The index is in development

Post Labor Economics
Get the daily brief

Methodology

Show the work

The combination rules are published. The index level is not. Names, directions, and the combination rule match the textbook figures, which are concept diagrams.

GIP

The full combination rules, sources, and gaps are in GIP Methodology v0.2. What follows is the short form. The Gross Intelligence Product is a family of indexes, not a dollar total. It is published as growth and as country shares. Advertising revenue is excluded.

Three equations, for country c at time t:

E is AI electricity consumed, in megawatt-hours, with crypto mining excluded. C is compute per megawatt-hour, from Epoch AI hardware trends. Q is capability per unit of compute, from the Artificial Analysis Intelligence Index. GIP Output is located where the power is drawn.

GIP Capacity is a second view of the same chain. It uses installed AI capacity in place of E.

GIP Consumed is world GIP Output times the country’s share of AI use. The use share averages the Microsoft AI Diffusion Report and the Anthropic Economic Index. Net AI exports are output minus consumption. A positive net means the country produces more than it uses.

Consumption and trade allocate output. They do not add to it. The same logic sits in the national accounts, where domestic use equals output plus imports minus exports. An earlier draft of this figure added imports and subtracted exports from output. That was wrong, and the figure does not do it.

A world index level is not published yet. When it is, the base and the country lines will be stated with the release. This page does not show a reading.

Figure 1. GIP: four index families, combined by rule FIG. 1 / FLOW SCHEMATICILLUSTRATIVEEAI electricity consumedMWH; CRYPTO MINING EXCLUDED1CCompute per MWhEPOCH AI HARDWARE TREND1aQCapability per unit of computeARTIFICIAL ANALYSIS INTELLIGENCE INDEX2ECQGIP Output(produced)LOCATED WHERE POWER IS DRAWNE × C × QΣ+−GIP ConsumedLOCATED WHERE AI IS USED3Share of world AI useMICROSOFT AND ANTHROPIC, AVERAGEDWorld GIP OutputSUM OF ALL COUNTRIES4Net AI exportsPRODUCED MINUS CONSUMEDPOSITIVE = NET EXPORTERGIP Capacity (second view)INSTALLED AI CAPACITY IN PLACE OF EEQ. 1 TO 3GIP Outputc,t​ = Ec,t​ × Ct​ × Qt​GIP Consumedc,t​ = World GIP Outputt​ × use sharec,t​Net AI exportsc,t​ = GIP Outputc,t​ − GIP Consumedc,t​Figure 1. GIP: four index families, combined by rule.AI electricity consumed in country c (family 1), compute per MWh from Epoch AI (1a) and capability per unit of compute from the Artificial AnalysisIntelligence Index (2) multiply into GIP Output, located where the power is drawn. GIP Consumed is world GIP Output times the country's share of AIuse (3), and net AI exports (4) are Output minus Consumed. As in national accounts, where domestic use equals output plus imports minus exports, tradeallocates output between countries and is never added to it. GIP Capacity runs the same chain on installed capacity. Published as growth and countryshares, not dollar levels; advertising revenue is excluded.
Figure 1. GIP: four index families, combined by rule. AI electricity consumed in country c (family 1), compute per MWh from Epoch AI (1a) and capability per unit of compute from the Artificial Analysis Intelligence Index (2) multiply into GIP Output, located where the power is drawn. GIP Consumed is world GIP Output times the country’s share of AI use (3), and net AI exports (4) are Output minus Consumed. As in national accounts, where domestic use equals output plus imports minus exports, trade allocates output between countries and is never added to it. GIP Capacity runs the same chain on installed capacity. Published as growth and country shares, not dollar levels; advertising revenue is excluded.

Open this figure at full size

Who exports intelligence

Figure 2 draws the net position. Both columns, GIP Output and GIP Consumed, sum to the same world total, so the nets sum to zero. The countries are hypothetical Country 1 to 6. Bilateral bands explain the idea. GIP records each country’s net. No public dataset yet maps cloud regions to users, so the figure does not pretend those flows are observed.

Figure 2. Who exports intelligence FIG. 2 / TRADE FLOW DIAGRAMILLUSTRATIVECountry 1128Country 1100Country 294Country 288Country 332Country 36Country 438Country 472Country 53Country 513Country 65Country 621GIP OUTPUT (PRODUCED)GIP CONSUMEDACBNET AI EXPORTSOUTPUT MINUS CONSUMEDCountry 1+28Country 2+6Country 3+26Country 4-34Country 5-10Country 6-16Exporters right of zero, importers left.Produced and used at homeExportedWIDTH = INDEX POINTSHYPOTHETICAL COUNTRIESAND VALUESFigure 2. Who exports intelligence.Left, each country's GIP Output (produced, located where the power is drawn); right, its GIP Consumed (located where AI is used). Both sides sum tothe same world total. Most output is used at home (A); export bands (B) carry the rest. Country 3 (C) produces far more than it uses, so it is a netexporter. Net AI exports equal Output minus Consumed. Countries and values are hypothetical: GIP publishes each country's net position, and no publicdataset yet maps bilateral flows.
Figure 2. Who exports intelligence. Left, each country’s GIP Output (produced, located where the power is drawn); right, its GIP Consumed (located where AI is used). Both sides sum to the same world total. Most output is used at home (A); export bands (B) carry the rest. Country 3 (C) produces far more than it uses, so it is a net exporter. Net AI exports equal Output minus Consumed. Countries and values are hypothetical: GIP publishes each country’s net position, and no public dataset yet maps bilateral flows.

Open this figure at full size

AI hours

The AI Labor Tracker will estimate the share of paid hours whose work product was produced by a model. Time with a chat window open does not count. That series is not published yet. Figure 3 is a concept diagram of the mechanism, not a reading of a quarter.

A cheaper model, or a model that takes a task, does not by itself cut hiring. In the competitive labor market of the introductory texts, demand for human hours shifts left only when automation’s displacement effect outweighs its productivity effect. Otherwise demand shifts right. The hatched area in Figure 3 is the change in the wage bill, w0L0 minus w1L1. Delta L is a distance, not an area.

Panel (b) follows Acemoglu and Restrepo. Tasks run from N − 1 to N. Labor’s comparative advantage rises with the task index. Machines do tasks below the frontier when they are cheaper there. Automation moves the frontier. New tasks raise N and give work back to people. The figure does not claim to be the CORE wage-setting and price-setting model.

Sources for the mechanism: Acemoglu and Restrepo, “The Race between Man and Machine,” American Economic Review, 2018. Acemoglu and Restrepo, “Automation and New Tasks,” Journal of Economic Perspectives, 2019. CORE Econ, The Economy.

Figure 3. Hours shifting from people to machines FIG. 3 / LABOR MARKET AND TASK DIAGRAMILLUSTRATIVE(a) The market for human hourswReal wageLPaid hours of human workE0​E1​w0​w1​L0​L1​SD0​D1​BACΔL(b) Which tasks machines doUnit cost, machine ÷ humaniTasks, ordered by human comparative advantage1ĨI0​I1​FRONTIERMACHINESPEOPLECHEAPER,NOT YETFEASIBLEN − 1NDr(i)Figure 3. Hours shifting from people to machines.(a) Competitive labor market. If automation's displacement effect outweighs its productivity effect, demand for human hours shifts left from D₀ to D₁(B) and the market moves from E₀ to E₁ (C): ΔL fewer paid hours at a lower wage. The hatched area (A) is the fall in the wage bill, w₀L₀ minus w₁L₁.If productivity gains dominate, demand shifts right instead. (b) Task framework after Acemoglu and Restrepo: tasks run from N − 1 to N, and r(i) ismachine over human unit cost. Machines do tasks below the frontier I, provided they are cheaper there (r below 1, tasks below Ĩ). Automation moves thefrontier from I₀ to I₁ (D). New tasks raise N and reinstate human work. Curves are schematic.
Figure 3. Hours shifting from people to machines. (a) Competitive labor market. If automation’s displacement effect outweighs its productivity effect, demand for human hours shifts left from D₀ to D₁ (B) and the market moves from E₀ to E₁ (C): ΔL fewer paid hours at a lower wage. The hatched area (A) is the fall in the wage bill, w₀L₀ minus w₁L₁. If productivity gains dominate, demand shifts right instead. (b) Task framework after Acemoglu and Restrepo: tasks run from N − 1 to N, and r(i) is machine over human unit cost. Machines do tasks below the frontier I, provided they are cheaper there (r below 1, tasks below Ĩ). Automation moves the frontier from I₀ to I₁ (D). New tasks raise N and reinstate human work. Curves are schematic.

Open this figure at full size

Inside a hall

Figure 4 is a cutaway of the power and cooling path, not a rendering of a named site. Utility power steps down through switchgear and transformers. The UPS carries the load until generators, tied through transfer switches, start. A coolant distribution unit separates the facility water loop from the rack liquid loop. Counts are simplified: a real hall holds hundreds to thousands of racks, and a GB200 NVL72 rack holds 18 compute trays and draws about 120 kW.

Figure 4. Inside an AI data center FIG. 4 / CUTAWAY ISOMETRICILLUSTRATIVEDETAIL A: COMPUTE TRAY, EXPLODED (GB200 NVL72 TYPE)Cold plates, lifted64 Blackwell GPUs2 Grace CPUsTray frontRear liquidquick disconnectsENLARGED, NOT TO SCALE1234578VIA TRANSFER SWITCHKEY1Utility substationHIGH TO MEDIUM VOLTAGE2Switchgear, transformersSTEP DOWN TO LOW VOLTAGE3UPS and batteriesCARRY THE LOAD UNTIL GENERATORS START4Backup generatorsFEED SWITCHGEAR VIA TRANSFER SWITCH5Compute hallRACKS FED BY OVERHEAD BUSWAY6Compute trayLIQUID-COOLED; SEE DETAIL A7Coolant distribution unitFACILITY WATER TO RACK LOOP8Dry coolersREJECT HEAT TO OUTSIDE AIRPower pathGenerator feedFacility water loopRack liquid loopCut sectionFigure 4. Inside an AI data center.Power comes from the utility substation (1) to switchgear and transformers that step it down to low voltage (2). The UPS (3) carries the load throughan outage until the backup generators (4), tied in through transfer switches, take over. Overhead busways feed the racks (5), and each compute tray(6) turns power into computation and heat. A coolant distribution unit (7) passes heat from the rack liquid loop to the facility water loop, and drycoolers (8) reject it to outside air. Counts are simplified: a real hall holds hundreds to thousands of racks, and a GB200 NVL72 rack holds 18 computetrays and draws about 120 kW. Not to scale.
Figure 4. Inside an AI data center. Power comes from the utility substation (1) to switchgear and transformers that step it down to low voltage (2). The UPS (3) carries the load through an outage until the backup generators (4), tied in through transfer switches, take over. Overhead busways feed the racks (5), and each compute tray (6) turns power into computation and heat. A coolant distribution unit (7) passes heat from the rack liquid loop to the facility water loop, and dry coolers (8) reject it to outside air. Counts are simplified: a real hall holds hundreds to thousands of racks, and a GB200 NVL72 rack holds 18 compute trays and draws about 120 kW. Not to scale.

Open this figure at full size

Files

Every line chart links a CSV in the repository. Writers drop the file next to the post. The build copies it to /data/posts/. See the README for the folder layout. Corrections, when there are any, are listed on the corrections page.

The daily brief

Get the daily brief

One chart, three numbers and their sources. Every weekday at 7 AM ET. Free.