# The Cost-Measurement Problem

Canonical URL: https://menfem.com/kb/inference-economics/concepts/cost-measurement-problem
Knowledge base topic: [Inference & Token-Pricing Economics](https://menfem.com/kb/inference-economics)
Frontier status: active
Tags: inference-economics, cost-methodology, gpu-utilization, dollars-per-petabyte, measurement

---

Every margin claim in this topic is a fraction. The numerator — revenue, price per token — is published, observable, and tracked to the cent by instruments like the [Token Price Index](./price-rung-persistence.md). **The denominator is measured almost nowhere — and until 2026-07-08, nowhere at all.** "Inference gross margin is 40-50%" (Evans, 2026-07-09) and "Anthropic runs 70%+" (Patel, attributed, 2026-07-10) are both statements about a cost basis that neither source defines, discloses a utilization assumption for, or states an as-of date on. This concept exists to make that hole visible, because it is upstream of every other question the topic asks.

**As of the 2026-07-23 compile the hole is *partly* closed: a measured cost-to-serve series now exists — for exactly one open model, on one SKU, on one date (§0 below).** It does not close the hole for any frontier/proprietary model, so the Evans-vs-Patel lab-margin debate stays unfalsifiable. But it converts the topic's central instrument from "price only" to "price plus one measured cost point," and it independently reproduces Patil's utilization thesis in dollars.

Two 2026 preprints attack the hole from opposite ends: one measures how badly the per-token cost unit fails, the other proposes replacing the unit entirely. A DigitalOcean engineering benchmark and an Epoch serving-capacity model then supply, for the first time, a *measured* dollar floor and the *physics* that sets it.

## 0. The measured cost floor — a POINT, not a series (added 2026-07-23)

DigitalOcean's engineering tutorial (Baranwal, 2026-07-08) does the one thing every other source in this topic cannot: it **measures cost to serve, in dollars, per million output tokens, on named hardware at a named date.** It is the topic's first genuine *cost* instrument.

**Measured configuration (as-of 2026-07-08):** a single NVIDIA H200 (141 GB) at **$3.44/GPU-hour** (DigitalOcean list), running **Llama-3.3-70B-Instruct, FP8**, on **vLLM 0.24.0**, 1,024-token in/out, swept across batch/concurrency:

| Batch / concurrency | Output tok/s (**measured**) | Cost / 1M output tokens (**derived** @ $3.44/hr) |
|---|---|---|
| 1 | 47 | **$20.32** |
| 8 | 351 | $2.73 |
| 64 | 1,656 | $0.58 |
| 128 | 2,140 | **$0.45** |
| 256 | 2,089 | $0.46 |

**The load-bearing fact:** identical hardware moves from **$20.32/M at batch=1 to $0.45/M at batch=128 — a ~44x cost swing from traffic shape / concurrency alone**, before any change of model, quant, GPU, or framework. This is an independent, dollar-anchored reproduction of Patil's $0.21–$15.25/M utilization spread — **different SKU (H200 vs H100), different rate, same mechanism.** Throughput is *measured*; the $/M figure is *derived arithmetically* from measured tok/s at the list rate.

The tutorial's own crossover: at the H200's peak throughput, **~73% sustained utilization → $0.643/M** (≈ break-even with DigitalOcean's own serverless at $0.65/M), and **40% util → $1.173/M** (~80% dearer than serverless). So for this config the own-a-GPU-vs-buy-serverless line sits near **~73% utilization** — below it, dedicated capacity loses to per-token serverless.

**The exact scope — this is why it is a point, not a series.** *One* SKU (H200), *one* model (open 70B, FP8), *one* framework (vLLM 0.24.0), *one* date (2026-07-08). The author states plainly: results will vary with any change of model, quant, GPU, framework version, or token lengths — "treat the measured numbers as a reproducible reference point for this specific configuration, not as universal performance guarantees." A single-H100 comparison at 236–335 tok/s (KV-cache-constrained, materially lower) is noted, confirming the H100→H200 generational gap alone moves cost.

**What it does and does not close.** It gives the KB a measured $/M-token *cost* floor for a 70B *open* model (~$0.45/M saturated, ~$0.64–$1.17/M at plausible utilization) — enough to sanity-check the shape of the utilization argument in dollars. It does **not** give a cost floor for any frontier/proprietary model (GPT-5.x, Claude, Gemini), which is what the Evans-vs-Patel margin debate is actually about. **That is now the topic's sharpest single gap** (see Open Questions and [frontier.md](../frontier.md)).

## 0a. The physics under the floor — Epoch's serving-capacity model (added 2026-07-23)

Epoch AI's serving-side model (Emberson & Sevilla, 2026-05-25) is the modelled counterpart to DigitalOcean's measured box: where DigitalOcean measures $/token on one GPU, Epoch models **how many tokens the world's chips can serve.** It matters here because it names the *physics that sets the cost floor* — and it is **calibrated to real experiments**, not projected in a vacuum.

- Inference is two stages with distinct bottlenecks: **prefill is compute-bound** (input tokens processed in parallel), **decode is bandwidth-bound** (output tokens generated sequentially, memory-bandwidth-limited). Stage time = max(compute time, bandwidth time).
- **Calibrated to 111 SemiAnalysis InferenceX runs** (Kimi K2.5): fitted **compute efficiency 65%, bandwidth efficiency 30%, per-step latency 5ms**. The ~30% bandwidth-efficiency reality is why decode is the expensive half.
- **Per-chip:** ~**400,000 tok/s** on a GB200 NVL72. **Global capacity:** ~500M–20B output tok/s (long context collapses it). **Capacity growth ~3.4x/yr vs demand ~10x/yr → compute crunch "likely near."**
- **Denominated in tokens/sec, NO dollars** — it does not convert capacity to $/token, so it complements rather than replaces the measured floor.

Two consequences for this page: (1) **decode being bandwidth-bound is the physical basis for Matsuoka's $/PB unit** (§2) — bandwidth-seconds are what you actually buy, and HBM scarcity gates exactly that; (2) a **3.4x-supply / 10x-demand gap is a mechanistic, cost-side reason list-price rungs can stay sticky** (scarce serving capacity rationed by price) — see [price-rung persistence](./price-rung-persistence.md).

## 1. The per-token cost figure is not a property of the hardware — it is a property of your traffic

Patil's concurrency-aware study (arXiv 2606.11690, 2026-06-10) surveyed public LLM cost calculators and found **every one treats GPU utilization as a fixed input** — user-entered, baked in as a preset, or silently assumed at 100% — never measured against the operator's actual load. That assumption is not a rounding error. It is the dominant term.

On **identical H100 hardware**, measured effective cost spans **$0.21 to $15.25 per million output tokens** (as of the June 2026 measurement window), computed at a **named input price of $6.98/GPU-hour** (Azure on-demand list, H100 NVL node $13.96/hr for 2 GPUs) — a **2.5-24x** underutilization penalty across ordinary enterprise loads of 1-10 requests/sec, rising to **36.3x near idle**. The driver is a single operator-controlled variable: offered request rate λ, which sets in-flight concurrency via Little's Law, and which **no open-source calculator exposes**. The bias has a clean closed form: a utilization-naive estimate **understates true cost by exactly 1/U**, which systematically over-sells self-hosting — worst for exactly the low-traffic workloads most likely to be evaluating it.

The finding replicates across silicon: repeating the core sweep on **A100 80GB PCIe (56 runs, $3.67/GPU-hr)** reproduces the load-driven spread at **7.0-11.4x**.

> **The $/GPU-hour input — FOUND (2026-07-23).** The prior version of this page flagged the dollar figures as inheriting an undisclosed input. The full text now resolves it: `P_GPU = $6.98/GPU-hour` (Azure on-demand list price for the H100 NVL node at $13.96/hr for 2 GPUs, co-tenant-amortized; §4.1 + Table 1), fed through `C_eff = P_GPU·10⁶ / (3600·Θ_achieved)`. The $0.21-$15.25 range is therefore anchored: it is measured throughput at Azure H100 NVL on-demand list. Carry the basis when citing the *level* (reserved/spot × 0.3-0.7×; dedicated non-co-tenant node roughly doubles the single-GPU figures). The **ratios** (2.5-24x, 36.3x, 7.0-11.4x) remain the assumption-light finding. Anchor examples: Mixtral 8x7B FP16 = $15.25/M @ λ=1 (> Claude Sonnet 4.6) → $0.87/M @ 25 rps; Qwen3-30B-A3B FP8 floor $0.21/M at saturation. *(arXiv's /pdf endpoint serves a defective 7-page render of the 26-page paper; full text read via the HTML5 render — see the raw file.)*

Patil also releases **`vllm-cost-meter`**, an open-source meter that attaches to a live vLLM server and reports real $/M-tokens against the operator's own traffic — which converts this from a critique into something a desk can actually run. See [vllm-cost-meter](../entities/vllm-cost-meter.md).

## 2. The unit itself may be wrong — $/PB instead of $/token

Matsuoka (arXiv 2607.07207, 2026-07-08) reformulates inference economics in **dollars per petabyte of bandwidth delivered ($/PB)**, arguing it is **model-agnostic for bandwidth-bound decode** in a way per-token pricing is not. The reasoning is structural: for memory-bandwidth-bound decode, what you are actually buying is bandwidth-seconds, and tokens are a downstream, model-dependent packaging of that. Two providers quoting the same $/Mtok can be selling very different quantities of the underlying resource.

This is a *rival unit*, not a refinement, and this KB records it as unresolved. Note the scope limit stated in the abstract: the model-agnosticism claim is made **for bandwidth-bound decode**. Prefill, long-context/cache-heavy, and reasoning-heavy workloads are not addressed in what has been read, and those are precisely the workloads whose share is growing.

*(The `hardware/` topic holds this preprint plus a desk close-read — see [hardware/raw/memory-scarcity-2607.07207-closeread.md](../../../hardware/raw/memory-scarcity-2607.07207-closeread.md). This page is the inference-economics lens only.)*

## 3. The demand denominator is also suspect

Matsuoka additionally lands a **measurement critique of demand**: public token trackers **overstate monetizable demand**, and **all pre-Q2-2026 projections predate the industry's shift from token maximization to token minimization**. That second clause is the sharper one — it puts an explicit expiry date on demand forecasts, and several of the demand figures this KB already holds (see [frontier.md → Evidence Located](../frontier.md)) sit on or near that line.

So the topic now has three separate measurement problems stacked: an unmeasured cost denominator (Patil), a contested cost *unit* (Matsuoka $/PB), and an inflated demand numerator (Matsuoka's tracker critique). A margin claim has to survive all three.

## The measured floor meets an observed price (added 2026-07-23)

The independent-provider matrix (Digital Applied, Q2 2026) gives the first observed list price that can be laid next to the measured floor: **Together's $0.65/M batch for a 70B open model sits right on DigitalOcean's measured ~$0.45–$0.64/M cost-to-serve** at high utilization. The cheapest observed list price is close to the measured cost floor — consistent with **thin-to-zero margin at the batch tier**, real margin only at SLA/specialty tiers. This is the first place in the KB where a measured cost and an observed price nearly touch; it is scoped to open 70B models and says nothing about frontier-model economics. See [independent inference providers](../entities/independent-inference-providers.md).

## Why this is the topic's load-bearing concept

MenFem's stated bet is that the edge sits in the intelligence + harness layer, not serving/silicon — and that **cost per task** is the discipline that proves or kills it. Cost per task is a cost figure. If the cost figure moves 36x on identical hardware depending on traffic shape, then *cost per task is not a number you can look up; it is a number you have to measure on your own load.* That is the whole argument for owning an instrument rather than citing a calculator.

## Key Claims

- **A single H200 ($3.44/GPU-hr, Llama-3.3-70B FP8, vLLM 0.24.0) measures $20.32/M output tokens at batch=1 → $0.45/M at batch=128 — a ~44x cost swing from traffic shape alone.** The topic's first MEASURED cost-to-serve series (throughput measured; $/M derived at list rate). *Evidence: moderate-strong — measured on a real box, but one SKU / one open model / one date; not a series* ([DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md))
- **Own-GPU-vs-serverless crossover for this config ≈ 73% sustained utilization ($0.643/M); 40% util → $1.173/M.** *Evidence: moderate (measured/derived)* ([DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md))
- **No measured cost floor exists for any FRONTIER / proprietary model — only for this one open 70B.** So the Evans-40-50% vs Patel-70%+ lab-margin debate remains unfalsifiable even though an open-model floor now exists. *Evidence: strong (absence is directly observable across the KB)*
- **Serving is prefill (compute-bound) + decode (bandwidth-bound); Epoch fits 65% compute / 30% bandwidth efficiency to 111 real InferenceX runs; capacity grows ~3.4x/yr vs demand ~10x/yr.** Decode's bandwidth-bound nature is the physical basis for the $/PB unit; the supply-demand gap is a cost-side reason rungs stay sticky. *Evidence: moderate — modelled but calibrated to measured runs; no $ stated* ([Epoch](../../raw/epoch-inference-compute-crunch-serving-capacity.md))
- **The cheapest observed independent-provider list price (Together $0.65/M batch, 70B) sits on the measured cost floor → thin-to-zero batch-tier margin.** *Evidence: moderate — observed list price × measured cost floor, both single-model, open 70B* ([Digital Applied](../../raw/independent-inference-provider-pricing-matrix-q2-2026.md), [DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md))
- **Every public LLM cost calculator surveyed treats GPU utilization as a fixed input, never measured against actual load.** *Evidence: moderate (preprint, single author, survey not enumerated in the abstract)* ([Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md))
- **Identical H100 hardware yields $0.21-$15.25 per million output tokens (as of 2026-06) purely from offered request rate, at a named input price of $6.98/GPU-hr (Azure H100 NVL on-demand list).** *Evidence: moderate — measured; the $/GPU-hour basis is now known, so the level is anchored, not floating; the ratio remains the stronger claim* ([Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md))
- **Underutilization penalty 2.5-24x at 1-10 rps; up to 36.3x near idle.** *Evidence: moderate (measured across 42 benchmarks)* ([Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md))
- **A utilization-naive cost estimate understates true cost by exactly 1/U.** *Evidence: strong (analytic identity, not an empirical claim)* ([Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md))
- **The spread reproduces on A100 80GB PCIe at 7.0-11.4x across 56 runs** — not an H100 artifact. *Evidence: moderate* ([Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md))
- **FP8 quantization benefits tested MoE architectures ~2.2-2.4x more than dense (+69 to +74% vs +31% peak throughput).** *Evidence: weak — author states n=3, broader validation needed; the dense-FP8 advantage inverts on silicon without native FP8 tensor cores* ([Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md))
- **$/PB of bandwidth delivered is proposed as a model-agnostic cost unit for bandwidth-bound decode.** *Evidence: moderate (preprint; asserted, scope-limited to bandwidth-bound decode)* ([Matsuoka](../../raw/memory-scarcity-open-models-ai-industry-restructuring-2026-2030.md))
- **Public token trackers overstate monetizable demand; all pre-Q2-2026 projections predate the token-maximization → token-minimization shift.** *Evidence: moderate (preprint claim; the underlying measurement critique is not reproduced in this KB)* ([Matsuoka](../../raw/memory-scarcity-open-models-ai-industry-restructuring-2026-2030.md))
- **Neither Evans' 40-50% nor Patel's 70%+ margin figure discloses a utilization assumption, a cost unit, or an as-of date.** *Evidence: strong (absence is directly observable in both sources as ingested)* ([Evans](../../raw/ben-evans-ways-to-think-about-token-pricing.md), [Patel](../../raw/dylan-patel-podcast-alpha-2026-07-10.md))

## Benchmarks & Data

*All figures carry their as-of date. Cost and price numbers decay fast — a figure without a date is not usable here.*

| Metric | Value | As of | Nature | Source |
|---|---|---|---|---|
| **Measured cost floor, 1 H200 (70B FP8), batch 128** | **$0.45 / M output tokens** | 2026-07-08 | **measured** throughput × $3.44/GPU-hr list | [DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md) |
| **Measured cost, 1 H200 (70B FP8), batch 1** | **$20.32 / M output tokens** | 2026-07-08 | **measured** throughput × $3.44/GPU-hr list | [DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md) |
| H200 concurrency cost swing (batch 1→128) | ~44x | 2026-07-08 | measured ratio | [DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md) |
| Own-GPU-vs-serverless crossover (this config) | ~73% util → $0.643/M | 2026-07-08 | measured/derived | [DigitalOcean](../../raw/measured-h200-token-serving-cost-digitalocean.md) |
| Per-chip serving throughput, GB200 NVL72 | ~400,000 tok/s | 2026-05-25 | **modelled**, calibrated to 111 InferenceX runs | [Epoch](../../raw/epoch-inference-compute-crunch-serving-capacity.md) |
| Serving capacity vs demand growth | ~3.4x/yr vs ~10x/yr | 2026-05-25 | modelled | [Epoch](../../raw/epoch-inference-compute-crunch-serving-capacity.md) |
| Cheapest independent-provider list (Together batch, 70B) | $0.65 / M output tokens | Q2 2026 | observed list | [Digital Applied](../../raw/independent-inference-provider-pricing-matrix-q2-2026.md) |
| Effective cost range, identical H100 | $0.21 – $15.25 / M output tokens | 2026-06-10 | **measured** (× $6.98/GPU-hr, Azure H100 NVL on-demand list) | [Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md) |
| Underutilization penalty, 1-10 rps | 2.5 – 24x | 2026-06-10 | measured ratio | [Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md) |
| Underutilization penalty, near idle | up to 36.3x | 2026-06-10 | measured ratio | [Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md) |
| Load-driven spread, A100 80GB PCIe | 7.0 – 11.4x (56 runs) | 2026-06-10 | measured ratio | [Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md) |
| Benchmarks in H100 sweep | 42 | 2026-06-10 | count | [Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md) |
| FP8 peak-throughput gain, MoE / dense | +69 to +74% / +31% (n=3) | 2026-06-10 | measured, low-n | [Patil](../../raw/beyond-per-token-pricing-concurrency-aware-cost-methodology.md) |
| Inference gross margin (industry) | 40-50%, training cost excluded | 2026-07-09 | **asserted, no method** | [Evans](../../raw/ben-evans-ways-to-think-about-token-pricing.md) |
| Anthropic gross margin | 70%+ | 2026-07-10 | **attributed, conflicted, no method** | [Patel](../../raw/dylan-patel-podcast-alpha-2026-07-10.md) |

## Conflicts on the Record

- **Unit conflict.** Per-token ($/Mtok) vs bandwidth-delivered ($/PB) are different units for the same economics. Matsuoka argues $/PB is model-agnostic for bandwidth-bound decode; the entire rest of this topic — Epoch's decline series, the Token Price Index, both margin claims — is denominated in tokens. **Nothing in this KB reconciles the two.** Recorded as open, not adjudicated.
- **Method conflict.** Patil's 36.3x utilization spread does not numerically contradict Evans' 40-50% or Patel's 70%+, but it undercuts both: a margin whose cost basis can move by more than an order of magnitude on identical hardware is not a single number. Both figures are best read as *unstated-utilization* estimates.
- **Demand conflict.** Matsuoka's "public token trackers overstate monetizable demand / all pre-Q2-2026 projections are obsolete" runs directly against the demand-side evidence this KB logged on 2026-07-17 (Goldman's 24x token-consumption forecast; Gartner's 5-30x agentic per-task multipliers) — see [frontier.md](../frontier.md). Preprint critique vs bank/analyst forecast; both moderate evidence; unresolved.

## Open Questions

- **[NOW THE SHARPEST GAP] What does a FRONTIER / proprietary model cost to serve?** The measured floor (DigitalOcean) is an open 70B on an H200. GPT-5.x / Claude / Gemini serving cost — the actual subject of the Evans-vs-Patel margin debate — is unmeasured anywhere in this KB. Until a frontier-model cost floor exists, both lab-margin figures stay unfalsifiable *even now that an open-model floor exists*.
- **Is the measured floor a POINT or can it become a series?** DigitalOcean gives one SKU / one open model / one framework / one date. No cost *time*-series; no reproduction on Blackwell GB200/GB300, TPU, Trainium, or custom ASIC; no long-context or reasoning-workload cost. Patil's `vllm-cost-meter` on real self-hosted traffic remains the cleanest way to build the series.
- ~~**What is the $/GPU-hour input behind Patil's dollar figures?**~~ **RESOLVED 2026-07-23: $6.98/GPU-hr** (Azure H100 NVL on-demand list, $13.96/hr node ÷ 2). The level is now citable with that basis attached.
- **Does the utilization spread hold on inference-optimised and non-NVIDIA silicon** (Blackwell-class, TPU, Trainium, custom ASIC), or is it an H100/A100-generation artifact? (A100 at $3.67/GPU-hr is confirmed; the spread narrows to 7.0-11.4x there.)
- **Is $/PB actually model-agnostic outside bandwidth-bound decode** — for prefill, long-context, and reasoning-heavy workloads, which are the growing share?
- **What would it take to measure MenFem's own cost per task honestly?** `vllm-cost-meter` on a live server against real traffic is the named instrument; a self-hosted run is the only way to convert this topic from citation to measurement.
- **How large is the token-tracker overstatement?** Matsuoka asserts direction, not magnitude, in what has been read.

## Related Concepts

- [Token Pricing & the Inference-Margin Question](./token-pricing-margin-question.md) — the umbrella question this concept destabilises: you cannot answer it without a cost denominator.
- [The Price-Decline Distribution](./price-decline-distribution.md) — measures *price*, not cost. This page is why that distinction matters.
- [The Depreciation Conveyor & Vintage Economics](./depreciation-conveyor.md) — the other half of Matsuoka: once you have a cost unit, whose cost is lower and why.
- [Price-Rung Persistence](./price-rung-persistence.md) — the sell-side price observable, unaffected by the cost-measurement problem (prices *are* published) and therefore the topic's most solid ground.
- [Cost Per Task (Cost-of-Pass)](./cost-per-task.md) — the unit cost-of-pass needs a per-attempt *cost* for; this page is why that cost is unmeasured and utilization-dependent, so cost-per-task inherits the missing denominator.
- [Independent Inference Providers](../entities/independent-inference-providers.md) — the observed list price ($0.65/M batch, 70B) that sits on the measured cost floor.

## Backlinks

*Pages that reference this concept:*
- [Token Pricing & the Inference-Margin Question](./token-pricing-margin-question.md)
- [The Price-Decline Distribution](./price-decline-distribution.md)
- [The Depreciation Conveyor & Vintage Economics](./depreciation-conveyor.md)
- [vllm-cost-meter](../entities/vllm-cost-meter.md)

## Changelog

- **2026-07-23 (compile)** — The gap this page names is now PARTLY CLOSED. Added §0 (DigitalOcean's measured H200 cost floor — $20.32/M @ batch 1 → $0.45/M @ batch 128, ~44x from traffic shape, one SKU / one open 70B / one date — the topic's first measured cost-to-serve series), §0a (Epoch's InferenceX-calibrated prefill/decode serving-capacity model — the physics that sets the floor, tokens/sec not $), and a cross-check that the cheapest independent-provider list ($0.65/M batch) sits on the floor → thin batch-tier margin. Recast the opener from "not measured by anyone" to "measured almost nowhere," and elevated the FRONTIER/proprietary cost floor to the topic's sharpest open gap (the open-model floor does NOT close the lab-margin debate). Added measured rows to the data table and cost-per-task + independent-providers to Related.
- **2026-07-23 (full-text touch-up)** — Patil 2606.11690. The $/GPU-hour input is FOUND — **$6.98/GPU-hr** (Azure H100 NVL on-demand list) — so the $0.21-$15.25 level is now anchored rather than floating; resolved the corresponding open question and updated the caveat, key claim, and data table. No structural change to the concept.
- **2026-07-22** — Created from two newly ingested preprints (Patil 2606.11690 measured-cost methodology; Matsuoka 2607.07207 $/PB unit + token-tracker demand critique). Establishes the topic's missing-denominator problem as a first-class concept and records the per-token-vs-$/PB unit conflict as open.

## Sources

- measured-h200-token-serving-cost-digitalocean
- epoch-inference-compute-crunch-serving-capacity
- beyond-per-token-pricing-concurrency-aware-cost-methodology
- memory-scarcity-open-models-ai-industry-restructuring-2026-2030
- independent-inference-provider-pricing-matrix-q2-2026
- ben-evans-ways-to-think-about-token-pricing
- dylan-patel-podcast-alpha-2026-07-10

---

Cite as: MenFem Knowledge Base — https://menfem.com/kb/inference-economics/concepts/cost-measurement-problem