AI: All that shines may well not be gold
Michael Roberts is an Economist in the City of London and a prolific blogger .
Cross-posted from Michael Roberts’ blog
Alphabet (Google) announced its earnings results yesterday. At first sight, the results were spectacular: revenue was up 24%, with revenue from its ‘cloud’ business up 82%. Earnings per share were $9.11, which is very high. But more than two-thirds of those earnings were based on booking $77bn from ‘unrealised’ gains from its purchase of shares in Anthropic, a major AI model firm. In other words, these ‘gains’ are just ‘on paper’ and depend on Anthropic shares staying up.
Google is one of the so-called ‘hyperscalers’, the mega tech companies that are ploughing huge of amounts of funds into AI, hoping and expecting it to lead to a sharp rise in profits down the road. Google’s investment in AI has pumped so much money into AI companies like Anthropic and OpenAI and into data centres to run AI with expensive ‘chips’ made by Nvidia, that its huge cash revenues are no longer covering the cost of investment. ‘Free cash flow’, as it is called, went negative for the first time in Alphabet’s history.
Google is increasing its AI investment massively this year, like its rivals, Meta, Microsoft and Amazon, to build AI infrastructure, with the four hyperscalers combined on track to spend more than $725bn in 2026. Before Wednesday’s results, Google had been seen as the hyperscaler best placed to withstand the AI arms race, with cash flows from its vast search business expected to cushion the financial pressure. But now, its cash burn to fund AI capex will require raising debt and/or issuing equity shares. Alphabet has already taken on nearly $100bn in debt, and in June it moved to raise about $85bn in its first share sale in more than two decades.
All this tells us that the AI stock market ‘bubble’, and that is what it is, is getting closer to bursting. The hyperscalers are hugely profitable, based on their existing businesses, with current earnings about 59 per cent above trend. But such is the size of the AI investment boom being conducted by these hyperscalers, that even these profits are being sucked into AI like water disappearing into the Sahara desert.
As argued in previous posts, t he US economy is one big bet on AI. The stock market value of the hyperscalers now accounts for roughly 40 per cent of the S&P 500’s total market capitalisation. If there is any sign that: first, they cannot sustain their current investment growth; and second, they are eating away their profits with no return on those AI investments, their stock values could turn south and take the whole stock market with them.
The US capitalist economy is balanced on the apex of AI. All will be well if: first, the AI models become in heavy demand and start to be used across companies in the US and globally, thus boosting profitability for the hyperscalers and eventually the rest of the corporate sector. Second, the generalised use of AI technology in all sectors of the economy also leads to a step change in the level of productivity, which can take the US economy into a new age of prosperity.
Before the advent of the AI boom, the US information technology had already become the driver of relatively faster economic growth in the US compared to the rest of the G7 economies during the Long Depression of the 2010s.
But in the 2010s, all the G7 economies, including the US, saw a slowdown in the growth of real GDP output, investment and productivity. Then, after the pandemic slump, inflation returned in most economies, threatening to introduce a new period of ‘stagflation’ (slow or no growth alongside faster and high inflation) not seen since the 1970s. Now the Iran war is accelerating energy price inflation. Central bank monetary policy failed to get economies going in the 2010s with low or zero interest rates and monetary injections (quantitative easing). And central bank monetary policy is failing to keep inflation from rising since 2020 using higher interest rates (ECB) and quantitative tightening (BoE). That’s because the only way economies can grow without inflation rising is by increasing growth in the productivity of labour.
The new Trump-appointed chair of the Federal Reserve expects a productivity boom from AI. Kevin Warsh says that “AI is perhaps the most significant change in our economy in my adult lifetime. AI will be a significant disinflationary force, increasing productivity and bolstering American competitiveness.” He reckons that productivity improvements from AI would drive “significant increases in real take-home wages. A one-percentage-point increase in annual productivity growth would double standards of living within a single generation.” He is right. Even just a 1% point rise in current US productivity growth over 20 years would not only keep inflation down, it would provide income and revenue that could end the deficits in government budgets, without raising taxes or making cuts in spending. Public debt to GDP ratios would fall.
But will AI deliver this step change in productivity? From 2005 through 2019, US productivity growth averaged about 1.5% per year. The pace remained similarly slow during and immediately after the pandemic (2020-2022). But US labour productivity has accelerated since 2022. Output per hour grew around 2.5% per year from the end of 2022 to the start of 2026, exceeding its pre-pandemic pace by 1 percentage point.
So is AI delivering? Most of the acceleration in labour productivity growth comes from faster growth of what mainstream economics calls total factor productivity (TFP). TFP is not something real that can be measured; it is just a mathematical residual from analysing the drivers of productivity growth. More workers working harder plus more machines working longer is what delivers most productivity from labour. TFP is the residual that is assumed to be from the impact of ‘innovation’ (eg AI).
But the rise in US TFP is not yet the result of AI application in the economy. According to economists at Barclays , AI adoption has been gradual and steady rather than rapid and transformative, with most households and businesses still reporting limited exposure to the technology. The rise in productivity growth since 2022 is really due to more intensive use of existing technology after the end of the pandemic, not the introduction of AI.
How do we know this? Well, Barclays economists looked at the St Louis Fed’s nationwide Real-time Population Survey (RPS) of working-age US adults. The most recent survey shows that, while AI has found its way into work routines for 45 per cent of respondents, the usage of AI in time is miniscule. Assuming an eight-hour workday, the average AI adopter has gone from roughly 20 minutes of usage in late 2024 to around 30 minutes by mid-year 2026, the RPS survey shows. And what’s not known is whether the workers are using their extra 2 per cent of work time per day to be more generally productive, or to verify and fix whatever the AI has produced, or to slack off!
Indeed, most surveys of AI adoption by companies show only modest progress. The US Census Bureau’s bi-weekly Business Trends and Outlook Survey shows just 21 per cent of businesses were knowingly using AI, 69 per cent reported no use, and 11 per cent weren’t sure either way! A Fed Board of Governors discussion paper finds that “productivity trends across all three levels have been relatively consistent over time, suggestive of micro-level productivity gains not adding up in aggregate.” It seems that ‘micro-level’ experiments typically measure task-level productivity, like the speed that a programmer generates code, rather than job-level or firm-level output. A 10% improvement on a task does not necessarily lead to proportional gains for a firm if adjustment costs or other bottlenecks lie elsewhere in the production process and erode the upstream productivity gains.
By running a series of complicated regression analyses on RPS survey data, the US Fed found that no evidence that AI’s doing anything positive: correlations between industry-level adoption rates and improvements in productivity are “not statistically distinguishable from zero”, it says. Any observable improvement “largely reflects persistent differences across industries rather than a measurable acceleration in productivity growth attributable to AI”.
However, the AI optimists remain just that – optimistic. They refer to the ‘J-curve’ that the productivity impact of new technologies generally follow. First, there is a slow, even negative, effect on productivity as companies invest heavily in the technology. And then the boom comes. Such a J-curve was seen in US manufacturing productivity growth, which fell in the mid-1980s and then, after the recession of 1991, accelerated sharply until the mid-2000s.
But even that optimistic view has its caveats. First, it seems that those workers exposed to AI are resisting its use. An April survey of 2,400 knowledge workers by AI firm Writer and Workplace Intelligence—both firms with commercial stakes in AI adoption — found the 29% of employees admit to actively sabotaging their company’s AI strategy . Among Gen Z (under 25) workers, that figure was 44%, up from 41% a year earlier. A separate WalkMe survey of executives and employees across 14 countries, conducted the same month, found that more than 54% of workers had bypassed their company’s AI tools in the past 30 days to do the work manually instead. The survey commissioners found that a “Fear of Becoming Obsolete” is driving much of this active and passive, and even passive aggressive resistance . Similarly, The Economist reported that AI usage among US workers, after an early spike, dipped as initial enthusiasm faded.
In the early 19th century at the start of the industrial revolution in Britain, a group of skilled weavers tried to resist the introduction of machine weaving in the early 19 th century by various means, including sabotaging and wrecking the machines. They were called Luddites. Now it seems that a form of Luddism has returned over AI. Just as the Luddites had a case about protecting their livelihoods, so do Gen Z workers now.
The AI optimists, Stanford’s Erik Brynjolfsson and ADP Research , are tracking 4.6 million workers across more than 730 occupations through their Canaries Dashboard . They find that jobs for workers aged 22 to 25 in AI-exposed occupations are shrinking more than 4% annually. Goldman Sachs analysis suggests information, professional services, insurance, and finance are best positioned for early productivity gains—but these are where the ‘sabotage’ surveys find the highest rates of resistance.
Goldman Sachs US economist Elsie Peng supports the J-curve thesis for AI. Her study of industry technology adoption found a modest drag for the first four years, statistically significant gains only after eight, and a peak impact of roughly 0.6 percentage points in year 12. If ChatGPT’s 2022 launch is the equivalent of the PC’s 1981 debut, that J-curve puts the productivity payoff arriving around 2030 at the earliest, and peaking around 2034. According to Peng’s analysis, significant labour productivity boosts from ICT didn’t show up until roughly 50% of businesses had adopted the technology. Not bought it, not piloted it. Actually adopted it into their core operations.
Peng’s ‘J curve’ echoes what the OECD’s economists argued some time ago, that transformative technologies historically take about 20 years from their breakthrough moment to deliver meaningful productivity gains at the macroeconomic level. If that pattern holds for generative AI, we’re looking at the early-to-mid 2030s or even later before the real payoff arrives. So there is a long way to go for generalised AI that can deliver a significant rise in labour productivity growth.
The step change in productivity growth that Fed chair Warsh puts all his hopes on still seems some way off. But what about profits from AI? So far, any revenue that the AI model companies are making is way, way short of covering the costs of AI R&D development and the building of data centres all over the US. And Goldman Sachs estimates that each dollar of ICT hardware investment requires at least another $1.70 of complementary “intangible” investment—software, data systems, and the hardest category to measure, organizational overhaul. There is now 15 times more data centre capacity than the demand to use them.
And debt is building up. B loomberg estimates that there’s over $500 billion in outstanding AI data centre debt. Nikkei Asia reported this week that Meta, Google, Amazon, Microsoft and Oracle have accrued around $1.65 trillion in outstanding debt in the last five years, with an additional hundreds of billions of dollars’ worth of “off balance sheet” debt, meaning that the corporate structure allows the company to not include it as part of its liabilities.
What revenues are being collected by the likes of OpenAI and Anthropic are just coming from the hyperscalers’ investments in their operations. There are no profits at all being made from the use of Chat GPT or Claude, partly because the AI companies are not charging the proper cost of using ‘tokens’ (compute units) to the users. So the AI companies are now trying to switch consumers from flat fees to usage-based pricing. But this has led to an exponential rise in the cost of using AI for companies, such that companies faced with ballooning AI bills are moving from“token maxxing” to ‘token rationing’. And US companies are switching to using ‘open-source’ AI models coming out of China that can nearly match the performance of the US models at a fraction of the cost.
First, there was DeepSeek that strikingly hit the industry back in early 2025. Now there is the newest Chinese AI model, Kimi K3, just released by the Beijing-based startup Moonshot AI. Chinese models are 112x cheaper than Anthropic per million tokens (ie. “barrel of intelligence”). One token costs $56 from Anthropic, $26 from OpenAI, $1.50 from Meta, $1 from xAI and Google, and $0.50 from the Chinese models. No wonder the proportion of tokens used by US firms that run through Chinese AI models is up to a record 58%. American companies now use Chinese AI models more than US-made ones.
But the optimists have not given up. Some suggest the so-called Jevons paradox that argues that increased efficiency (from AI) will lead to increased demand as unit costs of spending on AI falls. That will generate the profitability that AI companies are seeking and the stock market is hoping for. But only 2% of S&P 500 companies mentioned AI productivity during Q1 2026 earnings calls. And among those that did, the focus was overwhelmingly on cost savings rather than revenue growth. This raises serious questions about how the hundreds of billions in investment can be expected to turn into profits, let alone revenue.
The AI bet rests on two big assumptions. The first is that AI will be profitable – eventually. But just because a technology leads to a huge increase in productivity doesn’t mean it will generate strong returns. The second assumption is that there will be widespread demand for AI, and soon. AI adoption has risen, but it is still a long way from peak adoption. As above, Goldman Sachs estimates that it could take as long as 15 years; the OECD says 20 years. Can the current AI companies survive that long; can the stock market wait that long before the bubble bursts?
__________________________________________________________________
BRAVE NEW EUROPE is one of the very few Resistance Media in Europe. We publish expert analyses and reports by some of the leading thinkers from across the world who you will not find in state and corporate mainstream media. Support us in our work
To donate please go HERE
Michael Roberts – AI and productivity
Aggregated summary from an independent source. Read the original at BraveNewEurope.