AI Bubble 2026: Two Minutes to Midnight Across 26 Episodes

As of late May 2026, the AI bubble clock sits at 6:00 — six minutes to midnight — after a four-episode dovish swing from 1:15 that extended through episode 25 and then held flat through episode 26. The reversal was driven by Anthropic’s Mythos disclosure pulling the Federal Reserve into AI infrastructure planning, Paul Graham’s log-scale chart contextualizing AI capex against the US railroad bubble, and (most recently) the argument that open-weight models will plausibly hit Mythos-class cyber capability inside 6 months — making the AI sector geopolitically too-big-to-fail rather than a near-term burst risk. The load-bearing concern that pushed us toward midnight earlier in the year still hasn’t gone away: hyperscalers are spending $700 billion on AI infrastructure while direct AI revenue covers roughly 4% of that, and OpenAI just missed internal revenue targets per CNBC’s late-April reporting. But the comparison set has changed, and so has the math.

The Bulletin of the Atomic Scientists has maintained a Doomsday Clock since 1947, measuring how close humanity is to self-inflicted annihilation. We borrowed their metaphor for something less existential but more personally relevant: the AI market bubble. Every episode of ADI Pod since November 2025, we’ve closed the show with a segment called Two Minutes to Midnight: an assessment of how close the AI industry is to a correction, a pop, or whatever euphemism you prefer for “a lot of people losing a lot of money.”

Twenty episodes later, we have a dataset. Not a rigorous one (three engineers arguing over a metaphorical clock isn’t exactly the Bulletin’s methodology) but a longitudinal record of how the ground shifted beneath the AI economy in real time, as seen by practitioners who spend their weeks inside the code and their weekends reading the earnings reports. Here’s what the timeline looks like, and what the trend line says about where we’re heading.

The Framework: Reading the Clock

The original Doomsday Clock uses midnight as the moment of catastrophe. We adopted the same convention: midnight is when the AI bubble pops. The further from midnight, the more stable the market appears. The closer, the more signals we are seeing that valuations, infrastructure spending, and revenue projections are diverging from reality.

There are a few things the clock is not. It’s not a stock price prediction. It’s not financial advice; we are, as I’ve reminded listeners approximately forty times, filthy AI casuals, not analysts. It’s a vibes-based composite of the data points we encounter each week: funding rounds, layoffs, revenue disclosures, government contracts, product shutdowns, and the quality of cope in founder interviews. Think of it as a fear-and-greed index filtered through three software engineers who read too many newsletters.

The Timeline

Phase 1: Cautious Optimism (Nov 2025, Episodes 1-5)

The clock started close to midnight (0:20) on peak hype, then pulled back to 0:50 as early data tempered the initial alarm.

Episode 1 (November 7, 2025). We introduced the segment with the clock at roughly 0:20 (twenty seconds to midnight), meaning we thought a correction was quite close. The AI hype was deafening. An Oxford Internet Institute study had just found that only 16% of 445 AI benchmarks used rigorous scientific methods. The OpenAI CFO had implied the company was seeking government help meeting revenue targets. The vibes weren’t great.

Episode 2 (November 14). Dan brought in a Project Syndicate piece by William H. Janeway that introduced a framework I still think about: any financial bubble has both a focus (are the assets valuable?) and a locus (will the underlying technology deliver productivity gains?). The dot-com bubble had a technology that delivered but assets that were wildly mispriced. We also discussed a Ben Thompson argument that bubbles are beneficial, since cheap capital investment during manias creates lasting infrastructure. The fiber optic cables that bankrupted telecom companies in 2001 gave us AWS a decade later. Maybe the AI bubble builds something equally durable. We also flagged Perplexity as a potential first AI unicorn to fail.

Episode 3 (November 21). The clock stayed close to midnight. Thinking Machines was seeking a $20 billion valuation at nine months old. A detailed dot-com crash comparison article identified four structural weaknesses: too much spending, too much leverage, circular deals, and China. And then NVIDIA reported Q3 earnings, beat every top-line number, stock jumped 5% at open, then closed down 3.2%. I said on air that if there’s a moment marking the top of the market, that might be it.

Episode 5 (December 5). The first real clock movement: we moved from roughly 0:20 to 0:50. More optimistic. Martin Alderson’s analysis argued the AI data center build-out isn’t a clean analog to the telecom crash because the risk profiles are different. The Anthropic CEO went on the record about bubble talk, being cagey but cautioning about “players who might make a timing error.” Dan found pop-the-bubble.xyz, which used an LLM to predict February 24, 2026 as the burst date. We noted that AI adoption rates had started flattening for large firms; Microsoft’s agent revenue targets had been slashed roughly 50%. But the overall trajectory felt less dire than the initial hysteria. The clock moved back.

Phase 2: Tense Stability (Dec 2025 - Jan 2026, Episodes 6-12)

The clock held at 1:30 for nearly two months as strong enterprise adoption numbers balanced out increasingly creative financing schemes.

Episode 6 (December 12). Clock at 1:30. A significant move toward optimism. CoreWeave’s CEO was defending circular deals as “just working together,” which wasn’t confidence-inspiring. But OpenAI’s enterprise numbers were genuinely strong: 36% of US businesses using ChatGPT enterprise, and employees self-reporting 40-60 minutes saved per day. The enterprise revenue numbers gave us something real to weigh against the speculation. I noted the use of GPTs for institutional knowledge was a sticky moat that was harder to dismiss.

Episodes 7-9 (December 19 - January 9). The clock held at 1:30 through the holidays. Episode 7 replaced the bubble segment with a year-end retrospective. By episode 9 we were back, with no significant movement. The calm before the storm, as it turned out.

Episode 10 (January 16). OpenAI introduced ads in the ChatGPT free tier. Sam Altman had previously said ads were “for the desperate” and a “last resort.” The pivot was part of OpenAI’s goal to generate $20 billion in 2026 revenue to cover $1.4 trillion in infrastructure commitments. Twenty billion dollars in ad revenue is a staggering number, the kind that makes you wonder whether the people writing these business plans have ever actually sold an ad. That ratio ($1.4 trillion in obligations against a $20 billion revenue target) became one of the data points I kept returning to. Michael Burry of Big Short fame appeared in a conversation noting that “entire software PI is less than a trillion dollars” and the industry is “selling $400 billion of chips for less than $100 billion in end-user AI product revenue.”

Episodes 11-12 (January 23-30). The clock hovered. Brex’s AI pivot and subsequent Capital One acquisition at roughly 50% of peak valuation was a micro-case study: the fallen unicorn pattern. You could tell a hopeful story (AI transformation led to acquisition) or a dark one (company needed a buyer because the independent path was unsustainable). We discussed both.

Phase 3: Divergence (Feb 2026, Episodes 13-16)

The clock oscillated between 1:45 and 2:45, with peak optimism coinciding with mass layoffs, hundred-year bonds, and the emergence of the announcement economy.

This is where the data got interesting. The clock started oscillating, and the oscillation itself was the signal.

Episode 13 (February 6). Clock at 1:45. Three stories hit in one week. Anthropic was closing on a $20 billion round at a $350 billion valuation, just five months after raising $13 billion. Google was selling a rare 100-year bond to fund AI expansion, the first since Motorola in 1997 (right before everything went sideways). Oracle announced plans to raise $50 billion in debt to cover AI infrastructure obligations. Dan coined the term announcement economy for the pattern: non-binding memoranda of understanding announced as completed deals to drive hype cycles. The $100 billion Nvidia-OpenAI Stargate initiative, we noted, was never actually a legally binding agreement.

Episode 14 (February 13). The clock moved back to 2:15, more optimistic. This was counterintuitive given the prior week’s stories, but the reasoning was specific: Chinese Lunar New Year model releases were unimpressive, suggesting frontier labs had more competitive runway. But the data points themselves were darker than ever. A Where’s Your Ed At analysis revealed Anthropic’s actual margins were closer to 50%, not the 80-90% implied by headline numbers, once model training was properly attributed to cost of goods sold. Rahul broke down the SaaSapocalypse math: hyperscalers projected to spend $700 billion on CAPEX in 2026, but direct AI revenue covered only about 4% of that spend. The market, as Rahul put it, couldn’t make up its mind whether it was bullish or bearish. A bet that breaks either way.

Episode 15 (February 17). Clock stayed at 2:15. The numbers kept getting bigger and less connected to reality. Rahul cited analysis showing that anywhere from 20% to over 90% of 2025 GDP growth was attributable to AI-related capital expenditure. The S&P 500, dominated by tech giants, was dramatically outperforming the Russell 2000, a market concentration pattern that looked like a success-breeds-success feedback loop with no brake. Meanwhile, AWS went down for 13 hours due to an AI coding incident, which felt like a sign.

Episode 16 (February 24). Clock moved to 2:45, the most optimistic reading in the show’s history. Citadel Securities published a report arguing that AI’s recursive self-improvement doesn’t mean recursive adoption: organizational, regulatory, and physical constraints create S-curve adoption at best. Block laid off 45% of its workforce citing AI productivity gains and the market rewarded them for it (stock up 24%). A Substack post briefly rattled the S&P 500 with a doomsday scenario about white-collar jobs. The irony of maximum optimism coinciding with mass layoffs wasn’t lost on us.

Phase 4: The Turn (Mar 2026, Episodes 17-20)

The clock reversed sharply from 2:45 back to 1:15 as products started dying faster than narratives: Oracle retrenching, Sora shuttered, Stargate dissolving.

Episode 17 (March 3). Clock swung forward to 1:45. Oracle was cutting thousands of jobs as data center costs rose and OpenAI pulled out of their joint Stargate investment. AWS suffered another outage; this time AI-assisted Terraform nuked a production database, leading Amazon to require senior engineers to sign off on all AI-assisted changes. I floated the idea that Oracle, not the frontier labs, might be the canary in the coal mine. They’d taken on massive debt to build data centers using today’s architecture for tomorrow’s workloads. If Oracle stopped servicing that debt, it could bring the infrastructure house of cards down.

Episode 19 (March 17). Clock moved to 1:15, the closest to midnight since the earliest episodes. Jensen Huang projected $1 trillion in demand for Blackwell and Vera Rubin chips, a number so large it loops back around to comedy (Dr. Evil territory). And then OpenAI shuttered Sora, its AI video generation product, just six months after launch and four months after inking a three-year partnership with Disney. Not pivoting it. Not open-sourcing it. Shutting it down entirely, app and API. Rahul speculated they were cleaning the books ahead of an IPO. Dan was blunter: “Sounds like trouble.” If a flagship product from the most valuable AI company in the world can’t survive six months, the question of what’s sustainable starts to feel urgent.

Episode 20 (March 31). We took a two-week break. The clock held at 1:15.

Phase 5: The Snapback (Apr-May 2026, Episodes 21-26)

The clock reversed sharply from 1:15 to 6:00 over four episodes — the most optimistic stretch in the show’s history — as Anthropic’s Mythos disclosure reframed the systemic-risk math, Paul Graham’s railroad-capex chart contextualized current spend against a bigger historical bubble, Nathan Lubchenco’s open-weight cybersecurity argument pinned the sector as geopolitically too-big-to-fail, and a string of big-money dovish data points (Anthropic-Google $200B, Panthalassa $200M ocean DCs, Anthropic resolving its compute crunch via the SpaceX/XAI Colossus One deal) arrived without a matching bearish forcing function — then held flat through episode 26 as Cerebras’ IPO pop, Anthropic surpassing OpenAI on Ramp business data, and Andy Hall’s “Politics of Jobless Prosperity” arrived in roughly equal-and-opposite size with no clear forcing function in either direction.

Episode 21 (April 17, 2026). Clock moved to 2:45. Anthropic disclosed Mythos and Project Glasswing — a model so capable at vulnerability detection that Anthropic withheld release while the Federal Reserve and major infrastructure partners (Amazon, Apple, Cisco, CrowdStrike, Broadcom) patched first. The argument that moved the dial: if a single model release drives Federal Reserve coordination, the AI sector has tipped from “speculative” to “systemically important.” Even on a 25%-chance-of-step-change basis, the expected value of “this is a real industrial inflection” rose enough to dial the clock back.

Episode 22 (April 21). Clock moved to 3:30. Paul Graham’s log-scale tweet on US investment cycles became the load-bearing data point: AI capex sits at ~1% of GDP versus US railroad investment which peaked at ~10% of GDP, suggesting another ~9 points of GDP and roughly 40 years of headroom before the comparable railroad-bubble peak. Friction signals also got more concrete — Ars Technica’s satellite/drone analysis showed 40% of 2026 data center construction is delayed, and Epoch AI’s Stargate analysis reported 9 GW of planned capacity (roughly NYC’s peak demand) across nine sites running into 2028, with local opposition slowing several. Construction lag delays revenue but also delays capex — a self-cooling dynamic Phase 3’s spending-vs-revenue framing didn’t price in.

Episode 23 (April 28). Clock moved to 4:00 on a different line of argument again — guest Nathan Lubchenco’s claim that DeepSeek V4 (1.6T base / 49B active params with 1M context) is only 3-6 months behind frontier, which means open-weight models will plausibly hit Mythos-class cyber capability by late 2026. At that point, national-security stakes pin the sector in place independent of revenue economics — the bubble becomes “too-big-to-fail” rather than “burst-soon.” The bearish ledger had real weight this episode too: OpenAI missed internal revenue targets per CNBC’s April 28 reporting (Oracle dropped 5% in a day on the news), Toby Ord pegged frontier-agent costs around $350/hr for O3-class runs at ~50% task success (a human-professional rate at much higher reliability), and TechCrunch’s “two college kids raise $5.1M pre-seed for an AI social network” became the canonical “putting the letters AI in iMessage” example of late-cycle VC behavior. The dovish read continues to be “more runway than we thought,” not “no bubble.”

Episode 24 (May 8, 2026). Clock held at 4:00. Three signals to weigh and they pulled in opposite directions in roughly equal size. On the bearish side: Where’s Your Ed At broke down OpenAI’s projection that ChatGPT Plus drops from 44M subscribers to 9M — about an 80% collapse, the gap “made up” through cheaper tiers and an ad tier scaling 3M → 112M users. That is, more or less, OpenAI telling its investors that the flagship subscription business is going away. On the bullish side: DeepMind’s David Silver raised $1.1B for Ineffable Intelligence at a $5.1B valuation for a lab that has existed for a few months, and Scout AI raised $100M to train vision-language-action drone models for the Pentagon. The funding flood at the seed/Series-A end is the dovish signal I keep returning to: when capital is flowing into earlier and weirder bets, the cycle has runway. None of this is new in argument shape. It’s the same kind of evidence that moved the clock across episodes 21-23, restated. The snapback held without reversal.

Episode 25 (May 12, 2026). Clock moved to 6:00 — the new most-optimistic reading — on a week with no major bearish data point and several big-money dovish ones. Anthropic reportedly agreed to pay Google $200B for chips and cloud access, pushing the cumulative revenue backlog across Amazon, Google, Microsoft, and Oracle toward ~$2T — closer to 1990s-national-debt scale than 2000-telecom-debt scale. Panthalassa raised $200M to test floating AI data centers in the Pacific in 2026, a category-creation bet whose $200M seed-stage scale is itself the data point ($200M is now “side project money” in 2026 AI funding). On the demand side, Anthropic separately unlocked its compute crunch by getting access to SpaceX/XAI’s Colossus One supercomputer (220K NVIDIA GPUs) — Ars Technica reported Pro/Max token caps lifted and Claude Code peak-hour limits removed in the same week. And the Wall Street Journal reported Grok lost ~60% of its paid downloads (20M → 8.3M Jan-Apr), with paid penetration flat at 0.174% vs ChatGPT’s 6% and enterprise adoption at 7% vs Claude’s 48% — the first real “the market chose a loser” signal of this cycle. Tech investor Ben Pouladian’s framing in the WSJ piece — “OpenAI is Coke, Anthropic is Pepsi, Grok is RC Cola” — captures the shape of the differentiation. Two months ago this combination of signals would have looked like terminal-phase excess. In context, it reads as the snapback extending without a forcing function in either direction.

Episode 26 (May 19, 2026). Clock held at 6:00 — the first quiet hold after the four-episode swing. Three signals, all dovish-to-neutral and none load-bearing enough to move the dial. Cerebras raised $5.5B in IPO and the stock popped 108% on day one on the back of ~80× memory throughput vs comparable NVIDIA GPUs (via on-chip HBM that SemiAnalysis treats as the read-through on inference architecture). Anthropic surpassed OpenAI on Ramp business-card adoption at 34.4% — the first time a non-OpenAI lab has led that panel since it has been tracked, which is a “healthy competition” signal more than a bubble signal. And Andy Hall’s “Politics of Jobless Prosperity” (opening with FDR’s 1944 State of the Union) named a 2% unemployment jump as the political-stability line in the sand — finally identifying the macro forcing function the prior Phase 5 framings deliberately punted on. The shape of the hold is the read: four episodes net-dovish without a real bearish forcing function is approaching trend, but one clean bearish week is what would tell us whether the regime has actually shifted versus whether the cycle is just sleeping.

Clock Position Summary

What the Trend Line Says

If you plot the clock readings, you get something that looks less like a straight line and more like an EKG. The signal isn’t “the bubble is popping” or “everything is fine.” The signal is volatility itself.

The clock started near midnight (0:20), pulled back to moderate optimism (2:45 at the peak in episode 16), and then reversed sharply toward midnight again (1:15 by episode 19). The overall shape is a rise, a peak, and a decline. That’s the classic pattern of a market that rallied on hope and is now encountering friction from reality.

Three patterns stand out:

The infrastructure-to-revenue gap is the load-bearing number. Hyperscalers spending $700 billion on CAPEX while AI revenue covers 4% of it. NVIDIA projecting $1 trillion in chip demand against an entire software industry worth less than a trillion. OpenAI targeting $20 billion in revenue against $1.4 trillion in infrastructure commitments. Every time we looked at the raw ratios, they pointed the same direction. The numerators (spending, commitments, projections) are growing faster than the denominators (revenue, adoption, willingness to pay). That’s not a statement about whether AI works. It’s a statement about whether the current capital allocation makes arithmetic sense.

Products are dying faster than narratives. Perplexity flagged as a potential failure in episode 2. Sora shuttered in episode 19. Oracle retrenching. Brex acquired at half its peak valuation. Microsoft agent revenue targets slashed. Block laying off half its workforce. Each individual story has an explanation that isn’t “bubble popping” (strategic refocusing, efficiency gains, competitive pressure). But the cumulative pattern of flagship AI products and companies contracting while investment expands is the definition of a divergence that eventually corrects.

The announcement economy distorts the signal. Non-binding memoranda announced as deals. Projected demand announced as revenue. Fundraising rounds announced while previous rounds are still being spent. As Dan noted in episode 13, the Stargate initiative was never a legally binding agreement, and yet it moved markets and shaped narratives. When the gap between announcements and binding commitments widens, the market is running on press releases, and press releases aren’t balance sheets.

Where the Clock Is Now

As of episode 26 (late May 2026), the clock sits at 6:00 (six minutes to midnight) — the most-optimistic reading in the show’s history, after a four-episode swing from 1:15 and then a quiet hold through episode 26. Episode 26 was the first week after the snapback where nothing moved the dial in either direction — Cerebras’ IPO pop and Anthropic passing OpenAI on Ramp landed on the dovish ledger, Hall’s “Politics of Jobless Prosperity” named a real macro forcing function (2% unemployment) but didn’t deliver one. Four-plus consecutive episodes of net-dovish-to-flat data points are getting close to a trend, but we still want to see whether the snapback survives the first real bearish forcing function before treating it as a regime change. What changed across those three episodes: the Mythos disclosure pulled the Federal Reserve into AI infrastructure planning (a different category of capital than venture money), Paul Graham’s log-scale chart reframed the comparison set against US railroads, 40% of 2026 data center construction running behind schedule means the capex side of the worrying ratio is itself slowing, and the prospect of open-weight models hitting Mythos-class cyber capability inside 6 months pins the sector as too-big-to-fail. What hasn’t changed: the underlying revenue/CAPEX gap is still wide (OpenAI just missed internal revenue targets, Oracle dropped 5% in a day on the news), the announcement economy still distorts headline numbers, Sora is still shut down, frontier-agent costs are now ~$350/hr at ~50% task success per Toby Ord. The dovish reframing argues the bubble has more runway than we thought, not that it isn’t a bubble.

I want to be careful about what this does and doesn’t mean. The clock reflects the balance of signals we’re seeing, not a prediction about timing. Bubbles can run longer than anyone expects; the dot-com bubble had years of obvious overvaluation before the correction actually arrived. And the technology itself is real. AI coding tools genuinely save developers time. Enterprise AI usage is growing. The underlying capability isn’t vaporware.

But capability and market sustainability are different questions, and the market is answering the second one with increasingly creative financing rather than increasingly profitable products. When the most valuable AI company shuts down a flagship product after six months, when an infrastructure provider takes on $50 billion in debt to build data centers for a partnership that’s dissolving, when 90% of GDP growth traces back to a single sector’s capital expenditure, the clock ticks.

The telecom bubble built fiber that powers the internet today. The dot-com bubble built infrastructure that became AWS. Maybe the AI bubble builds something equally transformative. I think it probably does. But the people holding the inflated assets during the correction aren’t the people who benefit from the infrastructure afterward. They never are.

We’ll keep watching. The clock will keep moving. And every week, three engineers will sit down and argue about whether the data points from the past seven days made the hands move forward or back. If nothing else, we’re generating excellent training data for future AI historians.