AI Bubble Debate

The AI bubble debate is the running 2025–2026 argument over whether AI capital expenditure is economically sustainable — whether the spending now committed to compute, data centers, and power will be matched by revenue, or whether it will end in a financial reckoning. Competing camps anchor on different evidence. The most widely cited framing holds that roughly $500B in annual capex implies a need for about $2T in annual revenue by 2030, a gap of up to $800B that would have to be closed. Participants do not agree on whether the gap is closing, and the page presents the bull, bear, and "good bubble" positions as attributed arguments rather than resolving them.

Quantitative frame

The most-cited aggregate forecasts come from Bain and Morgan Stanley. Bain's framing pairs a ~$500B annual-capex requirement with a $2T annual-revenue requirement and an $800B shortfall, all dated to 2030; it also estimates 200 GW of incremental AI compute by 2030. Morgan Stanley projects $3T of cumulative AI-infrastructure spend through 2028. Apollo CEO Marc Rowan separately projects roughly $3T of AI-infrastructure capex through 2028, with private credit and specialist funds providing much of the financing (How the AI revolution has turbocharged M&A).

Power demand sits alongside the capex figures. NERC's long-term reliability assessment projects +224 GW of summer peak demand by 2035. US power consumption surged more in the past two years than in the prior fifteen combined; residential electricity prices have risen more than 40% since 2020 with continued increases; solar is meeting the majority of US load growth for the first time, though the US trails China and the EU on per-capita deployment (Source: apricitas.io).

A revenue-versus-depreciation comparison circulated by Adam Tooze on July 11, 2026, drawing on Bloomberg figures, found that global AI sales excluding China reached $25 billion in Q1 2026, exceeding the industry's estimated $21 billion in data-center and chip depreciation for the second consecutive quarter — while noting that depreciation still consumes over two-thirds of revenue (Source: adamtooze.substack.com).

The bull case

Bull-case proponents (including Sam Altman, Jensen Huang, and Dario Amodei) argue that compute produces capability, which produces value, an argument advanced in The Intelligence Age (Altman), Abundant Intelligence (Altman), and Machines of Loving Grace. They point to accelerating adoption documented in the Ramp AI Index, Anthropic Economic Index — March 2026: Learning Curves, and Stanford HAI AI Index Report 2026, and to historical parallels — railroads, electricity, and the internet all looked bubbly and all reshaped the economy.

Rogé Karma's Atlantic update "So, About That AI Bubble" (May 1, 2026) is presented as the clearest position-update from a previously bubble-sympathetic writer; Karma had authored a September 2025 Atlantic piece arguing the bubble case. His May 2026 reversal rests on Anthropic ARR growing from $14B to $30B in two months over Q1 2026 (Q1 growth he puts at 4× Google's during peak expansion years), Claude Code as a November 2025 inflection from chatbot-as-conversation-partner to agent-as-autonomous-worker, a METR study reversal in which the same researchers who showed AI made developers 20% slower in 2024 found 20% faster completion with current tools (an estimate Karma calls conservative because some heaviest AI users declined to participate without their tools), a mid-April 2026 Goldman Sachs survey of 40 software companies finding many "overrunning their initial budgets [for AI tools] by orders of magnitude," some spending up to 10% of total engineering labor costs, and demand-exceeds-supply mechanics (Anthropic limiting Claude Code peak hours, OpenAI scrapping Sora to free GPUs, and Nvidia's 2022-vintage fourth-best AI chip costing more today than three years ago). Karma's framing: "Six months ago, people arguing that AI was a bubble were pointing to real-world facts, whereas people arguing against the bubble hypothesis were making speculative promises about the future. Today, the roles have reversed." He summarizes the shift, citing Anthropic's Q1 2026 revenue trajectory (a ≥$5B ARR-equivalent run-rate) plus Claude Code adoption, as moving the bull case from "compute → speculative capability → speculative revenue" to "compute → measured capability → measured revenue," with the bear case now resting on speculation that future capex commitments will outrun measured-revenue growth (Source: theatlantic.com; see Karma source page). Ethan Mollick, quoted in the same context: "For years now, we've been in an era of chatbots that mostly say things. Now we've officially crossed into the era of agents that can actually do things." Karma cites an MIT 3,000-task white-collar benchmark study showing the AI completion rate rising from 50% (mid-2024) to 65% a year later, with authors projecting 80–95% by 2029.

SemiAnalysis (May 1, 2026) supplied what the source calls the strongest current bull-side data point: it estimated Anthropic ARR grew from $9B to over $44B year-to-date, with inference gross margins rising from 38% to over 70%, and argued AI labs are now capturing the bulk of stack value while Nvidia and TSMC have not yet repriced upstream (SemiAnalysis; Source: newsletter.semianalysis.com).

Kleiner Perkins chairman and Alphabet board member John Doerr, who backed Google and Amazon as early investments, argued in a WSJ interview (May 23, 2026) that generative AI is the largest technology "tsunami" of his career — larger than the personal computer, internet browser, iPhone, and cloud computing — that such waves arrive roughly every 13 years, and that the AI revolution is "if anything, underhyped," citing that 50% of Americans report using generative AI three years after ChatGPT's launch (Source: wsj.com).

The "good bubble" defense

A distinct sub-position holds that AI is a bubble and that this is acceptable (Even Silicon Valley Says That AI Is a Bubble — Lila Shroff (The Atlantic, March 2026)). Its canonical formalization is Huber & Hobart, Boom: Bubbles and the End of Stagnation (2024), which argues that "good bubbles" finance technologies — railroads, dot-com fiber, AI data centers — that disciplined capital would never underwrite. The pro-bubble defenders catalogued by Shroff are Hemant Taneja (General Catalyst), Jeff Bezos, Sam Altman, James Thomason, Marc Andreessen, Peter Thiel, Mary Daly (SF Fed), and Ben Thompson (Stratechery).

Two counters are advanced against the "good bubble" framing. On asymmetric incidence, Howard Marks (Oaktree) is quoted: "The investor doesn't say, 'Well, yes, I lost my money, but thank God it advantaged society.'" Bubble apologetics is described as easier for those with capital to lose voluntarily than for retirees with 401(k)s tied to AI stocks. Gita Gopinath (IMF) is cited estimating an AI crash could wipe roughly $35T in global wealth, and Carlota Perez calls the AI bubble "the eye of a much larger hurricane." On asset durability, the counter notes that unlike railroad tracks and fiber-optic cables (which last decades), computer chips obsolesce in years (Epoch AI's "GPU frontier lifespan" data), making the infrastructure-left-behind argument weaker for AI than for the historical analogs.

The bear case

Financial engineering and circular financing

The financial-engineering bear case is anchored by Ed Zitron ("Where's Your Ed At," Zitron), who argues Anthropic revenue claims are suspicious, OpenAI economics are broken, and agents are chatbots; and by Senator Elizabeth Warren's letter (Warren letter), which contrasts OpenAI's $1.4T in commitments with $20B ARR and warns of "systemic risk to the U.S. economy." Other named anchors are MIT NANDA — The GenAI Divide (State of AI in Business 2025) (95% of enterprise GenAI pilots yield no measurable P&L impact), METR — Measuring Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (AI coding tools make experienced developers 19% slower), and Why I Think AI Take-Off Is Relatively Slow (Cowen) (economic integration slower than boosters predict).

A distinct strand of the bear case concerns not the financing structure but the reliability of the adoption figures the bull case rests on. Consultant Nikhil Suresh's July 2026 essay argues that AI enthusiasm has degraded organisational decision-making itself, and that the reporting layer is systematically compromised: he reports 0% success across the AI projects his firm observed over eighteen months, and describes a coordination problem in which a vendor executive who contradicts a customer executive's productivity claims risks the contract, so no participant in the chain can report honestly (AI Mania Is Eviscerating Global Decision-Making (Nikhil Suresh)). He further argues that a substantial share of apparent AI projects are non-AI projects relabelled to pass an internal test, giving as an example a database migration completed by hand but reported upward as an AI success. If accurate, this would mean announced adoption and productivity figures overstate both. The essay is a first-person consulting account with deliberately anonymised sources, not a measured study.

Warren's letter names the circular financing pattern explicitly: OpenAI's $1.4T in committed spending includes a $250B Microsoft cloud purchase (Microsoft holds roughly 27% of OpenAI), a $100B Nvidia investment plus 10 GW of chips (Nvidia draws revenue from the same funds it provided), and $300B in Oracle cloud (Oracle takes on debt to build for OpenAI). The same pattern appears at CoreWeave, whose debt-financed data centers have OpenAI as a near-exclusive tenant. Circular financing — lab revenue from investors, with customers funded by the same investors — is treated as one structural bubble indicator (Circular Financing in AI), and the acquihire pattern at Microsoft and Amazon is treated as a mechanism that keeps lab capital on balance sheets while avoiding antitrust scrutiny (AI Acquihires).

Gary Marcus argued on April 30, 2026 that current Big Tech AI infrastructure spending may constitute "the greatest capital misallocation in history," noting his technical and economic warnings had been picked up by MarketWatch and HBO's Last Week Tonight (Source: garymarcus.substack.com). On May 28, 2026, in "Bad news for three of the biggest IPOs in history," Marcus highlighted FT modeling of best-case AI ROI and argued the projections put pressure on the planned IPOs of Anthropic, OpenAI, and SpaceX. The "best-case" qualifier is what makes the finding load-bearing: even under bull-case assumptions, the four largest hyperscalers come up negative (Source: garymarcus.substack.com).

At The Information's "Financing the AI Revolution" conference, panelists declined to publicly name market risks when asked; Ken Brown (April 29, 2026) framed the silence itself as a bear-case datapoint, comparing it to the 2006-vintage subprime decorum that preceded the global financial crisis (Source: theinformation.com).

Marcus returned to the theme after the June 5, 2026 selloff. In "AI's Black Friday" (June 6) he argued the day's losses signalled the generative-AI industry "may be in worse shape than people realize," reading SpaceX's decision to lease out GPUs to Google (110,000) and Anthropic (220,000) rather than hoard them as evidence the frontier-scaling thesis is weakening — that SpaceX is "waving the towel" by arming competitors. He characterized the prospective U.S.-government equity stake in OpenAI as a "bailout" and "crony socialism" rather than investment, arguing partial government ownership could erode global trust in American AI firms much as U.S. distrust constrains Huawei (Source: garymarcus.substack.com).

In "The month generative AI lost its mojo" (June 26, 2026), Marcus argued that generative AI "lost its mojo" over the course of June, citing reports that OpenAI was leaning toward delaying its IPO to 2027 amid doubts about its targeted near-$1 trillion valuation, month-long declines in Nvidia, Oracle, and SoftBank shares, weakness in SpaceX's debt, and a rising share of Chinese open models on platforms such as OpenRouter. He disputed the position that AI is not a bubble, contending that rising revenue does not by itself establish sustainable profitability (Source: garymarcus.substack.com). The argument runs against the bull-side "burden of proof has shifted" framing, which holds that measured revenue growth has moved the bull case off speculation; Marcus's counter is that revenue and profitability are distinct. On the same day, Cory Doctorow ("Criticizing the everything machine," June 6) argued that AI boosters' sprawling claims function as a rhetorical "Gish Gallop" that defeats compact criticism, and that AI's unit economics are deteriorating rather than improving (Source: pluralistic.net).

Official-sector financial-stability warnings

Economists and central bankers at the European Central Bank's symposium in Sintra, Portugal warned on July 1, 2026 about financial risks from the AI boom, pointing to rising debt issuance by AI hyperscalers and leverage among investors; the IMF's Tobias Adrian called leverage "on both sides" worrisome for financial stability. The Bank for International Settlements had said on June 28, 2026 that the AI spending surge risks reversing and tipping some economies into recession (Source: livemint.com). The warnings extend the official-sector engagement documented in AI Macro-Prudential Policy from AI capability to AI-investment leverage as the risk vector.

Consumer-side critique

Elizabeth Lopatto (Lopatto, The Verge, April 20, 2026) supplies a consumer-side bear case distinct from the financial-engineering arguments (source page). She argues that NFTs, the metaverse, VR/AR headsets, and now LLMs share a structural failure to solve a consumer-market problem — built to enrich VCs and operators rather than customers. Her central claim: "there is only really one customer for LLMs that can justify the massive cash incineration process required to build them: the US government." LLM consumer adoption, she argues, is real "as long as they remain free," with durable consumer use cases reducing to cheating on schoolwork, Google-Search substitution, enterprise data organization, and faster coding. She names "Silicon Valley incuriosity" as the cultural pathology (following the Andreessen "philosophical zombie" / no-introspection thread) that lets founders mistake an inability to imagine other lives for product-market fit, with a companion thesis: "the people who tell us AI will dominate our future and take our jobs are the people who are hoping that will be true." Two axes test the thesis: pricing (if OpenAI's $0 Apple bundle unwinds — the dev-log of May 16, 2026 notes OpenAI v. Apple legal-action prep — and consumer ChatGPT moves to a meaningful price floor, the "free-or-nothing" claim is tested) and revenue mix (rising frontier-lab dependence on government-contract revenue would confirm Lopatto; rising enterprise SaaS or consumer ARR would falsify). Lopatto's position sits on the bear side but from the consumer angle, distinct from Zitron's revenue/economics framing.

Consumer-adoption data points the same direction. Apptopia (April 30, 2026) reported that consumer chatbot daily active users had fallen in four of the prior five months, and OpenAI disclosed it had missed its goal of 1 billion ChatGPT users by year-end 2025; Friar publicly characterized demand as "a vertical wall," but the DAU data is treated as the harder bear-case anchor (Source: bigtechnology.com).

A business-model variant of the critique came from Cory Doctorow, who asked in a July 13, 2026 essay why AI companies sell labor-replacing systems to incumbents rather than competing with their customers directly — if the systems could really do the work, the labs would capture more value doing the work themselves — framing lab economics as a "go meta" pattern of selling tools to incumbents instead (Source: pluralistic.net).

Enterprise cost rationalization

Through May 2026 a stream of enterprise-customer signals emerged in which executives at large AI customers publicly questioned whether AI cost ramp produces proportional gains — described as the first enterprise-customer-led bubble signal, distinct from prior investor (Marcus, Zitron), policymaker (Warren), and pure-data (NANDA, METR) signals.

The WSJ reported (May 28, 2026) that senior technologists at Uber, Meta, Microsoft, Salesforce, and DoorDash described new programs to ration AI use, steer workers to cheaper homegrown tools, and track productivity return, after some enterprises hit their annual AI budgets in three months and saw token bills double or triple. Micro1 CEO Ali Ansari called the shift a "healthy swing away from tokenmaxxing" (Source: wsj.com). Axios surfaced corroborating cost-rationalization data on May 27, 2026: Microsoft began winding down some Claude Code licenses on May 22, 2026 (per The Verge, a move Fortune attributed to AI tokens becoming a material cost line), and Uber's COO said AI costs were getting "harder to justify," weeks after the company's CTO blew through his 2026 IT budget on AI usage (Source: axios.com).

Specific customer reports clustered in the May 25–26 cycle. Uber COO Andrew Macdonald (May 26) said Uber was "not seeing proportional productivity gains from increasing AI costs" and had already exhausted its annual AI token budget in a few months (Source: garymarcus.substack.com). Microsoft (May 25) discontinued Claude Code licenses partly on cost grounds, described as the highest-status enterprise AI-coding-tool de-adoption to date (The Verge, cited via Marcus: theverge.com). Starbucks (May 21) shut down an AI inventory-counting program nine months after deployment because it frequently miscounted and mislabeled items (Source: Reuters, May 21, 2026). The Target India head (May 25) reweighed AI-tool costs amid a shift to usage-based pricing (Source: Reuters, May 25). Marcus (May 26) wrote: "If enough other companies report the same, the bubble pops."

The Pragmatic Engineer (April 30, 2026) reported AI-agent token spending exploding across at least 15 tech companies over the prior two-to-three months, breaking budget assumptions and forcing finance and engineering teams to retool cost controls; Microsoft's same-day shift to consumption-based pricing for Copilot is read as ratifying the "seat pricing breaking" thesis (Source: newsletter.pragmaticengineer.com). The broader read is "rationalization, not elimination" — enterprise cost discipline compressing the bull-case revenue forecasts that anchor the $2T-by-2030 capex math.

The throttling pattern widened through early July 2026. 404 Media reported on July 2, based on internal Slack chats, dashboards, and emails, that companies including Atlassian, Adobe, Amazon, and Citi had begun throttling employees' AI use as providers shifted to per-token enterprise pricing: one company's AI spending tripled to more than $15 million a month, some firms cut off access to certain models entirely, and Adobe moved to end unlimited Claude access (Source: 404media.co). Per internal audio leaked June 24, 2026, Accenture was fighting industry-wide "soaring token spend" driven largely by non-engineers using models for tasks like converting PDFs into slides (Source: 404media.co). Tesla's $200-per-week employee cap and Uber's $1,500-per-month cap are documented at Transportation — AI Deployment.

Box CEO Aaron Levie, himself an active AI angel investor, argued in a May 27, 2026 X thread that "CEOs are uniquely prone to AI psychosis because they're sufficiently distant from the last mile of work that still has to happen to generate most value with AI," writing that executives "play with AI," build a prototype or generate a contract, and then leap to believing agents can do the work; he urged CEOs to use AI "a ton" so they "come out the other side with an appreciation for both the upside and the real work." The framing is noted as coming from an AI bull and locating the displacement-rhetoric error in CEO cognition rather than in AI capability (Source: techcrunch.com).

Productivity and deployment

Nick Lichtenberg's Fortune feature (May 25, 2026) anchors the May 2026 debate to two diagnostic phrases appearing in the same mouths (Lichtenberg — 'Nobody knows anything' and 'this time is different': the phrases that define — and haunt — the AI economy): "This time is different" — Templeton's "four most costly words in the annals of investing," codified by Reinhart & Rogoff (2009) — and "Nobody knows anything," William Goldman's 1983 Hollywood aphorism, picked up by Ethan Mollick in a New York Public Library address: "I spend my time talking to AI labs, famous people, I talk to CEOs all the time, and nobody knows anything. We're all making this up as we go along. So anyone who's like, 'We have the playbook' — they're lying to you."

On productivity, BofA's own bull-case math is described as implying a 0.66% productivity ceiling, with only about 0.1% realized economy-wide. Goldman Sachs (March 2026) found "no meaningful relationship between AI and productivity at the economy-wide level," even with median 30% productivity gains in customer support and software where AI is concentrated. The arithmetic behind the 0.66% ceiling: 20% of workplace tasks AI can transform × 23% cost-effective to automate × 27% labor savings × 50% labor share, before friction. The figure sits below the Acemoglu ≤0.66%-over-10-years anchor, using the bulls' own math.

Mollick supplies two structural bear-side observations. On KPIs: "The very nature of saying we need a 10% improvement constrains the kind of use cases that you see" — breakthrough applications replace processes rather than improving them and cannot be specified into existence by KPI, which reframes slow enterprise deployment as rational conservatism rather than a function of model inadequacy (see Ethan Mollick and Mollick's April 1, 2026 Economist "By Invitation"). On consulting arms: "It's weird that the AI companies are all now building their own consulting arms to do AI deployment. If the models are so good that you think they're going to destroy all white-collar jobs, shouldn't they also be able to help you deploy systems?" The Karma and Lichtenberg/Mollick pieces are described as the two highest-status syntheses of the May 2026 moment: Karma argues the bull case has migrated to "measured capability → measured revenue," while Lichtenberg and Mollick respond that the deployment problem (the consulting-arms tell and the 0.1% realized-productivity figure) keeps the bear case alive at the organizational layer.

Supply and compute constraints

Epoch AI's Luke Emberson and Jaime Sevilla published a calibrated model (May 25, 2026) estimating that the world's Blackwell GPUs (1.9M GB200 plus 1.5M GB300, roughly 40% of aggregate FLOP/s supply) can currently serve 500M–20B output tokens per second from a Kimi K2.6-class model depending on context length, against demand proxies growing roughly 10×/year versus capacity growth of about 3.4×/year — pointing to a near-term compute crunch, particularly for long-context agentic workloads, consistent with Anthropic's recent peak-hour Claude quota reductions (Source: epoch.ai; see Epoch AI). Even if demand justifies the buildout, the model implies a widening operational-throughput gap. Schumpeter (The Economist, May 2, 2026) declared "the AI supply crunch is here," arguing hardware-makers are failing to invest enough to match demand. Casey Newton ("Did xAI just concede the AI race?", May 7, 2026) framed the SpaceX–Anthropic Colossus 1 lease as xAI effectively conceding the frontier race, citing JLL data that 97% of global data-center capacity is occupied and more than three-quarters of new capacity is already committed (Source: platformer.news). Local-inference demand also surfaced as a supply signal: Apple's Mac mini and Mac Studio were supply-constrained for months on local-AI workload demand (April 30, 2026), which cross-cuts the bull-case "demand exceeds supply" thesis (Source: techcrunch.com).

Capex and market indicators

The capex floor became visible at roughly $725B for 2026, up 77% from $410B in 2025; the Q1 2026 hyperscaler earnings cycle (April 29–30) closed with combined Big Tech capex hitting a record $130B in Q1 alone (Source: ft.com). AI framing has also spread to non-tech equity issuance: sandwich chain Jersey Mike's mentioned artificial intelligence 22 times in its IPO filing published July 2, 2026, which TechCrunch presented as a measure of how far AI risk-and-opportunity language has diffused into non-tech listings (Source: techcrunch.com).

The NVCA-Pitchbook Venture Monitor, released July 8, 2026, projected that the SpaceX, Anthropic, and OpenAI public offerings "will generate more value than all U.S. VC-backed exits since 2000"; SpaceX had already gone public at a $1.77 trillion valuation, with Anthropic and OpenAI valuations pushing into the trillions (Source: techcrunch.com). Separately, Bloomberg reported on July 8, 2026 that Nvidia's market value had slid by roughly $1 trillion, returning its valuation multiple to pre-AI-boom levels (Source: bloomberg.com) — concentration-of-value and repricing datapoints that cut in opposite directions within the debate. Bloomberg reported on July 13, 2026 that Apple's roughly $600 billion market-value rally had been fueled by traders rotating away from AI trades (Source: bloomberg.com). The rotation continued that week: on July 16, 2026 an AI-centered sell-off sent the Nasdaq-100 down 1.6% and SpaceX shares below their $135 IPO price for the first time, closing at $131.11 (Source: New Developments Log/2026-07-16-2205-ai-developments.md), and on July 17 Apple overtook Nvidia as the world's most valuable company amid a deepening selloff in chip stocks (Source: reuters.com). Later on July 17, US and Asian equity markets fell after Moonshot AI's Kimi K3 announcement intensified concern that cheap Chinese open models could undercut the AI spending boom: the Nasdaq dropped 1.4%, the S&P 500 fell 1%, the Dow closed down 407 points, Taiwan's benchmark index lost more than 6%, and Japanese markets closed down 4% (Source: cnn.com). Citrini Research argued on July 20, 2026 that the selloff was a leverage-fueled momentum unwind rather than evidence of open-source cost competition, noting that aggregate token spend was still growing (Source: citriniresearch.com).

Snapshot — hyperscaler 2026 capex

Fortune's Eva Roytburg (April 30, 2026) noted that roughly half of Google's and Amazon's "blowout AI profits" derived from their Anthropic equity stakes rather than core operations, meaning the headline AI-margin story is partly Anthropic-revenue mark-to-market rather than core hyperscaler operating margin — a fragility the bull case typically does not flag.

Snapshot — equity and company prints

The late-May equity rally — the ~17% rally led by Nvidia and Microsoft, and the software sector's best month since 2001 with Snowflake and Okta surging on AI-revenue catalysts and usage-based billing — is read as a bull-side / measured-revenue marker consistent with Karma's "burden of proof has shifted" framing, with usage-based billing as the consumption signal the bear side's "tokenmaxxing rationalization" thesis (WSJ, May 28) predicts would compress. It does not resolve the bubble question: equity rallies are themselves a classic bubble symptom, the move runs against the same-window enterprise-customer skepticism cluster (Uber, Starbucks, Microsoft's Claude Code wind-down), and public-market enthusiasm for AI exposure remained intact even as buy-side cost discipline mounted. The Dell print is read as bull-side counter-data reinforcing the supply-constraint-not-demand-constraint framing; against it, the same-day Anthropic Series H is the bull-side anchor while Marcus's negative-ROI modeling is the bear-side anchor. Taken together, the May 28 cluster has the bear case arguing inside the bull case's arithmetic, while supply-side data continues to show demand pulling hardware faster than capacity can come online.

M&A and consolidation

The first multi-author FT synthesis of the AI-driven M&A boom (May 22, 2026) added structural data read mostly as bull markers (How the AI revolution has turbocharged M&A): a record $250B into private infrastructure funds in 2025 (S&P Global), flowing toward data centers, power generation, and digital networks; a $420B NextEra–Dominion utility merger framed as data-center power consolidation, with Constellation–Calpine ($27B) and BlackRock–AES/Allete ($39B combined) confirming a utility-consolidation wave; KKR's CoolIT exit at a 15× return ($270M to $4.75B in three years) in liquid cooling; Sandisk up more than 4,000% in market cap since its February 2024 IPO on AI-inference-driven storage rerating; an acquihire-as-antitrust-workaround pattern (Meta–Scale $14.3B, Nvidia–Groq $20B, Google–Windsurf $2.4B, Microsoft–Inflection, Amazon–Adept, Nvidia–Enfabrica); and a Blackstone–Google $5B "neocloud" greenfield justified as cheaper than acquisition. Lazard CEO Peter Orszag's "discretionary state capitalism" framing describes governments more actively directing capital and industrial policy, casting the Trump-era antitrust posture as a structural enabler of AI consolidation rather than a transient deregulatory blip. The FT synthesis also carried bear markers: Eric Schmidt was booed at a University of Arizona commencement ("There is a fear in your generation that the future has already been written, that the machines are coming, that the jobs are evaporating"), an NBC poll found AI less popular than ICE, and the closing FT framing noted that "some investors question the maths and the projected profits of the companies involved, evoking the memory of the dotcom bubble."

Public-equity filings and frontier-lab finance

SpaceX's S-1 (filed the evening of May 20, 2026) is described as the largest public-equity test of the AI-capex thesis to date. Across its three segments (Space, Connectivity, AI), SpaceX reported $18.7B total revenue and a $2.6B operating loss; the AI segment alone carried $4.0B revenue against $8.9B operating losses over five quarters and $20.4B infrastructure spending since the start of 2025. Stated TAM was $28.5T, of which AI is $26.5T (93%), with a mid-June offering expected to value the company above $1.5T. The bear reading: AI-unit losses-to-revenue of 2.2× and AI-infrastructure capex about 5× AI revenue. The bull reading: Musk is the second issuer (after OpenAI in May 2026) to file public-market documents predicated on a multi-trillion AI TAM (see xAI; Source: fortune.com).

CoreWeave's Q1 (May 7, 2026) showed $2B revenue (roughly doubled YoY), a $144M operating loss, and a $100B revenue backlog (up from $67B at year-end 2025) including new multi-year deals with Anthropic and Meta, with the stock falling about 9%. The backlog/active-capacity gap (1 GW active, 8 GW targeted by 2030) is described as the operative structural question — what either fills the bubble (compute genuinely scarce) or drains it (overbuild plus customer concentration) (Source: theinformation.com). The OpenAI–Broadcom $18B custom-chip deal hit a financing snag (May 7), with Microsoft not committed to buying roughly 40% of the new chips, exposing the difficulty OpenAI faces in replicating Anthropic's multi-supplier compute strategy (Source: theinformation.com). DeepSeek's $7.35B raise (May 8) — its first outside funding round and the largest by a Chinese AI company on record, with Liang Wenfeng writing the biggest check himself — is read either as Chinese frontier labs converging with US capex intensity (a bear marker on the "compute efficiency wins" thesis) or as a strategic attack on US-lab pricing power before commercialization (a bull marker on Chinese-lab confidence) (Source: theinformation.com).

Sebastian Barros (May 2, 2026) argued the four US hyperscalers (Microsoft, Google, Amazon, Meta) now spend more annual capex on AI infrastructure than the entire global telecom industry combined, framing Q1 2026 earnings as the end of the asset-light cloud era and the hyperscalers' shift toward capital-intensive, telco-like entities through 2030 (Source: sebastianbarros.substack.com).

Labor-disruption signals

Several headcount actions are tied to AI but read as dual-use evidence. Cloudflare cut 1,100 staff (one-fifth of its workforce, May 7, 2026), citing AI usage growth of more than 600% over the prior three months and a need to "architect our company for the agentic AI era," with the stock falling more than 15% after hours (Source: theinformation.com). Microsoft took a $900M charge (reported May 3, 2026) for its first voluntary retirement program (Source: GeekWire, May 3). ClickUp laid off 22% of staff (May 22–25) while spinning up roughly 3,000 internal AI agents and redirecting savings into "million-dollar salary bands" for top performers, with CEO Zeb Evans framing the goal as a "100x org"; a May 5 Gartner survey found roughly 80% of companies using autonomous AI have cut jobs but the reductions "are not necessarily translating into meaningful financial returns," a read placed beside ClickUp as the strongest current "agents replacing workers without producing returns" anchor (Source: techcrunch.com). Coinbase's 14% layoffs appear in the same dual-readable cluster.

Historical analogs

Three analogs recur. The 1870s railroad bubble produced transformative infrastructure alongside severe financial consequences for shareholders. The 1999–2001 dot-com bubble produced real transformation, a financial reckoning, and winners that emerged after the crash. For AI today, many observers (including Bain and Morgan Stanley) frame the economic challenge as real, with the open question being whether it ends in productive catharsis or non-productive waste.

Relation to policy and backlash

Casey Newton ("We may now know what kind of AI bubble this is," April 30, 2026) frames the government-control-of-AI trajectory (Pentagon–Anthropic exclusion, NSA Mythos testing, a White House executive workaround) as the bubble's potential resolution mechanism: even if commercial demand falters, government procurement provides a floor (Source: platformer.news).

Shroff's May 13, 2026 follow-on argument (The AI Backlash Could Get Very Ugly — Lila Shroff (The Atlantic, May 13 2026)) holds that bubble dynamics fuel the political backlash they enable: if the bubble pops, AI-blamed displacement (real or perceived) intensifies, comparable to Industrial Revolution rioting and machine-breaking during early-19th-century downturns. The only income cohort optimistic about AI in daily life is >$200K households (Quinnipiac, cited by Shroff), implying both bubble and backlash carry asymmetric distributional consequences along the same income lines (see AI-Driven Political Violence).

Companion macroeconomic frames include The Simple Macroeconomics of AI (Acemoglu: ≤0.66% TFP over 10 years), The 2028 Global Intelligence Crisis (Citrini Research: a rapid-displacement macro scenario), and Fed Gov. Barr — AI and the Labor Market Speech (May 2025) (the Fed hedging between incremental-progress and transformation scenarios).

The historical base rate the contrarian framings are anchored on is tracked at History of AI, along with the reason the analogy is contested: earlier contractions followed undelivered capability claims, whereas the current cycle involves systems in wide commercial deployment.

Relationships

• related: The Intelligence Age (Altman), Abundant Intelligence (Altman), Machines of Loving Grace (bulls); AI Is Really Weird, Sen. Warren Letter to OpenAI (2026-01-28), MIT NANDA — The GenAI Divide (State of AI in Business 2025), METR — Measuring Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, Why I Think AI Take-Off Is Relatively Slow (Cowen) (bears); Bain & Company — 6th Annual Global Technology Report (2025), NERC 2025 Long-Term Reliability Assessment (LTRA), RAND — AI's Power Requirements Under Exponential Growth (2025), Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? (quantitative anchors).
• related: Circular Financing in AI — circular financing as one structural bubble indicator.
• related: AI Acquihires — acquihire pattern as a mechanism keeping lab capital on-balance-sheet while avoiding antitrust scrutiny.
• related: So, About That AI Bubble — Rogé Karma (The Atlantic, May 2026) — Karma's May 2026 Atlantic burden-of-proof-shifted update.
• related: Even Silicon Valley Says That AI Is a Bubble — Lila Shroff (The Atlantic, March 2026), The AI Backlash Could Get Very Ugly — Lila Shroff (The Atlantic, May 13 2026), How the AI revolution has turbocharged M&A, Lichtenberg — 'Nobody knows anything' and 'this time is different': the phrases that define — and haunt — the AI economy, Silicon Valley has forgotten what normal people want (Lopatto, The Verge, April 20 2026).