All essays
RSS
FIELD NOTES · INFERENCE SYSTEMS ISSUE NO. 014
Verified vs. Alleged

The NVMe trick is real. The 85% margin is doing something else entirely.

A leaked investor call, a viral claim about DeepSeek's profitability, and a genuinely public storage architecture that's been sitting on GitHub since February. Untangling which parts of this story are engineering and which are arithmetic.

Reading time — 9 min Subject — KV-cache offload, 3FS, inference unit economics Confidence — mixed, see ledger
TIME TO FIRST TOKEN — 128K PROMPT, HIGH PREFIX REUSE● measured
Cold prefill
13.0s
Disk-cached
0.5s
Source: DeepSeek API — "Context Caching on Disk" release notes, Aug 2024. A ~26× reduction in first-token latency when the incoming prompt shares a cached prefix.
01 — The part that's public

DeepSeek really does cache KV state on SSD, and the repo is open

The "NVMe rumor" circulating after the leaked investor call isn't a rumor. DeepSeek has documented disk-backed prefix caching in its own API changelog since mid-2024, and shipped the storage layer underneath it as an open-source project, 3FS, in February 2025.

The mechanism is a distinction most explanations of LLM serving skip over. During live generation, the key/value state for the tokens actively being decoded has to sit in fast accelerator memory — there's no way around that; attention math happens on-chip. But a huge share of real-world traffic isn't novel tokens at all. It's the same system prompt, the same uploaded document, the same tool-scaffolding preamble, arriving over and over with only the last few hundred tokens changing.

For that reusable, "colder" portion of context, DeepSeek stores the KV blocks on large pools of NVMe storage rather than recomputing them, and pulls them back in when a new request shares an existing prefix. That's the entire trick: skip prefill, not skip compute.

request
├─ cached prefix → load compressed KV from SSD → mostly skip prefill // ~500ms
└─ new tokens → compute normally on GPU/NPU → generate output // full prefill

Two things make this practical rather than theoretical. First, DeepSeek's Multi-head Latent Attention (MLA) and its newer compressed-attention variants shrink each KV entry substantially compared to standard multi-head attention — smaller entries mean the SSD round-trip and network transfer are cheap enough to beat recomputation. Second, real traffic has absurdly high prefix overlap: multi-turn conversations re-send their own history every turn, RAG pipelines re-send the same retrieved documents, and agent frameworks re-send hundreds or thousands of tokens of scaffolding on every step.

02 — The hardware underneath it

3FS: the file system doing the fetching

3FS is a distributed file system purpose-built for the access pattern AI workloads actually produce — closer to the random-read demands of high-frequency trading (where DeepSeek's team has roots) than to conventional storage.

Public benchmark cluster — 180 storage nodes

16 × 14 TiB NVMe SSDs per node → 2,880 drives total
6.6TiB/s
Aggregate read, 180 nodes, under training load
40GiB/s
Peak KV-cache read throughput, single client node
3.66TiB/min
GraySort throughput on a 25-node Smallpond cluster

Figures as published in DeepSeek's 3FS repository and Open Source Week materials, Feb 2025. Not independently re-benchmarked here.

03 — What it doesn't explain

None of this, by itself, produces an 85% margin

Disk-backed KV caching lowers latency and cuts the compute cost of repeat traffic. It's a real economic advantage. But it's being asked to carry a number it never claimed to produce — and that number appears to have been mistranslated or over-interpreted on its way through the discourse.

What Liang Wenfeng reportedly said

According to summaries of the leaked May 2026 investor call, Liang framed DeepSeek's pricing philosophy around a payback target, not a margin: API pricing is set to recover the cost of the underlying compute hardware within roughly ten months. He described this explicitly as "reasonable profit," not profit maximization.

A ten-month payback target is a capital-recovery statement. Converting it into a gross-margin percentage requires an assumption the original claim doesn't supply: how long that hardware keeps earning after it's paid for itself. Here's what falls out under a few different assumptions:

AssumptionUseful lifeLife ÷ paybackImplied hardware-only margin
Conservative accelerator lifespan 3 years (36 mo.) 3.6× ≈ 72%
Commonly cited GPU lifespan 5 years (60 mo.) 6.0× ≈ 83%
"Six times profit," not revenue — (different claim) 7.0× ≈ 85.7%

The five-year case is the one that lines up closest with the "85%" figure now circulating — but notice what had to happen to get there. A payback period became a lifetime revenue multiple (life ÷ payback), and that multiple became a margin by assuming the hardware is the only cost in the denominator. None of electricity, networking, storage, staffing, R&D, or underutilization are in that calculation. And "six times the hardware cost in revenue" is not the same statement as "six times the hardware cost in profit" — the two produce different percentages, and it looks like the two got conflated somewhere between the original Mandarin remarks and the version now being repeated in English.

The honest version of this number

"DeepSeek can recoup its hardware spend in under a year, which implies very high hardware-only gross margins under standard depreciation assumptions" is a defensible reading of the leaked claim. "DeepSeek runs an 85% net margin" is not — it's a different, much stronger statement that the underlying numbers don't actually support.

04 — Chain of custody

Everything upstream of the technical claims is unaudited

It's worth being precise about what kind of document this is. The Liang Wenfeng call was a leaked, AI-translated transcript of a private investor meeting, not a filing, a press release, or an on-record interview. That doesn't mean it's false — but it changes how each individual claim inside it should be weighted.

Public / verifiable
Disk-backed KV/prefix caching, 3FS, MLA-style compressed attention, the TileLang-Ascend compiler backend. All documented in DeepSeek's own API release notes and open-source repositories, independently reproducible.
Derived / interpretive
The 72–86% "margin" figures. Arithmetically consistent with the leaked ten-month payback claim under stated assumptions — but the assumptions (useful life, cost scope) aren't Liang's; they're reconstructions.
Attributed, unverified
Specific revenue figures, the ~20,000 Hopper-equivalent compute estimate, and the claim of running two Huawei-chip clusters. These exist only inside an unaudited leaked transcript and have not been independently confirmed by DeepSeek, Huawei, or any third party.

The technical claims and the financial claims in this story don't carry the same evidentiary weight, even though they arrived in the same news cycle. It's reasonable to treat the storage architecture as fact and the margin figure as a plausible-but-unaudited estimate at the same time.

05 — The actual mystery

It's not one secret. It's five ordinary things, compounding

If there's a real puzzle here, it isn't "how does the NVMe trick work" — that part is published. It's how DeepSeek's inference costs got low enough for a ten-month payback claim to be plausible at all. That answer is almost certainly not a single undisclosed technique, but the multiplicative effect of several public ones running together.

01
Compressed attention (MLA and successors)

Shrinks each KV entry, which is what makes SSD-tier caching fast enough to bother with in the first place.

02
High prefix-cache hit rates

System prompts, documents, and agent scaffolding repeat constantly in production traffic — most tokens a server sees were already seen recently.

03
Disaggregated, NVMe-backed storage (3FS)

Moves the KV cache off scarce, expensive HBM and onto abundant, cheap SSD capacity without a fatal latency penalty.

04
Mixture-of-Experts routing

Activates a fraction of total parameters per token, cutting the compute (and therefore the accelerator-hours) each request actually consumes.

05
Low-precision inference, custom kernels, high utilization

None individually exotic — but stacked on top of the first four, they compound rather than add.

Every serious inference provider holds back some combination of scheduling heuristics, cache-eviction policy, and kernel-level tuning — that's normal competitive practice, not a hidden breakthrough. What's unusual about DeepSeek is less a secret ingredient than the willingness to publish most of the recipe while the margin conversation happening around it, in leaked-transcript form, runs several steps ahead of what's actually confirmed.

Read

The NVMe comment is directionally right, and not remotely secret — DeepSeek published the storage system. The 85% margin figure is a plausible arithmetic extension of a real claim (ten-month hardware payback) stretched past what that claim actually said, riding on a transcript nobody has audited. Treat the engineering as documented fact. Treat the economics as an informed guess wearing a precise-sounding number.