AI Bubble Tracker — Two Minutes to Midnight

The AI investment bubble refers to the widening gap between the hundreds of billions being spent on AI infrastructure — data centers, GPUs, model training — and the revenue AI products actually generate. This page tracks the financial signals, market events, and economic frameworks for understanding whether and when the AI bubble will correct, updated with each new episode of the ADI Pod.

Every episode of the ADI Pod ends with “Two Minutes to Midnight” — our Doomsday Clock-style segment tracking how close the AI investment bubble is to bursting. We borrowed the metaphor from the Bulletin of the Atomic Scientists because the dynamics are similar: the people building the thing are the worst judges of when it becomes dangerous, and the clock is a useful forcing function for honest assessment.

This page is the running record. Not predictions — data points, financial signals, and the frameworks for interpreting them.

The Clock

Current position: 1 minute 45 seconds to midnight (as of Episode 19, March 2026)

The clock has moved back from its closest reading of 25 seconds to midnight (Episode 3, November 2025). The retreat reflects a more nuanced read: AI revenue models are still lagging CAPEX, but the infrastructure being built has real long-term value and the Jevons paradox argument has gained weight. The bubble hasn’t popped — but the urgency has shifted from “imminent correction” to “slow grind toward reckoning.”

A note on methodology: the clock position is vibes, not math. It’s three practitioners reading the same financial reports and market signals as everyone else, filtered through the lens of “does this company’s business model make sense?” The clock is a conversation starter, not a financial instrument.

Disclaimer: The “Two Minutes to Midnight” clock is for entertainment purposes only. It is not financial advice, investment guidance, or a prediction of market outcomes. We are software developers who read earnings reports for fun, not analysts, economists, or fiduciaries. Do not make financial decisions based on a podcast segment named after an Iron Maiden song.

AI Investment vs. Revenue: The CAPEX Gap Explained

The AI bubble isn’t about whether AI is useful. AI is demonstrably useful. The bubble is about whether the investment in AI infrastructure is proportional to the revenue AI generates.

The numbers:

Gary Marcus framed it cleanly: “A trillion dollars is a terrible thing to waste.” Whether it’s wasted depends on whether revenue catches up to investment. So far, the evidence is mixed at best.

Timeline of Market Signals

A chronological record of the events we’ve tracked across 20 episodes, with links to every source we discussed.

Phase 1: Framing the Question (Nov 2025)

Episode 1 (Nov 28): The “Two Minutes to Midnight” segment debuts. OpenAI CFO Sarah Friar clarifies the company is “not seeking government backstop” — a statement that raises the question of why the clarification was necessary. Meanwhile, Google announces Project Suncatcher — a plan to put AI data centers in space, which tells you something about how far the infrastructure buildout has gone.

Episode 2 (Nov 28): Ben Thompson’s “The Benefits of Bubbles” provides the bull case: bubbles drive real innovation even if investors lose money. Project Syndicate explores “the AI Bubble’s Economic Fundamentals”. Counter: “Is Perplexity the first AI unicorn to fail?”

Episode 3 (Nov 28): Clock set at 25 seconds to midnight. “Nvidia didn’t save the market. What’s next for the AI trade?” — even NVIDIA earnings couldn’t sustain AI stock momentum. Mira Murati’s Thinking Machines seeks $50B valuation. “Boom, bubble, bust, boom. Why should AI be different?”

Phase 2: Adoption Skepticism (Dec 2025)

Episode 4 (Dec 5): OpenAI declares internal “Code Red” as Gemini gains 200M users in 3 months. AI adoption rates start flattening per Apollo Academy data, particularly among larger firms. Gary Marcus: “A trillion dollars is a terrible thing to waste.”

Episode 5 (Dec 12): Someone builds pop-the-bubble.xyz, a countdown site for the AI bubble. The Anthropic CEO weighs in on bubble talk. The question surfaces: “Are we really repeating the telecoms crash with AI datacenters?” And Microsoft’s attempts to sell AI agents are turning into a disaster.

Episode 6 (Dec 19): CoreWeave CEO defends circular AI deals as “working together” — AI companies buying from AI companies, inflating each other’s revenue numbers. OpenAI boasts an enterprise win days after the Code Red.

Phase 3: The Debt Reckoning (Jan-Feb 2026)

Episode 8 (Jan 9): “AI faces closing time at the cash buffet” (The Register). The NYT reports AI debt investors growing wary. NVIDIA acquires Groq. Meta acquires Manus. The takeaway we keep coming back to: “Not all bubbles are negative — technological bubbles can bring real efficiencies at the cost of investor capital.”

Episode 9 (Jan 16): CNBC surveys 40 tech leaders and analysts on the bubble question. The consensus is muddled, which is itself a data point. Jensen Huang says “We have to create prosperity for all, not just PhDs”.

Episode 10 (Jan 23): The PwC CEO survey drops: majority of CEOs report zero payoff from AI spending. OpenAI introduces ads in ChatGPT — a move that would have been unthinkable a year earlier when Sam Altman called ads “desperate” and “a last resort.” The Guardian: “AI companies will fail. We can salvage something from the wreckage.” Also: two AI researchers are now funded by Solana — when the crypto money arrives, you know things are getting weird.

Episode 11 (Jan 30): Microsoft CEO warns they need to “do something useful” with AI or lose “social permission” to burn electricity on it. TechCrunch asks: “A new test for AI labs: Are you even trying to make money?” Counter-narratives emerge: “What if AI is both really good and not that disruptive?” and “Are AI agents ready for the workplace?”

Episode 12 (Feb 6): Exponential View publishes “Inside OpenAI’s unit economics” — scrutinizing the gap between compute costs and subscription revenue. NVIDIA shares drop on reports that its OpenAI investment has stalled.

Phase 4: Infrastructure Strain and Job Market Signals (Feb-Mar 2026)

Episode 13 (Feb 13): Three bombshells in one week: Oracle plans $45-50B in debt/equity for AI data centers. Google issues a 100-year bond. Anthropic closes in on a $20B funding round. Om Malik names it: “The New Announcement Economy” — companies making massive spending announcements without corresponding revenue.

Episode 14 (Feb 20): “The AI Data Center Financial Crisis” arrives as a concept. The SaaSpocalypse Paradox surfaces: if AI kills SaaS, it destroys the very revenue streams funding AI development. Anthropic’s margins and COGS come under scrutiny. Hyperscaler CAPEX is explicitly compared to GDP levels.

Episode 15 (Feb 27): The circular economy concern intensifies — AI companies as each other’s primary customers. “The Number Is Going Up” covers the AI CAPEX spiral. Meanwhile, an AI coding bot took down Amazon Web Services — a reminder that AI-generated code has real operational risk.

Episode 16 (Mar 6): The most consequential week. Block lays off 45% of its workforce (~4,000 employees), citing AI productivity gains. Shares soar 24%. The market signal is unmistakable: Wall Street rewards AI-attributed headcount reduction. The Citadel Securities report on AI adoption S-curves drops. A Substack post about white-collar displacement literally rattles the S&P 500.

Episode 17 (Mar 13): Oracle cuts thousands of jobs as data center costs rise — an infrastructure provider bleeding money on the AI bet. We called this “the canary in the coal mine for the AI bubble, not the frontier labs themselves.” When the companies building the infrastructure can’t sustain the costs, the bubble’s physics become apparent regardless of how the model providers are performing. Amazon requires senior engineers to sign off on AI-assisted changes after AI-related outages.

Phase 5: Trillion-Dollar Projections Meet Product Failures (Mar-Apr 2026)

Episode 18 (Mar 20): Meta delays rollout of its “Avocado” AI model after performance concerns and may end up licensing Gemini — despite massive internal AI investment. The parallel to Apple’s trajectory (building internally, then licensing externally) is noted.

Episode 19 (Mar 27): The contradictions sharpen to a point. Jensen Huang projects NVIDIA’s Blackwell and Vera Rubin sales into the $1 trillion range — the largest demand projection in semiconductor history. Meanwhile, OpenAI shuts down Sora just 4 months after a 3-year Disney partnership — signaling either financial pressure or strategic failure. The overriding theme: “Accelerated FOMO in the Age of AI” — FOMO, not fundamentals, is driving market behavior.

Episode 20 (Apr 10): The Claude Code source leak dominates the episode. Copilot’s terms classify it as “for entertainment purposes only.” The gap between how AI tools are marketed and how they’re legally positioned widens. No bubble segment this episode.

Economic Frameworks for Understanding the AI Bubble

The Benefits of Bubbles (Bull)

Ben Thompson’s argument (Episode 2): every major technology transition involved a bubble. The dot-com bubble built fiber optic infrastructure that powered the next two decades of internet growth. The railroad bubble built transportation infrastructure that transformed the American economy. The AI bubble is building compute infrastructure that will have value regardless of whether current AI companies survive.

The counterargument: the dot-com bubble destroyed $5 trillion in market value and took a decade to recover. The infrastructure survived, but the investors and workers didn’t. “Real efficiencies at the cost of investor capital” is a polite way of saying “some people get rich, more people get burned.”

The Jevons Paradox (Bull)

Making a resource cheaper to use increases total consumption. If AI makes software cheaper to produce, we should expect more software, not less. More software means more demand for developers, infrastructure, and tooling. Sinofsky’s historical parallels are strong: PCs didn’t kill mainframes, spreadsheets didn’t kill accountants, ATMs didn’t kill bank tellers.

The counterargument: the Jevons paradox assumes the cheaper resource is a complement to human labor, not a substitute. Coal consumption increased because humans still ran the factories. If AI replaces the human in the loop, the paradox breaks down.

Workflow Automation Convexity (Bear)

Workflow automation convexity suggests that displacement will be sudden, not gradual. Long period of minimal impact, then a cliff. The current flat period — where AI automates individual tasks but not complete workflows — is the calm before the storm. When full-workflow automation arrives, the displacement happens faster than markets or workers can adapt.

The SaaSpocalypse Paradox (Bear)

If AI makes it cheap enough to build custom software, companies stop buying SaaS. If companies stop buying SaaS, SaaS revenue collapses. If SaaS revenue collapses, the SaaS companies that are AI’s biggest customers stop buying AI services. The AI industry’s revenue model partially depends on the survival of the industry it’s disrupting.

The Announcement Economy (Structural)

Om Malik’s framing (Episode 13): companies announce massive AI investments ($50B Oracle, $20B Anthropic, 100-year Google bond) and the market reacts to the announcement itself. The announcement becomes the product. Revenue is secondary to the narrative of investment. This dynamic sustains the bubble because each announcement validates the previous one — “if they’re spending $50B, there must be value here.”

The risk: announcement economies can sustain themselves as long as capital is cheap and optimism is high. When either condition changes, the cycle reverses fast.

Company Watch

OpenAI

The trajectory: from “ads are for the desperate” to running ads in ChatGPT. From a 3-year Disney partnership for Sora to shutting Sora down four months later. From a nonprofit mission to a planned IPO. Each step is individually rational. The pattern is a company under financial pressure making concessions it said it wouldn’t make.

The unit economics question (Episode 12) remains the central issue: does the subscription revenue from ChatGPT and the API cover the compute costs? The answer, as best we can determine, is no — not at the scale OpenAI operates. Ads and an IPO are monetization strategies for a company that hasn’t found product-market fit for its revenue model, even as it dominates usage metrics.

Anthropic

The $20B funding round reflects both AI investor appetite and Anthropic’s compute needs. The margins and COGS discussion from Episode 14 suggests that even frontier labs with strong technology face the fundamental CAPEX-to-revenue problem.

The Pentagon drama adds complexity: Anthropic’s ethical red lines (no mass surveillance, no autonomous weapons) are admirable but create business risk. When the Department of Defense threatens supply-chain risk designation for refusing certain contracts, the tension between values and survival becomes existential.

NVIDIA

Jensen Huang’s $1 trillion demand projection is either the most bullish signal in semiconductor history or the most bullish signal since Cisco’s CEO declared in 2000 that the internet would change everything right before Cisco lost 80% of its value. NVIDIA is simultaneously the best-positioned company in AI (everyone needs GPUs) and the most exposed to a correction (if GPU demand falls, so does the entire valuation thesis).

The Groq acquisition (Episode 8) hedges against one risk: alternative inference hardware. By acquiring Groq’s LPU technology, NVIDIA controls a potential competitor. Whether LPUs represent a genuine threat or a hedge depends on how inference costs evolve.

Oracle

Oracle’s trajectory is the clearest canary signal. Planning $45-50B in debt/equity for AI data centers (Episode 13), then cutting thousands of jobs as those data center costs mount (Episode 17). Oracle isn’t a frontier AI lab — it’s an infrastructure provider making a massive bet on AI demand. When the infrastructure providers are bleeding, the bubble’s physics become apparent regardless of what the model providers report.

Block

The 45% layoff with a 24% stock surge (Episode 16) is the market signal that matters most. Not because Block’s AI productivity claims are necessarily valid — but because the market reaction creates an incentive structure. Every CEO in America saw Block’s stock jump on AI-attributed layoffs. Whether the layoffs were genuinely AI-driven or conveniently AI-labeled, the incentive to follow the pattern is powerful.

Bull vs. Bear Scorecard

Key Indicators for the AI Bubble in 2026

Revenue-to-investment ratios. The gap between AI CAPEX and AI revenue is the single most important metric. When this ratio starts improving — not announced, but reported in earnings — the bull case strengthens. Until then, the bubble thesis stands.

Enterprise adoption beyond pilots. The PwC survey (zero payoff) measures current state. The Jevons paradox predicts future state. The transition point — when AI moves from pilot projects to production workflows — will be visible in enterprise software spending data.

The next Block. Whether other companies follow Block’s pattern of AI-attributed layoffs with stock surges will determine whether the pattern is an anomaly or a structural shift.

Infrastructure provider health. Oracle is the canary. Watch AWS, Azure, and GCP capital expenditure versus cloud revenue growth. If the infrastructure providers start showing stress, the foundation of the bubble is cracking.

OpenAI’s IPO pricing. The IPO valuation will be the market’s verdict on whether AI’s financial model works. An oversubscribed IPO at high multiples extends the bubble timeline. A disappointing IPO accelerates the reckoning.

Frequently Asked Questions

Is the AI bubble going to burst?

Something will correct. Whether it’s a sudden pop (like the dot-com crash) or a gradual deflation (like the crypto winter) depends on how quickly revenue models mature. Our current read: the CAPEX-to-revenue gap is widening faster than revenue models are improving, which favors correction over soft landing. But bubbles can sustain themselves longer than skeptics expect, especially when capital remains cheap.

Should I be worried about my job?

See our dedicated career impact guide. The short version: the Jevons paradox suggests more software demand. Workflow automation convexity suggests sudden displacement of specific roles. The safest position is one that combines domain expertise, architectural judgment, and the ability to direct AI agents — the skills that are furthest from full automation.

Is all the AI investment wasted?

No. Bubbles build infrastructure. The compute capacity being built will have value regardless of whether current AI companies survive. The question is who captures that value — the investors funding the buildout, or the companies that buy discounted infrastructure after the correction. Historically, the second group does better.

Why do you track this as software developers?

Because the AI bubble’s trajectory directly affects our tools, our jobs, and our industry. If the CAPEX-to-revenue gap closes, the tools get better and cheaper. If it doesn’t, some of these tools disappear and the companies behind them consolidate or fold. Either way, understanding the financial dynamics helps us make better decisions about what to invest our time learning and which platforms to build on.

How is the clock position determined?

Three people (Shimin, Dan, Rahul) looking at the same data and arguing about what it means. The clock is a conversation aid, not a financial model. It reflects our subjective assessment of the gap between AI investment and AI revenue, weighted by how many of the correction signals we’ve tracked are flashing. It’s vibes, but it’s informed vibes.

This tracker is updated with each new episode. Data points are sourced from publicly available financial reports, earnings calls, and industry analysis. Last updated April 2026.