Industries whose production is offshore carry an implausibly low intermediate share, and imports inherit their thin column

[WesIngwersen]

Split out from #850. That issue's negative intermediate-input targets were a build defect in the value-added block and are fixed there. This is the part that survives the fix, is in BEA's published data rather than ours, and has a direct consequence for published emission factors.

The pattern

A set of industries carry an intermediate share of gross output far below anything a production account should show, and the share keeps falling. Semiconductors buys 16.6% of its output in inputs at 2024 and sits at the 2nd percentile of all 402 industries. Pharmaceutical preparations falls from 38.6% to 18.3% on 201 billion USD of output. Automobiles falls from 70.1% to 47.2%.

Intermediate share of gross output, BEA's published value added, after the #850 fix:

industry	2017	2024	gross output 2024, million USD
325412 pharmaceutical preparations	0.386	0.183	201,288
334413 semiconductors	0.244	0.166	72,777
325414 biological products	0.388	0.185	58,575
336111 automobiles	0.701	0.472	26,834
325413 in-vitro diagnostics	0.464	0.287	30,231

Manufacturing's median intermediate share is 0.606 at 2017 and 0.566 at 2024, so these are not the sector moving together.

The reading

These are industries where the physical transformation happens abroad under contract while the design, the intellectual property and the resulting income stay onshore. The establishment is classified in manufacturing and its income is measured there; the fabrication it pays for is not a domestic intermediate purchase, so value added rises as a share of output without any corresponding input.

Semiconductors is the clearest case. BEA's gross output for the industry is 72.8 billion USD at 2024, while a single fabless firm classified in it books revenue near double that. The revenue is not in the industry's output, which is what tells you the classification and the production account have come apart.

The 2022 Economic Census contradicts the drift where it can see it. Cost of materials against receipts for semiconductor manufacturing establishments:

	2017	2022
receipts, thousand USD	51,830,460	62,435,106
cost of materials, thousand USD	21,839,435	28,959,532
ratio	0.421	0.464

Establishments that actually fabricate are buying more per dollar of receipts, not less, over exactly the years BEA's implied share falls. The census cannot see 2023-24, so it does not speak to the AI-chip years.

Why it matters for what we publish

The model builds one requirements matrix as the sum of the domestic and imported parts, so an imported input inherits the domestic column of its own commodity. For these industries that compounds:

1. the domestic column is thin, so an imported chip carries almost no upstream requirement; and
2. the column it does carry is on the US electricity grid, not Taiwan's or Korea's.

So the emissions of offshore fabrication are understated twice, and the affected commodities read artificially clean on total emission factor — which is the axis we prioritise on. This is a known and standard EEIO simplification, not a defect, but it interacts badly with exactly the industries whose production is least domestic.

Before any of this can be conditioned

There is a concept wedge to measure first. At 2017, BEA's intermediate inputs for semiconductors are below the census's cost of materials alone on the same denominator — 0.244 against 0.340 — even though BEA's measure also includes purchased services. Own-account research and development capitalisation, redefinitions and the producer-price basis all plausibly sit in that gap. Its 2017 cross-section has to be measured before any census ratio can be carried forward, the same way the shipments-to-output wedge was measured before #724 conditioned gross output.

Options

1. Accept BEA and document it. Cheapest. Leaves the affected commodities reading clean, and leaves the model comparison work needing a standing caveat.
2. Condition the intermediate share on the census, through the existing sector-conditioner registry in ec_go_adjustment but on the intermediate-input axis rather than the gross-output axis. Needs the 2017 wedge above. Reaches 2022 only; 2023-24 would chain BEA's own movement.
3. Handle it on the import side — give imported commodities a foreign rather than domestic requirement column for the worst-affected set. Larger change, and the honest place for the emissions, since the production really is offshore.
4. Split the affected industries into a fabrication part and a design part. Breaks the 402-industry schema; not Phase 2 work.

Option 3 is where the emissions actually belong and is worth scoping even if it does not land soon. Option 2 is the cheap partial and reuses machinery we already have.

Scope note

Do not scope this as a semiconductor problem. Pharmaceuticals is the larger dollar case and the same mechanism, and the screen should be run across all 402 industries — intermediate share against the 2017 benchmark, weighted by output — rather than on a named list.

[WesIngwersen]

Correction: the "why it matters" section is written on a wrong premise

Wes points out that bedrock builds the US model only. It is injected into a global model in the ceda repo, where the import block is split further across the countries the US imports from and the emission intensities are country-specific.

So this claim in the issue is wrong:

the column it does carry is on the US electricity grid, not Taiwan's or Korea's

That reads A = Adom + Aimp with a single set of intensities as if it were the published model. It is not — it is bedrock's intermediate representation, and the global model resolves imports by source country. An imported chip does not carry US grid emissions in what reaches users. Strike that half of the argument, and with it the "understated twice" framing.

What may still stand, and the question that decides it

The other half is about the shape of the requirements column, not the intensities. Semiconductors' domestic column sums to 0.166 per dollar of output, so the Leontief inverse propagates very little upstream from it, whatever intensities are attached downstream.

Whether that survives depends on something I have not checked in the ceda repo:

• if the global model substitutes the source country's own technology — its A columns — then an imported chip is produced by a Taiwanese semiconductor recipe and the thin US column never applies to it. This concern dissolves entirely.
• if the global model splits the import block by country and applies country-specific intensities but leaves the requirements structure US-shaped, then the thin column still governs how much upstream activity gets pulled in, and the concern stands in a weaker form.

Someone with the ceda repo should settle this before option 3 is scoped. If it is the first case, option 3 should be struck from the issue too — it would be proposing something the global model already does.

None of this touches the measurement finding, which is the substance of the issue and is unaffected: the intermediate shares in the table are BEA's published data, the 2022 Economic Census moves opposite to the drift, and pharmaceutical preparations remains the larger dollar case on the same mechanism.