GenAI frenzy: Are tech giants overbuilding data centres?

On the surface, this makes perfect sense: more models, data and computing power. But a growing chorus of researchers and technologists is beginning to ask a pointed question—what if we are preparing for a compute-heavy world that may never arrive?

Joe Tsai, chairman of Alibaba, raised this concern last week. “I start to see the beginning of some kind of bubble… I’m still astounded by the type of numbers that are being thrown around in the US about investing in AI," he said at a conference in Hong Kong (shorturl.at/CkHGE).

Also Read: The great AI reboot: Educators, techies and leaders all need to adapt fast

This isn’t idle scepticism. Recent breakthroughs in GenAI have hinted at a surprise trend: better models and less computing. DeepSeek, a Chinese model, recently stunned observers with its capabilities and efficiency. It achieved results comparable to GPT-4 on many benchmarks while running on significantly less hardware.

Around the same time, Qwen, a foundation model from Alibaba, demonstrated similar efficiency; it trained using optimization techniques that dramatically reduced energy consumption and dependency on high-end infrastructure. Other GenAI purveyors have responded with similarly ‘lean’ models. Tsai’s musings are alarming, but possibly also strategic, given its Qwen ambitions.

However, there is an underlying shift in how AI progress is being achieved. The initial wave of GenAI progress—spearheaded by OpenAI, Anthropic and others—relied on brute force: scaling up parameters, data and computing. This led to models that were impressive but power-hungry. Training GPT-3 required massive amounts of computing. GPT-4, by all accounts, was even more demanding. The tech industry’s response was to build hardware for this trajectory. If success depended on scale, then hyperscale was the way forward.

Also Read: Stay ahead of the curve in the age of AI: Be more human

This has profound implications for the current data centre boom. Amazon, Alphabet and Meta alone plan to spend up to $240 billion this year. Much of this is channelled into GPU clusters, power-hungry cooling systems and grid-scale energy contracts. While these are useful for current workloads, their utility may decline faster than expected.

There’s historical precedent for this kind of overinvestment. During the early 2000s’ dotcom bubble, telecom companies laid down vast fibre-optic cable networks, expecting insatiable internet demand. The demand did arrive—but years later. In the interim, many companies went bankrupt and their infrastructure was mothballed or sold cheap. The only winner was Indian IT outsourcing—when Western companies realized they could now cheaply move work to Indian workers using this overcapacity, causing the initial IT outsourcing boom. Only much later did content streaming, cloud computing and smartphones put that overcapacity to use.

The risk with AI data centres is that they are more specialized and less forgiving. A fibre-optic cable can be used over a long period; GPUs, on the other hand, have a short shelf life. A data centre optimized for dense GPU workloads may not adapt quickly to lighter and more widely distributed AI deployments. Power contracts signed in anticipation of 2025’s peak usage might become liabilities if models in 2026 need only a fraction of that capacity.

Also Read: The iPhone needs a new strategy to salvage Apple Intelligence

Of course, tech giants are not naive. Companies like Microsoft have scaled back some of their data centre lease contracts (shorturl.at/4XbhH). Still, some argue that even if model training becomes more efficient, true AI scale—serving billions of users—will still require vast infrastructure.

There’s also an environmental angle to consider. The AI industry’s hunger for computing has alarmed climate researchers. More efficient models are not just a cost-saving measure, but a climate imperative. If tech companies are serious about sustainability, they must pivot towards less resource-intensive approaches.

So why does the spending spree continue? Partly, it’s inertia. Budgets and build-outs planned one or two years ago are only now coming online. There’s also a competitive signalling game—if your rival is spending billions on AI, you don’t want to be seen standing still.

Also Read: AI chips: India must get its basics right to meet this great catch-up challenge

But most of all, it’s a bet that AI will remain compute-intensive and that infrastructure ownership will offer a strategic advantage. That bet may pay off in the short term, especially for companies with proprietary foundation models or operating massive inference platforms. But eventually, it’s possible that the most valuable AI companies won’t be those with the biggest data centres, but those with the smartest algorithms.

For now, bulldozers are digging, chips are arriving and the power meters are spinning. But the true frontier of GenAI may not be in silicon but in code—in training methods, model design and clever optimizations that do more with less.

If that’s the case, then today’s data centre boom could resemble a race to build superhighways only to discover that most people prefer flying.

The author is co-founder of Siana Capital, a venture fund manager.