The AI ROI Reality Check: From Infrastructure Hype to European Enterprise Value

For three years, the industry has operated on the assumption that demand for computing power would eventually catch up with investment. The next phase will be harder. Technology companies must persuade customers to pay for AI at scale, rather than simply demonstrate what their models can do. That is changing how they sell their products, whom they hire, and where they expect the next wave of growth to come from.

The cost of believing in the future

The scale of the investment makes the problem difficult to ignore. Amazon, Microsoft, Alphabet and Meta are on course to spend a combined $720bn–745bn on capital expenditure in 2026. Broader estimates for the technology industry put the figure at $971bn, rising to $1.3tn in 2027 and $1.4tn in 2028.

Much of that money is going into data centres, computing equipment and the infrastructure required to train and operate increasingly capable AI systems. These assets are expensive to build and require substantial financing before their commercial value can be established.

Inflation complicates the calculation. The September US consumer price figures have increased expectations that interest rates will remain higher for longer, putting upward pressure on bond yields and raising the cost of borrowing. For companies committing billions to infrastructure whose returns may take years to materialise, more expensive capital changes the economics of expansion.

Higher rates do not make AI less useful. They make the timing and scale of its commercial payoff more important.

An infrastructure project that looks attractive when financing is cheap can become considerably less compelling when the cost of capital rises and customers take longer than expected to adopt the technology. Data centres cannot be justified indefinitely by the prospect of future demand. Eventually, the investment must translate into revenue and cash flow.

The market reaction suggests that investors are becoming less willing to treat spending itself as evidence of future success. Barclays' downgrades of Dell and HPE, citing the absence of a substantial AI-driven acceleration despite growing server revenues, point to the same tension. Selling more hardware is not necessarily the same as capturing the economic value created by AI.

That distinction matters across the entire supply chain. The companies supplying chips, servers, and cloud capacity can benefit from the initial investment cycle, but their fortunes remain tied to customers' willingness to keep spending. If the financial returns disappoint, the pressure will travel backwards from corporate budgets to infrastructure providers.

The missing link between models and money

The industry's response is revealing. Rather than concentrating exclusively on larger models and more computing capacity, AI companies are building teams designed to make their products work inside ordinary businesses.

The role at the centre of this shift is the forward-deployed engineer, or FDE. These engineers work directly with client organisations, helping them adapt AI systems to existing software, data, security requirements and business processes. Their job is not simply to deliver a model or provide access to an application programming interface. It is to get a working system into production and make it useful enough to justify its cost.

That distinction sounds straightforward, but it exposes one of the industry's most persistent commercial difficulties. A capable model does not automatically produce a more efficient company. An organisation still needs to determine which tasks should be automated, how its systems should interact with the model, what safeguards are necessary, and whether the resulting changes improve performance.

The technical demonstration is often the easy part. Integration is where the expense, delays and organisational resistance emerge.

US job postings for forward deployed engineers exceeded 7,300 in the first nine months of 2026, compared with fewer than 200 in 2023, according to the figures attributed to labour-market analytics firm Lightcast. Median advertised salaries have risen above $188,000, against roughly $145,000 for traditional software engineers.

The premium reflects a role that combines technical expertise with the ability to navigate a client's operations. Engineers must understand not only what a model can do, but also how a business actually works and what would persuade its management to pay for the result.

For AI providers, this is a significant change in the relationship with customers. A cloud service can be sold remotely and measured through usage. An embedded engineering team requires a closer relationship, a better understanding of the client's internal processes and greater responsibility for the outcome.

The provider is moving closer to the customer's business. It is also taking on a more difficult commercial task: turning a general-purpose technology into a specific financial benefit.

Hiring an army to prove the returns

The scale of the new deployment programmes suggests that the industry sees this as more than a temporary adjustment to its sales strategy.

Amazon Web Services has committed $1bn to its forward-deployed engineering initiative, with teams organised into groups that undertake 45-day client projects. Microsoft has established a Frontier Company unit backed by $2.5bn and roughly 6,000 embedded engineers. OpenAI has launched a $4bn deployment company, initially staffed by 150 forward-deployed engineers. Anthropic has committed $100m to its Claude Frontier Academy, with a plan to train 10,000 engineers by the end of 2027.

The approaches differ, but the commercial logic is similar. Each company is investing in the capacity to bring AI into customers' operations rather than relying on customers to work out the implementation themselves.

This creates a paradox for investors. The industry is responding to concerns about excessive expenditure by committing still more money to deployment. Yet the new spending targets a different problem from the construction of data centres. Infrastructure creates the capacity to deliver AI; deployment teams are intended to create paying customers who can use it productively.

Whether this investment proves more efficient remains an open question. Embedded engineers are expensive, and a labour-intensive deployment model may be difficult to scale across thousands of clients. If every implementation requires extensive custom work, the economics could become less attractive than the technology providers expect.

There is also a question of ownership. Engineers who become deeply involved in a client's internal processes can help make a product indispensable, but they may also expose the limits of a generic offering. If every customer requires a different configuration, the provider must balance bespoke services against the efficiencies expected from software.

Still, the hiring boom points to a change in the industry's understanding of what it takes to sell AI. Model performance remains important, but it is no longer sufficient. Commercial success increasingly depends on whether a company can deliver an operational improvement that its customer can measure.

Europe's opportunity is not another American data centre

Europe faces the same commercial challenge, but from a different starting point. Its share of global data centre capacity is approximately 5%, compared with 75% in the US and 15% in China. The imbalance places European companies at a disadvantage in any contest that depends primarily on the ability to finance and operate enormous computing infrastructure.

The EU's €200bn InvestAI programme, including €20bn in grants for AI gigafactories, is intended to narrow the gap. Support for 19 AI factories and European Investment Bank backing form part of the broader effort to expand computing capacity and reduce dependence on foreign providers.

Yet Europe is unlikely to close the infrastructure gap simply by reproducing the American investment model. The capital requirements are formidable, and the US technology giants already possess enormous resources and established cloud businesses. Competing with them on raw computing power would require Europe to overcome both a financial disadvantage and a substantial difference in existing capacity.

A more plausible route lies in applications built around particular industries and European requirements.

The region's leading AI businesses span several distinct fields. France's Mistral AI raised €3bn in September 2026 at a valuation exceeding €21bn. Germany's Aleph Alpha focuses on sovereign enterprise AI, while DeepL and ElevenLabs operate in language and voice technologies. Helsing works in defence, Wayve in autonomous driving, and Isomorphic Labs in life sciences. Black Forest Labs specialises in visual AI.

These companies do not all follow the same business model, and their presence does not guarantee European technological independence. But they illustrate a strategy that places less emphasis on matching the scale of American infrastructure and more on developing capabilities for specific markets.

For European customers, particularly in regulated sectors, the question is not always which model is the most powerful. Control over data, compliance requirements and the ability to adapt a system to an existing operation can matter just as much. Sovereign AI becomes commercially relevant when organisations need to know where their data is processed, who controls the technology and whether it can be deployed within their institutional constraints.

NVIDIA's $10bn commitment to the European technology ecosystem provides another source of support for local scale-ups. But investment in European companies does not remove the underlying dependency on a global infrastructure market dominated by American providers.

Europe's opportunity therefore rests on a distinction between owning the entire technology stack and controlling the applications that matter to its economy. The former requires extraordinary capital. The latter requires technical expertise, access to industrial customers and the ability to solve problems that general-purpose systems do not address on their own.

The next test is inside the customer

The emerging division between American scale and European specialisation is not absolute. US providers are moving aggressively into enterprise deployment, while European companies still need access to computing capacity to develop and operate their products. Both sides face the same requirement: proving that AI can generate economic value beyond the technology sector itself.

The risk for infrastructure providers is that capacity may arrive faster than profitable demand. If companies continue expanding data centres without establishing convincing paths to returns, higher financing costs could make future projects harder to justify and expose the weakness of business models built on continuous capital accumulation.

For developers that sell models without a credible integration strategy, the danger is different. Customers may find that access to increasingly capable systems does little to change their operations. In that case, the commercial advantage could shift towards firms that understand the customer's business and can deliver a working solution, even if they do not own the underlying model.

The forward deployed engineer sits at the intersection of those two pressures. The role has become valuable because the industry has discovered that technological capability and economic productivity are not interchangeable.