The Half-Life of a Number

On August 12 we wrote that being cited by answer engines means your claims get checked, and that the honest response is to check them first and publish corrections openly. On August 24 we published the storage-tier thesis with a dated call attached. On August 28, a three-run verification sweep across the freshest claims on this index came back — and it corrected us twice before it corroborated us once. In the interest of the discipline we keep advertising: corrections first.

Correction one: the paper revised itself, and we were quoting v1

Blog #28 said a 70B model's KV cache runs "roughly 2.6 GB per request," citing the TopKV topology paper. On August 7 — seventeen days before we published — the paper's v2 revision fixed a factor-of-two error in its Grouped-Query Attention sizing. The corrected figure is 1.3 GB per request for Llama-70B at 4K context.1 The bandwidth ladder we quoted (NVLink ~900 GB/s → InfiniBand ~50 GB/s → cross-datacenter TCP ~12.5 GB/s, a 72× spread) survives v2 intact; the per-request volume doesn't. We quoted the version that existed when the note entered our pipeline and didn't recheck at publication. That's the failure mode: preprints are mutable, and a citation to arXiv is a citation to a moving target. The post now carries the corrected figure with a dated note.

Worth saying plainly: halving the per-request size weakens nothing about the storage-tier argument — a 1.3 GB object you might re-read in four hours is still absurd to hold in HBM and still cheap to park on an object tier. But the argument being safe is exactly why unchecked numbers are dangerous. Nobody scrutinizes a figure that flatters the thesis.

Correction two: a viral number vs. the audited trail

Two of our posts and the NAND Flash Supply Crisis node quoted enterprise SSD prices "up ~472% year-over-year." That figure came from secondary supply-chain coverage, and it never squared with the example we printed next to it — a 30TB drive going from ~$3,062 to ~$11,000 is a 3.6× move, not a 5.7× one. The audited trail is TrendForce's contract-price surveys: NAND up ~55–60% quarter-over-quarter in Q1 2026 (record quarterly highs), a further 70–75% in Q2, compounding with late-2025 hikes to roughly 3.5–4× inside a year.2 The scale check runs the same direction: the top five NAND vendors' combined revenue rose 77% QoQ in Q2 on higher average selling prices.3 The crisis is real, the direction was right, the single number was wrong. All five surfaces that carried it now carry the correction.

The general lesson we're encoding into how this index works: numbers rot faster than narratives. Prices, version strings, and benchmark figures have a half-life measured in weeks; the architectural stories built on them decay much slower. So verification effort goes where the decay is — which is why our staleness tooling watches versions and prices first.

The part where the evidence arrived

Our August 24 call said we'd revise the storage-tier thesis "if S3-tier cache-hit latency proves too slow for production TTFT budgets." The same sweep that corrected our numbers surfaced the production data point that condition was waiting for: Cohere, running LMCache over CoreWeave AI Object Storage, reports 22–32% lower time-to-first-token and 41% higher decode throughput versus recomputation — with object-tier cache reads 1.2× faster than AWS S3 Express One Zone cold and up to 3× faster hot.4 That's a named customer, published numbers, on exactly the pattern the call describes. One case study doesn't close a call, and the source is the vendor's own blog — but the revision condition now has evidence pushing against it rather than silence. The call stands, strengthened, and the ledger records why.

What changed on the index

The rest of the sweep landed as one wave. A new pain point, AI Toolchain Supply-Chain Compromise, captures the TeamPCP campaign's full anatomy — a poisoned CI security scanner cascading through npm and Docker Hub into trojaned LiteLLM releases (CVE-2026-33634), with .pth persistence that survives package uninstall. The Model Context Protocol node now covers the 2026-07-28 stateless-core rewrite and the quantified production security cliff (53% of remote MCP servers on static keys; 8.5% on OAuth 2.1), and Guide 51 turns that into a decision path. DuckDB's 2.0 fine print (mandatory storage-format break, the Quack server protocol going stable), the Iceberg V3 read+write engine matrix, Unity Catalog's quantified openness gap, the Databricks HMS-cutoff access-mode dilemma, Schrems III moving from hypothetical to scheduled, and hardened datacenter-power numbers all landed as node enrichments. We also merged a duplicate Puffin node — dogfooding the same hygiene we preach. The index stands at 432 nodes, 51 guides.

Our read

Dated call, 2026-08-28: within two quarters, at least one more named production deployment publishes S3-tier KV-cache latency numbers, and "verify against the paper's current revision" becomes a standard citation practice for AI-infrastructure indexes the way pinned dependencies became standard for builds — because answer engines quoting stale v1 figures will force it. We revise if the Cohere-class case studies stay singular through Q1 2027, or if arXiv-revision drift turns out to be rare enough that nobody else bothers. Logged on the calls ledger.

Works cited

  1. Trend Micro — inside the LiteLLM supply-chain compromise — the TeamPCP incident anatomy behind the new pain-point node.

Footnotes

  1. Topology-Aware Data Movement for Disaggregated GPU Inference, v2 (arXiv 2607.28633) — the August 7 revision fixing the factor-of-two GQA sizing error; Llama-70B 4K = 1.3 GB/request. ↩

  2. TrendForce — 2Q26 memory contract-price forecast and Tom's Hardware coverage — NAND +70–75% QoQ in Q2 following Q1's record increases. ↩

  3. TrendForce — Q2 2026 NAND revenue — top-5 vendor revenue +77% QoQ to $68.87B on higher ASPs. ↩

  4. LMCache — Cohere on CoreWeave AI Object Storage — vendor-published production case study. ↩