Journey in Garret

Power and Progress Review: What Would Make Big Tech Bargain?

Power and Progress by Daron Acemoglu and Simon Johnson — book cover, PublicAffairs

At a Glance

Persuasive on why technology doesn’t spread prosperity on its own. Less convincing on how to build the power to change that.

7/10

BOOK

Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity

AUTHORS

Daron Acemoglu & Simon Johnson

ORIGINAL PUBLICATION

PublicAffairs, 2023

EDITION READ

Korean translation (Saenggakui Him, 2023), 735 pages

READ

September–October 2025

Read it if

You want to examine who controls technological change, who benefits, and what bargaining power has to do with the answer.

Skip it if

You want AI stock picks, technical forecasts, or a short history with little discussion of politics and labor.

Convenience Is Only Part of the Story

An AI tool can summarize a report or translate a page in seconds. For the person using it, that is a convenience. For someone paid to do the same task, it may be a threat. Both reactions can belong to the same person.

That tension was already bothering me when I read Power and Progress in the autumn of 2025. I did not need convincing that AI was important. The larger promise, that its benefits would eventually reach everyone, was harder to accept.

Daron Acemoglu and Simon Johnson examine the distance between those two claims. Across a thousand years of technological change, they ask why expanding production sometimes improved ordinary lives and sometimes left most people with little to show for it. Their answer turns on who directs technology, what it enables workers to do, and who can demand a share of the proceeds.

The diagnosis is persuasive. The proposed remedy leaves me with a harder question: where will the power to carry it out come from?

More Output Does Not Settle the Distribution

1. A more productive company need not mean better-paid workers.

The authors call the comforting assumption that technological gains will spread through society the “productivity bandwagon.” But if a company can produce the same output with fewer people, why must it pay the remaining workers more? New tasks can increase workers’ productive contribution. Bargaining power can help them claim the resulting gains. Neither follows automatically from a larger output figure.

The Economic Policy Institute (EPI) makes the distinction visible in its productivity and compensation series. From the fourth quarter of 1979 to the second quarter of 2026, net productivity rose 93.7%, while the real hourly compensation measure rose 32.7%. Compensation increased; it did not keep pace. The chart describes that divergence, rather than isolating how much technology, labor institutions, or any other cause contributed to it.

01 / Power and Progress

Higher productivity. A smaller pay rise.

United States · 1948 Q1–2026 Q2 · 1948 Q1 = 100

From 1979 Q4 to 2026 Q2: net productivity +93.7%; real hourly compensation +32.7%. Compensation rose, but fell behind productivity.
Net productivityReal hourly compensation

Hover or tap to inspect. Keyboard: focus the plot, then use ← →, Home, or End.

Higher productivity. A smaller pay rise.Index10020030040019481960197920002026
Observation2026 Q2
Net productivity419.9
Real hourly compensation253.0
Productivity is economy-wide net output per hour. Compensation covers private-sector production and nonsupervisory workers: hourly earnings plus an estimated allowance for benefits, not the average pay of all workers. Both series use the same blended CPI. EPI interpolates pre-1964 quarterly compensation from annual data. The gap alone does not identify the causal contribution of technology, unions, or policy.
View the data table
PeriodNet productivity (index)Real hourly compensation (index)
1948 Q1100.0100.0
1948 Q2101.2100.8
1948 Q3100.8100.8
1948 Q4102.7103.1
1949 Q1102.2105.4
1949 Q2101.3106.6
1949 Q3104.7109.0
1949 Q4104.7110.3
1950 Q1108.9111.8
1950 Q2110.1112.4
1950 Q3112.1111.7
1950 Q4112.9112.1
1951 Q1111.5110.2
1951 Q2111.6111.2
1951 Q3114.4113.5
1951 Q4114.3113.7
1952 Q1114.3114.4
1952 Q2114.9115.8
1952 Q3115.3115.8
1952 Q4117.5117.3
1953 Q1119.9119.6
1953 Q2120.5120.6
1953 Q3120.2121.8
1953 Q4119.7122.6
1954 Q1120.3123.7
1954 Q2121.8124.5
1954 Q3123.7125.3
1954 Q4126.8127.8
1955 Q1128.8128.4
1955 Q2131.2130.2
1955 Q3131.3130.6
1955 Q4131.8132.3
1956 Q1132.4134.2
1956 Q2132.3134.5
1956 Q3132.3134.8
1956 Q4134.3135.8
1957 Q1135.9136.4
1957 Q2136.0137.5
1957 Q3136.4137.9
1957 Q4136.5138.5
1958 Q1135.0137.5
1958 Q2136.3138.0
1958 Q3139.6138.8
1958 Q4141.9140.0
1959 Q1143.4142.0
1959 Q2144.2143.0
1959 Q3144.5142.8
1959 Q4144.5143.3
1960 Q1148.0144.3
1960 Q2145.9145.6
1960 Q3146.2146.2
1960 Q4144.9146.5
1961 Q1146.6147.6
1961 Q2151.1148.4
1961 Q3152.8148.5
1961 Q4153.8149.1
1962 Q1155.7151.0
1962 Q2155.6151.6
1962 Q3157.4152.3
1962 Q4158.9153.3
1963 Q1160.7154.7
1963 Q2161.8155.8
1963 Q3163.8155.3
1963 Q4165.4157.2
1964 Q1167.4158.0
1964 Q2167.0158.6
1964 Q3169.6160.3
1964 Q4168.7160.4
1965 Q1171.1161.6
1965 Q2171.4162.4
1965 Q3175.0163.1
1965 Q4178.5164.8
1966 Q1179.3165.1
1966 Q2178.2165.4
1966 Q3178.7165.6
1966 Q4179.4165.7
1967 Q1180.1166.8
1967 Q2181.7168.5
1967 Q3181.5168.7
1967 Q4182.3169.2
1968 Q1186.3170.6
1968 Q2187.6171.1
1968 Q3187.9172.2
1968 Q4188.0172.9
1969 Q1189.7175.1
1969 Q2188.8175.1
1969 Q3190.4176.2
1969 Q4189.6177.2
1970 Q1190.1176.4
1970 Q2191.9176.8
1970 Q3195.3178.6
1970 Q4193.9178.9
1971 Q1200.0181.1
1971 Q2200.8182.1
1971 Q3203.2183.5
1971 Q4202.6184.9
1972 Q1204.7189.6
1972 Q2209.2191.6
1972 Q3211.0193.1
1972 Q4213.7193.9
1973 Q1214.8193.8
1973 Q2213.8192.1
1973 Q3211.4191.6
1973 Q4211.8189.9
1974 Q1207.3188.1
1974 Q2207.8187.6
1974 Q3206.0187.5
1974 Q4207.0186.6
1975 Q1209.3186.9
1975 Q2212.2187.5
1975 Q3214.3186.9
1975 Q4214.7186.6
1976 Q1216.6187.3
1976 Q2218.1188.7
1976 Q3218.1189.9
1976 Q4218.6190.7
1977 Q1219.4191.5
1977 Q2218.6191.8
1977 Q3221.1193.1
1977 Q4220.3193.5
1978 Q1219.5195.0
1978 Q2222.7195.0
1978 Q3222.5194.8
1978 Q4224.0195.4
1979 Q1221.7195.1
1979 Q2220.7193.0
1979 Q3218.9192.2
1979 Q4216.8190.7
1980 Q1215.7188.3
1980 Q2212.9188.2
1980 Q3212.8187.9
1980 Q4214.7187.7
1981 Q1217.9186.8
1981 Q2216.8187.3
1981 Q3219.2187.2
1981 Q4216.0186.5
1982 Q1215.4187.3
1982 Q2215.1187.8
1982 Q3214.9187.5
1982 Q4215.9187.1
1983 Q1218.7189.2
1983 Q2221.7188.9
1983 Q3223.0188.6
1983 Q4225.2188.5
1984 Q1225.5188.0
1984 Q2226.9187.7
1984 Q3228.2187.6
1984 Q4228.7187.4
1985 Q1229.9187.2
1985 Q2229.9187.3
1985 Q3232.5187.6
1985 Q4232.3187.2
1986 Q1234.2187.4
1986 Q2237.1189.4
1986 Q3237.9189.4
1986 Q4236.7189.3
1987 Q1234.5187.8
1987 Q2235.2187.0
1987 Q3235.1186.4
1987 Q4237.0186.4
1988 Q1237.6186.3
1988 Q2238.0186.4
1988 Q3239.0186.1
1988 Q4239.6186.7
1989 Q1239.7186.8
1989 Q2239.8186.0
1989 Q3240.8186.6
1989 Q4240.4187.1
1990 Q1241.6186.8
1990 Q2243.5186.6
1990 Q3241.7184.9
1990 Q4237.7183.8
1991 Q1238.9184.5
1991 Q2242.7185.7
1991 Q3244.3186.2
1991 Q4245.1186.4
1992 Q1249.0186.7
1992 Q2250.9187.4
1992 Q3253.5187.7
1992 Q4254.6187.0
1993 Q1253.4188.3
1993 Q2252.0187.5
1993 Q3252.8187.7
1993 Q4254.5187.8
1994 Q1255.6188.8
1994 Q2255.1188.2
1994 Q3253.8188.0
1994 Q4256.3188.2
1995 Q1255.0187.5
1995 Q2255.0187.2
1995 Q3255.0187.5
1995 Q4256.8187.8
1996 Q1258.3187.6
1996 Q2259.1187.5
1996 Q3259.0188.1
1996 Q4258.9187.9
1997 Q1259.3188.3
1997 Q2262.2189.4
1997 Q3264.4190.4
1997 Q4264.9191.6
1998 Q1265.7192.9
1998 Q2266.3194.3
1998 Q3269.5195.0
1998 Q4270.9195.8
1999 Q1273.8196.9
1999 Q2272.4197.2
1999 Q3273.8197.7
1999 Q4276.3198.0
2000 Q1273.2196.9
2000 Q2277.9198.1
2000 Q3277.8198.4
2000 Q4280.7200.0
2001 Q1278.4200.4
2001 Q2281.7201.6
2001 Q3282.2202.9
2001 Q4285.8205.0
2002 Q1290.7206.3
2002 Q2290.5206.3
2002 Q3292.0207.5
2002 Q4292.4209.1
2003 Q1294.2209.2
2003 Q2299.9210.6
2003 Q3304.0210.4
2003 Q4307.3210.0
2004 Q1307.3210.0
2004 Q2309.9209.8
2004 Q3311.6210.0
2004 Q4312.8209.3
2005 Q1316.2210.2
2005 Q2316.7210.5
2005 Q3315.9209.1
2005 Q4316.7209.0
2006 Q1319.1209.6
2006 Q2318.4210.1
2006 Q3316.6210.0
2006 Q4319.8212.6
2007 Q1321.4212.5
2007 Q2321.3212.9
2007 Q3323.2213.7
2007 Q4322.7212.7
2008 Q1319.2212.6
2008 Q2319.3212.2
2008 Q3316.9211.0
2008 Q4321.9217.6
2009 Q1327.4222.4
2009 Q2331.1222.5
2009 Q3333.7222.1
2009 Q4337.1222.1
2010 Q1338.7224.0
2010 Q2340.6224.9
2010 Q3344.6225.2
2010 Q4344.7224.8
2011 Q1341.0223.7
2011 Q2339.1222.1
2011 Q3338.2221.3
2011 Q4338.6221.4
2012 Q1340.1219.9
2012 Q2341.9220.4
2012 Q3339.9220.6
2012 Q4339.6219.8
2013 Q1342.4221.4
2013 Q2343.5223.1
2013 Q3344.8223.5
2013 Q4347.0224.0
2014 Q1344.2223.7
2014 Q2345.9223.5
2014 Q3349.0223.9
2014 Q4348.5225.2
2015 Q1352.1228.1
2015 Q2353.7227.8
2015 Q3354.7228.2
2015 Q4353.6229.9
2016 Q1354.6231.5
2016 Q2354.4231.1
2016 Q3355.4231.5
2016 Q4357.9231.4
2017 Q1357.8231.2
2017 Q2357.9232.4
2017 Q3361.4233.2
2017 Q4362.7232.8
2018 Q1364.0233.2
2018 Q2364.5234.1
2018 Q3365.5235.2
2018 Q4365.2236.7
2019 Q1367.3237.7
2019 Q2368.8238.3
2019 Q3372.1239.6
2019 Q4374.1240.0
2020 Q1373.2241.0
2020 Q2386.3253.7
2020 Q3392.8248.3
2020 Q4389.9248.0
2021 Q1394.5248.4
2021 Q2396.3247.7
2021 Q3394.9247.5
2021 Q4395.8246.1
2022 Q1391.6244.3
2022 Q2389.4242.0
2022 Q3388.8241.9
2022 Q4390.4242.5
2023 Q1391.9243.5
2023 Q2393.3245.1
2023 Q3398.3246.0
2023 Q4400.4247.2
2024 Q1401.1247.9
2024 Q2404.0248.8
2024 Q3406.5250.8
2024 Q4407.5251.2
2025 Q1406.9251.9
2025 Q2411.1253.5
2025 Q3415.0253.7
2025 Q4416.7254.7
2026 Q1418.3254.6
2026 Q2419.9253.0
EPI: The Productivity–Pay Gap
Source updated September 14, 2026. Fixed data snapshot; no live feed.

2. Technology has a direction, and someone chooses it.

AI can help a person handle more demanding work. It can also be used to eliminate that person’s tasks and cut costs. Investment decisions favor one path over another, and the interests of those making the decisions matter.

Acemoglu and Johnson distinguish “machine usefulness” (expanding human capabilities and creating productive tasks) from “so-so automation,” which transfers work to machines without delivering substantial productivity gains. Replacing a worker is not, by itself, evidence that a useful advance has occurred.

3. Sharing the gains requires a voice in the decisions.

This is why Acemoglu and Johnson emphasize unions, civil society, and democratic institutions as countervailing powers. People affected by a technology need some influence over how it is introduced and what happens to their working conditions. Government has a role in making that influence possible.

I agree with the need. I am less certain about the route from needing power to possessing it.

How Long Before Progress Reaches a Paycheck?

Early British industrialization is an uncomfortable example. Production expanded while many workers endured harsh conditions and waited decades for a substantial improvement in living standards. More wealth existed. Access to it remained narrow. Robert Allen’s account of “Engels’ pause” captures this lag between rising output and wages.

02 / Power and Progress

Output grew before wages caught up

Britain · 1770–1890 · each series rebased to 1770 = 100

Approximate reconstruction, not a raw statistical dataset. This reproduces the lines in Allen’s Figure 1; it does not resolve the historical debate over their measurement.
Real GDP per workerReal earnings

Hover or tap to inspect. Keyboard: focus the plot, then use ← →, Home, or End.

Output grew before wages caught upIndex10015020025017701800183018601890
Observation1890 · approximate figure recovery
Real GDP per worker≈ 256
Real earnings≈ 244
Recovered mechanically from the vector paths in Allen (2009), Figure 1; cropped from 1770–1913 to 1770–1890 and separately rebased. The output series is GDP per worker, not per person. The earnings series is not hourly total compensation and has no annual-hours adjustment. No new interpolation, smoothing, or extrapolation was added. Later research has reconsidered the size and interpretation of the gap.
View the data table
PeriodReal GDP per worker (index)Real earnings (index)
1770 · approximate figure recovery≈ 100≈ 100
1771 · approximate figure recovery≈ 100≈ 96
1772 · approximate figure recovery≈ 100≈ 93
1773 · approximate figure recovery≈ 100≈ 94
1774 · approximate figure recovery≈ 100≈ 94
1775 · approximate figure recovery≈ 100≈ 96
1776 · approximate figure recovery≈ 99≈ 99
1777 · approximate figure recovery≈ 99≈ 97
1778 · approximate figure recovery≈ 99≈ 100
1779 · approximate figure recovery≈ 99≈ 104
1780 · approximate figure recovery≈ 99≈ 102
1781 · approximate figure recovery≈ 100≈ 100
1782 · approximate figure recovery≈ 100≈ 101
1783 · approximate figure recovery≈ 101≈ 101
1784 · approximate figure recovery≈ 102≈ 101
1785 · approximate figure recovery≈ 102≈ 104
1786 · approximate figure recovery≈ 103≈ 107
1787 · approximate figure recovery≈ 103≈ 106
1788 · approximate figure recovery≈ 104≈ 106
1789 · approximate figure recovery≈ 105≈ 104
1790 · approximate figure recovery≈ 106≈ 104
1791 · approximate figure recovery≈ 106≈ 106
1792 · approximate figure recovery≈ 107≈ 108
1793 · approximate figure recovery≈ 107≈ 109
1794 · approximate figure recovery≈ 108≈ 108
1795 · approximate figure recovery≈ 109≈ 102
1796 · approximate figure recovery≈ 109≈ 111
1797 · approximate figure recovery≈ 110≈ 118
1798 · approximate figure recovery≈ 111≈ 119
1799 · approximate figure recovery≈ 111≈ 110
1800 · approximate figure recovery≈ 112≈ 92
1801 · approximate figure recovery≈ 113≈ 96
1802 · approximate figure recovery≈ 114≈ 110
1803 · approximate figure recovery≈ 114≈ 112
1804 · approximate figure recovery≈ 115≈ 108
1805 · approximate figure recovery≈ 115≈ 109
1806 · approximate figure recovery≈ 116≈ 113
1807 · approximate figure recovery≈ 117≈ 117
1808 · approximate figure recovery≈ 117≈ 111
1809 · approximate figure recovery≈ 118≈ 107
1810 · approximate figure recovery≈ 118≈ 105
1811 · approximate figure recovery≈ 119≈ 106
1812 · approximate figure recovery≈ 120≈ 100
1813 · approximate figure recovery≈ 121≈ 98
1814 · approximate figure recovery≈ 122≈ 110
1815 · approximate figure recovery≈ 122≈ 117
1816 · approximate figure recovery≈ 123≈ 105
1817 · approximate figure recovery≈ 124≈ 102
1818 · approximate figure recovery≈ 125≈ 102
1819 · approximate figure recovery≈ 126≈ 104
1820 · approximate figure recovery≈ 127≈ 108
1821 · approximate figure recovery≈ 128≈ 110
1822 · approximate figure recovery≈ 128≈ 114
1823 · approximate figure recovery≈ 128≈ 113
1824 · approximate figure recovery≈ 129≈ 108
1825 · approximate figure recovery≈ 130≈ 106
1826 · approximate figure recovery≈ 130≈ 109
1827 · approximate figure recovery≈ 130≈ 107
1828 · approximate figure recovery≈ 131≈ 110
1829 · approximate figure recovery≈ 131≈ 108
1830 · approximate figure recovery≈ 132≈ 110
1831 · approximate figure recovery≈ 133≈ 109
1832 · approximate figure recovery≈ 134≈ 115
1833 · approximate figure recovery≈ 136≈ 121
1834 · approximate figure recovery≈ 137≈ 123
1835 · approximate figure recovery≈ 138≈ 129
1836 · approximate figure recovery≈ 140≈ 126
1837 · approximate figure recovery≈ 141≈ 123
1838 · approximate figure recovery≈ 142≈ 124
1839 · approximate figure recovery≈ 144≈ 118
1840 · approximate figure recovery≈ 145≈ 120
1841 · approximate figure recovery≈ 146≈ 124
1842 · approximate figure recovery≈ 148≈ 126
1843 · approximate figure recovery≈ 149≈ 139
1844 · approximate figure recovery≈ 151≈ 135
1845 · approximate figure recovery≈ 152≈ 139
1846 · approximate figure recovery≈ 154≈ 132
1847 · approximate figure recovery≈ 156≈ 122
1848 · approximate figure recovery≈ 157≈ 138
1849 · approximate figure recovery≈ 159≈ 142
1850 · approximate figure recovery≈ 160≈ 144
1851 · approximate figure recovery≈ 162≈ 145
1852 · approximate figure recovery≈ 164≈ 150
1853 · approximate figure recovery≈ 167≈ 143
1854 · approximate figure recovery≈ 169≈ 136
1855 · approximate figure recovery≈ 172≈ 137
1856 · approximate figure recovery≈ 174≈ 140
1857 · approximate figure recovery≈ 176≈ 142
1858 · approximate figure recovery≈ 179≈ 151
1859 · approximate figure recovery≈ 181≈ 153
1860 · approximate figure recovery≈ 184≈ 142
1861 · approximate figure recovery≈ 187≈ 141
1862 · approximate figure recovery≈ 185≈ 145
1863 · approximate figure recovery≈ 188≈ 152
1864 · approximate figure recovery≈ 192≈ 159
1865 · approximate figure recovery≈ 199≈ 161
1866 · approximate figure recovery≈ 202≈ 155
1867 · approximate figure recovery≈ 196≈ 148
1868 · approximate figure recovery≈ 200≈ 152
1869 · approximate figure recovery≈ 196≈ 160
1870 · approximate figure recovery≈ 206≈ 162
1871 · approximate figure recovery≈ 215≈ 163
1872 · approximate figure recovery≈ 215≈ 165
1873 · approximate figure recovery≈ 216≈ 172
1874 · approximate figure recovery≈ 223≈ 180
1875 · approximate figure recovery≈ 224≈ 181
1876 · approximate figure recovery≈ 224≈ 181
1877 · approximate figure recovery≈ 221≈ 182
1878 · approximate figure recovery≈ 224≈ 184
1879 · approximate figure recovery≈ 214≈ 190
1880 · approximate figure recovery≈ 233≈ 186
1881 · approximate figure recovery≈ 238≈ 191
1882 · approximate figure recovery≈ 242≈ 192
1883 · approximate figure recovery≈ 245≈ 192
1884 · approximate figure recovery≈ 240≈ 203
1885 · approximate figure recovery≈ 233≈ 210
1886 · approximate figure recovery≈ 232≈ 212
1887 · approximate figure recovery≈ 239≈ 218
1888 · approximate figure recovery≈ 245≈ 224
1889 · approximate figure recovery≈ 255≈ 236
1890 · approximate figure recovery≈ 256≈ 244
Allen (2009), Figure 1 — author’s PDF · Crafts (2022): reassessing real-wage growth
Historical estimates recovered from a published figure. Rounded readouts avoid spurious precision.

The historical estimates are contested; later work by Nicholas Crafts revisits the wage and productivity evidence. The figure here is an approximate reconstruction of Allen’s published graph, not a newly assembled set of raw statistics. The broader question survives those qualifications: what happens to the people waiting for progress to reach their paychecks?

In Acemoglu and Johnson’s account of the shared prosperity after the Second World War, rising demand for labor mattered alongside institutions that gave workers a stronger claim on production. What made the difference was that people had acquired the ability to bargain.

My review of This Time Is Different looked at the borrowing that sustains a technological boom. Power and Progress asks what remains unresolved even if the investment succeeds. A useful technology, a profitable company, and a better life for its workers are three outcomes. One does not guarantee the next.

What Would Bring a Platform to the Bargaining Table?

In key parts of twentieth-century manufacturing, unions had a concrete source of leverage. When workers stopped a critical stage of production, expensive equipment sat idle and the loss could spread through the factory. The 1936–1937 Flint sit-down strike, which helped secure GM’s recognition of the United Auto Workers, shows the force of that position. Organization, law, and political support mattered too. But workers could impose a cost their employer could not easily ignore.

People exposed to an AI platform do not all share an employer or a production line.

A translator, a designer, and an office worker may face related pressures while working for different clients and companies. The software that could reduce demand for their work may also be the tool they need to remain competitive. If one person stops using it, the immediate inconvenience may fall mostly on that person. A common anxiety does not yet amount to a capacity for common action.

The conflict also runs through each of us. As consumers, we welcome cheap, convenient AI. As workers, we worry about what it will do to the price of our labor. As citizens, we may dislike the concentration of power while choosing those companies’ services every day.

That is more than an awkward inconsistency. Giving up a useful service imposes a cost now; the benefit of a less concentrated market is uncertain and shared with everyone else. Individual choices can therefore be quite reasonable while reinforcing the dependence that makes collective resistance difficult.

This conflict predates AI. What matters here is how dependence on the same platforms extends across occupations without bringing those people into a common employment relationship. Nor is the platform’s power confined to the workplace. In its April 2024 assessment, the UK Competition and Markets Authority raised concerns about large firms’ positions across computing, data, models, and routes to market. Switching one application may leave much of that structure intact.

03 / Power and Progress

Big Tech spans several layers of AI

Selected relationships in CMA’s April 16, 2024 report · 15 nodes, 20 typed connections

A 2024 snapshot, not a map of today’s contracts. Investment, compute provision, data access and model distribution are different relationships. Lines do not measure transaction size or control.
InvestmentCompute provisionData-access agreementModel distributionIn-house model development

Select a name, click a node, or focus a node and press Enter. On narrow screens, swipe horizontally across the diagram.

Selected AI relationships documented by CMA in April 2024Google → Anthropic: InvestmentAmazon → Anthropic: InvestmentMicrosoft → OpenAI: InvestmentNvidia → Cohere: InvestmentGoogle → Anthropic: Compute provisionAmazon → Anthropic: Compute provisionMicrosoft → OpenAI: Compute provisionReddit → Google: Data-access agreementShutterstock → Meta: Data-access agreementMeta → Meta / Llama 2: In-house model developmentOpenAI → Microsoft Azure: Model distributionAnthropic → Amazon Bedrock: Model distributionAnthropic → Google Vertex AI: Model distributionCohere → Amazon Bedrock: Model distributionMistral AI → Microsoft Azure: Model distributionMistral AI → Amazon Bedrock: Model distributionMistral AI → Google Vertex AI: Model distributionMeta / Llama 2 → Microsoft Azure: Model distributionMeta / Llama 2 → Amazon Bedrock: Model distributionMeta / Llama 2 → Google Vertex AI: Model distributionLarge technology companiesModel developers and data providersCloud distribution of modelsGoogleAmazonMicrosoftNvidiaMetaRedditAnthropicCohereOpenAIMistral AIMeta / Llama 2ShutterstockAmazon BedrockMicrosoft AzureGoogle Vertex AI

All companies: 20 selected relationships

A selection from relationships documented in the CMA report, not the entire partnership network. Arrows indicate the direction of the named relationship, not a common measure of ownership. The diagram does not establish unlawful monopoly power.
View the relationships and report references
FromToRelationshipReport reference
GoogleAnthropicInvestmentp.31, Figure7 (investment + partnership)
AmazonAnthropicInvestmentp.31, Figure7 (investment + partnership)
MicrosoftOpenAIInvestmentp.31, Figure7 (investment + partnership)
NvidiaCohereInvestmentpp.33-34, para.2.59, footnote158
GoogleAnthropicCompute provisionp.32, para.2.59, footnote149
AmazonAnthropicCompute provisionp.32, para.2.59, footnote149
MicrosoftOpenAICompute provisionp.32, para.2.59, footnote149
RedditGoogleData-access agreementp.33, para.2.59, footnote150
ShutterstockMetaData-access agreementp.33, para.2.59, footnote150
MetaMeta / Llama 2In-house model developmentpp.21-22, Figure3 (Meta’s Llama 2)
OpenAIMicrosoft AzureModel distributionp.21, Figure3
AnthropicAmazon BedrockModel distributionp.21, Figure3
AnthropicGoogle Vertex AIModel distributionp.22, Figure3
CohereAmazon BedrockModel distributionp.21, Figure3
Mistral AIMicrosoft AzureModel distributionp.21, Figure3
Mistral AIAmazon BedrockModel distributionp.21, Figure3
Mistral AIGoogle Vertex AIModel distributionp.33, para.2.59, footnote151
Meta / Llama 2Microsoft AzureModel distributionpp.21,33, Figure3 / footnote151
Meta / Llama 2Amazon BedrockModel distributionp.21, Figure3
Meta / Llama 2Google Vertex AIModel distributionp.22, Figure3
CMA: AI Foundation Models Technical Update Report, pp. 21–22, 31–34 (April 16, 2024). Relationships may since have changed.

So who organizes these people, around which demand, and with what leverage? Bargaining with an employer over the terms of AI adoption and challenging a platform’s market position are connected tasks, but they require different points of intervention.

The book’s last chapter does offer policies: tax rules that stop favoring capital over labor, breaking up or regulating the largest tech firms, public research on AI that helps people, and stronger unions and civic groups. A later paper with David Autor, Can We Have Pro-Worker AI?, develops a similar agenda. These are concrete proposals. My difficulty comes one stage earlier: assembling the force needed to pass and enforce them against beneficiaries of the existing arrangement.

A list of necessary policies is not an explanation of how the people who need them acquire the power to win them.

The authors’ answer is historical. In the Progressive Era, journalists, reformers, and voters changed how Americans saw the trusts, built organizations, and turned that pressure into antitrust law. A new story about technology, one built to serve people instead of replacing them, is meant to do the same work today. My doubt is that narrative and policy are themselves shaped by money, and lobbying already filters what reaches a legislature.

There is evidence that bargaining over AI can work. In its 2025 case studies of social dialogue, the International Labour Organization (ILO) describes worker representatives influencing its introduction. The study also finds stronger results where employers’ exit options are constrained and collective worker voice is supported. Those conditions are the difficult part. Successful bargaining in some workplaces does not establish how fragmented groups can gain comparable influence over a platform on which they all depend.

Where Does the Power to Reform Come From?

Walter Scheidel’s The Great Leveler makes Acemoglu and Johnson’s optimism harder to sustain.

Scheidel connects history’s most dramatic reductions in inequality with mass-mobilization warfare, transformative revolution, state collapse, and lethal pandemics. His historical argument puts destruction alongside distribution: concentrated fortunes and entrenched power have often been shaken by events that broke much else with them.

Read together, the books illuminate different parts of the postwar settlement. Acemoglu and Johnson emphasize how bargaining institutions shared economic gains. Scheidel asks what gave those institutions such extraordinary momentum. War destroyed assets, but it also expanded taxation and state intervention and changed what could be demanded of the wealthy. Institutions and shocks meet here. Explaining what good institutions achieved does not fully explain the conditions that made them politically possible.

Scheidel is not claiming that peaceful reform accomplishes nothing. In a 2020 interview, he acknowledges improvements without catastrophe, including in Scandinavia and Latin America. His pessimism concerns their scale: could those changes rival the great compressions of inequality produced by exceptional historical shocks?

That question becomes pointed when applied to AI. Firms with capital, computing capacity, data access, and distribution networks also have resources to resist rules that would weaken their position. Can gradual reform reverse that concentration when its opponents already possess so much of the power needed to shape reform itself?

Changing power requires changing institutions; changing those institutions already requires power. That is the circle I could not find a convincing way out of in Power and Progress.

That is no reason to welcome catastrophe. A society in which everyone is poorer has not necessarily improved because the rich lost more. War and collapse offer no guarantee that the next holders of power will be better. They make the peaceful alternative more urgent—and the explanation of how to achieve it more demanding.

Read Next

The Great Leveler, Walter Scheidel. A challenging companion to Power and Progress: read one for the institutions that can change distribution, and the other for the historical conditions that allowed those institutions to acquire force.

Final Thought

Power and Progress persuaded me to distrust the optimism that technology will make everyone better off on its own. I remain unconvinced by the further optimism that people can acquire the power to turn that technology toward the outcomes they want. That is why I share its diagnosis so strongly and still hesitate at its remedy.

In

, , , , , , , , ,

One response

  1. This Time Is Different Book Review — Reinhart & Rogoff on Financial Crisis

    […] review of Power and Progress takes up the question that remains even if the investment succeeds: who shares in the […]

    Reply

Leave a Reply

Your email address will not be published. Required fields are marked *

Menu

Markets Books Films About Contact

Contact

excalibur9118@gmail.com

Socials

© 2026 Journey in Garret. All rights reserved.