AI’s three clocks: technology, adoption and capital

A technology can be changing the world while the companies financing it are having a very difficult year.

That is the thought I keep returning to as I watch the AI investment cycle. A falling share price, a strained balance sheet and a useful product can all exist at the same time. We saw it with the internet. We should keep that experience in mind as we decide where to invest, what to build and how to help people use AI.

In my earlier article, I wrote about technology diffusion: the work of bringing a new capability into ordinary life. The investment question is what happens when that work moves at a different speed from the capital paying for it.

Three clocks, running at different speeds

The first clock is technology. A new model, chip or network makes something possible that was previously difficult or expensive.

The second is adoption. People have to find a useful task, learn a new habit, connect the tool to their work and trust the result.

The third is capital. Investors price in future growth. Companies raise money and build capacity. Lenders expect interest and repayment on an agreed timetable.

These clocks do not automatically stay together. Investors can correctly predict a large future market and still overpay, build too much or borrow on terms that leave too little room for delay. A lender’s repayment date does not move simply because customers need another year to change their habits.

Did the internet really take fifteen or twenty years?

It depends on where we start the clock and what we mean by adoption. The internet predates the Web. CERN made its Web software freely available on 30 April 1993. People were communicating and buying online well before the market peak in 2000.

But turning the Web into a reliable, always-on part of everyday life took longer. Pew’s historical U.S. survey series gives a useful illustration:

From the open Web in 1993 to widespread home broadband in 2008–2010 is roughly fifteen to seventeen years. That is a useful perspective, rather than a universal waiting period. It describes one country, one form of access and one stage of adoption.

Think of the layers involved. A website needed a customer with a connection. Online shopping also needed secure payment, a dependable delivery service and confidence that an order would arrive. Video calling needed sufficient bandwidth at both ends. Mobile services needed suitable devices, networks and interfaces. The invention of a protocol did not deliver all those things at once.

Diffusion continues after a technology becomes familiar. The U.S. Census Bureau reported that e-commerce accounted for 17.1% of retail sales in the second quarter of 2026, on a seasonally adjusted basis. Online retail is hugely important without accounting for every purchase. That statistic does not capture all the internet’s influence on commerce, including research before an in-store purchase.

The fibre was real. The financial timetable was wrong.

The dot-com crash and the telecom bust were related, but they were not the same event. Nor did either have a single cause.

A 2003 study published by the Richmond Fed describes the surge in telecom investment and the overbuilding of long-distance fibre networks. Demand was increasing, but forecasts were too optimistic. Improvements in transmission technology also meant more capacity could be carried on the fibre already installed. Supply expanded from both construction and better equipment.

Regulatory uncertainty and the 2001 recession added to the problem. In the wider crash, business models, financing conditions and accounting failures mattered too. Saying that people simply refused to use the internet misses much of the story.

Debt made the mismatch painful. Contemporary congressional testimony on Global Crossing described borrowed money funding network construction, excess capacity and falling prices. The company entered bankruptcy in January 2002 with reported debts of $12.39 billion. That testimony also raised accounting concerns; this was not a clean experiment in slow adoption.

The basic investment lesson is straightforward. If competing networks provide more capacity than customers will buy at the expected price, prices fall. Lower prices may help the internet spread, while leaving the original lenders and shareholders with losses. A useful asset for society can be a poor investment at its original cost.

Amazon shows both sides of the story

Amazon’s 2000 shareholder letter reported that its shares had fallen more than 80% since the previous letter. Yet sales had risen from $1.64 billion in 1999 to $2.76 billion in 2000, and customers from 14 million to 20 million.

The market price collapsed while customer use grew. Amazon subsequently became a company valued in trillions of dollars; its 2025 annual filing alone records more than $2 trillion in voting stock held by non-affiliates at 30 June 2025.

That is evidence of what a survivor could build, not evidence that every company in the boom deserved to survive. Looking back only at Amazon hides the businesses that ran out of money. Being right about the internet was insufficient. The business still needed customers, an operating model and the ability to fund the journey.

Another milestone became clear earlier this year. Amazon, whose online store opened in July 1995, reported $716.9 billion in net sales for the year ended December 2025. That edged above Walmart’s $713.2 billion in total revenue for the year ended January 2026. Building that scale took roughly thirty years.

It is a striking milestone in the long development of internet-enabled business. It does not mark full adoption of online shopping: the reporting periods differ by a month, Amazon’s total includes AWS and advertising, and Walmart itself sells online. The internet has reshaped both companies. The lesson is how long building commercial scale can take, while adoption and reinvention continue.

The AI infrastructure parallel is already visible

Today, specialist cloud providers—often called neoclouds—assemble chips, data-centre space, power and networking, then rent computing capacity to customers. They help meet genuine demand. They also take on commitments before all the revenue arrives.

October’s market moves show how quickly confidence in that financing can change. Reuters reported on 14 October that Nebius and CoreWeave shares fell 15.1% and 9.5%, respectively, amid interest-rate concerns. A rebound followed on 19 October. Those are dated market events, rather than proof that an entire technology has failed.

CoreWeave’s own numbers show demand and financial pressure together. Its second-quarter 2026 results reported revenue of $2.575 billion, up from $1.212 billion a year earlier, alongside $640 million in net interest expense and a $626 million net loss. Its June 2026 quarterly filing reported $35.6 billion of total indebtedness.

A growing customer bill does not automatically leave enough cash to pay for equipment, electricity, interest and replacement hardware. Contracts help, but their value depends on delivery, customer payment and the terms. A backlog is not money already in the bank.

The fibre comparison has limits. A buried cable and a rapidly changing AI chip have different economics. Better chips, more efficient models or lower rental prices can put pressure on older equipment. Power availability and construction delays can hold back capacity even where customers are ready to pay. We should not assume that every strained AI project is suffering from a lack of demand.

I see the recent correction as a warning about financing and expectations. Whether we call it a mini-bubble bursting matters less than understanding which commitments can survive a less generous market.

Anthropic shows why this wave is different

Demand is already substantial. Anthropic’s 2 November 2026 S-1/A filing reports:

That is approximately thirteen times the revenue in the comparable nine-month period: recognised revenue, rather than an annualised run rate.

The filing estimates the economic value of global knowledge work in 2025 at more than $40 trillion. That is broader than an addressable market for AI spending. It also says spending with Anthropic averaged less than 0.1% of total expenses among its publicly traded Fortune 500 customers in Q3 2026. That measures spending depth in those customers, not AI-market penetration.

Revenue is not profit. My reading is that a large opportunity remains, with uncertain pricing, timing and value capture. These figures establish fast commercial growth; they do not establish what every provider or investor will earn.

AI can spread faster—and still need help

We should not turn the internet’s history into a prediction that AI must wait fifteen years. AI arrives through phones, browsers and cloud services that people already have. In that respect, it benefits from the infrastructure and habits built during earlier waves.

A study by Alexander Bick, Adam Blandin and David Deming, revised in February 2025, found that nearly 40% of U.S. adults aged 18–64 were using generative AI by late 2024. Overall uptake was faster than for PCs or the internet at comparable stages after their mass-market launches.

Trying a tool, using it every week and redesigning a business around it are different stages. A person may generate a useful answer in seconds while their employer spends months agreeing data access, checking reliability and deciding who approves an action. Rapid consumer uptake and slower organisational change can coexist.

Instinct: making the technology easier to reach

Instinct is a useful example of the kind of product I mean. Its website describes a personal assistant that people can text or call, with tasks such as ordering groceries or booking a train. Those are the company’s product claims, rather than results I have independently tested.

The interesting idea is the starting point: a person has something to get done. They should not need to understand the model underneath, learn a complicated interface or become expert at writing prompts.

That is a route to diffusion. It moves the technology closer to an existing habit and a recognisable need. Sustaining adoption will still depend on reliability, sensible permissions, control over consequential actions and a business model that covers the cost of serving the user. An easier interface is valuable; it does not remove the economics.

Where IJK wants to build and invest

At IJK, we are cautious about heavy capital expenditure whose return depends on uninterrupted funding, high utilisation and favourable future prices. That is not a blanket rejection of infrastructure. It is a reason to examine who pays, when cash arrives and what happens if the plan slips.

I am particularly interested in businesses that help turn powerful technology into useful adoption: a better interface, a dependable workflow, access to the right information, practical training or distribution that reaches people who would otherwise struggle to get started.

These businesses have risks too. An application may depend on another company’s models, face rising computing costs or lose customers when a larger platform adds a similar feature. We still need to ask whether customers return, pay and receive enough value to support a durable business.

Our work with UBI gives us a practical way to learn. Developing DeeJay has involved everyday tasks, agreed definitions, human review and feedback from the people doing the work. In education, helping a colleague use a tool responsibly is itself part of creating value.

Through our investments and AI implementation work, we want to support more of that process. The measure should be a real improvement: a customer served better, a teacher with more time for students, or a smaller company able to do work it could not previously manage.

My conviction is that AI will become deeply woven into ordinary life. The route will include failed companies, repriced investments and changes of direction. A correction can change who owns the infrastructure and what it costs without ending the demand for useful intelligence.

The opportunity I find most compelling is helping the adoption clock catch up: bringing the technology within reach, making it dependable and giving people a reason to keep using it.