Important notice

Lindy Capital is a private investment company. It is not open to external investment, and nothing on this website is an offer, invitation, solicitation or advice.Not open to external investment.

Lindy Capital is a private investment vehicle. It invests capital belonging to its own group entities and does not accept, seek or manage money from outside investors. It is not open to external investment, and there is no current intention to open it.

Nothing on this website is an offer, invitation, solicitation or recommendation to invest in Lindy Capital or in any security, fund, property or other financial product. This website does not advertise, promote or offer any financial product or service.

Lindy Capital does not provide financial product advice, investment management or any other financial service to the public, does not hold an Australian Financial Services Licence, and is not a registered managed investment scheme.

Any discussion of investments, holdings or strategy on this site describes Lindy Capital’s own approach and decisions only. It is general in nature, may be out of date, and must not be relied upon as advice or as the basis for any investment decision. Anyone considering an investment should obtain independent professional advice.

Important noticeLindy Capital is a private investment company. It is not open to external investment, and nothing on this website is an offer, invitation, solicitation or advice.Not open to external investment. Nothing here is an offer or advice.Read the full noticeClose notice

Lindy Capital is a private investment vehicle. It invests capital belonging to its own group entities and does not accept, seek or manage money from outside investors. It is not open to external investment, and there is no current intention to open it.

Nothing on this website is an offer, invitation, solicitation or recommendation to invest in Lindy Capital or in any security, fund, property or other financial product. This website does not advertise, promote or offer any financial product or service.

Lindy Capital does not provide financial product advice, investment management or any other financial service to the public, does not hold an Australian Financial Services Licence, and is not a registered managed investment scheme.

Any discussion of investments, holdings or strategy on this site describes Lindy Capital’s own approach and decisions only. It is general in nature, may be out of date, and must not be relied upon as advice or as the basis for any investment decision. Anyone considering an investment should obtain independent professional advice.

Important information page →

Is the AI Bubble Real? Why the AI Buildout Is Just Getting Started

Is the AI bubble real? We think the sceptics are wrong. Chat was phase one. Agentic AI needs far more compute, and that is where the next wave of demand begins.
Abstract financial data visualization on a dark screen, representing strategic analysis.

Every few months the same headline returns: the AI bubble is about to burst. The spending is too large, the returns are too thin, and this is the dot-com boom all over again. We think that view is wrong, and wrong for a specific reason. It judges AI by its first phase, when the next phase is already here and needs vastly more computing power than the first.

Here is how we see the AI buildout, why we believe it is only just getting started, and where we think the next enormous wave of demand comes from.

Why people say AI is a bubble

The sceptics are not making things up, and their case deserves a fair hearing:

  • The spending is enormous. Microsoft, Alphabet, Amazon and Meta have guided to roughly US$720–745 billion of combined capital expenditure in 2026, most of it on AI data centres, chips and power. Alphabet alone expects US$195–205 billion, more than double its 2025 figure.
  • Returns at the customer level look patchy. An MIT study widely reported in late 2025 found that around 95% of organisations using generative AI had seen no measurable return, and only about 5% had AI integrated into their workflows at scale (report).
  • It rhymes with history. In the late 1990s telecom companies laid far more fibre-optic cable than the internet needed. Much of it sat unused — “dark fibre” — after the dot-com crash, and investors lost fortunes.

If AI were only a chatbot that people opened a few times a day, we would share some of that concern. But chat was never the destination. It was the on-ramp.

Phase one: AI adoption through chat

The first phase of AI, from the launch of ChatGPT in late 2022 onwards, was about adoption. Hundreds of millions of people learned to ask a model a question and read the answer. Businesses ran pilots. Software companies bolted chat windows onto their products.

A chat exchange is light work for a data centre. A person types a prompt, the model writes a few hundred words back, and the conversation pauses while the human reads, thinks and decides what to do next. The human is the bottleneck, which caps how much compute each person can consume.

Even so, phase one grew at a remarkable pace. Google has said it processed about 9.7 trillion tokens (the small chunks of text AI models read and write) a month in 2024, about 480 trillion a month in 2025, and more than 3.2 quadrillion a month by May 2026 (Shacknews). That is the adoption curve. The demand curve that matters for the next decade is steeper.

Phase two: agentic AI gets the work done

The current phase is agentic AI: AI agents that don’t just answer a question but complete a task. You give an agent a goal, such as “reconcile these invoices against the bank feed”, “find and fix the bug in this code” or “research these five suppliers and draft a comparison”, and it plans the steps, uses tools, checks its own work and keeps going until the job is done.

This is the shift from AI as something you talk to, to AI as something that works for you. For businesses that means agents handling customer service, coding, compliance checks, data entry, scheduling and analysis. For individuals it means personal agents that book, buy, organise, research and manage on their behalf.

Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% in 2025 (Gartner). That is an extraordinary rate of change inside the world’s software.

Why AI agents need massive amounts of compute compared with chat

This is the part the bubble argument misses. An agent completing a task uses far more computing power than a person chatting, for four reasons:

  • Many steps instead of one. A single task can involve dozens of model calls: planning, searching, calling tools, reading results, checking, retrying and summarising.
  • The context snowball. At every step the agent re-reads everything that came before — the instructions, documents, tool outputs and its own earlier reasoning. The work gets heavier as the task gets longer.
  • Reasoning before acting. Modern models “think” through problems before they answer. In early 2025 Nvidia’s Jensen Huang said reasoning AI already needed around 100 times more computation than the models available when ChatGPT launched.
  • No human bottleneck. Agents don’t pause to read, sleep or take lunch. A business can run hundreds of them in parallel, around the clock.

Estimates of the gap vary with the task, but they all point the same way. Gartner analysis puts agentic AI at roughly 5 to 30 times more tokens per task than a standard chatbot exchange, and research on coding agents has found gaps of up to 1,000 times compared with simple code chat (summary of estimates).

Now multiply. More people and businesses using AI, times more tasks handed to AI per user, times many more tokens per task. Each of those factors is growing at once, so the result compounds. That is where the next enormous wave of compute demand comes from.

AI bubble or real demand: abstract data visualisation of growing AI compute usage

The evidence: demand is outrunning supply

If the AI buildout were a bubble, we would expect to see idle capacity and falling prices for compute. We see the opposite.

  • Microsoft reported Azure growth of 43% for the June 2026 quarter, with its commercial backlog of contracted future revenue up 84% year on year to US$678 billion (Blockspace).
  • The large cloud providers have repeatedly said they are capacity constrained: they could sell more AI computing if they could build and power it faster.
  • The bottlenecks have moved down the supply chain to power, grid connections, cooling and land — including here in Australia, where operators are competing for sites and energy around Sydney and Melbourne to build new AI data centres.

Token prices are falling fast, and that matters too. Cheaper AI doesn’t shrink demand; it makes new uses economic. When something useful gets cheaper, people tend to use far more of it — economists call it the Jevons paradox. Agents that were too expensive to run last year become routine this year.

AI bubble vs dot-com bubble: what is different

The dot-com comparison is the strongest card in the bubble argument, so it is worth being precise about where it breaks down:

  • Utilisation. After 2000, a large share of the new fibre went unused for years. Today, AI computing capacity is being used as fast as it can be installed.
  • Who is paying. Much of the telecom build was funded with debt by companies that had little revenue. Most of today’s AI infrastructure is being built by some of the most profitable businesses in history, largely out of operating cash flow.
  • Real customers. The demand comes from businesses and individuals paying for the service, backed by contracted commitments, rather than from speculative traffic forecasts.

None of this means every AI stock is a bargain, or that the path will be smooth. Prices of individual companies can run far ahead of their businesses, and some will. Saying the buildout is real is not the same as saying any price is a good price.

What could prove us wrong

We try to hold our views with discipline, so we keep a list of things that would change our mind:

  • Agents failing to deliver reliable results at scale, so businesses pull back after pilots.
  • Efficiency gains so large that the same work needs far less hardware, faster than new uses appear.
  • Power and grid constraints that delay projects and strand capital.
  • Rising debt among the more speculative builders and data centre developers, or circular financing between suppliers and customers.
  • Valuations that already price in a perfect outcome.

We watch these closely. For now, the evidence is pointing the other way.

How we think about AI infrastructure investing at Lindy

We invest for decades, not quarters, and we prefer businesses that win whether or not any single AI model or app comes out on top. In practice that means owning the companies that build and run the infrastructure, and those that turn AI into products used by billions of people.

Our listed holdings include Alphabet, Amazon, Meta and Microsoft, which own much of the world’s AI computing capacity and the platforms agents will run on; Broadcom, which designs custom AI chips and the networking that connects them; and Vistra, a US power producer, because every new data centre needs electricity. You can read more about what we own and our long-term investing approach.

Our conviction rests on a simple observation. Chat taught the world to use AI. Agents put AI to work. Work is far more compute-hungry than conversation, and the world has barely started handing it over. The people calling the top of an AI bubble are, in our view, measuring the first phase of something much larger. For more on why we think durable trends deserve long holding periods, see the Lindy effect, explained.

This article reflects Lindy Capital’s own views as at October 2026 and is general information only. It is not financial advice or a recommendation to buy or sell any security. Lindy Capital may hold positions in the companies mentioned. Figures are drawn from public sources and company reports and may change. Past performance is not a reliable indicator of future performance.