On 10 December 2025, Cisco Systems closed at $80.25 - finally clearing its dot-com peak from 27 March 2000. And it only took twenty-five years, eight months, and thirteen days. Adjusted for CPI, the stock still trades at roughly half its real 2000 high; meaning anyone who bought at the top has waited a quarter of a century to break even in nominal dollars, and in actual purchasing power, they’re not there yet.

Cisco was arguably one of the most important companies of the dot-com era; routers and switches were the picks and shovels, and they sold the infrastructure everyone else had to buy. They were profitable, and they had actual revenue and customers. The boxes they shipped were as close to indispensable as networking gear got.

And the stock returned a generation of negative real performance.

Hold that fact in your head.

NVIDIA closed last week at a market cap hovering around $5 trillion. The stock trades at more than forty times trailing earnings and an enterprise value of about twenty-four times sales. In a year, the company has added more market value than the entire UK FTSE 100. Retail investors hold it in their pension accounts and their ISAs. Every algorithmic momentum strategy holds it, and so does every passive index fund.

The stock has started behaving like a vote on the future itself.

But if history is any guide, the short version of that future is this: a decade of expected real returns hovering near zero, possibly negative.

The dominant infrastructure provider in a capex supercycle returns less than people think over the decade following peak euphoria, sometimes by enormous margins. There’s never been a clean exception, and there’s no theoretical reason there should be one now.

So: the base rate.

The first great infrastructure capex cycle in the modern industrial era was the American railway boom. Between 1868 and 1873, more than 33,000 miles of track were laid. Steel rails, the air brake, signal telegraphy, federal land grants etc. It was the most ambitious physical buildout the still-young country had ever attempted. And those railways did remake the economy; New York to Sacramento compressed from six months to seven days. The integration of national markets that followed turned Sears, Roebuck and Standard Oil into empires.

But the railways themselves went bankrupt. The Panic of 1873 wiped out more than a hundred railway companies, and the Panic of 1893 took out roughly a hundred and fifty more. By the early twentieth century, the sector was a graveyard of bondholder workouts and federal receiverships. Over the following century, railway equity returns underperformed broad indices so consistently that they’re still considered textbook case of infrastructure value destruction.

The customers got rich, but the builders got hosed.

The 1920s electrification cycle produced the same outcome. Utility holding companies became the most fashionable equities in the world. They traded at multiples that made their cash flows look ridiculous, and once the crash started in 1929, electric utility holding companies lost more than ninety per cent of their value by 1932. Electrification was real; it changed agriculture and manufacturing.

The industrial value was generational, but the shareholders got hosed.

In the 1970s and 1980s it was mainframes. IBM was the indispensable compute provider of the post-war era. By the late 1960s it commanded the moat NVIDIA enjoys now; and through the 1980s and 1990s, IBM equity returns were a slow-motion disaster.

Between 1996 and 2000, telecommunications capex rose by more than seventy-five per cent cumulatively. Global Crossing, MCI WorldCom, Lucent, Nortel, AT&T, Qwest, and dozens of others laid roughly 80 million miles of fibre optic cable on the pitch that they were going to wire the planet. They had real customers, and Government policy was at their backs. Analyst models justified pretty much any P/E ratio you wanted.

Eighty-five per cent of the fibre sat dark for a decade. The cost of bandwidth fell by ninety per cent, and the telecom sector collapsed by ninety-two per cent. Twenty-five years on, the index still hasn’t recovered. Nortel went bankrupt. Global Crossing went bankrupt. Lucent had to be absorbed into Alcatel and then again into Nokia. The fibre that survived became the substrate Netflix and Facebook and Google and Amazon used to extract trillions in value.

The customers won, but the builders got hosed.

Cisco was supposed to be different; they were the picks-and-shovels play, a real business with real margins, selling real equipment to real customers. At its 2000 peak, Cisco traded at a P/E above two hundred and an enterprise value to sales ratio above thirty. The argument for those multiples was that Cisco was indispensable, that internet capex would compound for decades, and that nobody else could deliver what Cisco delivered.

All of that turned out to be true. The stock still hasn’t recovered in any real term.

The shale gas mania of 2010 to 2014 produced the same pattern at smaller scale. Fracking worked, the geology was real, the energy independence was real, and the equity returns for the upstream producers were a serial-killer chart. But Chesapeake Energy went bankrupt. Whiting Petroleum went bankrupt. Shale capex peaked, then collapsed, and then collapsed again.

The technology worked. The shareholders…well.

You understand.

You can run this exercise as far back as you want; canal mania in 1840s Britain, the bicycle craze of the 1890s, and the automobile boom that produced more than two hundred US car companies in the 1920s and a Big Three controlling ninety-four per cent of the market by 1955. Aviation in the 1930s. Photovoltaic cells in the 2000s. Each cycle had a real innovation and a defensible thesis for why the dominant infrastructure provider would compound shareholder value indefinitely; each cycle ended with that provider returning less than US Treasuries over the subsequent decade.

Capex cycles almost always overshoot; the economics of competitive bidding for productive capacity all but guarantee that capacity gets built past the point where the marginal unit earns its cost of capital. Builders capture the rents during the build phase, when demand outstrips supply. Once supply catches up, the rents collapse. The customers absorb none of the capex risk; they capture the durable value because they get the productive asset at clearance prices.

In its most recent quarter, four direct customers accounted for sixty-one per cent of NVIDIA’s total revenue; and just two customers accounted for thirty-seven per cent. The same disclosure existed at Cisco around 2000, where the largest customers were the very telecoms that subsequently went bankrupt. Concentration like this is what the late innings of a capex bubble look like; by that point, only a handful of buyers can write cheques big enough to matter, and those buyers know it.

Those four customers are projected to spend somewhere between $700 billion and $725 billion on capex in 2026 alone. To put that in perspective: US telecom capex peaked at roughly 1.2 per cent of GDP in 2000. AI hyperscaler capex is now on track to match that share in a single year, compressed into a sliver of the telecom cycle’s duration and concentrated among four buyers.

Fibre, once laid, lasted forty years. Railways, once built, lasted a century. NVIDIA’s H100 was state-of-art two years ago and is now a generation behind Blackwell. Blackwell will be a generation behind Rubin. Rubin will be a generation behind whatever follows. Hyperscalers are buying assets they’ll write down in three to five years. The capex is enormous and the asset life is short; that math can work for a few years, but it can’t work for a decade.

Google’s TPUs already handle a meaningful share of internal workloads. Amazon has Trainium and Inferentia, Meta has MTIA, and Microsoft has Maia. Broadcom is forecasting $100 billion in AI chip revenue by 2027. TrendForce expects ASIC-based AI servers to hit roughly twenty-eight per cent of shipments in 2026. The CUDA moat matters less for inference than it does for training, and the buyers know this; they’re spending billions to escape the supplier they currently depend on. Every one of them has board-level pressure to reduce NVIDIA’s share of their cost base.

This is the dynamic that killed Cisco’s premium. Customers built their own gear, standards commoditised the function, and white-box switching ate the margin. Cisco the company survived, but Cisco the stock didn’t.

When does the multiple compress on NVIDIA?

In Q1 or Q2 of 2027, one of the four mega-customers will say, on a quarterly earnings call, that its capex growth rate is moderating. They won’t call it a slowdown; it’ll be “rationalisation” or “improving capital efficiency.” That same quarter, MediaTek’s CoWoS wafer allocation from TSMC will have ramped from 20,000 in 2026 to 150,000 in 2027, which is when custom silicon becomes a credible second source. Broadcom’s AI revenue will be tracking towards its $100 billion target; and the market will see, for the first time, that the substitution is starting to happen at scale.

NVIDIA’s forward earnings multiple compresses from twenty-five times to somewhere around twelve to fifteen. Earnings themselves keep growing, possibly substantially, for another year or two. But the algorithmic strategies that bought NVIDIA on momentum will sell it on momentum. Passive flows that pushed it up via S&P 500 weight will pull it down via the same mechanism; and the stock falls fifty to seventy per cent over twelve to eighteen months.

The compounding math does the rest. Even if NVIDIA grows revenue at ten per cent annually for the following decade, an outcome that would still make it one of the great businesses of the era, and one of the greatest of all time, the de-rating from a premium multiple to a normal one eats every cent of that growth and most of the inflation. Actual returns through 2035 settle somewhere between zero and slightly negative.

This is, broadly, what the historical base rate looks like for the dominant infrastructure provider in any capex supercycle. The stock returns nothing while the underlying economy gets remade.

Half of you reading this think I’m wrong because NVIDIA is different. Cisco didn’t have CUDA. Railways didn’t have Jensen. IBM didn’t have a software moat compounding into robotics and simulation and the foundation models that will define the next decade of computing.

Maybe.

I’ll own that possibility.

I’d be a fool not to.

The other half will think I’m wrong because the bubble is bigger than I’m describing. The multiple compression is already overdue, and the customer concentration plus asset depreciation plus ASIC substitution mean the de-rating starts this year.

Also maybe.

What both halves should hesitate before believing is that NVIDIA’s stock will compound from here at the rate it has compounded so far; because the historical base rate for that outcome is roughly nil. Every previous capex supercycle has ended with the dominant provider returning less than cash for the decade following peak. There’s never been an exception, and there’s no theoretical reason there should be one now beyond “trust me, bro.”

NVIDIA’s customers will get artificial intelligence. And their shareholders will get to find out what twenty-five years of waiting feels like.

Selfonomics is published by my lab, Studio Self.

We make tech legible.

Work with us

My fourth month of unemployment. I'm fresh off an interview with a senior technical role advertised with a decent compensation range whose upper end would let me make median rent, and whose lower range would not. At the end of an excellent interview, I am informed that said compensation range, contrary to the Base Compensation tag in the job description, is in fact all inclusive, and that the role actually tops out below the living wage threshold for a three adult household for its base compensation.

I feel it's highly relevant to note that this was a role where I would own the entire technology estate for an employer. Every switch, every firewall, every database, every server, every phone, every laptop, every cloud, every badge system. All data, everywhere, at any time.

Everything.

For an employer in the public safety industry.

This is not the first time that compensation has been mismatched to the role, but it is the first time it's happened with a role so critically important to a company playing such an outsized role in as critical an area as public safety.

Which got me thinking: have employers just lost the plot on what compensation is actually for? What its intended function is? Because it seems to me that everyone thinks compensation is merely payment for labor and nothing more, a number to be driven down by any means necessary in order to keep more for those at the top.

And oh my god it is so much more than that.

We live in a society. This society has been arranged around using currency to purchase necessities, because some people decided that necessities are not guaranteed. You acquire currency by either investing currency you already have to compound it through the labor of others, or you labor in order to earn less currency than investors do.

Is that a gross oversimplification? You betcha, but I'm really trying to stay on topic and not fill this with reaction GIFs.

Author's Note: I failed.

Anyway, everyone needs currency to afford everything. Rent costs currency, food costs currency, electricity costs currency, water and sewage and garbage and healthcare and childcare and education all cost currency. Only after your essentials have been paid for, can you use excess currency to save for tomorrow - buying a home, or a car, or a retirement if you're really lucky.

Now at some point, the humans involved in handing out currency decided that too many people were living too nicely. The thinking was simple: trading time for currency in the form of labor was a sucker's game, and those with currency deserve more currency because they already had currency. Companies in particular deserved more money than the people who worked for the company, at least according to Powell.

Thus labor was reframed into those terms: wages were merely a payment to those who did labor, and labor was only to be paid the minimal amount possible in the marketplace for said labor, and not one cent more, regardless of external forces (like the cost of living). Minimum wages are bad, because wages should only ever go down relative to inflation and productivity, never up. Only those who own business deserve more currency, because they're the real workers.

Again, a gross oversimplification coming from an openly biased dinosaur. That's not the point.

The point is that wages aren't meant to merely compensate labor; they're also meant to protect the company.

I'll just be blunt: wages are also protection money. They're not just compensation for doing your job, they're compensation for not burning down warehouses, not going on strike, not sabotaging workloads, and not unionizing in the first place. It's the longest unspoken social contract dating back to pre-history: you pay me to live, or else.

Occasionally, employers will call labor's bluff. If twenty-thousand years of history is any indication, their temporary wins are always undone by the sheer ratio of workers to wealthy owners, though not before employers provoke and employ violence first. We are going through such a phase right now, if the above links are any indication.

Employers haven't paid their dues to labor in decades. Labor, kind as we are, have cut back on our lives as much as we're able to. We've given up homeownership, we're dropping out of healthcare, we're begging from food pantries, we're taking on gig work, and we've seen none of the wage gains from productivity our elders enjoyed. It's gotten so bad that, with workers having sacrificed our very ability to move up the socioeconomic ladder, the economy has gone K-shaped.

Despite this, employers seem to think there's more room for workers to yield. I say this because despite median wages being unable to afford median homes, I'm still finding employers offering lower and lower pay for work on their job descriptions. This isn't just economics anymore, this is a security risk, and employers are playing with fire.

Take your CPA, for instance, median pay of $81,680 a year and at moderate exposure to AI displacement according to Anthropic. These are people who know your books better than you. Where every penny goes, and where every penny can be silently swindled. They also audit your fancy new AI financial workflows, making corrections when it goes off the rails, and know where all your financial skeletons are buried.

Do you really want to pay your accountant so little that they can't make rent or buy a home? Do you really want them going to a food bank instead of a grocery store? They're excellent at judging risk, and know exactly what that pay gap is worth to them if the timing is right. Maybe it's blowing the whistle before you can fix an issue and leading to a costly investigation, maybe it's sharing your supply chain costs with a competitor for a higher paying job, or maybe it's committing outright embezzlement that they're sure your fancy AI tooling will miss.

Tech work, the last bastion of Middle Class employment, isn't doing any better. Take Computer and Network Architects, with a median pay of $130,390 per year; or the Computer Systems Analysts, median pay of $103,790 per year; or the Systems Administrators, with a median pay of $96,800 per year. These groups have the keys to your systems, your data, your endpoints, your real estate. They can see and do anything with a keystroke, including destroying billion-dollar businesses. The reason for the comparatively high wages was their comparatively high degree of trust.

Instead, employers outsourced to MSPs, then offshored overseas, then on-shored to underpaid and exploited H1Bs shackled to their employer's temporary sponsorship, then briefly hired North Korean spies, and are now attempting to replace the technical workers outright with AI that routinely drops production systems. All the while they lay off workers by the tens of thousands, over and over again.

I feel like there's an example of the consequences of not treating your technical experts with respect in popular media...

I'm not Dennis Nedry, but I've worked with folks like them before. Brilliant minds who can debug complex architectures and systems, who pour their lives quite literally into the work because they have a passion for it...and increasingly are all too willing to burn it to the ground when they feel slighted. Spend time around actual engineers and the like in most orgs, and you'll see patterns of ill-health: smokers, drinkers, chewers, vapers, over-eaters, out-of-shapers, poor posture, bags under eyes, thinning and greying hair, high amounts of stress, messy desks. All signs of humans sacrificing their own health for their employers, prioritizing work over life, overworked to an early grave.

Most folks aren't as egalitarian as I am, and as someone who has sacrificed physical, mental, emotional, and psychological health to the field for over fifteen years, I sympathize with where my peers are coming from. Most people aren't wired to "do no harm" no matter what, which means most people are a huge security risk if they're undercompensated.

Thing is, undercompensation isn't limited merely to your specialists and senior workers. Fast food workers will slow down lines to give themselves breathing room due to understaffing, and retail workers won't put in the added effort of store maintenance when they can't even maintain a roof over their heads. Office workers doing more menial tasks aren't going to follow through on security best practices if they're more worried about how to pay the electric bill this month while also affording insulin. Your contracted-out security staff aren't likely to pay close attention to camera feeds since they know they'll be replaced in three months before benefits kick in. Your MSP or offshored technical staff won't be invested in your long-term success when their KPIs only cover ticket counts and response times, and their competitors are already preparing to underbid at renewal anyhow.

The workers have been incredibly clear about their problems for twenty years, now, especially the younger cohorts. Employers haven't wanted to listen, believing one more technical control or one more AI system will finally give them the permanent, unassailable leverage they need to keep all the money and fire all the workers.

I don't really have a positive way to end this. This is a warning, another canary in an increasingly smoke-choked mine. We're at the point that workers are quite literally burning down infrastructure and engaging in violence against leadership, and the response from those who can change things - our politicians, our corporate leaders, the investor class that's richer than ever in human history - don't really seem to give a fuck. There's this thick tension in the air between workers scrambling to survive, and monied classes who feel the demands of the workers are wholly unreasonable.

History paints a pretty clear picture of how this ultimately ends, but for what it's worth, I still feel like I should at least try to warn folks about the consequences of undercompensation.

Failing to pay your workers the money they need to live is breaking the social contract. It's the single biggest security vulnerability in your organization, and I promise you that there is not, and never will be, a technological control that can protect against it.

You gotta pay up, or you're going to get burned down.

An addendum.

AI is making it faster and easier to brute-force security vulnerabilities at a time when open source is falling apart due to lack of funding and successors. Major companies are firing engineers to replace them with AI tooling, then hiring them back at lower pay packages when the AI fails, but still holding the AI Sword of Damocles over their heads. Software is expanding rapidly at a time when employers seek to eliminate the technical professionals who ensure their safety and prosperity, who can translate institutional processes and knowledge into cost-effective infrastructure.

Housing prices are up. Rent is up. Utilities are up because of AI datacenter builds. Food costs are rising due to global conflicts instigated by America. So too are energy prices, tariffs, inflation, and interest rates.

You, the employer, have a decision to make: do you start raising wages, working with policymakers to immediately address affordability, cease arbitrary layoffs, invest in worker futures, and promote regulatory schemes that reign in the worst myopic excesses of your peers for society's collective benefit?

Or do you take up smoking cigarettes while sitting inside a warehouse of loose gunpowder and dynamite, with a mob of torches and pitchforks right outside?