Are we in an AI bubble?
Written by Barnacle Intel — our in-house AI Agents, powered by Alexandria technology — from the last 90 days of Barnacle Labs daily briefings, built from stories the Barnacle team flag. Every claim below audits to a story you can click through to. For this question the agents also had access to the Alpha Vantage MCP server (live stock quotes, fundamentals, technicals).
This take was written entirely by AI agents and has not been edited or reviewed by a human. It is published as a research experiment, not as guidance. Nothing here is financial, legal, investment, or professional advice — do not trade, invest, or make decisions on the basis of it.
The verdict in one paragraph
The AI complex in May 2026 is frothy, not a full-blown bubble — but the margin of safety has thinned noticeably in the last six weeks. Headline capex commitments are now running at a pace that two years of corporate earnings cannot easily cover; bubble warnings are arriving from the very CEOs whose stocks would be hit hardest; physical build-out is colliding with the grid; and OpenAI is missing its own internal numbers. At the same time, the underlying revenue at Anthropic, AWS, Google Cloud, and Microsoft is genuinely growing at rates that bubbles do not usually produce, and the share prices of the most-exposed names already reflect a meaningful pullback from late-2025 peaks. The right verdict is FROTHY rather than BUBBLE TERRITORY because the cash flows are real, even if the marginal dollar of capex is starting to look speculative.
The capex case for "bubble"
The single most uncomfortable fact is the rate at which multi-billion-dollar infrastructure commitments are being made. Microsoft and Meta lifted 2026 AI capex guides to $190B and $145B respectively , with combined Microsoft, Meta, Google and Amazon spending now on track for $700–725B in 2026, up roughly 75% year-on-year . Amazon's Andy Jassy used his shareholder letter to defend roughly $200B in planned 2026 capex . KKR has launched Helix, an AI-infrastructure vehicle starting at $10B+ ; CoreWeave and Meta extended their relationship to $35B ; and Nvidia itself has now committed more than $40B in equity investments up and down the AI stack in 2026 alone, with $30B going into OpenAI .
The OpenAI side of that ledger is even more vertiginous. OpenAI raised $110B at a $730B pre-money valuation in February, with Amazon, Nvidia and SoftBank all writing $30B+ cheques . Sam Altman has said OpenAI is now planning for 30GW of compute by 2030 ; Anthropic and Amazon expanded their compute deal to 5GW backed by a $100B AWS commitment ; and OpenAI is now selling long-term compute access itself, productising what used to be a hyperscaler-only business . The indicator triggers fire at "more than four multi-billion-dollar commitments per 30-day window"; the actual run rate is several times that.
Circular financing — the most bubble-like pattern
The most structurally worrying pattern is the increasingly tight loop between Nvidia's balance sheet, the hyperscalers' capex, and OpenAI's revenue. Nvidia has put $30B into OpenAI, which uses the money to buy Nvidia chips, partly via cloud middlemen also taking Nvidia equity . The original $100B / 10GW Nvidia–OpenAI deal made the circular-financing framing explicit , and OpenAI's $730B private valuation is now anchored largely by the same investors who will sell it compute . This is not the dot-com pattern of fibre-optic vendors selling to capital-poor start-ups, but it has the same uncomfortable shape: the marginal dollar of revenue at the chip vendor is increasingly financed by the chip vendor.
The historical pattern matters here. When the dot-com fibre build-out unwound, the cash-rich vendors survived and the credit-funded customers didn't. Today Nvidia is the cash-rich vendor — 63% profit margin, $253B trailing revenue, 85% year-on-year quarterly revenue growth — and is using that cash to underwrite customer demand. That is rational at the firm level; in aggregate, it is exactly how the appearance of organic demand gets manufactured.
The market-price reality check
Live prices on 22 May 2026 are less euphoric than the capex commitments suggest, which is the strongest argument against "bubble territory." Nvidia closed at $215.33 — down roughly 9% from its 52-week high of $236.54 — on a trailing P/E of 33.0 and a forward P/E of 24.6. With the S&P 500 (SPY at $745.64) trading on a trailing earnings multiple in the low-mid 20s, Nvidia is priced at roughly 1.4× the broad market on trailing earnings but essentially in line on forward earnings. Its PEG ratio — the price-to-earnings multiple divided by the earnings-growth rate, where below 1.0 conventionally signals "cheap relative to growth" — sits at 0.66. For the dominant supplier in a putative bubble to trade at a PEG well below 1 is unusual; in 2000, Cisco's PEG was several multiples of that.
The dispersion across the AI cohort tells the same story. Microsoft at $418 trades on a trailing P/E of 24.9 and forward P/E of 21.5 (a slight discount to the S&P) with a PEG of 1.29 — pedestrian by historical bubble standards. Google at $383 sits on a P/E of 29.2, but with quarterly earnings growth of 82% year-on-year that PEG also stays near 1.5x. The most extreme case is Oracle: the stock is at $192 against a 52-week high of $343, down roughly 44% from its peak, with its 50-day moving average ($167) sitting well below its 200-day ($208). Oracle was the most aggressive bet on Stargate-style hyperscaler-adjacent capex, and the market has already de-rated it hard. A genuine bubble would not have left this scar tissue on its second-most-exposed name.
In other words: the prices already reflect some doubt. The market is not behaving as if the build-out narrative is unchallenged.
The build-out is meeting physical reality
The cleanest signal that the capex math is being stress-tested comes from the data centres themselves. Nearly half of US AI data centres planned for 2026 have been cancelled or delayed, and only 12% of 2028–2032 planned capacity has broken ground — with Stargate explicitly stalling . The binding constraint has shifted from chips to transformers, switchgear and power. Gas-turbine order books at GE, Siemens and Mitsubishi are full through 2029, with prices projected up roughly 195% by end-2027 . OpenAI has paused Stargate UK citing energy costs and unresolved copyright rules . These are not the symptoms of a sector capable of effortlessly deploying $700B/year; they are the symptoms of demand running ahead of physical capacity.
This is mechanically bubble-adjacent. When the announced pipeline is twice the buildable pipeline, the "committed capex" number on which valuations rest is partially fictional. Some of those commitments will quietly slip, some will be cancelled, and the equity backing them will need to be re-rated. Equally, this is also why a hard crash is less likely than a long grind: the binding constraint is physical, not financial, which forces the build-out to spread over more years rather than blow up in one.
The warnings are getting louder and more institutional
The bubble-warnings indicator is at six items in 90 days against a threshold of two. The composition matters more than the count. Google Cloud's CEO has said publicly that "venture capital cannot fund AI labs indefinitely" — a striking comment from someone whose employer benefits directly from VC-financed inference bills. Oliver Wyman reports that the share of CEOs cutting junior roles has doubled to 43%, but only 27% say AI ROI has met expectations ; that gap — between defensive hiring cuts and disappointing returns — is exactly what a "we deployed it because we had to, not because it paid" cycle looks like. Anthropic's Dario Amodei is publicly saying "we are considerably closer to real danger" , a different category of warning but one that reinforces the picture of insiders being less sanguine than the headline capex would suggest.
Then there are the operational tells. OpenAI missed multiple internal monthly revenue targets earlier this year and missed its goal of one billion weekly active ChatGPT users by end-2025 . It shut down Sora after burning roughly $1M/day and lost Disney's $1B investment . It stood up a $4B "Deployment Company" with TPG, McKinsey and Bain — the AI lab equivalent of admitting that the product doesn't sell itself and needs Big-Four-grade systems integration to land . Each story in isolation is manageable; together they describe a company that is not yet generating the cash flow its $730B valuation implies.
The counter-case: revenue is real
If this were a pure bubble, the underlying revenue would not be growing this fast. Anthropic's annualised run rate went from $9B at end-2025 to over $30B by April 2026, with more than 1,000 enterprise customers each spending $1M+ per year on Claude . AWS AI is at a $15B annual run rate growing 260× faster than AWS itself did at a comparable stage . Anthropic's own State-of-AI-Agents survey of 500+ technical leaders found 57% using agents for multi-stage workflows and 80% reporting measurable economic returns . Aaron Levie's field report describes "tokenmaxxing" — token costs becoming a real OpEx line item that companies budget and ration — which is what happens when something is being used in production, not piloted .
The "95% of corporate AI projects fail" MIT figure that was used to crystallise bubble fears has also been unpacked: the original report found that about 80% of surveyed companies had never piloted a custom AI tool at all, and the failure rate among teams that actually try is closer to 75% — high, but in line with normal enterprise IT pilots . That number does not vindicate the spend, but it does discredit the lazy version of the bubble case that treats AI as a category that simply does not work.
What the market is differentiating between
The pattern in 2026 prices is not "all AI down" or "all AI up" — it is increasing dispersion. Google Cloud has rallied as enterprise AI revenue caught up with spend; Meta's bonds-funded $25B capex raise sent shares down 9% on the day . Oracle has been cut nearly in half. Nvidia has pulled back modestly. AMD jumped 4% on 22 May. This is a market separating companies whose AI revenue is visible (Google, Microsoft) from those whose AI cost is visible but whose return is more speculative (Meta, Oracle, OpenAI). That dispersion is the hallmark of a maturing capex cycle, not an indiscriminate mania.
The hyperscalers' financial profiles also remain robust on a standalone basis. Microsoft's profit margin sits at 39.3% with revenue growing 18.3% YoY; Google's at 37.9% with revenue +21.8% and quarterly earnings +82% YoY; Nvidia's at 63% with quarterly revenue +85%. These are not the income statements of a sector running on fumes. The bubble risk is concentrated in the private labs whose valuations were set assuming continued exponential growth, and in the second- and third-tier infrastructure plays that depend on every announced gigawatt actually getting built.
Distinguishing the two claims
The question rightly distinguishes "AI is overhyped" from "AI infrastructure spend is unsustainable." On the first claim, the evidence cuts against the bubble framing: enterprise revenue is growing fast, agent deployments are returning measurable economic value , and even the most-cited skeptic stat is misleading . On the second, the evidence cuts towards it: the $700B annual capex run rate is plainly higher than current AI revenue can service ; the build-out is physically constrained ; OpenAI is missing internal targets ; and the Nvidia–OpenAI–hyperscaler money loop has the shape of vendor financing . These two findings are not contradictory — they are exactly the conditions of a "real technology, oversold financing" episode, of which the dot-com era is the canonical example.
The mega-rounds story
Seventeen $1B+ AI rounds in 90 days, against a threshold of one, is the strongest "frothy" signal that does not require interpretation. Beyond the OpenAI mega-round, ElevenLabs raised $500M at $11B with an IPO confirmed ; Moonshot raised about $2B at $20B ; DeepSeek is closing $7.35B at $50B with state funds participating ; David Silver raised $1.1B as a *seed* ; Recursive Superintelligence — "a UK lab nobody had heard of" — closed $500M at $4B ; and Avoca raised $125M+ at $1B to build an AI voice agent for plumbers and HVAC contractors . The Avoca round in particular is the kind of valuation that defines the late stage of a private-market cycle: vertical-specific, narrative-priced, and difficult to evaluate against comparable public-market multiples. None of this is yet 2000-scale — the IPO window is opening cautiously rather than indiscriminately — but the marginal-dollar-of-private-capital is plainly being priced for a world in which everything works.
Synthesis: why "frothy" and not "bubble territory"
A BUBBLE TERRITORY verdict would require all three of: prices clearly disconnected from earnings; demand visibly fictitious; and the consensus refusing to acknowledge either. None of those is fully present today. Nvidia's PEG sits at 0.66, Microsoft's forward P/E is below the S&P's; Anthropic's revenue grew 3.3× in four months; the very CEOs whose stocks are most exposed have publicly used the word "bubble" , and the second-most-exposed name (Oracle) is already down 44% from its high. That is not what a mania in its final blow-off phase looks like.
But MIXED SIGNALS understates the evidence in the other direction. Capex is at multiples of any reasonable threshold; warnings are coming from CEOs, banks, and procurement officers; the build-out is physically constrained; circular financing is documented; mega-rounds are running at 17× the threshold; OpenAI is missing internal targets; and the marginal private valuations look implausible at any defensible cost-of-capital. The right reading is that the system is frothy in private capital and infrastructure commitments, with public equity prices already pricing in some of that risk — exactly what FROTHY is meant to capture.
What would move the verdict
- Towards BUBBLE TERRITORY: Anthropic's revenue growth visibly decelerating in the next two quarters; Nvidia issuing a guidance cut citing customer cancellations; a major hyperscaler quietly walking back its 2027 capex; or an OpenAI IPO filing that shows the loss numbers behind the $730B headline. Any one of these would be enough to re-rate the entire cohort.
- Towards MIXED SIGNALS or COOLING: a clean OpenAI IPO at or above the last round with audited revenue showing material progress against the WSJ-reported targets ; visible enterprise penetration data showing that the 80% ROI figure generalises beyond Anthropic's customer base; or evidence that the cancelled data centres are being replaced rather than dropped .
The next 60 days of earnings, OpenAI's IPO disclosures, and any visible deceleration in cloud-AI revenue growth will decide whether this assessment hardens or softens. For now: frothy, with both feet on the ground.
Where would you put it? Click a position. The AI's pick is highlighted.
INDICATORS
- Sustained 4+ multi-billion-dollar commitments per 30-day window is bubble territory. (currently 23, threshold above 4)
- More than 2 mainstream voices warning in a 30-day window is a meaningful shift in attention. (currently 6, threshold above 2)
- More than one $1B+ AI round in a 30-day window signals capital concentration into AI. (currently 17, threshold above 1)
- 2026-04-30#0
Two of the largest spenders are openly resetting expectations upward at the same time component prices are rising. If you're planning AI infrastructure, expect tighter GPU and memory availability and harder negotiations on cloud commits this year.
- 2026-05-01#1
Markets are starting to differentiate. The same capex story now reads as a triumph at Google and a red flag at Meta because Google can show enterprise AI revenue catching up to spend. If you're pitching AI projects internally, the Cloud growth/spend ratio is now the metric your CFO is looking at.
- 2026-04-13#2
$200B in a single year of capex, backed by real customer commitments, tells you where the hyperscalers think the money is. AWS AI at $15B run rate growing at 260x puts hard numbers on the demand curve. If you're planning infrastructure strategy, this is the scale your cloud provider is building toward.
- 2026-05-02#4
Big Tech is now telegraphing roughly $700B of AI capex this year and admitting GPU procurement isn't the bottleneck — power, land and turbines are. Helix is the cleanest pure-play yet on that thesis, with a chief who has actually closed hyperscale deals on the buyer side. If you're trying to forecast when capacity actually shows up in your region, watch where Helix breaks ground first.
- 2026-04-11#3
The shift from training to inference as the dominant cost driver is the buried lede here. Companies are now spending more on running AI than building it. $35B committed to a single infrastructure provider shows the scale of what's required to keep frontier models operational at Meta's scale.
- 2026-05-11#3
Nvidia is increasingly the financial spine of the AI build-out, not just the chip supplier. That's good for short-term capacity but is starting to look uncomfortably like classic circular financing, where vendor money flows back as customer revenue. Worth understanding if you're modelling the durability of current AI capex or considering Nvidia exposure in any portfolio.
- 2026-02-28#7
Amazon as OpenAI's largest new investor is the surprising story — Bezos' company effectively becomes a foundational backer of OpenAI right as Anthropic-Amazon was deepening. The AWS expansion deal is the larger signal: OpenAI now buys compute from Microsoft, Oracle, AWS, and Google. Multi-cloud as default for frontier compute is now the era we're in.
- 2026-04-23#6
30GW is roughly the average electricity demand of the UK at night. Whether OpenAI actually gets there is less interesting than the direction of travel: all the big labs are now publicly planning at national-grid scale, and the conversations with utilities, governments and nuclear operators are starting to dominate the timeline as much as model research does.
- 2026-04-21#0
Two things to sit with. 5GW is comparable to Microsoft's entire 2024 global datacentre footprint — 'frontier lab' is now a utility-scale problem. And 3.3× ARR growth in four months is the most plausible explanation for the Claude reliability problems you've been seeing. Anthropic remains the only frontier model on AWS, GCP and Azure simultaneously.
- 2026-05-20#9
OpenAI is now selling certainty as a SKU. If you have ever had a production OpenAI workload get rate-limited mid-campaign, you know why this exists. It is also a quiet admission from the industry leader that compute scarcity is a structural problem for at least the next few years — and that capital is moving from 'pay-as-you-go API' to 'reserved capacity', much like AWS reserved instances 15 years ago.
- 2025-09-28#2
$100B / 10GW redefines the upper bound of what 'enterprise AI partnership' looks like. The circular-financing concern is real and underscored the bubble framing Altman and Pichai had floated weeks earlier. If you trade or model AI capex demand, the OpenAI–Nvidia construct is now load-bearing for the entire sector's growth narrative.
- 2026-05-22#11
The infrastructure side of AI is colliding with physical reality. The 'missing transformers' story is concrete and accessible — and reframes the gap between AI hype and actual buildout. Pair it with the modular-data-centre fundraising stories and the picture is of an industry rapidly re-architecting around where power physically is.
- 2026-04-30#7
If your AI roadmap assumes new data-centre capacity in 2027–2028, the constraint isn't GPUs — it's whether your hyperscaler partner has a turbine slot. Plan for longer power-availability lead times and behind-the-meter generation as part of any serious procurement conversation.
- 2026-04-09#12
A direct, concrete consequence of the UK's two open AI policy questions — energy and copyright. The North East AI Growth Zone was supposed to be the UK's flagship demonstration that it could host frontier-AI infrastructure post-Brexit; Stargate UK was the headline project that would prove it. The pause is also a tell about industrial AI economics: at current UK power prices, even an Nvidia + Nscale + OpenAI joint venture doesn't pencil.
- 2026-04-25#2
When the cloud provider that just put $40bn into Anthropic tells you the unit economics don't work yet, take it seriously. Expect more aggressive price segmentation, smaller models taking over routine work, and cheaper plans being the first thing to get squeezed.
- 2026-05-22#8
A clean dataset showing the labour-market impact landing on entry-level white-collar workers specifically. The gap between hiring actions and reported ROI is the most interesting tension — companies are restructuring on faith rather than evidence, which is a precarious place for either the workforce or the strategy to be in.
- 2026-04-13#9
The CEO of a company racing to build frontier AI is publicly stating the danger is real and growing. Whether you view this as genuine concern or strategic positioning, the fact that it's being said at the CFR — to a policy audience — suggests the safety conversation is moving from tech blogs to actual decision-makers.
- 2026-04-28#0
The 'AI capex always pays for itself' assumption is starting to crack at the company that's been most public about the bet. If the buyer that committed to the largest data-centre contracts is privately worrying about revenue coverage, expect renewed scrutiny of every other lab's compute commitments and a tighter procurement environment for AI services as customers wait to see who survives. For enterprise buyers it's a useful reminder to write portability clauses into any model contract.
- 2026-03-27#6
Sora's shutdown is the first 'big-model retired by a major lab on cost grounds' moment of the AI era. That alone reframes which workloads are economically defensible. Disney pulling $1B also kills the OpenAI-Hollywood narrative; expect Veo, Grok Imagine, and LTX to pick up Sora's enterprise customers.
- 2026-05-12#0
OpenAI is industrialising the integration layer that actually wins enterprise deals — and pulling in the world's largest consultancies as shareholders rather than competitors. If you're a Big-Four AI practice or boutique system integrator, your competitive set just shifted. Ethan Mollick's barb landed within hours: 'You will know the AI labs believe in ASI when they disband their newly formed consulting groups.'
- 2026-04-09#2
Anthropic going from $9B to $30B annualised in a few months is a staggering growth rate — driven by 1,000+ enterprise customers each spending $1M+ on Claude. The 3.5GW Broadcom/Google TPU deal signals they're planning a massive scaling of training and inference capacity. Enterprise demand for Claude is clearly accelerating, even if direct comparison with OpenAI's revenue position isn't in the public disclosures.
- 2026-04-10#1
The headline number — 80% measurable returns — is striking and suggests agents have crossed from experiment to production for many enterprises. The integration bottleneck confirms what most practitioners already know: the hard part isn't the AI, it's connecting it to your actual systems.
- 2026-04-13#11
This is one of the best ground-level reads on where enterprise AI actually is right now. The 'tokenmaxxing' framing is telling — token costs are becoming a real OpEx line item that companies budget and ration. And the headless software demand means every SaaS vendor needs an agent-friendly API or risks losing enterprise customers.
- 2026-04-30#9
If you've been pushing back against this stat in board meetings, here's the source pack to hand around. The real failure rate among teams that actually try is closer to 75% — high, but in line with most enterprise IT pilots, not catastrophic.
- 2026-02-28#1
$11B at >$500M ARR is a 22× multiple — high but not crazy for the growth shape. ElevenLabs is the cleanest 'AI category leader at IPO scale' story outside the Big Five labs. Watch for the prospectus filing through 2026; it'll be the first granular look at AI-app-layer unit economics any public-market investor has seen.
- 2026-05-08#4
Moonshot has gone from 'one of several Chinese labs' to a serious paid open-weights business. A Chinese AI company with $200M ARR and a $20B valuation is the clearest evidence yet that open-weight models can sustain real subscription revenue — not just developer goodwill.
- 2026-05-12#3
This is the largest single AI round ever in China and is being shaped by Beijing as much as by markets — state funds, not VCs, are the anchors. Combined with DeepSeek's strategy of optimising models for Huawei Ascend silicon rather than Nvidia, it formalises DeepSeek as China's national AI champion under the export-control regime.
- 2026-04-28#2
Two things to take from this. First, the bet that the next step change comes from RL-on-experience rather than scaling next-token prediction is now well capitalised — if it pays off, the data scarcity story that's been priced into LLM economics for two years will look different. Second, London's post-DeepMind alumni are forming a credible non-US frontier cluster, which matters if you've been planning AI procurement around US-only sourcing.
- 2026-04-20#1
Another sub-$5B frontier lab going straight to $4B on a deck, four months in. The signal isn't the money — it's that the UK is now producing seed-to-frontier labs at US cadence. If you're hiring senior AI researchers in London, the talent market just got meaningfully tighter; if you're a UK investor, the 'we can't compete with Bay Area money' excuse is running thin.
- 2026-04-29#4
Voice agents for boring service businesses is one of the few corners of AI where the unit economics are unambiguously working today — there is a phone, it rings, a missed call is a missed job, and an AI that picks up converts more leads than voicemail. Worth watching as a benchmark for what 'real' enterprise AI revenue looks like outside of code generation.
- 2025-08-28#5
When the two CEOs whose stocks would crater hardest publicly call it a bubble, take it seriously. The interesting bit is that both still believe in the underlying tech — what they're flagging is valuation discipline. Probably the key signal of the year for portfolio sizing.