Recent
August 10, 2026
A data-driven analysis that grants every AI bull case its premise (the models work, they keep improving, every demo ships) and asks a narrower question: who actually pays the roughly $1.2 trillion a year of token revenue the buildout needs, rising toward $2.5 trillion by 2031, and do they ever get the money back?
The answer is payroll, and no. Every dollar of token spending traces back to a wage either replaced or never hired, and in a competitive market the savings get competed through to customers rather than kept by the company that bought the tokens, a dynamic visible in Jamie Dimon's own words on JPMorgan's earnings call and in Nordhaus's finding that innovators historically keep only about 2.2% of the value they create.
Core argument:
• The tab is real: Goldman Sachs models $765B of AI capex this year rising to $1.6T by 2031; covering just the hardware (not profit) requires roughly $1.18T of compute revenue now and $2.5T by 2031.
• Demand has to be untangled from supply financing itself: circular deals between hyperscalers, labs, and chipmakers can't prove that an outside customer, spending only its own money, will pay. It's the same problem that inflated telecom vendor financing in the late 1990s.
• The only pool big enough is wages: money saved by replacing paid work, or by underpricing the humans a new AI-built business would otherwise have hired. That pool is worth tens of trillions worldwide, dwarfing the ~$0.7-1.4T global software market.
• Competition erases the buyer's gains: the heaviest token buyers are the ones least able to keep the savings, because competitive pressure passes cost reductions straight to customers, leaving "insurance-shaped" defensive spending as the most fragile part of the current boom.
Includes 3 interactive Plotly visualizations on capex by company, the hardware payback hurdle, and the GPU price-to-cost spread, plus a depreciation stress-test model with a concrete date already on the calendar.
May 29, 2026
A data-driven analysis of the AI capital loop: hyperscalers fund model labs, labs spend the money back on chips and cloud, private valuations rise, and those markups can flow into public-company earnings. The missing link is broad proof that AI buyers are turning the tools into measurable revenue, margins, or output per worker.
The loop is real, but that does not automatically make it fake. Vendor financing and strategic supply commitments are normal in capital-intensive buildouts. The risk is narrower: AI spending and valuation gains are visible now, while end-customer productivity remains uneven and hard to audit.
Core thesis:
• The first links are documented: Anthropic's valuation and revenue growth, hyperscaler compute commitments, and investor markups are already showing up in the financial system.
• The final link is still weak: broad productivity should show up as revenue per employee, margin expansion, or output per hour. The current macro and enterprise ROI evidence is mixed.
• The verdict depends on conversion: if enterprises turn AI into redesigned workflows and measurable P&L gains, the circularity looks like an infrastructure buildout. If not, the loop reprices.
Includes 3 interactive Plotly visualizations mapping the capital loop, contrasting evidence strength, and showing the gap between AI supply-chain validation and downstream productivity proof.
October 21, 2025
A data-driven analysis of whether OpenAI's $852 billion valuation, $20.9 billion operating loss, governance chaos, and product proliferation resembles Google's path to dominance or WeWork's cautionary tale, with Snapchat's product confusion thrown in for good measure. Updated July 2026 with verified 2025 financials.
OpenAI's verified 2025 numbers ($13.1 billion in revenue against a $20.9 billion operating loss, a ~160% loss-to-revenue ratio that is worse than WeWork's ~105% at peak) make the WeWork comparison sharper rather than softer, even as revenue nearly tripled year over year. API pricing collapsed 83% in 14 months, users can't distinguish between model versions, and the company hemorrhages safety researchers to competitors after restructuring from non-profit to for-profit.
Core findings from financial analysis:
• Worse-than-WeWork burn: $20.9B operating loss on $13.07B of 2025 revenue ≈ 160% loss-to-revenue, versus WeWork's ~105% at its peak
• Pricing collapse: GPT-4 API pricing fell 83% (from $30 to $5 per million tokens) in just 14 months due to competition from Claude, Gemini, and free Llama models
• Product chaos: 5+ major model versions in 18 months (GPT-4, 4-Turbo, 4o, o1, o3) with unclear differentiation; users literally can't explain which to use
• Talent exodus: 6+ key researchers departed in 12 months, mostly to Anthropic, citing "safety culture has taken a backseat to shiny products"
• Valuation disconnect: ~38x trailing revenue at the October 2025 valuation of $500B (repriced to $852B in March 2026) while losing over $20B a year; compare Google at IPO (10x, profitable) or successful tech IPOs (7-10x)
With 3 interactive data visualizations comparing OpenAI's trajectory to WeWork's financials, tracking API pricing collapse, and documenting product proliferation timeline. Includes steel-man counterarguments and falsifiable predictions for what would prove this analysis right or wrong.
October 13, 2025
An empirical analysis of how labor market failures can create economic incentives for gang participation across American cities. Synthesizes evidence from economics, criminology, and urban sociology while separating direct gang evidence from broader crime evidence.
The evidence is strongest for a labor-market effect on youth crime and violence, with more limited direct evidence on gang membership itself. During the Great Recession and the COVID-19 shock, youth unemployment rose sharply and many cities saw violence concentrate in already disadvantaged neighborhoods. The pattern is not mechanical, but it is persistent: where legitimate work is scarce, gangs can become more attractive.
Key findings from rigorous research:
• Some gangs function as alternative labor markets. In one well-documented Chicago case, gang "foot soldiers" earned about $3.30/hour in late-1980s/early-1990s dollars, though many also held legal jobs.
• Employment interventions can reduce violence. A randomized Chicago summer-jobs program reduced violent-crime arrests by 43% over 16 months, though that is not the same as directly measuring gang exit.
• The causal evidence is uneven. Unemployment-crime studies find clearer effects for property crime than violent crime, and gang-specific data is harder to isolate.
• Six mechanisms operate simultaneously: opportunity cost reduction, income replacement, status provision, network effects, time allocation, and institutional deterioration.
• The policy implication is clear: enforcement alone cannot solve gang problems rooted partly in joblessness. Employment, mentoring, education, and focused violence prevention need to be part of the public-safety toolkit.
With 5 interactive D3.js visualizations illustrating recession impacts, city-level patterns, age effects, mechanisms, and policy cost-benefit comparisons. Includes academic citations from Becker, Freeman, Levitt & Venkatesh, BLS data, and recent RCT evidence.
October 10, 2025
An in-depth exploration of why automation creates winners and losers in the labor market, and what we can do about it. The dominant narrative about automation tends toward simplicity: either technology devastates worker earnings or lifts all boats. Both views miss the empirical reality.
This comprehensive analysis advances the Heterogeneous Impact Thesis: automation creates winners and losers not randomly, but according to predictable patterns related to task composition, skill complementarity, and market power dynamics.
Key findings from the empirical evidence:
• Task composition matters more than skill level. Non-routine tasks experience less wage pressure than routine tasks.
• Geography determines outcomes. Detroit's wages fell 6.3% as robot adoption accelerated, while San Francisco's grew 18.2% over the same period.
• Institutions matter enormously. German autoworkers' wages rose 18% with automation; American autoworkers' fell 2%. Same technology, different institutional context.
• Worker bargaining power is decisive. In unionized firms, automation raised wages 4.7%; in non-union firms, production workers saw just 0.8% gains.
With interactive D3.js visualizations showing robot density by country, task polarization trends, and within-firm wage effects. Includes 50+ academic citations and policy implications for the AI transition.
December 13, 2025
A data-driven analysis tracking four empirical indicators historically associated with democratic backsliding: economic distress (unemployment), declining national pride, demographic change rates, and institutional trust. Drawn from peer-reviewed research on the drivers of populism and democratic erosion.
Key finding: The composite index currently registers at 76.8/100 ("elevated risk"), higher than both the 1980 stagflation (70.7) and 2009 recession (72.0). Unlike previous crises, this is not driven by economic factors; unemployment is just 4.4%. Instead, the primary drivers are historic lows in national pride (58% combined proud, vs. 87% in 2001) and institutional trust (22%, vs. 77% in 1964).
Current Reading: 76.8 / 100 (Elevated Risk)
Includes important methodological caveats: correlation ≠causation, proxy limitations (demographic change ≠attitudes toward immigration), and factors not captured (institutional resilience, elite behavior, information environment). With 12 academic citations and 5 interactive Plotly.js visualizations.
October 8, 2025
What conditions create fertile ground for extremist movements? Historical analysis from the 1920s to today reveals a consistent pattern: extremism rarely arises from a single cause. Instead, four key ingredients combine to create conditions for democratic backsliding.
The four ingredients that precede populist/fascist turns:
1. Economic Crisis (~33% of effect): Financial crises trigger unemployment, debt, uncertainty → risk-seeking behavior. Meta-analysis of 36 studies confirms causal effect.
2. Status Threat (major driver): Perceived loss of national standing or group dominance. Stronger predictor than economics in 2016 US election.
3. Cultural Backlash (major driver): Demographic change, immigration → anxiety and scapegoating. Cultural concerns can drive economic discontent (reverse causation).
4. Institutional Weakness (accelerant): Young democracies, mainstream party failure → trust collapse.
The pattern: Weimar Germany, Greece 2009-2015, Hungary 2006-2014 all show these ingredients combining. When all four align during acute crises, extremist parties surge from 0.3-2.6% to 7-37% within years. Between 1920-1939, Europe's democracies fell from 24 to 11, a 54% collapse.
October 7, 2025
Americans are living through a confusing economic moment: low unemployment, rising wages, controlled inflation. Yet millions feel financially crushed. The answer lies in a fundamental bifurcation of the economy that aggregate statistics completely miss.
Analysis of four major spending categories (housing, healthcare, education, food) from 1980-2024 reveals a stark pattern:
Investment goods (housing, healthcare, education) have pulled away from wages: housing's price-to-income ratio is up about 40%, healthcare costs 73% more relative to wages, and education 43% more. Meanwhile consumption goods (food) have become about 13% MORE affordable.
You can afford to survive, but thriving keeps getting more expensive. Food is cheap; the things you need for economic mobility (buying a home, getting healthcare, earning a degree) have pulled steadily away from wages.
Quality-adjusted (homes are about 35% larger), the size-adjusted decline in housing affordability is modest, roughly 4%, but the unadjusted ratio is what buyers actually face, and you can't opt out: zoning laws limit smaller homes, you can't buy "1980s healthcare" at 1980s prices, and credentialism means you can't skip expensive degrees.
July 11, 2025
Twelve charts that reveal how economic rewards have shifted from workers to capital owners, machines, and foreign competitors.
A comprehensive analysis of BLS and FRED data from 1947 to present reveals a genuine shift in economic rewards away from labor toward capital, with labor's share of national income declining from roughly 50% to 44%:
From 1947 to 2024, labor's share of national income has fallen from roughly 50% to 44%, about six percentage points (closer to four after adjusting for self-employment income), representing hundreds of billions of dollars annually shifted from workers to capital owners, landlords, and corporate profits.
This was caused by a combination of technological displacement, globalization (particularly the China Shock), weakening labor institutions, and tax policies that increasingly favor capital over labor [...]
We've also discovered that current tax incentives actively encourage businesses to replace workers with machines, creating a self-reinforcing cycle of automation.