Are We In An AI Bubble – Or Not

Bynaylorcpa_qr2uwn

I’ve heard this story before. The last time a “this changes everything” technology rolled through finance and operations, I was the one in the room translating the slide deck into the production environment. The internet, RPA, the cloud, and now AI all share the same dynamic: the technology is real, the timeline is wrong, and the people who pay the price are usually the practitioners — not the people who signed the contract. So when operations leaders ask me “are we in an AI bubble?”, I think the more useful question is the one underneath it: what moves can I make this quarter that I won’t regret regardless of which scenario plays out?

There’s a reason that question is harder to answer this time than it was during RPA. Toby Stuart, writing in HBR in April 2026, put a name to something I’ve been trying to articulate: the AI fog. AI’s rapid advance has created extreme opacity about the short-term future — not just normal uncertainty, but a condition where the probability distributions themselves are unknowable. His framing: when you can see 30 years ahead, you build skyscrapers and railways. When you can see only months ahead, you pitch a tent and buy a bicycle. The fog is the meta-problem sitting underneath both the bubble question and the implementation question. And right now, the fog is dense.

Three forces, one room, zero visibility

I started and led a finance and accounting process-improvement group through the last big “everything changes now” cycle — RPA. I watched as vendors and consultants asked executives to sign seven-figure contracts on the strength of a thirty-minute demo. I watched the practitioners underneath them spend the next two years figuring out which 12% of those processes actually worked. The technology was real. The transformation was real, eventually, for some firms. But a lot of money and a lot of careers were affected in the gap between the slide deck and the production environment.

Today’s AI moment is bigger. The slope is steeper, the capex is larger, the timelines are tighter, and the macro backdrop is unstable in ways the RPA cycle wasn’t. But there’s one difference that makes this cycle harder to navigate than anything I’ve seen before: we can’t see far enough ahead to reliably price the bet. During the RPA cycle, you could at least model a three-year ROI and have a defensible argument with it. Today, the range of plausible outcomes for a given process, a given role, or a given company spans from “strategically critical” to “structurally displaced” — and the probability distribution between those endpoints is genuinely unknowable. The fog isn’t a metaphor. It’s the operating condition.

So the question I keep getting from operations leaders is some version of: are we in an AI bubble? And underneath that: what am I supposed to be doing about it?

We are at the intersection of three forces that don’t usually arrive together. First, a financial bubble dynamic in AI capex — an investment cycle the market is pricing as if every link in the chain (compute, frontier model profitability, enterprise ROI) will work out on a fast timeline. Second, an implementation/displacement question — what actually happens to the economy if AI succeeds at the pace the ownership class is pricing in. Third, a macroeconomic instability backdrop — debt service pressure, institutional fragility, and a 2026–2028 window that Ray Dalio is now publicly calling a “heart attack” risk for U.S. Treasury markets. Each of these on its own is manageable. Together, wrapped in the fog, they raise a different kind of question for the operations leader: how do I make moves I won’t regret regardless of which scenario plays out?

1. The bubble question

Let me be clear up front. I’m not asking whether AI will change finance and operations work. It will. The internet changed work too. The bubble question is not about the technology — it’s about the timing of cash flows between today’s investment and tomorrow’s profits. And the fog makes that question materially harder than prior cycles.

Here’s what the market is currently pricing in. Big Tech AI capex is on track for $650–$700 billion in 2026, with capex as a percentage of sales for the hyperscalers projected to hit 34% this year and 37% by 2028 — already above the dot-com peak of 32% in 2000. The financing has visible circular patterns: Nvidia’s $100B commitment to OpenAI funds data centers that OpenAI fills with Nvidia chips; Microsoft, Google, and Amazon invest in OpenAI and Anthropic, who turn around and rent compute back. This isn’t fraud — it’s how a closed ecosystem looks when it’s the only game in town for venture capital. But the critics’ argument deserves a careful read: in Ed Zitron’s estimates, ~70–80% of Microsoft’s AI revenue is OpenAI paying for Azure, and ~80%+ of Amazon’s AI revenue is Anthropic paying for AWS.

And the enterprise ROI on AI is, charitably, unproven. MIT’s NANDA initiative found in late 2025 that about 95% of enterprise GenAI pilots failed to deliver measurable P&L impact. Goldman Sachs’ “Tracking Trillions” report (May 2026) projects $7.6 trillion in AI buildout through 2031 — and finds that most businesses report AI costs more than it saves.

Here’s where the fog concept adds something the standard bubble framing misses. The standard bubble critique asks whether the cash flows will materialize — that’s a risk question, where you assign probabilities and make a bet. What we actually have is uncertainty — a condition where the probability distribution itself is unknown. The difference matters in a specific, structural way. Equity valuations are built on a terminal value assumption: that we believe the business is durable far into the future. At current public market multiples, terminal value accounts for 60–80% of a typical company’s total market capitalization. If the AI fog brings into question the long-term viability of a company’s core product or service — and for most software-based businesses, it does — then we must also question whether it will earn its terminal value. The valuation doesn’t collapse on a specific negative event; it collapses on the simple inability to price the durability bet. That is a structural fragility, not a calculable risk. Any business whose competitive advantage is mostly software-, process-, or content-based is now vulnerable to this dynamic. That is not a small category.

The dot-com analogy holds in a specific, narrow way. The internet was real, transformative, and accurately predicted to change everything. Most of the companies financing the buildout still went to zero. Pets.com was wrong about the company; right about the future. The bubble question for AI is not “is the technology real?” It is “are these specific cash-flow assumptions going to survive contact with a multi-year integration timeline, inside a fog that makes terminal value structurally unreliable?” If even one link in that chain — data center completion rates, frontier model gross margins, enterprise ROI, or hyperscaler capex appetite — breaks, the rest can unwind quickly.

The honest framing: I don’t know when. But I would not bet the firm on the soft-landing scenario.

2. The implementation question — if AI succeeds

This is the part that gets less airtime. Let’s suspend the bubble question and ask: what happens if AI does succeed at the pace the investors are pricing in?

Citrini Research’s “2028 Global Intelligence Crisis” thought exercise lays out the bear case: an intelligence displacement spiral where AI replaces enough white-collar workers that consumer demand craters, real estate corrects, and the engine of demand that the whole system runs on stops turning. Their key quote: “a Claude agent can do the work of a $180,000 product manager for $200/month.” If that’s even partially true, the math of professional services changes fast.

The fog adds a dimension to this that isn’t in the bear-case literature: the chilling effect on human capital investment. It’s not just that AI might replace workers — it’s that the uncertainty about which workers, when, doing what is already freezing the decision to invest in human capital. Who commits to a medical residency if they can’t articulate what a doctor even is in 2035? Who enrolls in an MBA if the firms that historically absorbed MBAs can’t confidently forecast their hiring needs three years out? The fog isn’t a future problem; it’s already distorting how individuals and organizations commit to long-duration investments in people. The data is beginning to reflect it: Anthropic’s Economic Index (March 2026) found no systematic rise in unemployment for highly AI-exposed workers through 2025, but with one telling caveat — hiring of younger workers has slowed in exposed occupations. That’s the fog’s chilling effect showing up in the labor data before the displacement does. The training pipeline is the early warning signal, not the unemployment rate.

That’s the bear case. The bull case — which I think gets too little credit in the doomer literature — comes from looking at what historically happens to demand when the price of a service collapses. Alex Emas’s “What Will Be Scarce” framework, expanded by NLW at the AI Daily Brief, makes the case that the lump-of-labor fallacy has never held through a technology transition. Spreadsheets ate the bookkeeping clerk; they created the financial analyst and the FP&A function. Lower cost of delivery brought entirely new buyers into the market — small businesses that couldn’t afford a $5,000 design engagement or a $3,000 legal review become buyers at $500 and $300.

What I take from this: both cases deserve serious weight. The bear scenario assumes a pace of displacement that is historically fast — the difference between a 2-year timeline and a 10-year timeline is the difference between a depression and a managed transition. The bull case is not wishful thinking; it is the baseline outcome of every prior technology transition on record. But it is also not guaranteed, and “it worked out before” is not a strategy. The categories that survive in either case are the ones with human premium: accountability, trust, translation, relationship, behavior change. CPAs sign returns. Someone has to own it when the AI gets it wrong. That doesn’t disappear in any plausible scenario.

And finance and accounting are uniquely behind — CFO Connect’s 2026 State of AI in Finance report ranks F&A last among business functions in AI deployment maturity. That is simultaneously the biggest vulnerability and the biggest opening for individual practitioners.

3. The macro backdrop

The bubble dynamics and the displacement dynamics are not unfolding in a stable macro environment. That’s the part that doesn’t fit comfortably in a vendor pitch deck — and the fog makes macro navigation materially harder.

Ray Dalio’s framing in May 2026 was the bluntest he has used: an “economic heart attack” in U.S. Treasury markets, with debt service consuming a growing share of the federal budget and the 2026–2028 window being the highest-risk corridor he has identified. The CBO projects roughly $19–20 billion per week in federal interest payments in 2026. Dalio’s interconnected-risks frame names AI explicitly as one of five forces converging on the same window; debt is just the one with the hardest math.

I’ll leave the institutional commentary to the brief version: the policy-response capacity available today is visibly reduced from what it was in 2008 or 2020. You don’t have to agree with anyone’s politics to observe that institutional fragility is a real variable in this calculation.

The reason this matters for the AI question is straightforward. The resilience of the financial system is a precondition for the AI capex story to play out smoothly. If credit tightens, if Treasury markets dislocate, if a recession lands in the same window the data center debt comes due, the circular financing structure becomes much less forgiving. Pension funds are now exposed to AI data center debt through private credit vehicles (Apollo/Athene, Blackstone); that is not a market that can absorb a shock quietly. And in a fog, the feedback loops between “the market senses trouble” and “the investment thesis unravels” can move faster than a 10-year DCF model suggests.

So what is my message for the reader?

The standard advice in uncertain environments is “be cautious.” That’s not useful. The better frame — borrowed from how venture capital has always thought about unknowable futures, and now applicable to every operations practitioner navigating the fog — is to optimize for optionality. When you can’t see far enough ahead to build the skyscraper, you don’t freeze and you don’t guess. You make the smallest commitment that buys you information and the right, but not the obligation, to follow on with more. That is not timidity. It is disciplined positioning for a condition the fog makes inevitable.

Don’t try to time the bubble. Practice resilience. I’m not smart enough to call the top, and neither is anyone else. The question isn’t “should I bet on the AI trade?” — it’s “what moves can I make this quarter that I won’t regret in either scenario?” That is a much smaller, much more answerable question.

The RPA echo is the most useful lens you have. RPA was real. It also impacted a generation of process-improvement careers because the practitioners didn’t separate the technology (which worked) from the implementation environment (which mostly didn’t). The same gap is opening in AI right now. Vendor demos run on clean master data and pre-trained models; production runs on the data and processes you actually have. Plan for 2–4× the vendor’s stated implementation timeline.

Become the practitioner who delivers measurable ROI on the tools that exist today. Not the tools that are coming. Not the agents that will be reliable next year. The ones you have a license for, this week. That capability is durable across both bubble and non-bubble scenarios. It is also, per the Section.ai data, what only ~15% of knowledge workers can actually do.

Build “translation” skill aggressively. The single most underrated human-premium category is the ability to take a messy business problem and convert it into something an AI can actually do. The C-suite cannot do this. The vendor cannot do this. The IT team usually cannot do this. The operations practitioner who can is going to be more valuable in both the displacement scenario and the muddle-through scenario.

Stage-gate your AI spend. The fog has made the 10-year ROI case for most enterprise AI investments structurally unreliable. The better question isn’t “what is the return on this investment?” — it’s “what is the smallest commitment we can make now that buys us information and the right to follow on?” Treat AI spend as a series of option purchases, not a capital allocation decision. If you can’t define in advance what “success” looks like for a given AI initiative — in measurable terms — you don’t have an investment thesis, you have a hope. Token-based billing (already underway in 2026, with GitHub Copilot leading) is making this easier; treat AI like a utility and build measurement into the contract from day one.

Build your sensing system. The fog doesn’t lift automatically — you have to actively monitor it. At the practitioner level, this means carving out dedicated time (not just following the news) to track frontier AI capabilities and translate them into operational implications for your specific function: What new capabilities are coming? Which of your current processes are now ripe for automation? Which skills may have a shorter shelf life than you planned? The firms that navigate this best will be the ones that saw the transitions coming before the market forced them. This isn’t a full-time job at the individual practitioner level, but it is a habit that pays compounding returns.

Diversify vendor exposure. If any one of the frontier model providers, the major data center developers, or the leading hyperscalers takes a credit hit in the 2026–2028 window, your workflow risk concentration matters. Build skills in two ecosystems, not one. Keep your own context, prompts, and notes in a portable form — one reason I’m bullish on file-over-app knowledge systems for practitioners is exactly this portability.

Build relationships and accountability around the AI work. The seven human-premium categories — relationship, embodied presence, trust, accountability, translation, behavior change, provenance — are the part of your value that survives both a bubble pop and a successful AI displacement scenario. If you can credibly own the outcome of an AI-assisted analysis in front of a client or a CFO, you are doing the part of the job that doesn’t compress.

If you’re not sure where to start, here’s the short version.

This week:

  1. Pick one AI tool you’re already licensed for that you’re underusing. Block two hours. Run a real work task through it and write down — before you start — what “success” would look like.
  2. For one process you own, define what measurable ROI would look like if AI handled part of it. Don’t let the vendor define this for you.
  3. Start a plain-text prompt library. Any prompt that works — save it somewhere you own, outside the vendor’s UI.

This quarter:

  1. Build your translation skill deliberately. Take five messy, real problems from your work and practice converting each into a structured AI task. This is the capability gap the C-suite and the vendor can’t close for you.
  2. Get hands-on in a second AI ecosystem. If your firm is all-in on one provider, spend time in another. Portability is a hedge that costs nothing to build now.
  3. Audit your AI workflow concentration. Know which processes depend on which vendors, and have a rough exit path in mind if one changes pricing, terms, or goes dark.

I don’t think the answer to “are we in an AI bubble?” is yes or no. I think the more useful answer is this: there is enough probability of a sharp correction in capital markets that you should not stake your career on the soft-landing scenario, AND there is enough probability that AI delivers on a long-horizon transformation that you cannot afford to opt out of building the skills.

The fog is not a temporary condition. The opacity is structural — it will persist as long as the pace of capability change stays faster than our institutions can measure and price. Operating well in it means making optionality your default position. Not because you’re paralyzed, but because the visibility genuinely doesn’t support 10-year bets right now. Make commitments small enough to reverse, build skills that survive across scenarios, keep your sensing systems on.

That’s not a hedge. That’s the actual position. “How do we get involved in driving the train of AI while attempting not to get run over by it?” That’s the question for the next two years. The work to answer it is concrete, available today, and largely invisible from inside the demo.

If you’ve lived through one of these cycles already, you already know the pattern. If you haven’t, the next twenty-four months are going to be the most useful classroom you’ll ever get.

What’s the one AI workflow you’re putting into production this quarter — and how would you know if it actually delivered ROI? Reply or comment. I’m collecting these for a follow-up piece on what “real ROI” looks like at the practitioner level.

Sources

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.