat transformer attention like a religion," he said, as something not to be touched.

The second chart: design work, cheaply

The other chart plotted quality against cost on an open benchmark for design work — Mostaque's gloss: "how do you make pretty websites?" He described V4.1-Flash as a 500-gigabyte model, small enough to sit on a desktop Mac, which on his reading of the benchmarks beats leading closed models while being twenty times faster and twenty times cheaper. On design tasks, he said, it is now the best in the world apart from Astra.

He also flagged the shape of the thing. DeepSeek's announcement describes a 552-billion-parameter mixture-of-experts model — one that activates only a fraction of its parameters for any given token — with an asymmetric causal encoder–decoder: eight billion active parameters for reading input, sixteen billion for generating output. Reading and writing, in other words, are no longer served by the same machine. "This will be a nightmare for the western inference providers to actually do," Mostaque said. He summarized DeepSeek's posture as: "your margin is my opportunity."

That led to the sharpest complaint of the segment, aimed at Anthropic. One panelist called it "scandalous" to see Anthropic's model sitting below something so small and cheap on the chart, and said its visual reasoning is "weak and anemic" next to Astra and the Chinese open-weight models. Benchmarks focused on front-end and interface design, he argued, should be a wake-up call for the company to take computer vision seriously. DeepSeek's own release adds native visual understanding to the model.

What it cost, and what it borrowed

Mostaque finished a calculation on air: "This model costs $10 million to train." The immediate question from the panel was how much of the training data came from distillation — learning from the reasoning traces produced by Western models. Mostaque said the ratio holds fairly steady: roughly 10 to 15 times cheaper to be the fast follower if you are using the frontier's reasoning traces.

He then qualified his own point. DeepSeek's write-up, he said, credits simulated data environments rather than borrowed traces, and the company attributes the gains to new pretraining methods and larger-scale reinforcement-learning post-training. His comparison stood anyway: against what he put at a billion-dollar training run for Astra, this is a model slightly behind at ten million "plus maybe a bit of data borrowing."

The message to large companies was that the window keeps reopening. Each time a Chinese open-weight release closes the gap, a Western release reclaims the lead a few weeks later — and then it closes again. "Every time there's a three-week lead, something else comes out that gives you the opportunity to get in the game and compete again," one panelist said, applying the point to big corporations and sovereign countries alike.

The practical consequence arrived on a schedule. DeepSeek said requests to its older V4-Pro model would be routed to Flash, at Flash prices, from 14 September until a V4.1-Pro arrives — with off-peak rates at half the peak price. Whatever the charts eventually mean for HBM orders, the cheaper model is already what the older one's customers get.

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

Graylin: cheaper AI could undermine the debt funding data centers

Alvin Graylin argues that AI can become more useful while earning less for the companies financing its infrastructure. His warning centers on cheaper models and local computing weakening cloud revenues, just as NVIDIA proposes financing platforms intended to mobilize more than $500 billion of outside capital.

5 min read

Why renting a three-year-old NVIDIA chip got 22% more expensive in a month

On Moonshots, the panel picked apart a rental index showing H100 prices rising 22% in a single month to $3.28 per GPU-hour. Dave called it a reversal of a lifetime of chip depreciation; Emad Mostaque explained why better models make the same old Hopper worth more; and the warning for companies was that the compute they assume will be there later is already sold out.

5 min read

Astra tops one leaderboard and trails another — the panel reads it as a computer-use model

OpenAI's GPT-6 Astra nearly saturates the interactive ARC-AGI-3 benchmark and leads Epoch AI's composite capability index, yet sits third on Artificial Analysis's suite, behind Claude Fable 5.1 and Muse Spark. On Moonshots EP #286, the panel works through what each ruler measures — and argues that Astra's real target was doing tasks with fewer output tokens, so a model can drive a desktop at conversational speed.

9 min read

Memory, not GPUs: the shortage that could redesign AI hardware

On Moonshots, Peter Diamandis reported back from meetings with SK hynix and Solidigm leadership with a claim that memory, not compute, now limits AI. The panel argued that a changed workload and a supplier industry scarred by past busts are pushing prices up faster than factories can respond — and that the fix may be new chip designs, including etching model weights into silicon, rather than simply paying more.

8 min read