• Transformational tech often fails to deliver extraordinary long-term returns;
• Tech advantages quickly commoditizes into standard baseline expectations;
• Strong unit economics and capital returns determine long-term investment success

Earlier this week, I wrote that every generation experiences its own technological revolution.

Railroads, electricity, radio, and the internet all changed the world in ways that were almost impossible to appreciate while they were unfolding.

Artificial Intelligence (AI) is in the early stages of following the same path.

It"s rewiring technology – giving humans new superpowers.

I say this with some bias (having worked in the space for the past decade) — the excitement feels justified.

But if I remove my engineering hat – as investors – we should avoid confusing excitement for what is ultimately valuable.

Here I like to draw on my own investing mistakes from the dot-com bubble.

Consider networking company Cisco (CSCO).

In the late 1990s – it briefly became the most valuable company in the world – as it supplied much of the networking equipment to enable the internet"s connectivity.

Think of it as the Nvidia (NVDA) of the day.

I bet (big) on Cisco as it was the quintessential "picks and shovels" play. The numbers were phenomenal:

• Revenue: Scaled roughly 9.3x – from ~$2B in 1995 to nearly $19B by 2000. CAGR of ~57%.
• Earnings: GAAP net income climbed more than 5x to 6x – from $421M to over $2.6B

Cisco (FY1994–FY2000)

Fiscal Year	Revenue	Net Income (GAAP)
1994	$1.24 Billion	$315 Million
1995	$2.03 Billion	$421 Million
1996	$4.10 Billion	$913 Million
1997	$6.45 Billion	$1.05 Billion
1998	$8.49 Billion	$1.35 Billion
1999	$12.17 Billion	$2.10 Billion
2000	$18.93 Billion	$2.67 Billion

I wasn"t buying a business that simply added "dot com" to its name – this was a business producing real cash at gross margins north of 65%.

Of course, the internet went on to exceed almost everyone"s expectations. However, Cisco"s shareholders (like me) did not.

I made one of the poorest investment decisions of my life circa 1998 — where I watched my holdings lose ~90% less than three years later.

And whilst the stock would recover its share price loss some 25 years down the track – I had moved on. I rode the stock all the way up and down.

Two things I took away:

• My mistake wasn"t believing in the underlying technology (the demand and need was insatiable);
• I confused what was exciting with what was valuable.

I assumed that extraordinary tech automatically translated into extraordinary shareholder returns.

That said, I consider myself fortunate the price I paid for admission (as we all pay a price in the investment game) – was very early in my career (mid 20s).

Learning from your mistakes is where the gold is.

Innovation vs Economics

When I observe companies like (not limited to) Google investing massive sums to build out AI infrastructure—it"s tempting to assume that if a technology transforms the economy, the companies building that technology must also become extraordinary investments.

However, as my Cisco story shows, it"s never that simple.

Some of the most important technologies ever invented generated surprisingly ordinary returns for the businesses that first commercialised them.

Let"s compare Cisco with the growth in Nvidia using a 6 year window (as there are some parallels)

Cisco (FY1994–FY2000) vs. NVIDIA (FY2019–FY2025)

Metric / Company	Cisco (FY94–FY00)	NVIDIA (FY19–FY25)
Time Horizon	6 Years	6 Years
Starting Revenue	$1.24 Billion (FY94)	$11.72 Billion (FY19)
Ending Revenue	$18.93 Billion (FY00)	$130.50 Billion (FY25)
Revenue Total Growth	15.23x (+1,423%)	11.13x (+1,013%)
Revenue CAGR	57.44%	49.43%
Starting Net Income	$0.31 Billion (FY94)	$4.14 Billion (FY19)
Ending Net Income	$2.67 Billion (FY00)	$72.88 Billion (FY25)
Earnings Total Growth	8.48x (+748%)	17.60x (+1,660%)
Earnings (Net Income) CAGR	42.80%	61.29%

• Top-Line: Cisco"s top line compounded slightly faster (57.44% CAGR) than NVIDIA"s (49.43% CAGR) — multiplying total revenue over 15x.
• Bottom-Line: While Cisco compounded earnings at 42.80% CAGR to reach $2.67 billion, NVIDIA leveraged its pricing power and market position to compound net income at 61.29% CAGR, exploding nearly 18x to over $72 billion.

Cisco was an incredible business — printing vast amounts of cash. But looking back, I incorrectly blurred technological importance with long-term economic value.

But let me be clear – this is not to say NVDA will follow the same path as CSCO.

It may not. However, I would be willing to bet there is a very high probability their margins will not be sustained.

As I explained in my previous post – margins like what Cisco experienced typically don"t last for long.

One of the great things about capitalism is it will attract (fierce) competition.

For example, investors flood the lucrative space with money seeking similar returns, new competitors launch, existing players expand capacity, and supply scales up rapidly. We are seeing this play out in the chip sector today.

But when supply eventually outpaces demand, the technology begins to commoditize.

Customers—who suddenly have many alternative vendors to choose from—start buying based on price rather than brand.

Once a product commoditizes, the intense price-cutting war begins.

From there, margins erode and the massive amounts of capital required to stay in the game stop generating attractive profits.

Even though society benefits enormously from the cheaper, more abundant technology, the businesses themselves are left earning ordinary, or even dismal, returns on the capital tied up in their operations.

Every business, regardless of how sophisticated its technology might be, eventually confronts this exact economic reality: it must earn an attractive return on the capital invested (ROIC).

As Warren Buffett pointed out in his 1987 Berkshire Hathaway Shareholder Letter:

"By itself, this figure (earnings) says nothing about economic performance. To evaluate that, we must know how much total capital—debt and equity—was needed to produce these earnings."

Put simply, generating raw revenue and net income is only half the equation.

Long-term investment returns are decided by businesses that can protect their profits from aggressive competition and maintain consistent, strong returns on the capital they plow back into the ground.

As an aside, evaluating how well a company defends those returns over the long-term is one of the pillars of the "quality score" framework I shared in my last post.

A Personal Story on Innovation…

During my decade working at Google, an important observation was that innovative breakthroughs didn"t remain breakthroughs for very long.

Around 2018, we were bringing 3D models to the web as a way to transform product visualization.

The goal was simple:

As humans we live in a 3D world. However the information we typically interact with is in 2D (e.g., text, photos and video). Therefore, if we were to make more of the information we consume in 3D – it could help people answer visual and spatial questions more easily. For example – questions such as:

• How does this shoe look on my foot?
• What does this watch look like on my wrist?
• Does this table or chair fit in my lounge room? etc etc.

But several large technological challenges stood in our way (and this list is not exhaustive):

• Lightweight, interoperable 3D models didn"t exist at (Google) scale for web consumption – limited only to sophisticated gaming environments.
• Phones in 2018 were quite limited in how they could perceive your environment (e.g., lacking depth sensors; sufficient chips to experience AR);
• Creating 3D assets was incredibly expensive – where there was no proven guaranteed return on investment on the content.

Fast forward a few short years – we now find text-to-3D generation, alongside tools like Meshy, NVIDIA"s research models, and 3D Gaussian splatting that can instantly convert 2D video frames into navigable environments.

What"s more, these images are very difficult to distinguish from a product studio photo. Take a look at this example from Adidas and its popular SAMBA shoe.

What"s more, with 3D models exploding, this content can now training models for AI – generate and infinite array of (product) angles required foir learning.

This was (and is) and incredible piece of innovation (which is now widely available)

However, not long after our product teams started integrating these capabilities into surfaces like Search, competitors would respond surprisingly quickly with similar (or improved) technology (e.g., Amazon, Meta, Snap and others)

What was a breakthrough one year was considered a baseline the next.

Each push forward had a remarkably short half-life.

And whilst we kept advancing the technology every quarter, the expectations of what we needed to do advanced just as quickly.

After 30+ years working in tech – I"ve consistently observed that innovation creates temporary advantages, and markets have a habit of turning those advantages into baseline expectations.

And when I look at AI, I see the same pattern on a macro scale.

Every week brings another leading frontier model, another multi-billion-dollar data centre announcement, another hyperscaler increasing capital expenditure just to remain competitive (what I call the Red Queen Race), or another benchmark showing impressive gains.

Estimates of trillions in total capital investment over the coming years feature prominently in market commentary, and the argument you typically hear is that demand is insatiable, tech companies have to spend to meet it, and from there, those suppliers must inevitably become fantastic businesses.

Sure, the demand is real.

Companies such as Microsoft, Google, Amazon and Meta will remind you of this with every earnings report. Very few will suggest otherwise.

However, the more important question is whether the unit economics makes sense.

And whether it was the overbuilding of fibre-optic cables in the late 1990s; or brutal margin compression in building out America"s railways – society will certainly benefit from the technology (as it will with AI)

However commoditisation and falling prices is something investors need to monitor.

Will the (AI) Returns be There?

On the subject of unit economics – this brings me to what analysts like Ben Evans have been pointing out.

Evans argues that today"s market for AI inference is operating under unusually distorted conditions.

Demand has risen much faster than compute capacity can be built.

As a result, this has created a genuine supply shortage, while hundreds of billions flow into data centres (and the like).

Evans challenges the economics by asking what happens when (less so if) the cost of generating AI tokens falls dramatically over the next decade.

In other words, whilst not explicitly calling it out, Evans is citing the Cisco example.

Coming back to the capex boom in fibre during the 1990s:

• Bandwidth became cheaper, storage became cheaper, and computing power became cheaper.
• Each time, demand expanded far beyond what anyone initially imagined because entirely new applications suddenly became economically viable.
• But lower prices also test the economics of the companies supplying that demand.

Imagine two industries experiencing identical growth.

One operates behind durable competitive advantages, where customers rarely switch providers, returns on incremental capital remain consistently high, and competition stays limited. The other continually attracts new capital whenever margins improve, capacity expands, and customers increasingly choose suppliers based on price rather than differentiation.

Both can generate extraordinary demand.

However, it remains highly doubtful whether both will continue to generate extraordinary shareholder returns.

And whilst the technology baseline continues to improve (per my 3D/AR innovation story) — its not implied that translates into improved unit economics.

I don"t pretend to know whether frontier models like Claude, Gemini, ChatGPT or Llama (it"s a long list) ultimately become commodity infrastructure (like electricity); or whether a handful of companies develop genuine pricing power.

Another way to see this…

It could be the case that a "Google, Meta, Microsoft or Amazon" have incredibly powerful moats that they can sustain their margins long into the future (that"s the bet being made)…

But if you"re simply asking who has the best "AI benchmark" this month, you"re asking the wrong question, because that benchmark will quickly become tomorrow"s new baseline.

While metrics like context windows, reasoning models, and parameter counts are important, they count for very little if they fail to capture sustainable economic value.

From where I sit — the more interesting question is whether businesses (and consumers) continue paying for additional intelligence once there are perhaps "ten" companies offering something that is good enough.

One might ask:

• How many customers genuinely require the most capable frontier model rather than one that is simply sufficient?
• Will the frontier continue advancing quickly enough to justify ever-larger compute investments, or will efficiency improvements gradually compress pricing?
• Does competition narrow to a handful of dominant players, or do frontier models become increasingly interchangeable?
• And how much of the economic value ultimately resides inside the model itself versus the software, workflows, and industry-specific products built on top of it?

I don"t have the answers to any of these questions (only theoretical guesses).

But it"s the answers to these questions I will be watching.

Putting it All Together

This post is designed to help you think about the investment side of this shift.

During the early days of the internet, very few people predicted cloud computing, digital advertising, smartphones, or streaming.

They correctly identified the technology and the expected demand—they simply couldn"t foresee where the economics would eventually settle.

It turns out that companies like Google, Amazon, Meta, Microsoft and Apple captured most of the economic value (which was built on top of the internet)

But will they tomorrow?

Whenever investors encounter a transformational technology, there is a natural temptation to search history for perfect analogies, comparing AI to electricity, semiconductors, or mobile networks.

Each comparison contains useful lessons, but none provides a definitive answer.

History doesn"t repeat with sufficient precision to predict outcomes, but it reminds us of a vital truth: transformational technologies do not automatically produce transformational investments.

• Sometimes the infrastructure captures most of the value;
• Sometimes the applications do; and sometimes the
• Companies that ultimately dominate haven"t even been founded yet.

That said, uncertainty isn"t something to fear – it"s simply the price investors pay for participating in periods of profound technological change.

I don"t lose sleep over who has the best AI model today.

But I am concerned about the economics that will remain after today"s supply shortages ease, competition intensifies, and abundant capital begins doing what it has always done.

Regards
Adrian Tout