Are we in an AI bubble – and what kind of bubble is it?

AI is driving one of the biggest technology buildouts we’ve ever seen. Chipmakers, data center operators, and cloud providers are pouring hundreds of billions of dollars into GPUs, memory, storage, and power. Stock prices in AI-related names have gone vertical. The obvious question: are we in a bubble?

The more interesting question: even if this is a bubble, what exactly is inflated – AI infrastructure, corporate earnings, or the value of human labor itself?

Inflation is cooling, but money is still expensive

Before looking at AI directly, it helps to understand the macro backdrop. Inflation, at least by some measures, is easing. Traditional core PCE is still above the Federal Reserve’s 2% target, but alternative measures like the “trimmed mean” PCE (which strips out the most extreme price moves) are trending closer to 2–2.5%.

Even so, interest rates remain high. The long end of the yield curve has backed off a bit, but consumer borrowing costs are still elevated. Auto loans and credit cards are seeing delinquency rates climb to levels last seen around the global financial crisis. Student loans and auto loans are at or near series highs for late payments, and mortgage delinquencies are ticking up from very low levels.

In other words, households are under pressure. Personal income growth is flat, but spending is still rising – likely funded by expensive credit. That’s important context: the AI boom is happening in a world where money is not free and consumers are already stretched.

The stock market is partying like nothing can go wrong

Despite the macro stress, equity markets – especially tech and small caps – have been on a tear. Major indexes like the NASDAQ (via QQQ), small caps (IWM), and the S&P 500 (SPY) are sitting at or near all-time highs. One-year returns in some segments are eye-popping:

• QQQ: well over 100% over three years and more than 120% over five years
• IWM: beating the S&P on a one-year basis
• SOXX (semiconductor ETF): ~61% in three months, ~173% in one year

Individual chip and storage names are even more extreme. Some memory and storage stocks have posted 900%–4,000% returns over the last year. Cybersecurity and SaaS names have run 50–80% in just three months and are heading into earnings at all-time highs, priced for perfection.

Implied volatility is low, signaling that options markets are pricing in very little risk. The message from equities: there’s nothing to worry about.

AI capex is exploding – and so are earnings

Under the surface of GDP data, AI is already showing up in a big way. In recent US GDP revisions, a huge share of equipment investment growth came from information processing equipment and software – exactly the categories tied to AI data centers and infrastructure.

Estimates suggest that by around 2026, annual AI-related capital expenditure could reach $450–$550 billion. That includes:

• Silicon: GPUs, memory, storage, and related chips
• Equipment: servers, networking gear, cooling systems
• Structures: the physical data centers and power infrastructure

Yet current AI revenue run rates are only around $70 billion. That’s roughly a 7:1 ratio of capex to revenue. Earnings look fantastic right now because companies are recognizing all of the new AI revenue, but only a small slice of the associated costs – the depreciation – has hit the income statement yet.

For example, if we assume $500 billion in AI capex per year, and spread that across typical depreciation schedules (about six years for silicon, 15 for equipment, 30 for structures), the first year might only show around $54 billion in depreciation expense against hundreds of billions in gross margin. That’s why profits look so strong.

But depreciation accumulates. If companies keep spending at that pace, the annual depreciation bill rises sharply in later years. At some point, the depreciation schedule can catch up and start to overwhelm the incremental AI revenue – unless that revenue grows even faster.

The four classic stages of a tech capex bubble

To understand where we might be, it helps to use a simple framework for capex super-cycles. Historically, these cycles have four main stages:

Stage 1: The trigger

Every big cycle starts with a credible new general-purpose technology or a major demand shock. AI clearly qualifies as a general-purpose technology: it can, in principle, touch almost every information and knowledge-based task.

ChatGPT and similar models proved that scaling up these systems works. As models got bigger, their capabilities improved in a way that felt discontinuous to many users. That created a belief that AI demand could be effectively unlimited – from chatbots to AI agents to embedded AI in physical systems.

Early on, the bottleneck was GPUs. You simply couldn’t get enough high-end chips. Now, as GPUs become more available, other constraints are emerging: memory, storage, power, wafer capacity, and even physical sites for data centers.

Stage 2: The acceleration or land grab

Once the opportunity looks real, the land grab begins. Being late is framed as existential. Companies tell themselves they must move now or be left behind, and capex budgets start to explode.

In this phase:

• High early returns and fear of missing out (FOMO) pull in a flood of capital.
• Capital is allocated based on competitive positioning, not careful ROI analysis.
• Financing is abundant: equity, debt, and vendor financing all pile in.
• The “picks and shovels” players – chipmakers, memory vendors, data center builders – post explosive revenue growth and fat margins.

That’s exactly what we’re seeing today in names like Nvidia, Micron, and the big cloud providers. Their valuations have run the hardest, and their guidance keeps pushing capex higher.

Stage 3: Peak and saturation

In the peak phase, capacity finally catches up to – and then overshoots – near-term demand. That can happen in two ways:

• Supply keeps rising while demand growth slows.
• Demand actually falls while supply is still ramping.

Key signs of this stage include:

• The return on new capital starts falling, but spending continues because of competitive pressure.
• Capex-to-sales ratios hit extreme levels.
• Investors and analysts start asking: “Where is the return on all this spend?”
• Financing structures become more aggressive and circular – special purpose vehicles (SPVs), vendor financing, and off-balance-sheet debt.

We already see many of these footprints. ROI is becoming part of the conversation. Companies are quietly questioning their AI token budgets, and some are discovering that AI usage is more expensive than the humans they hoped to replace – at least for now. Yet capex budgets are still being revised upward, which is very much a Stage 2 behavior.

That’s why it’s reasonable to say we’re straddling Stage 2 and Stage 3: the land grab is still on, but the warning signs of saturation are appearing.

Stage 4: The bust and digestion

The bust comes when either demand disappoints relative to the capacity built, financing tightens, or both. Because data centers and chips are largely fixed costs, operators cut prices to fill capacity. Utilization falls, pricing collapses, and returns on invested capital plunge.

In this phase:

• Orders don’t just slow – they fall off a cliff as new buildout stops.
• Weaker players go bankrupt or get acquired.
• The suppliers that rode the boom the hardest (GPUs, memory, storage) get hit the hardest on the way down.

If we reach this stage in AI, companies like Nvidia and major memory vendors could see demand dry up for new capacity while the industry digests what’s already been built.

Off-balance-sheet AI: SPVs and circular financing

One of the clearest signs that we’re in late-stage acceleration/early saturation is the rise of complex financing structures around AI infrastructure.

A common pattern looks like this:

• An SPV (special purpose vehicle) is created to build and own a data center.
• The SPV raises tens of billions in debt from private credit and asset managers (PIMCO, BlackRock, Apollo, Blue Owl, etc.).
• A big tech company (Meta, Oracle, an AI startup) signs a long-term lease for the capacity.
• On paper, the tech company records lease payments, not debt – keeping its balance sheet cleaner.

Recent examples include:

• Meta’s ~$30 billion data center project in Louisiana, financed via an SPV with $27 billion in debt and $3 billion in equity, none of which appears as debt on Meta’s balance sheet.
• Oracle’s OpenAI-related facilities in Texas, Wisconsin, and New Mexico, financed through multiple SPVs with multi-billion-dollar loan packages.

These SPVs can even receive investment-grade ratings, despite their sponsors being high-yield credits. On top of that, there are already signs of “bonds backed by bonds backed by SPV bonds” – very reminiscent of pre-2008 mortgage securitization structures.

This kind of circular, off-balance-sheet financing is a classic Stage 3 footprint. It allows the boom to continue longer than it otherwise would, but it also concentrates risk in opaque corners of the credit market.

Is AI a bubble? It depends on the market you think it’s eating

Whether we’re truly in a bubble – and how dangerous it is – depends heavily on what you think AI’s real total addressable market (TAM) is.

Scenario 1: AI is mostly eating IT budgets

Global IT spending is around $1.8 trillion, with roughly $1.4 trillion tied to software. Even in an aggressive scenario, AI might only replace or absorb about $300 billion of that spend in the medium term. You’ll never get IT budgets to zero; you’ll always need people and infrastructure.

Now compare that to the current AI capex trajectory: roughly $500 billion per year chasing a $300 billion slice of IT budgets. That’s a problem.

Under this scenario:

• Demand ultimately falls short of the hype.
• AI revenue growth fails to keep up with the rising depreciation schedule.
• Financing structures get tested as cash flows disappoint.
• The bubble is in AI infrastructure and earnings, and it eventually bursts.

Using the earlier depreciation example, if we shorten GPU lifetimes to four years instead of six, annual depreciation on a $500 billion capex run rate jumps to around $75 billion – already more than current AI revenues. By 2027, with another year of similar capex, total annual depreciation could easily exceed $100–150 billion. AI revenues would need to nearly double just to cover that non-cash expense, never mind generate attractive returns.

Scenario 2: AI is eating the global cognitive labor market

There’s a very different way to look at AI’s TAM. Instead of thinking about IT budgets, think about labor budgets.

Global GDP is roughly $120+ trillion. Historically, about 50–60% of that goes to labor, which means global labor income is on the order of $60+ trillion. A rough estimate is that around a third of that – say $20 trillion – is cognitive labor: knowledge work, analysis, writing, coding, design, and so on.

In this framing, AI isn’t just selling tools to workers; it’s selling the work itself. If AI can replace even 10% of that cognitive labor, that’s a $2 trillion market. If it can reach 15–20%, the prize is $3–4 trillion.

Suddenly, $500 billion per year in AI capex doesn’t look so crazy. In this scenario:

• The bubble isn’t in AI infrastructure – it’s in the wage bill for cognitive workers.
• As AI gets cheaper and more capable, economic pressure to substitute away from human labor intensifies.
• The “crash” lands on labor: wages, employment, and demand, not on GPUs and data centers.

The key condition is that the cost per successful AI outcome must fall decisively below the cost of a human doing the same task – not just a little cheaper, but 20–30% cheaper or more. Only then does large-scale substitution become irresistible.

So far, inference costs for a given capability level have been falling fast. GPT‑4-level performance reportedly dropped from about $20 per million tokens in late 2022 to around $0.40 – roughly a 50x reduction. If that trend continues, AI work could become dramatically cheaper than human work for many tasks.

Under this scenario, we are not in an AI infrastructure bubble. We’re in the early innings of a long, painful repricing of human cognitive labor. That has huge implications for the future of work and inequality, a topic we explore more deeply in our guide to AI and the future of work.

The token paradox: cheaper tokens, bigger bills

There’s a complication: falling token prices don’t automatically mean lower AI bills.

As we move from simple chatbots to AI agents and multi-step workflows, the number of tokens consumed per task explodes. Gartner data suggests that agentic workflows can require 5–30x more tokens per task than basic chat interactions. A single user request can trigger 10–20 separate LLM calls.

This creates an “inference cost paradox”:

• Per-token prices fall dramatically.
• But usage (tokens per task) grows even faster.
• Total token spend keeps rising, especially if vendors don’t fully pass through cost savings.

Some enterprises are already feeling this. There are reports of companies burning through their entire annual AI token budgets in just a few months. Others are discovering that, at current prices and usage patterns, AI is more expensive than the human workflows it was meant to replace.

That’s why the right metric to watch is not cost per token, but cost per resolved outcome. The key question is: does the total AI cost to successfully complete a task fall below the fully loaded human cost (salary, benefits, overhead) before the depreciation schedule forces painful write-downs?

What could trigger the correction?

Regardless of whether you believe in Scenario 1 (IT budgets) or Scenario 2 (labor budgets), some kind of correction is likely. The difference is where the pain lands.

If AI is mostly eating IT budgets

In this case, the correction looks like a classic tech capex bust:

• AI revenues disappoint relative to capex and depreciation.
• SPVs and other financing vehicles start showing credit stress.
• Capex budgets flatten or get cut as ROI questions dominate.
• Chipmakers, memory vendors, and data center builders see orders collapse.

Historically, big cloud buildouts have topped out at around a 2.4:1 capex-to-revenue ratio. Today’s AI buildout is closer to 7:1. If that ratio doesn’t start to shrink meaningfully, it’s a warning sign.

If AI is eating the labor market

In the labor scenario, the correction still happens – but it shows up in employment and wages rather than in capex first.

Early signs to watch include:

• Hiring freezes in AI-exposed cognitive roles (rather than mass layoffs at first).
• Slower job finding rates for younger workers in fields like programming, design, and content.
• Pressure on human-services firms whose business model is selling hours (staffing, BPO, some consulting and IT outsourcing).

The Dallas Fed has already noted that employment declines for under-25s in AI-exposed fields are driven more by hiring freezes than layoffs. That’s exactly what the start of labor budget deflation looks like: firms quietly stop hiring rather than firing.

At the same time, we may see a “productivity paradox.” Individual workers who use AI tools become much more productive, but at the organizational level, measured productivity is flat because companies have to hire additional people to deploy and manage the AI systems. Until AI can truly replace roles end-to-end, its net impact on jobs may be roughly neutral – a view supported by some forecasts through 2026.

For a deeper dive into how this might reshape employment, see our analysis of whether we’re already in the foothills of an AI singularity.

What to watch next

We’re still early enough in the AI buildout that the party can continue for a while. But there are several key indicators that will tell us which scenario is playing out – and how close we are to a turning point.

1. Cost per successful outcome vs. human cost

Track whether the all-in cost of AI per resolved ticket, document, code change, or customer interaction falls decisively below the fully loaded human cost. When that happens at scale, substitution pressure on labor will intensify.

2. The capex-to-revenue gap

Right now, AI capex is roughly seven times AI revenue. Watch whether:

• AI revenues start growing faster than capex, shrinking that ratio.
• Or capex keeps accelerating while revenue growth slows, widening the gap.

A move toward something closer to 2–3:1 would look more sustainable. Staying at 7:1 or higher is a classic bubble signal.

3. Depreciation policy changes

If major cloud providers or AI infrastructure players shorten their assumed useful life for GPUs and related equipment (for example, from six years to four), that’s a red flag. It means higher annual depreciation expense and could be an early sign that management expects hardware to become obsolete faster than previously thought.

4. Free cash flow and SPV stress

Keep an eye on:

• Free cash flow trends at hyperscalers and AI infrastructure companies.
• Credit spreads and ratings on AI-related SPVs and securitizations.

If SPVs start to show stress or refinancing becomes difficult, the entire financing machine that’s supporting the buildout could slow down abruptly.

5. Labor market signals in AI-exposed roles

Finally, watch hiring and wages in AI-exposed cognitive roles:

• Are job postings shrinking?
• Are entry-level opportunities drying up?
• Are wages stagnating or falling relative to other fields?

If AI really is going after the $20 trillion cognitive labor market, these signals will eventually turn sharply – even if headline unemployment stays low for a while.

So, are we in a bubble?

There are strong signs that we’re in at least a partial bubble:

• Stock prices in AI infrastructure names have gone parabolic.
• Capex is running far ahead of current AI revenues.
• Complex, off-balance-sheet financing structures are proliferating.
• Earnings are inflated by front-loaded revenue and back-loaded depreciation.

At the same time, AI is a real, general-purpose technology with the potential to reshape a huge share of the global economy. That’s exactly the kind of technology that has historically justified massive, painful capex cycles: railroads, electrification, telecom, the internet, and now AI.

In the end, the answer may be: yes, we are in a bubble – but the more important question is where the bubble is. If it’s in AI infrastructure, we’ll see a classic tech bust and a long digestion phase, followed by a new wave of application-layer winners that rent cheap, overbuilt capacity. If it’s in wages, the adjustment will be slower but far more disruptive for workers and governments.

Either way, this won’t resolve in a few quarters. The buildout can easily run through the next couple of years. For now, the party is still on – but it’s worth keeping one eye on the exits.