Georgieva’s AI stability warning needs more than three bare numbers

Three figures arrived for 1 September: 2739, 0.196 and 28.801. They are meant to describe the machinery around artificial intelligence — INFO_SECTOR_EMP, ELEC_PRICE and SEMI_PPI — but they arrive without a trend line, a unit in one case, or a comparison.

Kristalina Georgieva has said artificial intelligence pushes growth. She has also said it threatens global financial stability. Put together, those remarks frame the problem facing investors: the same technology can raise output and concentrate risk, and the market may not know which effect is dominant until after capital has moved.

The missing context matters. A level of 2739 for INFO_SECTOR_EMP cannot be read as expansion or contraction unless it is set against prior months. ELEC_PRICE at 0.196 needs a currency, a market and a baseline. SEMI_PPI at 28.801 needs an index base and a history. The numbers are precise; the story around them is not.

Design capture is plausible, but not yet proven

The more ambitious figure is that 68% of value added in artificial intelligence goes to companies that hold the design. If true, that would mean the economics of AI resemble a platform or patent model more than a commodity manufacturing chain: scarce intellectual property takes the margin, while others supply labour, power or capacity.

The numbers are precise; the story around them is not.

But the figure is not tied to a named report, a year, a methodology or a definition of value added. It also sits beside a broader assertion that huge AI profits are concentrated in one hand, which compresses a possible oligopoly into a single holder. Those are different statements. One is about market structure; the other is rhetorical.

A third argument says the production burden is located in developing countries. That, too, needs a map. AI production includes chip fabrication, server assembly, data-centre construction, power supply, model training, annotation and maintenance. Each stage has different capital intensity, different geography and different margins. Without segment-level data, the burden argument is a slogan, not a ledger.

Stability is the harder variable

Georgieva’s stability warning is the part that can move policy. Financial-stability risk does not require AI profits to be unjust; it only requires leverage, correlated bets or sudden repricing. If lenders, landlords, utilities and funds are all exposed to the same build-out, a slowdown in AI investment can travel beyond the sector.

But again, the mechanism needs numbers. Are data-centre projects financed with debt? Are power contracts priced off the 0.196 electricity figure? Are semiconductor prices, at 28.801, rising enough to squeeze downstream buyers? Without those links, the warning remains a direction, not a magnitude.

What to watch instead

The practical test is whether the three instruments can be turned into a time series. INFO_SECTOR_EMP should show whether AI-related employment is broadening or peaking. ELEC_PRICE should show whether compute demand is feeding through to power costs. SEMI_PPI should show whether hardware input costs are falling, stable or spiking.

Until then, the strongest available facts are narrow: Georgieva has made both growth and stability statements, and three instrument levels exist for one date. The 68% design-share figure, the developing-country production burden and the one-hand profit argument remain unverified. For a market trying to price AI, that distinction is the difference between a thesis and a rumour. The next useful disclosure would be a sourced breakdown of AI value added by function, owner and geography. It would also say whether design, fabrication, power and labour are being counted in the same currency.