Seven trillion six hundred billion dollars. That is the colossal amount that technology companies plan to invest in artificial intelligence infrastructure between 2026 and 2031, according to Goldman Sachs. A figure that exceeds the combined GDP of Japan and Germany. But the investment bank is beginning to wonder: will this titanic bet ever find profitability?
Market euphoria around AI is reaching dizzying heights. OpenAI valued at 900 billion dollars, Anthropic at 965 billion, SpaceX at 1750 billion. These figures defy all traditional economic logic for companies that still struggle to demonstrate sustainable business models. Goldman Sachs Global Institute is sounding the alarm: never in recent history has a sector mobilized so much capital with so little visibility on returns.
The Essentials
- 7,600 billion dollars in AI investments planned between 2026 and 2031, more than the combined GDP of Japan and Germany
- Current valuations of OpenAI (900 bn$), Anthropic (965 bn$) and SpaceX (1,750 bn$) exceed the market capitalization of century-old industrial giants
- Goldman Sachs identifies a growing gap between technological promises and evidence of economic profitability
- Investment per employee in AI exceeds that of other emerging technology sectors by 40 times
An Investment Without Historical Equivalent
The scale of the sums involved exceeds anything the global economy has ever known. To contextualize these 7,600 billion dollars, Goldman Sachs recalls that the Marshall Plan for the reconstruction of Europe after 1945 represented, in constant dollars, approximately 200 billion. The entire American space program since the 1960s cost only 650 billion in current value.
This investment frenzy is concentrated in three segments: specialized chips (2,400 billion planned), data centers (2,100 billion) and transmission networks (3,100 billion). Nvidia, which supplies 80% of AI processors, displays a market capitalization of 3,200 billion dollars, surpassing the entire global automotive sector.
But Goldman Sachs points out a troubling statistical anomaly: investment per employee in pure AI companies reaches 15 million dollars, compared to 350,000 dollars in traditional tech and 80,000 dollars in manufacturing. This disproportion suggests either revolutionary productivity to come, or a speculative bubble of unprecedented scale.
Revenues Fall Short of Promises
Despite investor enthusiasm, evidence of profitability remains sparse. OpenAI, the sector leader with ChatGPT, generates approximately 3.4 billion dollars in annual revenues for a valuation of 900 billion — a price-to-sales ratio of 265, where Amazon, in its most speculative early days, never exceeded 20.
Goldman Sachs’ analysis reveals that 73% of companies using AI observe no measurable improvement in productivity after 18 months of use. Use cases that generate significant revenues remain confined to niches: machine translation, medical image recognition, logistics optimization. Far from the promise of general economic transformation.
Microsoft, which invested 13 billion in OpenAI, struggles to monetize this alliance. Revenues from its AI services (Copilot, Azure AI) represent less than 2% of its total revenue. Google spends 30 billion a year on AI research but sees its traditional advertising margins stagnate in the face of competition from TikTok and other platforms.
This tension between massive investments and limited returns echoes the early days of the internet in the 1990s, but at an incomparable financial scale. The difference: Amazon and Google took fifteen years to reach profitability. AI investors promise results in three to five years.
The Industry Bets on Deferred Network Effects
Defenders of these valuations advance a sophisticated argument: AI would follow a non-linear adoption curve, where benefits would appear suddenly once a critical usage threshold is reached. This theory of “deferred network effects” would justify current massive investments.
Jensen Huang, CEO of Nvidia, claims that “every dollar invested in AI today will return a hundred dollars tomorrow.” He relies on the example of graphics chips, long confined to video games before becoming essential to scientific computing and cryptocurrency mining.
This strategy is not without precedent. Telecommunications followed a similar pattern: investments in fiber optics in the 1990s seemed absurd until the explosion of streaming and cloud computing twenty years later. Similarly, 4G network investments took a decade to be justified by the emergence of smartphones and mobile applications.
But Goldman Sachs identifies a crucial difference: telecommunications infrastructure served obvious human needs (communicating, moving, consuming). AI, despite its technical achievements, still seeks its truly indispensable mass applications. AI increases wages without destroying jobs, but has not yet revolutionized overall productivity.
Warning Signs Multiply
Several indicators concern Goldman Sachs analysts. The first: the extreme concentration of the sector. Four companies (Nvidia, Microsoft, Google, OpenAI) control 85% of the AI value chain, from the chip to the final service. This situation recalls industrial monopolies from the early twentieth century, but with multiplied market volatility.
The second signal concerns energy consumption. AI data centers already consume 4% of global electricity, a figure that could reach 12% by 2030. This growth calls into question the economic and environmental sustainability of the sector. OpenAI estimates that training its next GPT-5 model will cost 500 million dollars in electricity alone, not counting infrastructure.
The third warning point concerns talent. The AI sector employs approximately 350,000 highly qualified people worldwide for a market capitalization of 8,000 billion dollars. This apparent productivity per employee (23 million dollars) exceeds that of the oil industry (3 million) or pharmaceuticals (1.2 million), sectors nonetheless renowned for their barriers to entry and technological rents.
This statistical anomaly suggests either an unprecedented productivity revolution, or massive overvaluation. Goldman Sachs leans toward the second hypothesis, noting that AI companies massively externalize their real costs (energy, raw materials, manufacturing labor) while capturing financial value.
The Reality Test Approaches
The next eighteen months will determine whether this AI bubble follows the path of the internet (brutal correction followed by real growth) or that of telecom 2000 (prolonged collapse). Goldman Sachs identifies several decisive tests ahead.
The first test concerns enterprise adoption. Despite announcements, only 12% of Fortune 500 companies have integrated AI into their critical processes. Most limit themselves to experiments or cosmetic uses (chatbots, automatic summaries). True adoption would require massive complementary investments in training, organizational change and security.
The second challenge concerns regulation. The European Union is preparing an AI Act that could drastically limit the use of certain algorithms. China is developing its own champions (Baidu, ByteDance) to reduce its dependence on American models. This geographic fragmentation would complicate the emergence of a unified and profitable AI market.
The golden age of industry could well depend on AI’s ability to move beyond the stage of technical demonstration to create measurable economic value. Goldman Sachs remains cautious: “Technological history is dotted with promised revolutions that took decades to materialize, and sometimes never materialized at all.”
The banking institute concludes its analysis with a warning: if only 10% of planned AI investments find profitability by 2031, the sector will experience a correction on the order of 70 to 80%. A scenario that would make the current AI explosion the largest speculative bubble in modern history, surpassing the internet bubble of 2000.
One question remains: unlike previous bubbles, AI has unprecedented political and strategic support. The United States, China and Europe see it as a matter of national sovereignty. This geopolitical factor could artificially prolong the investment phase, delaying but amplifying the economic reality test.