The $2.7tn AI offset reaches 0.1% into the countries paying the energy bill

The split does not add up to the whole. High-income economies take 96% of the measured labour-cost savings, middle-income economies 0.6% and low-income economies 0.1% ^[9], leaving 3.3% unassigned in the figures as reported ^[1]. Applied to a $2.7tn base ^[7], the low-income share works out at roughly $2.7bn a year spread across every country in that bracket, and the middle-income share at about $16bn ^[2]. The same base implies combined GDP of about $79tn for the 86 countries covered ^[3].

The estimate comes from five waves of the Anthropic Economic Index, a record of Claude conversations from January 2025 to February 2026 ^[8]. So the geography being measured is partly the geography of one vendor's user base, and a concentration index near 1 in Tanzania, where the authors find nearly all the value pooling in a small professional group ^[11], is describing access as much as productivity. The United States reads 0.44 on the same index ^[11].

The diffusion that is visible in the data runs vertically rather than across borders. The usage-weighted average wage of AI users fell 5.5% over 13 months as adoption reached lower-paid work, while aggregate gains doubled ^[12], and the annual figure has gone from $1.2tn to $2.7tn, a 2.25x rise ^[13][4]. That is real movement down the wage ladder inside the countries that already hold 96% of the total, and it does nothing to the cross-country distribution.

Now set the productivity claim against the macro table in the same account. Saved time worth 3.4% of GDP ^[7] sits beside a forecast survey that cuts 2026 growth to 2% from 2.2% and lifts consumer inflation to 3.2% from 2.6%, with job growth marked down from 64,500 to 45,000 ^[14]: 0.2 points off growth, 0.6 points on inflation, and about 30% off hiring ^[6]. Time saved is not output booked, which is the plain reason one number can be enormous while the other is shrinking. Treat the table as indicative, though. The account does not say whose survey it is ^[19], and it offers a 27% January inflation reading as context without naming the economy it belongs to ^[18].

Georgieva's incidence claim, that parts of Asia and Sub-Saharan Africa carry more of the energy pain than their European counterparts ^[4], is asserted here without supporting figures. It is also the half of the tug-of-war with no counterweight in the AI numbers.

The fund is on record against its own optimism as well. A May report warned that advanced models could drive more cyberattacks and threaten financial stability, citing models able to find and exploit software flaws without expert operators ^[16], and Georgieva said the IMF is "very keen to see more attention to the guardrails that are necessary to protect financial stability in a world of AI" ^[17].

Her constructive scenario is that expensive energy pushes governments toward diversification and efficiency ^[6]. That is a decade-length answer to a bill arriving monthly, and even the projected fall in WTI to $79.66 by year end, about 17.5% below the $96.57 reference close ^[15][5], unwinds only part of the move.