AI Capital Cycle ⑦ — The One Number

July 12, 1987. Robert Solow wrote one line in a New York Times Book Review. "You can see the computer age everywhere but in the productivity statistics." That single line was cited as the "Solow paradox" for thirty years.

Solow's paradox pointed to the fact that 1970s-80s US firms invested astronomically in IT, yet macro productivity statistics actually slowed. The productivity boost didn't appear in macro data until the mid-1990s (1995-2004). Brynjolfsson and Hitt's follow-up: "$1 of IT capex correlated with $12 of market value" — meaning the other $11 had to be intangible complementary investment (process redesign, training, organizational change) for IT to translate into productivity. *Productivity arrives, on average, 10 years after IT investment*.

This was the enduring mystery of the 1999 dotcom cycle. Everyone knew the internet would change industries. But *when* that change would show up in macro productivity — could only be known ex-post.

2026 is different. *That mystery can now be measured quarterly*. The Anthropic Economic Index is that instrument.

February 10, 2025. Anthropic published the first *Economic Index* report. The Clio (Claude Insights and Observations) engine analyzes Claude.ai conversations (~1M) and maps them — to *20,000+ occupation tasks in the US Labor Department's O*NET database*. Quantitative measurement of how much Claude is used in which task of which occupation. Privacy preserved (Anthropic employees do not access raw conversations).

This is not a simple usage statistic. *It is a quarterly time series of AI penetration by occupation*. The first such measurement made possible since the 1987 Solow paradox.

Report progression:
- Report 1 (Feb 2025): ~36% of occupations have 25%+ of tasks handled by Claude. 4% of occupations have 75%+. *57% augmentation vs 43% automation*
- Report 2 (mid-2025, Sonnet 3.7 data): automation directive 27% → 39%. For the first time in API data, automation (49.1%) surpassed augmentation (47%)
- Report 3 (Sept 2025, geographic): 150+ countries, 50 US states. AUI (AI Usage Index): Singapore 4.6x, Canada 2.9x (per capita expected). India 0.27, Nigeria 0.20. Within US: Washington DC 3.82x, Utah 3.78x. *In India coding is 50%+ of all Claude tasks*
- Report 4 (Jan 2026, Economic Primitives): introduces 5 primitives — task complexity, human skill, AI skill, autonomy, success rate. Cumulative — *49% of jobs have had 25%+ of their tasks performed using Claude*. Computer & math 35.8% (observed), office/admin 34.3%, business/finance 28.4%, sales 26.9%, legal 20.4%
- Report 5 (Mar 2026, Learning Curves): experienced users (6+ months) achieve 10% higher success rates on identical tasks than new users. Top-10 tasks fell from 24% → 19% of all conversations (diversification). Top-5 US states' share fell 30% → 24% (Aug 2025 → Feb 2026)

The single most striking number: Computer programmers — 74.5% (highest single occupation). Customer service 70.1%. Data entry 67.1%. Medical records 66.7%.

If AEI measures "how much Claude is used in which occupation," Stanford's Erik Brynjolfsson measured a more direct question — *does AI usage actually affect employment*.

"Canaries in the Coal Mine?" (Brynjolfsson, Chandar, Chen — Stanford Digital Economy Lab, August 2025). Analyzed ADP payroll data covering 25M+ US workers. Core finding:

- Workers age 22-25 in AI-exposed jobs — employment -13% since late 2022
- After controlling for firm-level shocks: new entrants in AI-exposed jobs saw *-16% relative employment decline vs experienced workers*
- Mechanism: AI affects the labor market not by mass layoffs but by not hiring

*This is the real shock*. Anthropic's data shows *AI usage*. Brynjolfsson's data shows *that usage translates into actual employment decisions*.

BLS data goes one step further:
- NY Fed (2026): Computer Science graduate unemployment 6.1%, Computer Engineering 7.5% — *higher than philosophy*. Overall recent-graduate unemployment averaged 5.3%, peaked 9.3% (highest non-pandemic since 2014)
- Bay Area 2025: information sector -4,500 jobs (-4%), Bay Area tech layoffs ~40,000
- SF job postings -37% vs Feb 2020 (Oct 2025)
- UC Berkeley Enrico Moretti: "The number of jobs being created in AI *is not enough to offset* the job losses at traditional Big Tech"

Specific AI-cited layoffs:

| Company | Count | Date | AI link |
|---|---|---|---|
| Salesforce | 4,000 + <1,000 | Sept 2025 + Feb 2026 | Agentforce/Einstein Copilot |
| Microsoft | 15,000+ | 2025 | 30% code AI-written; >40% engineers |
| Meta | ~8,000 | 2025-26 | AI infra cost pivot |
| IBM | ~8,000 HR | 2024-25 | AskHR chatbot (94% handled) |
| Klarna | 700 → reversed | 2024 → 2025 | AI customer service (reversed) |
| Duolingo | 10% contractors | 2024 | GPT-4 translation |

Combined: 130,000+ in 2025 alone.

In Memo 1 we began with — Cisco reclaiming its $80 all-time high after 25 years on December 10, 2025. Over those 25 years, Cisco's business:
- FY00 → FY25 revenue $18.9B → $56.7B (3.0x)
- FY00 → FY25 GAAP EPS $0.36 → $2.61 (7.25x)
- Operating cash flow $10B+ every year

*The business didn't fail*. Revenue 3x, EPS 7x. But — *in real terms still -50%*. $80 in 2000 and $80 in 2025 are not the same value after inflation.

This is the coldest lesson of the entire series.

If the circuit collapses like Lucent, the company itself disappears. But even in the *Cisco scenario where the circuit stops but the business survives* — the investor must wait 25 years for recovery. Over those 25 years the business triples but the multiple compresses from 150x → single digits. *Business success and stock success are different things*.

In the 2026 AI cycle, *we don't yet know whether NVIDIA becomes Lucent or Cisco*. But — even in the Cisco scenario — if an investor who buys at $80 today *takes 25 years to recover that price*, then over those 25 years the AI business may grow explosively but at the portfolio level the return is *zero*.

Cahn's $600B Question update (Sequoia Dec 2025): 2026 hyperscaler capex $700-725B vs 2025 enterprise AI revenue $37B (Menlo Ventures). Gap $500B+. Even with 3x revenue growth (~$110B est.), the gap remains $600B+. *Current capex pace far outruns revenue pace*.

This is the quantitative evidence for Marks's "investor behavior" bubble (see Memo 3).

The same data is starting to show in Korea's labor market.

Korea Employment Information Service (2025-2035 Outlook): of 170 occupations analyzed, 93.4% maintain or grow employment. Only 6.6% (12 occupations) show slight decline — cashiers, bank tellers, design/editing assistants. On the surface it looks calm.

But *the developer market has already seen massive change*.
- In 2025, Kakao, Line, Coupang, Baemin (delivery), Daangn, Toss — new SWE hiring *zero or near zero*. Only Naver maintained routine hiring.
- Naver new hires: 838 (2021) → 258 (2024). *3x decline*.
- Kakao new hires: 994 (2021) → 314 (2024).
- Kakao publicly stated: "will restrict new hires for roles AI can replace."
- One anonymous mid-size Korean SW company: for the first time in 25 years, *hired zero new developers in 2025*.
- Korean junior developer hiring — *less than half* of 2021 levels. Peak decline expected -77%.

KDI research: AI productivity effects concentrated in *younger cohorts* (men 30-44, women 15-29) + *college-educated workers*. Negative effects on employment and wages in these groups. KDI estimates that *by ~2030, 90% of jobs could have 90%+ of tasks automatable*. Korean firms 69.2% consider AI competency in hiring decisions.

*The surface calm and the collapse of the new-hire market happen simultaneously*. The pattern Brynjolfsson found in the US ("not mass layoffs but no new hiring") is the same in Korea — *only faster and deeper*.

Korean society is feeling the end of "just be good at coding and you'll be hired" — *faster than* the Anthropic Economic Index's quarterly data.

This series, across 7 memos — on the surface — covered very different subjects.

1. The circuit of capital (NVDA→OpenAI→MSFT→NVDA)
2. The model duopoly and Claude Code
3. The IPO wave, what differs from 1999
4. Empire's periphery (HBM·CoWoS)
5. The next bottleneck (optical & networking)
6. When watts beat chips (power)
7. The one number (Anthropic Economic Index)

But *all 7 memos collapse into one question*. Will the AI capital cycle's $700B+ capex be recouped? For it to be recouped, the revenue generated by that capex must — *actually displace labor*. Real enterprise customers must pay real money, and real tasks of real jobs must be substituted.

In 1999, we could only know the answer to that question *ex-post*. That's why the Solow paradox stayed valid for 30 years. Whether internet capex was recouped — only became clear in macro data *years later*.

2026 is different. The Anthropic Economic Index measures that question quarterly. Claude Code's revenue is validated quarterly. Cisco's 25-year lesson is mapped against fresh data. Occupation penetration, new-hire data, Bay Area layoffs — *all visible in real time*.

So the series's final proposition is simple:

*We do not know. Whether the AI capex cycle will be recouped or not. Whether it becomes Cisco's 25 years or Lucent's -99%. Whether NVIDIA is the cycle's real rich or Lumentum is. Whether SK Hynix keeps memory power or Samsung recovers via HBM4.*

*But for the first time — we are seeing. Quarterly.*

*If the 1999 Solow paradox was a mystery only solved 30 years later, the 2026 AI cycle is — a fact measurable every quarter. That is the real difference of this cycle.*

The single number we should watch is — the next quarter's AEI penetration rate. If accelerating, real labor substitution is happening. If stalled or decelerating — the Cisco 25-year scenario truly begins.

That is the series's single watch metric. Every other data point is — merely *supporting evidence or refutation* of this one number.