NLT143 AI Economy Index: editorial brief
NLT143 RESEARCH. Research cutoff 18 November 2026. Product name: NLT143 AI Economy Index, Report 001 (an economic intelligence report series; "Index" names the series, not a single composite score). Larger project: The American AI Dividend.
The question
Where is AI creating economic value in America, who receives it, what does it cost, and what would make the gains broadly shared?
The thesis under test
"AI can become a broad-based U.S. growth factor when improvements in capability and cost translate into productive adoption, better services, business expansion, good jobs, higher real incomes, and consumer benefits."
The site tests each link of that chain separately and rates the evidence at each link. The story is designed to remain useful if the evidence disagrees with the thesis.
Corrected starting claim
Not: "Anthropic measures the jobs of AI users, and software leads the economy."
Instead: "In Anthropic's sample of U.S. Claude.ai conversations (May 2026), 21.1% were classified to computer and mathematical tasks, the largest of 22 occupational task groups. This describes activity on one platform, not users' occupations or the economic effects of all AI."
The widely shared 21.7% figure was not reproduced from the 2026-06-26 release (U.S. April 21.9%, U.S. May 21.1%, global May 23.8%).
What the evidence supports today (narrowest defensible conclusions)
- Capability and cost: strong. Training compute keeps rising (Epoch); private data-center construction ran at a $96.5B annual rate in Sep 2026 (Census C30, preliminary), up from $1.6B in Jan 2014; Ramp customers' blended price per million tokens fell 14.8% in a month.
- Adoption: strong, but definition-bound. Census BTOS (representative): 24.2% of employer businesses used AI in the prior two weeks (collection from 19 Oct 2026, ±1.2 pts at 95%). Ramp (vendor sample): 56.8% of its customers paid for AI in Oct 2026. Measured Census use roughly doubled when the question was broadened on 17 Nov 2025 (10.0% to 17.3%), so the level is partly a measurement choice.
- Diffusion is uneven. Census: 22.8% of firms with 1 to 4 employees vs 44.1% with 250+. Sector range from 6.0% (agriculture) to 45.0% (information). Accommodation and food (12.1%) and transportation (12.6%) trail.
- Workflow: moderate. Platform telemetry shows what tasks people bring to AI; U.S. Claude.ai conversations split roughly evenly between automation-style (49.6%) and augmentation-style (50.4%) patterns.
- Output: weak. One firm-level study (Ramp/Revelio) finds high-intensity adopters grew headcount 10.2% relative to later adopters, concentrated in Information; low-intensity adopters show nothing. No sector-level output change is attributable to AI with the evidence assembled.
- Welfare: not established. No assembled evidence ties AI to real wages, consumer prices or service quality at scale.
- Jobs. A net national count of AI-created jobs is not currently established by the evidence assembled. The data-center operations industry (NAICS 518) has edged down since early 2023 even as construction spending soared; the jobs are in construction and upstream manufacturing (SIEPR models about 304,000 supported at the 2026 run rate).
Tone
Hopeful, exacting, respectful of workers and host communities. Beautiful because clear. No robots, no glow.
Permitted wording for headline claims
- C01 (verified). In Anthropic's sample of U.S. Claude.ai conversations (May 2026), 21.1% were classified to computer and mathematical tasks, the largest of 22 groups. This describes activity on one platform, not users' occupations or the effects of all AI.
- C02 (not reproduced). Use the release value with its period (21.1%, U.S., May 2026) rather than 21.7%.
- C03 (verified). Among businesses that use Ramp, 56.8% paid for an AI product in October 2026. This is not a national adoption rate.
- C04 (verified). When Census broadened the AI question in November 2025, measured use jumped from 10.0% to 17.3%. The two series are shown separately and never joined.
- C05 (verified with caveats). In a linked sample of Ramp customers, the heaviest AI spenders grew headcount about 10% relative to later adopters over two years, driven by Information firms. Lighter adopters showed no change. This cannot be scaled to national employment.
- C06 (verified). SIEPR models about 304,000 U.S. manufacturing jobs supported by data-center construction at 2026 investment levels (range ~150,000 to 450,000+ if investment is 50% lower or higher). It is a modeled estimate, not a count of new hires.
- C07 (verified with caveats). Employment in the industry Census uses for data centers rose from 306,000 (2016) to 501,000 (2023). That industry includes hosting and data processing, and it has edged down since 2023 on BLS monthly data.
- C08 (verified). In late October 2026, 87.3% of U.S. businesses reported no change in headcount from AI over six months; 6.6% reported an increase and 2.5% a decrease (self-reported, one survey cycle).
- C09 (verified). Penn Wharton's model (Sept 2025) projects a 1.5% higher U.S. GDP level by 2035. A projection, not an observation.
- C10 (verified). Acemoglu (2024) bounds the 10-year TFP gain at 0.66% (under 0.53% after adjusting for task difficulty).
- C11 (unverified). McKinsey (2023) estimated global annual value potential of $2.6 to $4.4 trillion across modeled use cases (not re-verified this pass; never shown as U.S. GDP).
- C12 (reported, not independently verified). No Memphis/Shelby County incentive agreement for Colossus was identified in the records reviewed so far; the Chamber has said there were none. Tax records have not been obtained.
- C13 (reported, not independently verified). xAI said Colossus 2 would add 200 to 400 jobs (company claim, not verified).
- C14 (verified). Private data-center construction ran at a $96.5 billion annual rate in September 2026 (preliminary), up from $1.6 billion in January 2014.
Open gaps
- OEWS occupation history 2016-2024 (radiologists, nurses, teachers trends). bls.gov flat files return HTTP 403 to automated clients without a contact address
- Full SEC issuer universe (Layer A). sec.gov company_tickers_exchange.json returned 403
- AI-specific revenue for any issuer. Not tagged in XBRL
- Memphis tax, PILOT, MLGW contract, permit PDFs. Not retrieved
- OpenAI enterprise and economic reports. HTTP 403 to automated retrieval
- McKinsey figure re-verification. Site refused automated retrieval
- Ramp token data by model and token type (input/output/cached). Only provider-level shares published; cached-token chart values not in payload
- Ramp reuse terms. Not reviewed
- Sector outcome studies (health, transport, retail) with measured outcomes. Not yet assembled
- EIA electricity sales and prices near data centers. EIA API requires a key
- Net national AI-created jobs. No credible economy-wide identification