Inteligência artificial: a maior aposta económica da história dos EUA . Robert´s Blog

A economia dos EUA continua a expandir-se mais rapidamente do que as outras economias do G7, mas o principal fator é o enorme investimento em modelos de IA, centros de dados e todos os chips e tecnologias relacionados com IA. 

O PMI composto dos EUA (medida da atividade económica) subiu para 58,4 em agosto, contra 56 em agosto, a maior expansão da atividade do setor privado desde julho de 2021 e marcando o quarto mês consecutivo de crescimento acelerado. Os ganhos foram impulsionados pelo setor de serviços (incluindo serviços de informação), com o maior aumento na produção em mais de cinco anos, enquanto o setor transformador também acelerou. Os novos pedidos cresceram no ritmo mais rápido desde abril de 2022, enquanto as contratações no setor manufatureiro foram as mais fortes desde fevereiro de 2021.

Em junho, comentei que a IA era apenas "uma grande transação para a economia dos EUA ". Mas agora, em setembro, isso parece ser um eufemismo. A expansão da IA ​​está a caminho de se tornar a maior aposta econômica da história dos EUA, superando em muito os investimentos feitos para financiar outros grandes projetos de infraestrutura no passado, como as ferrovias no século XIX ^, o sistema rodoviário no século XX ^e a internet no século ^XXI .

Analistas estimam que os investimentos de capital em cinco dos chamados hiperescaladores — Alphabet , Amazon.com , Meta Platforms , Microsoft e Oracle — chegarão a US$ 4,2 trilhões nos quatro anos que terminam em 2029, segundo a FactSet. Os gastos com data centers são maiores do que os gastos com canais, ferrovias e rede elétrica combinados, projetados para totalizar US$ 10,3 trilhões de 2025 a 2032, de acordo com novas estimativas da Brookings Institution. Isso representa uma média impressionante de 3,6% do PIB por ano. Nunca antes a economia dos EUA foi tão dependente da expansão de um único setor.

Até julho, foram gastos 37 bilhões de dólares em centros de dados privados, sendo que a maioria ainda não está em operação.

Em contrapartida, nos Estados Unidos, os gastos com construção privada em todos os outros segmentos — casas, prédios de apartamentos, centros comerciais e assim por diante — ficaram cerca de US$ 46 bilhões abaixo dos níveis do ano anterior nos primeiros sete meses deste ano. 

Segundo estimativas do LinkedIn, o investimento em IA criou 750 mil novos empregos desde 2023. E esses empregos são bem remunerados: o salário anual médio para vagas relacionadas à IA anunciadas no LinkedIn gira em torno de US$ 180 mil, em comparação com US$ 80 mil para todos os empregos.

Above all, the AI investment has led to huge gains in stock-market wealth. As of Q2 2026, US stock and mutual fund holdings came to $63 trillion, according to the Federal Reserve—nearly double the amount at the end of 2022. Most of this increase in financial wealth has gone to the already rich, as working people own little stocks or bonds. 

Foreign investors are piling into US assets.  They now hold a record $39 trillion in US equities and bonds, up since 2022.. This is keeping the US dollar relatively strong and driving up stock prices.  The wars in Ukraine and Iran encourage foreigners to shift their assets to the US to take advantage of the boom.

At the same time, demand for equipment that goes into data centres like memory chips is driving up costs for tech products. Import prices on computers, peripherals (such as hard drives) and semiconductors were 20% higher in August than a year earlier. These high import prices are in turn putting upward pressure on the costs of consumer goods, such as iPhones and gaming consoles, and contributing to general inflation. 

But here is the problem. The gap between hyperscaler spending and cash flow is widening fast.  Capital expenditures at Amazon, Meta, Microsoft, and Alphabet are projected to exceed $1 trillion in 2027 for the first time. At the same time, combined ‘free cash flow’ (ie money from profits in existing businesses) is projected to fall below $100 billion. A year ago, free cash flow was around $200 billion, while capex was $300 billion. Now, AI spending is accelerating at the same time as the cash available to fund it is disappearing.

The bigger this gap becomes, the more the hyperscalers need to rely on debt and equity markets to finance their AI spend.

The issue is that if AI spending fails to generate sufficient returns (profits), the stock market could take sharp turn downward as investors bail out.  US stock market prices are massively overvalued relative to existing earnings. The trend ratio of stock market prices to earnings per share (called the CAPE ratio) is above the level just before the 2008 financial crash and nearly at the level just before the dot.com bust of 2000.

Will profits come through? Research by Fathom Consulting shows that for the multitrillion-dollar AI boom to turn a profit, it would need the AI-related sales of the tech companies involved to rise by $600-800bn within the next two years. But the consulting firm Panmure Liberum calculated that current CAPEX and revenue forecasts through 2030 imply a negative internal rate of return on invested capital for Alphabet, Meta, Microsoft, and Oracle.

So either the hyperscalers significantly reduce their capital spending on AI to levels that generate a reasonable profit on capital already invested or by some miracle they deliver massive profitablity from a huge future increase in demand for AI products.  If they cut spending, that would signal to investors that AI is not delivering and they would sell off accordingly.  A crash would ensue.  So they must keep spending more and more.

At the same time, what companies can charge for AI computing costs (tokens) is falling fast. The LLM Token Expenditure Index, which tracks the market price companies pay for AI model output, has fallen to just $0.97, its lowest level since the index was created late last year and more than 50% below its summer peak. Token prices are collapsing as cheaper models, open-source Chinese competitors and falling training (inference) costs make AI usage increasingly cheap. That is eroding revenue growth for the AI labs, making it more difficult to meet the bills for AI infrastructure spend.

The AI labs (OpenAi and Anthropic) continue to claim they will soon make big profits and so the hyperscalers will eventually get their share of the booty. But much of these claims are based on dubious profit estimates. AI-related investment gains increasingly flatter earnings, with so-called “other income” (contracts with other AI firms) rising to 54% of pretax income.

Indeed, the AI companies are keeping their heads above water only through what is called  ‘circular financing’ where one firm lends funds to another and the latter then claims it has made a profit. Sona Asset Management have mapped the AI universe and catalogued the interconnections between major players. Everybody is depending on everybody else to deliver.

 

Sona also found that AI firms’ revenue is almost completely tied to the capex decisions of one or two other AI players. This is a systemic bust in the making.

A key question is whether AI is actually going to deliver a step-change in US labour productivity that could boost economic progress for a generation. The AI lab, Anthropic, wants to issue shares worth $100bn to the public in November (thus valuing the company at $2trn!).  To build up its case, it published a report in which it claimed that if AI really takes off, US GDP could rise by 32% by 2030(!), that’s annual growth in GDP of up to 15% (against current US growth at 2.5% at best).

This is wild nonsense that assumes that AI works in boosting productivity growth as every company in the US adopts AI agents and tools to run their businesses, while sacking millions of workers who are no longer needed.

Historically, automation has historically proceeded at roughly 2% of tasks per year for two centuries, without ever pushing growth much above 2%. Past so-called ‘general-purpose technologies’  took decades to diffuse even after the technology itself worked. For example, electrification took around 40 years to show up in factory productivity. Similarly, Comin and Mestieri’s study of technology adoption across countries documents average adoption lags of around 45 years, and still 7-18 years for more recent technologies.

AI adoption appears to be much faster than that of the PC or the internet, but adoption is only the first step: measured productivity typically first falls while firms make the necessary complementary investments, the “productivity J-curve” of Brynjolfsson, Rock and Syverson. Even the most bullish insiders are noticing this, for example Sam Altman who recently conceded: “I think I was wrong about a few things, but one of them, in terms of the speed, one of them is the economy just has so much inertia. […] we’ve all been too ambitious on timelines […] Society and the economy will adapt more slowly.”

And remember most work is physical. An AI capability explosion is first and foremost an explosion in cognitive capabilities. But only around a third of the economy consists of work that can be done on a computer (Epoch AI’s remote-work piece). The rest of GDP is produced in mines, on construction sites, in kitchens, hospitals and care homes. Automating two thirds of all tasks by 2035 therefore requires robots that do a large share of physical work. These robots would need to be designed, manufactured, installed and maintained by the billions within a decade. While progress is certainly happening, robotics development is slower than software development and robots that can do a wide range of physical tasks at close to the cost of a worker are still not on the horizon.

So far there is scant evidence of economy-wide productivity gains; in fact, total factor productivity (a measure of productivity from new technologies) is falling below trend.

It’s true that AI adoption by companies is picking up, at least among service companies. In just two years, AI usage has risen from 25% to 61% among service firms and from 16% to 51% among manufacturers.  But only 17% of employees in services and 7% in manufacturing are actually using AI regularly in their work.

Como admitiu o Federal Bank de Nova York : “ A próxima etapa será muito mais importante, pois, por enquanto, a IA ajuda principalmente as pessoas a escrever, resumir, programar ou analisar mais rapidamente. Amanhã, agentes e ferramentas especializadas poderão lidar com fluxos de trabalho inteiros. Nesse momento, o impacto na produtividade e no emprego poderá realmente atingir outro patamar. O uso está explodindo, enquanto os gastos corporativos permanecem relativamente limitados, mas o preço da inteligência continua caindo. Isso é muito otimista para a difusão da IA ​​e para a produtividade, e bem menos otimista para a monetização de todos os participantes do ecossistema.” Em outras palavras, o crescimento da produtividade pode eventualmente aumentar, mas à custa da lucratividade, à medida que os preços da IA ​​para uso caem: uma contradição clássica do capitalismo.

E ainda existem falhas inerentes à própria natureza dos Grandes Modelos de Linguagem de IA. Há dois anos, pesquisadores de Oxford e Cambridge provaram que todo Grande Modelo de Linguagem treinado apenas com conteúdo gerado por IA desenvolve um distúrbio irreversível. Cada novo modelo aprende apenas com o que o anterior escreveu. Na nona geração, o novo modelo já havia esquecido do que estava falando. Pergunte a ele sobre torres de igrejas medievais e ele responderá com uma lista de lebres. Eles chamaram isso de "colapso do modelo". Cada vez que alguém publica a saída de uma máquina, o próximo modelo se torna um pouco mais mediano.

Os sinais de alerta de uma crise iminente continuam se multiplicando. Uma tendência de alta nas taxas de juros em meio à queda dos preços dos títulos pode ser o fator desencadeador da crise. Ou pode ser um IPO fracassado (Anthropic, OpenAI), ou a crescente concorrência dos chamados modelos LLM de "peso aberto" da China e de outros países, reduzindo os lucros. A maior aposta da história econômica dos EUA é extremamente arriscada.