Is AI a Bubble Like Telecom? A Founder's Filter

August 31, 2026 · 13 min read

The telecom boom laid fiber that sat dark for a decade; social media threw off cash almost from the start. Same wave of technological optimism, opposite outcomes for owners. The difference wasn't the technology — it was how fast spending turned into recurring free cash flow. That's the filter worth carrying into every AI conversation.

Roughly once a month, some version of the same question surfaces among founders who have recently sold a business and are now deciding where to put the proceeds. The AI build-out, with its historic sums flowing into data centers, specialized chips, and power infrastructure, is said to echo the late-1990s telecom boom. Fiber was laid across continents and under oceans. Much of it sat unused for years. Revenue lagged the capital deployed, debt service became impossible for many players, and equity was destroyed — even though the underlying technology eventually proved essential.

The parallel is worth taking seriously. But the useful version of the argument is more precise than a blanket warning. The same era that produced the telecom wreckage also produced a second wave of businesses — social media platforms — that scaled users and began generating cash on a timeline that allowed free cash flow to appear quickly. Both waves rode the same broad technological shift. The outcomes for owners diverged sharply. And the reason they diverged has almost nothing to do with who had the better technology.

It comes down to one question: whose spending converts into recurring free cash flow — the money a business has left over after paying to keep the lights on and to keep growing — and how soon? That question is the filter. Not "who has the best model?" Not "which company is spending the most?" Those are technology questions. The one that matters for owners is a cash question.

Two booms, two very different endings

Telecom: right about the future, ruined by the timeline

The mechanics of the telecom episode are worth sitting with, because the lesson lives in the details.

In the late 1990s, a belief took hold that internet traffic was about to explode. That belief was essentially correct — traffic did explode, eventually. What followed was one of the largest infrastructure build-outs in modern corporate history. Carriers and equipment suppliers raised enormous sums and buried vast quantities of fiber-optic cable across countries and under oceans, on the assumption that demand would arrive on the same schedule as the construction.

It didn't. Demand grew, but far more slowly than the build assumed. Much of that fiber sat dark — installed, the capital already spent, but carrying no traffic and therefore earning nothing.

Here is the part that owners should sit with, because it contains the whole lesson. Much of that build was financed with debt — borrowed money that has to be serviced on a fixed schedule, whether or not the revenue shows up. The fiber could wait patiently for demand. The interest payments could not. When the cash to service the debt failed to arrive on time, a number of once-celebrated companies collapsed, and enormous amounts of invested capital were destroyed.

Notice what did not fail: the technology worked, and the demand did eventually appear. Investors still lost fortunes — because the cash the assets produced arrived years after the bill for building them came due, and the debt in between did the killing. Being right about the future is not the same as being paid for it on a timeline that keeps you solvent. That gap — between when you spend and when the cash comes back — is where owners get ruined. Leverage is what turns a survivable wait into a fatal one.

This pattern is not unique to fiber. The nineteenth-century railroad booms and the early build-out of the electrical grid rhyme with it closely: massive upfront investment, real and lasting infrastructure, and a long list of original investors who never saw the returns because they had borrowed against a future that arrived a decade late. The assets endured. Many of the owners did not. The businesses that eventually captured lasting free-cash-flow growth were those that developed paying customers and efficient operations around the infrastructure — not those that simply supplied the physical layer.

The businesses that scaled a few years later looked almost nothing like this. A social platform did need real infrastructure — server farms are not free — but adding its next hundred million users cost very little relative to the revenue those users generated. The business model was already working as the build happened: usage and monetization arrived together, more or less concurrently.

That's the crucial mechanical difference. When a telecom company laid a cable between two cities, that cable earned exactly zero until a paying customer sent data across it — a speculative bet on demand that hadn't materialized yet. When a social platform added servers, they filled with users almost immediately, and those users were already being monetized through advertising. The spending converted into cash as it happened, which meant these companies could fund much of their own growth out of the cash they were already generating — without leaning on the dangerous levels of debt that doomed telecom.

Once leading platforms reached scale, incremental usage produced incremental revenue without proportional increases in capital spending. That created a self-reinforcing loop: cash flow funded product improvements, which attracted more usage, which generated more cash flow. The compounding was visible and relatively fast.

Not every social-media venture succeeded, of course. Many burned capital chasing growth without a clear path to durable revenue and eventually faded. The distinction that mattered for long-term owners was whether the business model turned engagement into predictable, high-margin revenue that exceeded ongoing costs — and produced growing free cash flow. Two business models, one era, one lesson. One required vast upfront capital and a long, uncertain, often debt-financed wait for the cash to justify it. The other required comparatively little capital and turned growth into cash almost in real time. Both rode the same real technological wave. The economics of owning a piece of each were opposites.

Free cash flow is the only honest scorekeeper

The lens that cuts through the noise in moments like this is the same one that cuts through it in every other moment: long-term returns are driven by the growth of a business's free cash flow over time. Own quality, durable, cash-generative businesses, let that cash flow compound across decades rather than quarters, and the returns tend to follow. It sounds almost too plain to be useful. But its plainness is exactly what makes it valuable in a period of enormous excitement — because excitement rewards every other kind of measurement first.

Think about how a boom feels from the inside. The numbers that get quoted are user counts, model benchmark scores, chips ordered, gigawatts of power secured, capital committed. Every one of those can be genuinely staggering and still tell an owner nothing about whether they will ever see a dollar of durable profit. The telecom era had spectacular metrics too — miles of cable laid, network capacity multiplied many times over. The figures were real. They simply weren't cash.

Free cash flow is unforgiving in precisely the way excitement is not. It asks the questions the story would rather skip. After a business pays to build and maintain everything it needs — the data centers, the chips, the power, the engineers — how much is actually left? Is that leftover amount growing? And is it recurring — does it show up again next year without the entire enormous bill having to be paid all over again to earn it?

That last question is where the telecom-versus-social-media contrast really bites. A business that must re-spend heavily every few years just to stay in place is running hard to stand still. A business whose spending builds something that keeps paying, year after year, with far less fresh capital required to sustain it, is compounding. Two companies can report identical revenue and be completely different animals underneath — and the difference only becomes visible when you follow the cash instead of the headline.

So which one is AI?

Here honesty is required, because the satisfying answer — "it's telecom" or "it's social media" — is the wrong answer. AI isn't one thing. It's a stack of very different businesses sitting at very different points on this spectrum. The useful move is not to judge "AI" as a category but to sort the layers by how their spending converts into cash.

The capital-heavy layer

Some of what's being built right now closely resembles the telecom story in shape. The physical foundation of AI — the data centers, the specialized chips, the power to run them, the cooling to keep them from melting — requires spending on a genuinely historic scale. Annual capital expenditures across the sector have reached levels that would have seemed implausible not long ago. This is real infrastructure with real economics, and some of it will prove enormously valuable.

But it carries the telecom warning label, and one that's arguably sharper. Fiber, once buried, degrades slowly; a dark cable could sit dormant for years and still work perfectly when demand finally arrived. The specialized chips at the heart of AI are not passive glass. They lose competitive value quickly as newer, faster versions arrive, which means much of this spending is not a one-time cost but a treadmill: to stay current, you keep buying. That compresses the timeline. There is far less luxury to wait a decade for demand to catch up, because the asset itself doesn't wait.

An owner looking at this layer should be asking the timing-and-financing question, not the technology question. Not "is this powerful?" — obviously it is — but "how many years until the cash it produces exceeds the cash it keeps demanding, how much of the build is financed with debt that won't wait, and what has to stay true the whole way through?"

The layer that already converts

At the other end sit businesses that were already durable, cash-generative franchises before this wave began, and are now folding AI into what they already do. A business with an established product, real pricing power, and customers who pay again every year does not need AI to eventually justify a decade of speculative spending. It can adopt AI to widen its advantage, lower its costs, or deepen how much its customers rely on it — and see the benefit in its cash flow relatively quickly, because the cash engine is already running.

This is the quieter, less headline-friendly part of the story, and it's often where the more durable economics live. The businesses least likely to be described as "AI plays" may be the ones best positioned to turn AI into recurring free cash flow — precisely because they were converting spending into cash long before the current excitement, and are simply adding a sharper tool to a machine that already works. In prior infrastructure waves, the greatest value often accrued not to the lowest-cost builders of the physical layer but to companies that controlled customer relationships, data advantages, or distribution that made their cash flows harder to displace. The same logic applies here.

The uncertain middle

Between these sits a large, noisy middle: newer businesses selling AI-enabled products, many with impressive growth and real revenue, but without settled answers to the questions that ultimately separate who compounds from who fades. Do customers stay, or is the cost of switching to a competitor close to zero? Is the pricing durable, or does it collapse the moment a larger rival offers the same capability for less? Does each dollar of revenue arrive with healthy margins — or does it require an almost equal dollar of spending on the underlying computing power to deliver it?

A demo cannot answer these questions. Only time and cash flow can. The instinct many feel when "bubble" enters the conversation is a binary choice — commit heavily or stand aside entirely. Both treat AI as a single undifferentiated exposure. The more useful posture is to apply the conversion filter to each specific opportunity: where does this business sit on the spectrum from pure infrastructure to established cash engine, how quickly does its spending convert into recurring cash, and how much fresh capital must be deployed simply to maintain position?

The cost-of-capital thread that runs through every cycle

There is a second thread worth pulling, and it connects straight back to the telecom lesson — because it is the same mechanism.

Much of what amplified losses in the telecom period was not the technology failing but the financing structure. When revenue lagged and the cost of borrowing rose — or simply stopped falling — patience evaporated. Investors shifted preference toward cash today rather than cash far in the future. Companies that had borrowed against a distant payoff found the market unwilling to keep extending that bet. The result was not a slow fade but a rapid collapse of equity value, even for assets that were physically real and eventually useful.

The same dynamic applies whenever capital markets re-price the value of waiting. A period of rising interest rates — the cost of borrowing money — tends to compress how much investors will pay today for cash flows that are years away. Businesses whose entire investment case rests on a future payoff feel this acutely. Businesses already generating strong and growing free cash flow feel it far less, because they are delivering now rather than promising later.

For long-horizon owners, the implication is not a timing call — trying to predict when rates move is not a game that rewards most players. It is a structural preference: favor businesses whose economics do not require the market to remain patient indefinitely. A company already generating substantial and growing free cash flow is less hostage to shifts in borrowing costs or sentiment. A company whose story depends on years of further investment before meaningful cash appears carries far greater sensitivity to those shifts. The through-line remains the same across every cycle: durable growth in free cash flow is both how returns compound over decades and what allows owners to maintain positions through periods when sentiment reprices the future.

This is, ultimately, what the telecom-versus-social-media comparison is really about. It is not a prediction about whether AI will prove transformative — it almost certainly will, in some form, over some timeline. It is a reminder that being right about the technology has never been sufficient. The owners who compounded wealth across the last several technology waves were not necessarily the ones who identified the winning technology earliest. They were the ones who identified the businesses whose spending was already converting into durable, growing cash — and held those businesses long enough for the compounding to do its work.