In the late 1990s, the surest way to lose money in technology was, paradoxically, to play it safe. The logic seemed impeccable: forget the frothy dot-com stocks with no revenue; invest instead in the companies laying the pipes. The internet plainly needed more bandwidth, and telecoms operators were building it. Capital poured in. WorldCom, Global Crossing and their peers raised billions. The return on capital, the market assumed, was all but guaranteed.

I had a front-row seat. As the head of Citigroup’s equity capital markets business in Australia, I watched the machine from the inside. Citi was a powerhouse in telecoms financing, propelled by Jack Grubman, its star telecoms analyst (later disgraced, but at the time the most influential voice in the sector). The deals were enormous. The conviction was total. Telecoms was the “safe” trade.

Then came the multiplex moment.

A hundred times over

Dense wavelength division multiplexing, or DWDM, was not the sort of technology that made newspaper front pages. It was, in essence, a way of splitting a single strand of fibre-optic cable into dozens of separate channels of light, each carrying its own data stream. The effect was extraordinary: existing fibre could suddenly carry more than a hundred times the data it had carried before.

For the telecoms operators and their investors, this was catastrophic. The entire investment thesis had rested on the assumption that vast quantities of new fibre needed to be laid to meet surging demand. DWDM blew that assumption apart. The pipes already in the ground were suddenly, spectacularly, sufficient. The supply-side shock was total. Returns on the billions already deployed collapsed. The telecoms bust arrived, and with it, the deeper unravelling of the dot-com bubble itself.

What remained was dark fibre: mile upon mile of unused telecoms capacity, available at fire-sale prices.

The killer app emerges from the wreckage

The pattern that followed is the one worth studying most carefully. By 2002, the market had all but given up on the internet. Yet it was precisely in this period of maximum pessimism that the real winner emerged.

Google had two advantages that, in combination, proved decisive. The first was a business model of breathtaking elegance: the per-click ad auction. It was a business model that monetised intent, scaled with usage, and did not require subscription friction. The second was timing. Google bought dark fibre at distressed prices, acquiring on the cheap the infrastructure it needed to offer search to the world for free. The surplus capacity created by the telecoms bust became the foundation of the most profitable business model the internet had yet produced.

I saw the power of that model before most. After leaving Citi, I found myself running an internet company called Dark Blue Sea in 2001, about as unfashionable a career move as one could make at the time. Dark Blue Sea operated a pay-per-click ad auction, a clone of the model pioneered by GoTo.com, which later became Overture before being acquired by Yahoo. Even in 2001, amid the smoking ruins of the dot-com crash, the business was remarkable. Revenue arrived with a directness and a predictability that was unlike anything else in the internet economy. I could see, in the unit economics of every click, that the commercial model for the internet was not merely viable – it was extraordinary.

This was before Google had a revenue model. My experience at Dark Blue Sea was a key source of my personal conviction that the internet would not merely recover but transform the economy, at a time when that view attracted more pity than agreement.

History rhymes

Consider, now, the present landscape. The AI boom has its own infrastructure layer: the large language model providers, the data centre operators, the chip manufacturers. Nvidia, the hyperscalers and their peers are the WorldComs and Global Crossings of this cycle. They represent the “safe” way to play the AI revolution. Capital is pouring in on the same seductive logic: never mind which AI application wins, invest in the picks and shovels.

The question that should trouble every investor in this trade is a simple one: where is AI’s multiplex moment?

It probably has not arrived yet, not fully. The release of DeepSeek, the Chinese AI model that achieved striking results at a fraction of the usual cost, was an early tremor. Infrastructure stocks wobbled briefly before recovering. But the underlying vulnerability remains. The gap between today’s AI infrastructure and what is theoretically possible is, by any measure, enormous.

As Elon Musk has observed, the human brain is a piece of meat that runs on roughly 20 watts. Today’s data centres consume megawatts to perform cognitive tasks that a three-year-old manages effortlessly. The implication is not that AI will fail, but that the current approach to delivering it is almost certainly not the final one. There is vast scope for breakthroughs in efficiency; in architecture, in chip design, in training methods, in inference optimisation, and any one of these could reprice the entire infrastructure layer overnight.

AI’s multiplex moment may not arrive as a single event. It may come as a series of advances that compound, each one shaving another order of magnitude from the cost of intelligence. The effect, however, would be the same: a supply-side shock that destroys the return-on-capital assumptions underpinning today’s buildout, floods the market with cheap compute, and creates the conditions for an entirely new generation of companies to emerge.

The opportunity on the other side

If the dot-com analogy holds, the most important companies of the AI era have probably not yet been founded. The Google of AI will not be the company that built the biggest data centre. It will be the company that finds the killer app: the business model so elegant and so powerful that it captures the value of cheap, abundant intelligence the way Google captured the value of cheap, abundant bandwidth.

What might that killer app look like? The honest answer is: if it were obvious, it would already be priced in. But history suggests the shape of it:

•         a business model where the value capture is immediate and measurable,
•         a product where “intelligence” is not a feature but the core engine,
•         distribution that scales without proportionate cost,
•         and unit economics that improve as the system learns.

This is why, as an angel investor, I find the present moment so compelling. I am not looking for the next Nvidia. I am looking for the founders and the ideas that will flourish if and when the AI multiplex moment arrives. I am looking for the capital-light companies that will build extraordinary businesses on top of infrastructure whose cost is collapsing. The companies that will turn today’s dark fibre equivalent into tomorrow’s dominant platforms.

The angel opportunity in that environment will be spectacular. When infrastructure becomes radically cheaper, the barriers to entry for startups fall with it. The vintage of companies founded in the years immediately following the dot-com bust – when bandwidth was cheap and business models were being reinvented – produced some of the greatest returns in the history of venture investing. There is every reason to believe the same dynamic will play out in AI.

The safe trade is never as safe as it looks. The real opportunity is always on the other side of the correction, in the hands of founders who see what the wreckage makes possible.

moocoo.vc