Companies are spending hundreds of billions of dollars to develop artificial intelligence models, hoping to establish technological advantages that could translate into lasting commercial returns.
But as competitors release increasingly capable models within months of one another, investors face a difficult question: how sustainable is a strategy built around continuously spending more to stay ahead?
The rapid pace of innovation is creating opportunities for AI developers to attract users, win enterprise customers, and reshape established industries.
However, the short shelf life of technological leadership raises questions about whether the revenue generated by successive model upgrades can justify the cost of developing them.
Meta’s Muse has emerged as a prominent example of the commercial potential of AI.
More than 10 million people in the US and Canada had downloaded the Muse app, Chief Executive Mark Zuckerberg said during the company’s earnings call last week.
The milestone is notable because ChatGPT’s app took about three times as long to reach 10 million downloads, according to data from market-intelligence firm Sensor Tower.
Anthropic’s Claude Cowork, meanwhile, triggered a sell-off in software stocks in February as investors worried that AI could undermine traditional software business models.
These developments demonstrate the technology’s ability to attract users and challenge established industries.
Yet user adoption and market disruption do not necessarily translate into sustainable profits for the companies financing the underlying technology.
As Anthropic filed its S-1 registration earlier this week, it revealed an operating loss of $693 million in the first half of 2026.
Results of big tech companies like Meta, Alphabet, and Amazon also revealed huge cash spending for AI buildout.
The heavy cash expenditure has resulted in stocks plunging on Friday as investors rotated out of artificial intelligence-related stocks.
As the AI race intensifies, it becomes more important for AI providers to generate sustainable revenue and not just release the most advanced model.
AI spending is rising faster than the certainty of returns
The scale of investment required to compete in the AI market is becoming increasingly difficult to ignore.
Goldman Sachs estimates that the largest US AI hyperscalers are on track to spend $800 billion in capital expenditure in 2026, with consensus forecasts pointing to $1.1 trillion in 2027.
The bank expects capital expenditure to increase further in 2027 and potentially exceed consensus estimates, although the pace of growth and the size of positive surprises are expected to diminish compared with recent quarters.
According to Goldman Sachs’s equity analysts, hyperscalers will need to generate approximately $300 billion in annual AI revenue over the next few years to break even on their investments, based on average annual capital expenditure in 2026 and 2027.
There are reasons for optimism. Hyperscaler cloud revenues accelerated sharply in 2026, annualising approximately $70 billion above the pre-AI trend in the second quarter, while announced revenue backlogs exceeded $1.5 trillion.
Nevertheless, these figures do not eliminate the central uncertainty: how much of the revenue growth will translate into returns that adequately compensate companies and their investors for the cost and risk of the investments?
Alexander Lis, chief investment officer at investment firm SDV, believes the industry will need a substantial increase in sales to justify its spending.
“It is going to take massive additional revenues to justify the current aggregate CapEx levels of ~ $1.4 trillion per annum: from $5 trillion to $8 trillion in incremental sales (factoring in EBITDA margin and ROI assumptions),” Lis told Invezz.
He warned that the problem could emerge when investors begin to doubt whether such incremental sales are achievable.
For now, growing usage has helped compensate for falling token prices, he added.
But Lis cautioned that “Should AI adoption decelerate, massive CapEx would start to be a problem.”
When a better model becomes a short-lived advantage
The financial challenge is not simply that AI development is expensive.
It is that a technological lead may not last long enough to justify the investment required to achieve it.
A company can spend billions developing a flagship model, only for a competitor to release a more capable or cheaper alternative a few months later.
The first company’s technological advantage may consequently have limited value if customers can switch to a rival offering comparable performance at a lower price.
A Bloomberg report citing The Epoch Capabilities Index said the top spot among AI models changed 20 times since 2024.
A model has only remained as the frontier model for about seven weeks. That’s a very short time to be numero uno after spending billions of dollars.
Ilman Shazhaev, Founder and CEO of Dizzaract, argues that the economics of pursuing a technological lead have already become questionable as a strategy for maintaining a competitive advantage.
“The core misconception behind 2026 capital spending of roughly $765 billion is mistaking high use for high value,” Shazhaev told Invezz.
Even if lower prices encourage greater usage and keep computing clusters busy, high utilisation does not automatically translate into attractive margins.
Even if prices plunge and demand rises enough to fill clusters, a power plant runs at full capacity for its whole life and still makes utility margins. Meanwhile, hardware efficiency, algorithmic compression, and open-weight models matching closed frontier performance within four months trigger what I call the “Fixed-Capability Collapse.” This cuts the cost of serving a given level of intelligence roughly tenfold every year.
He argues that unless frontier workloads continue to require exponentially larger, tightly connected computing clusters, some of today’s infrastructure could become economically outdated by 2030.
The risk is particularly significant because AI hardware does not retain its economic value indefinitely.
Shazhaev estimates that AI accelerators could become obsolete within three to five years, while new chip generations arrive approximately every 12 months.
This creates a potential mismatch between the accounting treatment of expensive infrastructure and its actual economic usefulness.
Shazhaev estimates that the gap between accounting depreciation and economic reality could reach $176 billion between 2026 and 2028.
Open-ended models eating into closed-ended models’ lunch
Another threat for the top-end models is the cheaper open-ended models, which seem to be catching up to frontier models’ performance within months.
Chinese developers and open-weight models are adding another layer of competition, challenging the pricing power of companies developing proprietary, closed models.
As the performance gap narrows, customers have more reasons to consider alternatives that deliver comparable results at a fraction of the cost.
A Bloomberg News analysis found that Moonshot’s Kimi K3 could build the same e-commerce website as Anthropic’s Claude Fable 5 for one-fourth of the cost.
The comparison shows how cheaper alternatives could erode the commercial advantage of expensive frontier models, especially for customers whose workloads don’t require the most advanced capabilities.
The competitive pressure also extends beyond pricing.
Open-weight models allow developers to access and adapt models’ underlying weights, potentially giving businesses more flexibility in how they deploy AI.
That could make them particularly attractive in markets where customers have limited budgets.
AI routing company OpenRouter said the five most popular models on its platform were Chinese, suggesting that affordability and accessibility can influence adoption alongside technical performance.
For US AI developers, the risk is that continued investment in building the most capable models may not translate into a proportionate commercial advantage if customers increasingly choose cheaper alternatives.
The pressure is already influencing pricing decisions at leading AI laboratories.
In September, OpenAI and Anthropic introduced less expensive models as competition from open-weight alternatives intensified.
OpenAI added GPT-6 Sol and GPT-6 Luna to its GPT-6 family, cutting API prices by 50% compared with GPT-5.6 promotional pricing.
Anthropic, meanwhile, introduced Claude Opus 5.5, describing it as a more token-efficient model that would cost around 40% less to run than its Opus 5 predecessor.
These moves suggest that even established frontier-model developers are under pressure to improve cost efficiency rather than compete on performance alone.
Lower prices could help expand adoption and attract more customers, potentially supporting revenue growth through higher usage volumes.
However, they could also reduce the revenue earned per unit of AI processing, making it more difficult for companies to recover the costs of developing and operating their models.
Big Tech has a funding advantage, but capital alone may not be enough
The ability to finance the AI race is unevenly distributed.
Analysts said the large technology companies have established revenue streams, access to public markets, cloud businesses, and existing customer relationships.
These advantages give them more flexibility to absorb high research and infrastructure costs than smaller, independent AI developers.
Lis said large technology incumbents are in a stronger position to fund the race because of their access to public markets. However, he questioned whether substantial investment necessarily creates a sustainable competitive advantage when rivals can offer comparable technology at much lower prices.
Shazhaev similarly argues that established technology companies are better equipped to absorb frontier-model training costs.
He added that “but as raw intelligence becomes commoditized, proprietary behavioral data and distribution matter far more than capital or hardware.”
In his view, control over computing hardware can provide a temporary advantage, but it cannot necessarily prevent competitors from reproducing similar capabilities through other models.
Pure-play AI laboratories face a different challenge.
Without diversified businesses generating substantial cash flows, they may have to rely more heavily on venture capital, strategic investors, and other financing sources.
Shazhaev pointed to the large funding requirements of leading AI laboratories and the possibility that public listings could shift more of the financial risk onto public-market investors.
What are sustainable revenue avenues for AI companies?
If the advantage from a superior model is temporary, companies need other ways to protect their returns.
During the last technology revolutions, the greatest beneficiaries were the ones that figured out how to use technology to strengthen their products, improve productivity, or distance themselves from competitors (like streaming and gaming companies vs. telecoms during the internet boom). From this point of view, proprietary data sets and strong distribution networks matter more than access to cheap capital.
Rather than spending continually to outperform every rival on model benchmarks, a company can build products around customer needs, integrate AI into established workflows and use specialised information to deliver results that competitors cannot easily replicate.
Shazhaev makes a similar case, arguing that “companies escape the leapfrog cycle when they treat models as parts they can swap and build their edge in layers that stay scarce: workflow integration, simulated testing worlds, low-cost distributed delivery, and smart allocation that keeps customers.”
He draws a distinction between a high valuation or a temporary revenue surge and a durable business.
“Fast revenue helps short-term fundraising, but when open-weight models deliver frontier performance four months later at a fraction of the cost, customers churn unless your software is deeply woven into their daily work.”
The same principle applies to AI applications in financial services, healthcare, cybersecurity and industrial automation, where specialised knowledge, proprietary data and regulatory requirements may help companies establish defensible positions.
For investors, this means evaluating whether AI spending creates measurable economic benefits rather than treating model performance or user growth as sufficient evidence of commercial success.
The race to develop superior models may also create opportunities for businesses that support AI deployment rather than compete directly in model development.
Ryan Kirkley, chief executive and co-founder of Global Settlement Network, said “I think the economics become questionable when companies are spending billions to maintain a technological advantage that lasts six months, particularly if that advantage doesn’t translate into customers willing to pay a premium.”
He expects a handful of companies to continue developing frontier models while the broader commercial opportunity expands into the infrastructure supporting AI applications.
Michael Heinrich, chief executive and co-founder of 0G Labs, similarly argued that “As intelligence gets cheaper, the scarce thing becomes proof: that the model you paid for is the one that ran, that your data stayed private, and that an agent did what it was told. I expect durable value to sit with whoever can supply that proof at scale.”
The post As the AI model race heats up, who will come out on top? appeared first on Invezz