AI valuations: bubble debate sharpens as $650bn of capex meets $50-70bn AI revenue

Newsorga's short answer to the question of whether AI company valuations are justified or whether US equity markets are in an AI bubble is: we are in something that looks much more like a bubble than not, but it is a different kind of bubble than 1999-2000, and the resolution will not look like the dotcom crash. The longer answer requires working through the numbers that are actually framing the May 2026 debate among central banks, chief economists, bulge-bracket sell-side desks and a small number of high-conviction value-oriented buy-side investors who have started moving capital away from the Magnificent Seven and into other corners of the market.

The valuation stack at a glance

Three numbers anchor the AI valuation conversation in Q2 2026.

One: OpenAI is being marked at approximately $852 billion following its March 2026 funding round of $122 billion — by some distance the largest single funding round in the history of venture capital. To put that number in context, $852 billion is roughly the 2024 annual revenue of Walmart and roughly the entire market capitalisation of Mastercard.

Two: Anthropic is marked at approximately $380 billion, following its own roughly $30 billion raise in early 2026. Anthropic reported 2025 revenue in the low single-digit billions on aggressive growth trajectories.

Three: Q1 2026 venture funding for AI companies in aggregate hit approximately $300 billion — a quarterly record by a margin that breaks every prior comparison series. That is not a typo. In a single three-month window, AI startups absorbed more capital than the entire 2019 global venture-capital industry deployed across every sector.

Layered on top of these private-market marks are the public-market exposures: Nvidia at a market capitalisation north of $4 trillion, the Magnificent Seven (Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, Tesla) representing more than 30 percent of the S&P 500 by weight, and the top 10 holdings of the index accounting for approximately 40 percent of market cap — both figures documented by Apollo chief economist Torsten Slok in a chart book that explicitly framed the 2026 concentration as exceeding the 2000 dotcom peak. Slok's written formulation: the current AI episode has "swells larger than the IT bubble a quarter-century ago."

The capex-versus-revenue gap

If valuations are the symptom, the underlying issue is the gap between spending and monetisation. The four major US hyperscalers — Microsoft, Alphabet, Amazon and Meta Platforms — are committing approximately $650-700 billion of capital expenditure across 2026, broken down roughly as follows:

• Amazon (AWS): ~$200 billion
• Microsoft: ~$190 billion
• Alphabet (Google Cloud): ~$185 billion, with projections rising toward $250 billion in 2027
• Meta: $60-65 billion

That 2026 total represents an approximately 60 percent increase on 2025 levels and is roughly 2.4 times the four firms' combined 2022 capex of about $135 billion. The vast majority is for AI infrastructure — Nvidia GPUs, custom ASICs (Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia, Meta's MTIA), data-centre construction, power-supply contracts, networking and cooling.

Against that $650-700 billion of spending, the four hyperscalers reported approximately $50-70 billion of AI-attributable revenue across 2025 and are tracking toward something in the $100-130 billion range for 2026, depending on how aggressively each firm defines 'AI revenue.' The industry-wide global data-centre revenue base is approximately $250 billion annually — less than the four-hyperscaler 2026 capex alone.

Newsorga's decomposition of what this means for invested-capital returns: to justify a 15 percent return on invested capital — what the equity market is implicitly pricing into the Magnificent Seven as a group — the 2026 capex vintage alone would need to generate roughly $100 billion of incremental annual AI-attributable operating profit simply to cover depreciation at acceptable hurdle rates. That requires approximately doubling combined hyperscaler AI revenue every 12 months at industry-leading operating margins, sustained over the depreciation life of the assets. It is mathematically possible. It is not historically common.

The free-cash-flow collapse

The capex push is already showing up in free cash flow numbers, which matter because FCF is the eventual source of dividends and buybacks that public-market shareholders ultimately receive.

• Combined hyperscaler FCF fell from approximately $237 billion in 2024 to approximately $200 billion in 2025, with further sharp declines forecast for 2026.
• Amazon's trailing-twelve-month FCF has collapsed by approximately 95 percent year-over-year to roughly $1.2 billion, with the firm itself guiding to negative FCF of $17-28 billion in 2026.
• Alphabet's FCF is projected to fall to approximately $8.2 billion in 2026 from $73.3 billion in 2025 — a roughly 90 percent decline.
• Microsoft is the only one of the four whose operating cash flow is currently expected to exceed its AI capex in fiscal year 2026. Alphabet, Amazon and Meta are all capex-negative — i.e. they are funding the AI build-out by drawing down balance sheet, raising debt, or both.

Where the BIS comes in

The most consequential institutional intervention on the AI valuation question came from the Bank for International Settlements (BIS) in BIS Bulletin No 120, published in May 2026 by Iñaki Aldasoro, Sebastian Doerr and Daniel Rees under the title "Financing the AI boom: from cash flows to debt." The bulletin's three key takeaways are worth quoting in full:

First: "Investment related to artificial intelligence is surging — both in nominal amounts and as a share of GDP — and currently accounts for a substantial share of economic growth." This is the macro point. AI capex is not just a sectoral story; it is now visibly contributing to US GDP growth in a way that is showing up in real-time GDP-nowcasting.

Second: "The size of anticipated investment needs will require firms to shift the source of financing from operating cash flows to debt, with private credit playing a rapidly increasing role." This is the financing-architecture point. The AI build-out is moving off corporate balance sheets and into private credit vehicles — Blackstone, Apollo, KKR, Ares, Blue Owl, Sixth Street, Carlyle — and into the public debt markets. Newsorga's read: this matters because private credit risk transmission to the wider financial system is structurally different from balance-sheet risk on cash-flow-rich tech firms.

Third — and the most important sentence in the entire bulletin: "While macroeconomic and financial stability risks from the AI boom appear moderate, the boom's sustainability hinges on AI firms meeting high earnings expectations. The fact that equity prices have run far ahead of debt market pricing underscores this tension." That last clause is the analytical core. Debt markets — which are populated by professional credit analysts who price probability of default and recovery rates rather than narrative momentum — are pricing AI infrastructure debt at materially less aggressive valuations than equity markets are pricing the same firms' stock. When debt and equity disagree on the same balance sheet, equity is historically the one that turns out to be wrong.

The circular-financing structure

The single most-discussed mechanical feature of the 2026 AI capital landscape is the circular financing loop linking Nvidia, OpenAI, Microsoft, Anthropic and the smaller frontier labs. Bloomberg's 2026 graphics feature mapped the structure clearly:

• Microsoft has invested $13 billion-plus in OpenAI.
• OpenAI has agreed to purchase approximately $250 billion of cloud services from Microsoft over the term of their partnership.
• OpenAI has separately agreed to purchase approximately $100 billion of Nvidia chips.
• Nvidia has invested in OpenAI, Anthropic, CoreWeave, Crusoe, Lambda, xAI and several smaller labs, often in deals that are explicitly contingent on the recipient buying Nvidia GPUs.
• CoreWeave raised billions from Nvidia, used much of the proceeds to buy Nvidia GPUs, and is now the largest customer of Nvidia by some measures.
• Anthropic has cloud commitments to both Amazon (AWS) and Google Cloud, both of whom are equity investors.

Newsorga's editorial framing of this loop: it is not, in itself, evidence of fraud or wrongdoing. It is the standard pattern of a fast-moving capex-heavy industry where capital, customers and suppliers are densely intertwined. What it does mean is that reported revenue at the labs is not independent of reported capex commitments at the hyperscalers and chip vendors. A meaningful share of the OpenAI revenue line items that justify the $852 billion valuation is, in accounting substance, Microsoft and Nvidia capital being recycled back into the system. If the underlying demand from third-party enterprise customers does not eventually replace this circular flow, the valuations collapse simultaneously across the loop.

Jensen Huang, Nvidia's CEO, made an unusually candid comment on CNBC on March 4, 2026 that captures the saturation point: a $30 billion Nvidia investment in OpenAI "might be the last" of that scale. The acknowledgement matters because it signals that even Nvidia — the entity sitting at the centre of the loop and benefiting more than anyone from the circularity — sees the cycle approaching its limits.

The dotcom comparison — and why it's imperfect

The 1999-2000 dotcom bubble is the obvious historical comparison, and the AI episode shares structural features with it: extreme equity-market concentration, capex-heavy infrastructure build-out, narrative-driven valuations on companies with limited or no current profits, retail-investor euphoria, and a peer-pressure dynamic that forces every corporate board to claim AI exposure regardless of underlying business fit.

But there are three important differences that make the bubble shape different, even if the eventual correction is severe.

One: the underlying firms are profitable. The 2000-era poster children (Pets.com, Webvan, Cisco's capex customers, the alternative carriers like WorldCom and Global Crossing) were either entirely unprofitable or building infrastructure whose end demand never materialised. The 2026 AI-exposed mega-caps — Microsoft, Alphabet, Meta, Amazon, Nvidia — are, in absolute terms, the most profitable companies in human history. Nvidia's most recent fiscal-year net income is comfortably over $100 billion. The capex is being funded out of real free cash flow, even as it eats most of that cash flow.

Two: there is some underlying revenue. Microsoft is generating tens of billions of incremental cloud revenue identifiable as AI-Attributable. Meta reported 24 percent year-over-year ad-revenue growth in Q1 2026 with 6 percent price-per-ad gains — both substantially attributable to AI-driven ad-targeting improvements. AI revenue is not zero; it is just smaller than the capex.

Three: AI has end-user pull that the alternative-carrier telecoms of 1999 never had. Hundreds of millions of consumers are now actively using ChatGPT, Claude, Gemini, Perplexity, Copilot, Cursor and an expanding family of AI-enabled products. 93 percent of professional developers report using AI coding tools, according to surveys cited in independent analysis by Philipp Dubach — even if measured productivity gains remain modest at around 10 percent. The consumer surplus is real even if the producer surplus is contested.

What could cause a real bust

Three identifiable triggers could turn the current valuation regime into a 1999-style correction.

First: the inference price collapse. Epoch AI research, summarised by Philipp Dubach, shows AI inference prices declining at a median rate of approximately 50 times per year for equivalent performance. If that rate persists, the commoditisation of frontier-model inference will compress margins for everyone in the value chain — OpenAI, Anthropic and the hyperscalers selling GPU-hours alike. AI would become the new rebar: indispensable, ubiquitous and unprofitable.

Second: a single capex pull-back by one of the four big hyperscalers. The current arms race is in part driven by competitive game theory — none of the four can pull back capex without ceding share. But if Meta, for example, were to materially cut its AI capex in response to a specific quarter of disappointing ad-revenue growth or Reality Labs capital reallocation, the others might follow, and the circular-financing loop would unwind quickly.

Third: an enterprise-adoption failure. Independent research has been documenting consistently that 85 percent of enterprise AI projects fail to reach production. If that ratio does not improve materially through 2026-2027, the enterprise-software thesis that justifies the demand side of the build-out breaks down. The Q4 2025 sell-off in software equities — the "SaaSpocalypse," with the IGV software ETF falling 32 percent despite 17 percent aggregate earnings growth — was an early signal of this risk.

Why this still doesn't end like 2000

Even with the bubble framing, Newsorga's working view is that the resolution does not look like a sharp NASDAQ-2000-2002-style -78 percent crash. The structural reasons:

• The mega-cap AI-exposed firms have real cash flow, large return-of-capital programmes that limit downside, and balance sheets that comfortably survive a 30-40 percent drawdown in their stocks without operational impairment.
• AI capex spending, even if it is over-funded relative to current revenue, is creating durable physical infrastructure — data centres, GPU fleets, power grids — that has value irrespective of whether the current cohort of frontier labs achieves the implied AGI business case.
• The AI demand side, unlike the 1999 end-state, has end-user pull from hundreds of millions of paying consumers and enterprise customers. The product itself is not a phantom.
• Private credit — the financing channel the BIS is most worried about — has lock-up structures that prevent the kind of synchronised forced selling that turned the 2008 collateral markets into a one-way trade.

The realistic correction scenario, in our base case, is a 30-50 percent drawdown in the most AI-exposed names (Nvidia, CoreWeave, Palantir, the smaller chip names, the most aggressive AI-exposed software companies), a 15-25 percent drawdown in the mega-cap hyperscalers, and a 60-80 percent decline in the unprofitable thematic AI small-cap complex. Private valuations — OpenAI at $852 billion, Anthropic at $380 billion — would likely re-price by 40-60 percent in any subsequent funding round, with secondary market discounts visible considerably earlier.

The honest editorial position

Are AI valuations justified? On the basis of the numbers above, not in aggregate. The $650-700 billion of 2026 capex against $50-70 billion of current AI revenue, the circular Microsoft-Nvidia-OpenAI financing structure, the BIS observation that equity has run ahead of debt, the Apollo documentation of unprecedented index concentration, and the Bank of England / IMF warnings of approaching dotcom-level valuations together describe a market that is pricing one of the most aggressive consensus future-earnings paths in capital-market history.

Are we in a bubble? Yes, in the technical sense that prices are above any reasonable fundamental valuation under a wide range of assumptions about the future of AI revenue. But bubbles can persist for longer than is conventionally assumed, and the AI capex cycle is still in the building phase rather than the collecting-customer-cash-flows-on-built-infrastructure phase that will eventually settle the question.

Newsorga's practical guidance for retail readers thinking about portfolio decisions: treat AI exposure as a high-conviction directional bet, not as a passive index allocation. Recognise that buying the haystack — the S&P 500 — is, in 2026, an active overweight to a small group of AI-exposed mega-caps rather than the broad-market diversifier it was 10 years ago. Apply normal risk-management discipline: position-sizing, drawdown tolerance, time horizon, liquidity preferences. And read BIS Bulletin No 120 — at 8 pages, it is the most efficient summary available of what the central-bank community is actually worried about, and the institutional view is, on balance, that the equity market is mispriced relative to debt and that the resolution will come within the next 24-36 months.