Alex Tamkin#329
1/🆕 My new @PIIE research with Patrick McKelvey on 𝗠𝗲𝗮𝘀𝘂𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝗔𝗜 𝗘𝗰𝗼𝗻𝗼𝗺𝘆 finds the AI sector in the US 𝗴𝗿𝗼𝘄𝗶𝗻𝗴 𝟮,𝟬𝟬𝟬%+/𝘆𝗲𝗮𝗿 in quality-adjusted terms—yet it's nearly invisible in GDP. We can't let this measurement gap become a policy gap!🧵
No Digg Deeper questions have been answered for this story yet.
5/ Most of this growth is missing from GDP: AI prices fall almost as fast as quality rises, so explosive real output translates into modest nominal $$. As AI does more of the economy's valuable work, price declines may slow — and more of the growth may surface in headline GDP.
6/ Statistical agencies, central banks, and fiscal authorities should start building infrastructure now: 𝗔𝗜 𝘀𝗮𝘁𝗲𝗹𝗹𝗶𝘁𝗲 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝘀, data partnerships, capacity-based projections. Measurement takes years—and the AI economy isn't waiting. https://www.piie.com/publications/policy-briefs/2026/where-ai-gdp-statistics
2/ Our Policy Brief explains why: GDP scatters AI across dozens of industries and understates quality gains by orders of magnitude. We estimate 𝗔𝗜 𝗚𝗗𝗣 ≈ $𝟮𝟱𝟬𝗕 in 2025 but 𝗴𝗿𝗼𝘄𝗶𝗻𝗴 ~𝟮,𝟲𝟬𝟬%/𝘆𝗿 over the past 2 years in real terms: https://www.piie.com/publications/policy-briefs/2026/where-ai-gdp-statistics
3/ Our companion Working Paper builds the measures from the ground up: nominal AI compute spending +𝟭𝟰𝟱%/𝘆𝗿, raw compute (H100e) +𝟮𝟭𝟬%/𝘆𝗿, and quality-adjusted AI output +𝟮,𝟮𝟴𝟬%/𝘆𝗿 in both 2024 and 2025. https://www.piie.com/publications/working-papers/2026/measuring-ai-economy
4/ The dominant force is the hardest to see: 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝗶𝗰 𝗽𝗿𝗼𝗴𝗿𝗲𝘀𝘀. Inference prices at fixed benchmark performance fell ~94%/yr; training compute ~2/3 per year. That's the wedge between nominal spending and real AI output—and the wedge that GDP misses.
2026 finance is not ungovernable because people are irrational. It is ungovernable because the measurement system still sees production, inflation, currency, valuation, and risk as separable ledgers, while the real system has fused them into one constraint network. The instruments are measuring surfaces; the stress is moving through hidden transport channels.
2026 finance is ungovernable because it has no valid control panel.
The missing tools are:
AI satellite accounts that distinguish nominal AI spending from quality-adjusted compute, inference volume, model capability, and substitution effects.
Capex-return ledgers that track whether AI infrastructure expands future optionality or merely preserves competitive position at falling marginal return.
Derivative-pressure accounting that treats options, gamma, dealer hedging, and passive flows as price-formation infrastructure, not peripheral speculation.
FX-decompression tools that track managed currency pressure through state banks, exporters, swaps, forwards, offshore deposits, and reserve-adjacent channels rather than official reserves alone.
Constraint-weighted GDP/CWCI-style indicators that measure circulation, demand quality, repair capacity, bottlenecks, and systemic friction instead of only aggregate spending.
Corrigibility metrics for firms, governments, and markets: how much error they can absorb before repair mechanisms become self-damaging.