Every time another tech giant announces billions of dollars in spending on AI infrastructure, the same question comes to mind: are we witnessing the next great technological revolution, or simply watching another investment bubble inflate? It's an understandable concern. History has shown that transformative technologies often attract more capital than they can ultimately justify. But after looking through the latest developments in data-centre construction, power infrastructure and AI demand, I think the picture is more balanced than the headlines suggest. The risk of overbuilding is certainly real, but for now, the industry still appears to be constrained more by its ability to build than by a lack of demand.
AI Data Centre Demand: Why Headlines Don't Tell the Whole Story
One reason the debate has become so heated is the constant stream of headlines about cancelled or delayed data-centre projects. At first glance, these stories appear to support the argument that demand is weakening. In reality, the picture is far more nuanced. Many of the projects encountering difficulties are being developed by newer entrants with limited experience navigating the complexities of permitting, procurement and electricity connections. By contrast, the major hyperscalers continue to complete most projects close to schedule while simultaneously planning additional facilities.
Building a modern AI data centre is considerably more complicated than constructing a conventional commercial building. Developers must secure land, obtain permits, order specialised electrical equipment, negotiate grid connections and coordinate long supply chains that often stretch across multiple countries. Missing a single step can delay an entire project by months. Experienced operators understand these timelines, while less experienced developers frequently underestimate how long each stage requires. As a result, isolated cancellations should not be interpreted as evidence that the broader expansion has stalled.
Power Infrastructure Bottlenecks & Grid Constraints
Perhaps the greatest constraint facing the industry today is not computing hardware, but electricity. AI workloads require enormous amounts of continuous power, and electrical grids in many developed countries were never designed to accommodate such rapid growth in industrial demand. Connecting a new data centre to the grid can take several years, creating a significant bottleneck even after construction has been completed.
This challenge has encouraged developers to rethink how data centres are powered. Rather than waiting for permanent grid connections, many projects are increasingly adopting behind-the-meter generation, where electricity is produced directly alongside the facility. Containerised gas turbines, for example, can often be installed within six to twelve months, providing temporary power while permanent grid infrastructure is completed. Although these solutions increase costs, they allow operators to begin generating revenue much sooner, making the trade-off worthwhile for projects where speed is critical.
The composition of future electricity generation is also evolving. Solar energy and battery storage are expected to contribute significantly over the next few years, while natural gas is likely to play an increasingly important role later this decade, particularly in the United States where fuel prices remain comparatively low. Nuclear energy continues to attract attention, but most proposed projects remain more than a decade away from making a meaningful contribution. Many current agreements simply guarantee that power will be purchased if future reactors are successfully completed, rather than committing to near-term electricity supply.
Supply Chain Delays & The Cost of Speed in AI Construction
Power is only one of many bottlenecks. Transformers, switchgear, cooling equipment, networking hardware and specialised construction services all remain in short supply. Ironically, these relatively ordinary components can delay projects just as effectively as shortages of advanced AI chips. Because every month of delay represents lost revenue, developers are often willing to pay substantial premiums to secure scarce equipment. This willingness to overpay has pushed costs higher throughout the supply chain and reinforced the importance of modular construction techniques that reduce installation times.
Modern data centres increasingly rely on factory-built components that can be assembled off-site before being delivered for installation. Mechanical and electrical systems are becoming more standardised, reducing construction time and improving consistency. While modularisation cannot eliminate permitting delays or equipment shortages, it does reduce the amount of work that must be completed on-site, allowing projects to come online more quickly.
Is the AI Boom a Tech Bubble? Evaluating Hyperscaler Capex
From an investment perspective, perhaps the most important question is whether soaring capital expenditure reflects rational expansion or speculative excess. Spending alone tells us very little. What matters is whether the underlying businesses continue generating strong revenue, profits and cash flow. At present, many AI-related businesses are doing exactly that. Hyperscaler Cloud divisions are reporting exceptionally strong growth, with rising margins supporting continued investment in additional capacity.
This distinction is critical. Rising capital expenditure accompanied by genuine profitability is fundamentally different from spending that relies solely on investor optimism. Previous technology bubbles were characterised by impressive revenue growth that failed to translate into sustainable cash flow. Today's leading hyperscalers, by contrast, continue to generate substantial earnings while reinvesting those profits into expanding AI infrastructure. Although this does not guarantee every project will succeed, it suggests that current investment is being driven by real demand rather than speculation alone.
Key Risks for AI Investors: Debt, Memory Cycles, and Global Competition
Debt remains another area investors should monitor carefully. While the largest technology companies are borrowing more than they have in the past, overall leverage remains modest compared with many industrial businesses.
More significant is the increasing use of joint ventures, infrastructure funds, long-term leases and take-or-pay agreements to finance new projects. These arrangements distribute risk but can also obscure the industry's true level of financial exposure. Similarly, vendor-backed financing, where equipment suppliers help customers secure funding, introduces another layer of complexity. These structures are not yet large enough to represent a systemic threat, but they could become more concerning if they continue expanding over the coming years.
Semiconductor supply chains present another challenge. AI processors rely heavily on specialised high-bandwidth memory, an industry that has historically experienced pronounced boom-and-bust cycles. Strong demand is currently producing exceptional profits, but memory markets have rarely remained elevated indefinitely. Manufacturers are expanding capacity with support from long-term customer commitments, which may reduce volatility in the near term, yet the industry remains inherently cyclical. Investors therefore need to balance attractive valuations against the possibility that today's unusually strong earnings will eventually normalise.
Other spoilers
Beyond infrastructure, competitive dynamics may ultimately determine whether today's investment delivers attractive returns. Chinese AI developers are rapidly improving the quality of open-source models, particularly for everyday tasks such as coding assistance, document summarisation and image recognition. Even if leading Western AI systems remain technically superior, widespread availability of capable free alternatives could place downward pressure on pricing. Combined with China's strengths in manufacturing, energy expansion and industrial deployment, this represents a meaningful long-term risk to AI profitability across the sector.
Another concern frequently raised by investors is technological obsolescence. If each new generation of AI chips rendered previous hardware worthless, the economics of data-centre investment would deteriorate rapidly. Fortunately, evidence so far suggests a more gradual pattern. Older hardware continues to generate meaningful revenue, albeit at lower rental rates than the latest processors. Rather than becoming stranded assets overnight, previous generations appear to retain useful economic lives, reducing the risk that billions of dollars of infrastructure will suddenly become obsolete.
Where I Think We Are in the Cycle
When I step back and look at the evidence, I don't think we're at the point where the AI infrastructure boom has clearly gone too far. Could we eventually reach that stage? Absolutely. Every major technological wave has experienced periods of excessive investment, and AI is unlikely to be the exception. But today's environment still looks fundamentally different from the final stages of previous bubbles.
The biggest constraints remain physical rather than financial. Electricity is scarce. Critical equipment is difficult to source. Construction capacity remains stretched. Meanwhile, the companies driving most of this investment are still generating strong profits and cash flow, rather than relying solely on optimistic forecasts. For me, those are far more important indicators than the occasional cancelled project or alarming headline.
That doesn't mean investors should become complacent. Profitability, AI adoption, electricity availability, financing structures and infrastructure spending will all determine whether today's investments ultimately deliver attractive returns. But for now, I believe the evidence points to an industry that's still building enough capacity to satisfy demand—not one that's racing far ahead of it.
