Is there an AI bubble? And is there still value? Both can be true.

Ever since ChatGPT burst onto the scene in late 2022, investors have been grappling with the same question: Are we living through an AI bubble?

My answer is that two things can be true at the same time:

• Yes, we are in an AI bubble

• Yes, there is still value in parts of the AI ecosystem

This is not a contradiction. It is how most technological revolutions unfold.

Technological revolutions often begin with a surge of speculation that funds the infrastructure needed for decades of real value creation. The dot-com era is the clearest example: a bubble that wiped out hundreds of unprofitable startups but simultaneously gave rise to Amazon, Google, and the fibre-optic backbone of the modern internet. The railway booms of the 1800s followed a similar path.

In every major technological shift, bubbles and enduring value appear side-by-side. AI is no different.

AI: A transformational technology with real economic pull

Let’s start with the obvious: AI is not a fad.

For anyone studying futurists, technologists, or simply the capex budgets of the hyperscalers (Microsoft, Amazon and Alphabet), it is clear that AI is a foundational technology that will reshape productivity, decision-making, and entire industries.

One of the clearest signals of demand is how difficult it is for these companies to secure enough AI GPUs. Despite spending hundreds of billions across data centres, networking, and chips, supply still lags demand. When the companies best positioned to forecast real-world AI usage cannot obtain enough GPUs, that tells you something powerful about this trajectory.

Hyperscaler demand is not speculative. It is already here.

And where demand is visible, supply will follow. Over time, today’s imbalance will ease not because AI demand slows, but because capacity eventually catches up. That is good news for the companies providing that capacity: NVIDIA, TSMC, ASML, Google, and SK Hynix.

But not all AI capex is created equal

Where the picture gets murkier is in the non-hyperscaler part of the ecosystem.

This becomes clear when you look at CoreWeave, Oracle, the emerging Stargate project, and OpenAI’s reported US$1.4 trillion of proposed infrastructure deals, investments where demand is hoped for rather than demonstrated.

It is also important to note that these headline multi-billion-dollar agreements are not guaranteed spending. Much of what OpenAI has signed with Oracle, Nvidia, and others sits in remaining performance obligations rather than firm commitments. If model demand does not scale as expected, these obligations can be delayed or may never materialise.

Recent dealmaking, whether it is Sam Altman’s trillion-dollar hardware ambitions, Oracle’s infrastructure push, or CoreWeave’s rapid expansion, raises a simple question:

Who actually has visibility into demand?

Hyperscalers can see it. Their customers are already training models, inference workloads are scaling, and they sit on the world’s richest data reservoirs.

Outside the hyperscalers, however, the spending looks increasingly speculative. When you are building supply before you see clear demand, you run a real risk of overcapacity.

This divergence is one reason the “AI bubble” narrative has weight. Not because AI is not real, but because capital is not being allocated with discipline everywhere.

The model wars: From secret sauce to commodity

Another unusual feature of this cycle is how quickly the moat around AI models has evaporated.

It was only six years ago that Microsoft invested US$1 billion into a little-known company called OpenAI, and only a few years later, OpenAI became the clear leader.

Google, despite inventing the transformer architecture that underpins modern AI, did not fully appreciate its potential until it saw ChatGPT go viral. Alphabet’s early stumbles made it look like a giant unable to commercialise its own breakthroughs.

Fast forward to today and we are in a completely different world:

• Google’s Gemini models are leaner, more efficient, and competitive with OpenAI’s best

• Anthropic continues to produce highly capable models

• Microsoft, anchored by OpenAI and supported by Llama, offers a broad suite of options

• Amazon’s Bedrock lets customers choose from multiple top models, making switching simple

The implication is clear: models have become commoditised. And in a commoditised world, the bottleneck is no longer model architecture. It is the data.

Data Gravity: The only durable moat

Andy Jassy put it succinctly: “data has gravity.”

Models can be replicated. Chips can be manufactured. Infrastructure can be leased. But data does not move easily. It stays where it is generated, and it becomes exponentially more valuable when combined with domain context.

Long-term value will flow toward companies that have:

• proprietary, high-quality datasets

• distribution that collects user behaviour at scale

• products that embed AI deeply into workflows

This is why the hyperscalers remain so well positioned. They host the data, they process the data, and increasingly they control the infrastructure that underpins how it flows through storage, compute, and AI systems.

So where does that leave investors?

In every bubble, there are two types of capital:

• real capital that earns a return

• speculative capital that eventually evaporates

AI is no different. A bubble is clearly forming, which is unsurprising given the amount of capital being pledged without real customer demand behind it.

Jeff Bezos captured this dynamic well at Italian Tech Week some weeks back:

“When people get very excited as they are today about artificial intelligence, for example, every experiment gets funded. Every company gets funded. The good ideas and the bad ideas.”

But alongside the speculation, there is unmistakable long-term value where demand is genuine, supply is constrained, and moats are tied to data and infrastructure.

This is why companies like NVIDIA, TSMC, ASML, Alphabet, and SK hynix still have compelling investment cases. They sit at chokepoints where:

• demand is visible

• economics are improving

• scale generates competitive advantage

These are not the AI tourists. They are the ones selling the picks and shovels in the largest compute buildout in history.

Final Thoughts

AI is both bubble and boom. It is speculative and deeply real at the same time. That is what makes this moment so fascinating and so challenging for investors.

The opportunity is not in guessing which speculative bets will survive. It is in owning the companies with real demand, real economics, and real moats.

The bubble will eventually fade. The value created by the winners will not.