Leevi SAARI

PhD candidate, University of Amsterdam, EU Policy Fellow, AI Now Institute

Since the launch of ChatGPT in November 2022, the tech intelligentsia has been locked in an elaborate dance between bulls and bears. The bulls manifest a technological cornucopia brought by AI just over the horizon. The bears point to the mundane facts on the ground: staggering capital burn, plateauing capabilities and a persistent lack of profits.

The arc of reason is bending towards the bear case. A motley crew of venture capitalists, investment bankers, Nobel-winning economists and even tech CEOs themselves are converging on the same reality: the business case for current large-scale AI is missing. However, this has spurred a,  debate between optimistic and pessimistic bears. While both agree that bubble is real, optimists argue, that it still does yield long-term productive benefits – much like the railway manias of the 1800s, which left behind a crucial infrastructure long after the initial speculators were burned. The pessimists point to the Dutch tulip craze of the late 1600s, where the mania suddenly evaporated, leaving behind only a gust of hot air.   

Both perspectives, however, share the same core flaw: they treat a ‘bubble’ as an event. This implies a singular reckoning; a thundering ‘pop’ that shatters the present. Instead, I suggest we see this moment as a process: a source of continuous change, where political and economic interests hurl history forward with violence. And we are already in the midst of this moment.

The bear case

Scepticism regarding large-scale AI models is grounded in the lack of organic demand, adverse unit economics and the overstretched supply chain. Despite clenching capital expenditures, the organic, valuable end-user demand remains hard to find. Last August, a study by the Massachusetts Institute of Technology rattled markets by noting that 95 per cent of generative AI efforts are failing. Late March, OpenAI closed off Sora, its video-generation engine, due to extreme costs yielding only meagre profits. The search for customers willing and able to pay the full cost of these services remains elusive.

One of the key problems is unit economics. Unlike traditional software, where an additional user costs practically nothing, every new large language model query has a fixed marginal cost. This shift makes AI companies look less like high-margin software firms and more like capital-intensive utilities or infrastructure providers. This also means that, to make the service affordable for customers and not bleed market share to competitors, companies are burning cash. For example, the current market leader in AI coding, Anthropic, is estimated to gift its customers 25 dollars for every dollar spent. OpenAI is expecting to lose 115 billion by 2029, before making a single dime of profit.

The primary justification for flaunting these actuarial niceties has been the march toward artificial general intelligence (AGI), a term sometimes used to describe AI capable of replacing all economically valuable human work. Achieving that would make traditional financial calculations quaint. In May 2024, the CEO of OpenAI, Sam Altman, noted he “doesn’t care” whether the company will spend 5 or 50 billion a year, as the resulting AGI will be “totally worth it”. Yet, in 2026, AGI remains elusive. The model capabilities are beginning to plateau. ChatGPT’s new editions have been underwhelming. On 9^thof April 2026, OpenAI tightened the usage for existing subscriptions to contain expenses. According to popular benchmarks, leading AI labs are converging to a similar capacity 

Hence, the competition is not any more who is the first to conjure the ‘frontier’, godlike model from the depths of neural networks. Rather, it has returned to the capitalist fundamentals: who has the most entrenched distribution, the infrastructure, the best product – and the deepest pockets to fund all this.

Beyond economic and technical limits, AI expansion is hitting physical limits. AI models live in a sprawling, complex and largely physical network that includes chips, energy grids and raw materials. These resources are stretched to the breaking point. In a surreal example, the shortage of gas turbines for data centres has led some firms to repurpose aircraft engines to generate power on-site. The tension is increasing by the day, with makeshift patches covering the leaking cracks. Systemic shocks, such as increase in gas prices due to conflict in Iran, might stretch the system past its limits

The boulder that rolls 

However, the contradictions do not mean that the momentum comes to a sudden halt. Capital has gravity. Trillions worth of capital investment will find its profit, for better or worse. Strategic thinking complicates the situation. AI is now a geostrategic imperative, with states entangled in the current buildup. Tech giants are locked in fear of missing out, willing to absorb massive losses to avoid falling behind. Hence, instead of waiting for a white flash and a pop on the horizon, the boulder is already rolling downhill. Looking around, we can already see the trail it has left behind.

We see it in market consolidation, where sweltered startups are absorbed by the incumbents. We see it in the billions of venture capital money, masking the lack of demand. We see a self-dealing economy where chip giants fund startups that, in turn, use that capital to buy the giants’ own chips. We see it in the race for retail, where the AI startups seek to list on public markets and offload their gory balance-sheets to the hands of the excited masses, absolving their founders. We see its traces in the extraordinary investment rounds for AI startups explicitly not working on large language models. And we see it in the merging of public and private interests, as states shift to a role of backers that resuscitate their national AI champions as a matter of geopolitical imperative. 

The messy remaking is upon us

The large-scale AI bubble is not waiting to burst; it is already mangling the world. As the cycle matures, the real story is not the ‘pop’, it’s the messy and ongoing transformation occurring while we wait for it. And the deeper the contradictions, the more violent the hurling will be – the more extreme the distribution between winners and losers and the more aggressive bending of the reality to accommodate the convulsions of this super cycle. And in that bending, societies will be remade.

Photo credits: Shutterstock/gualtiero boffi