This Obviously is an AI Bubble. The Math Says So
A Quantitative Case for the Obvious.
Editor's note: This article is an updated and significantly expanded version of 'The Current AI Investment Landscape,' originally published December 9, 2025.
Between 2023 and 2026, the world’s largest technology companies will have committed approximately $1.6 trillion to artificial intelligence infrastructure. By 2027, annual spending alone is projected to exceed $1 trillion. These are not projections from optimistic analysts or AI evangelists. They are the companies’ own guided figures, disclosed in earnings calls and SEC filings.
The bull case is straightforward: AI is a transformational technology, and the companies building its infrastructure are rationally securing dominance before their rivals do. Cloud revenue is growing fast. Google Cloud surged 63% year-on-year in Q1 2026, AWS 28%, Azure 40%. Backlogs are building. Demand appears real.
But the math tells a more complicated story.
J.P. Morgan estimates the industry needs to generate $650 billion in new annual revenue just to achieve a 10% return on the infrastructure being built. Current AI-attributable revenue, even under generous assumptions that credit all incremental cloud growth to AI, sits somewhere between $50 and $150 billion annually. That is a gap of between 4x and 13x, depending on how optimistic your assumptions are. The Bain projection is starker still: $2 trillion in annual AI revenue required by 2030.
Meanwhile, the financial structures underpinning this buildout are becoming harder to defend. Depreciation schedules set in 2023 are already outdated, understating real asset depletion by an estimated $176 billion between 2026 and 2028. Over $662 billion in lease commitments sit off balance sheets, quietly undermining the argument that this expansion is conservatively self-funded. At Amazon, capital expenditure has grown from $12 billion in 2017 to $151 billion in the last twelve months, now exceeding its operating cash flow entirely and pushing free cash flow into negative territory for the first time.
We are optimistic about the long-term future of AI. The technology is real, the productivity potential is real, and some of the revenue growth is real. This article does not argue that AI will fail. It argues something more specific and more verifiable: that the capital being deployed today is so far ahead of the revenue it can plausibly generate, under any reasonable set of assumptions, that the current investment cycle has the defining characteristics of a capital misallocation bubble. The math makes that case more clearly than any narrative can.
The Unprecedented Scale of AI Capital Expenditure
The Monetization Gap: Assessing the Disconnect Between Investment and Returns
Financial Engineering and Structural Fragility
The Energy Constraint: From Narrative to Numbers
“Valuations Are Not that High”. Yet.
Conclusion: When the Numbers Don't Add Up
1.0 The Unprecedented Scale of AI Capital Expenditure
The global technology industry is in the midst of an unprecedented capital expenditure boom, centered entirely on building the foundational infrastructure for Artificial Intelligence. What began as an aggressive but arguably defensible investment cycle has, by 2026, crossed into territory that demands serious scrutiny. The world’s largest technology firms are committing capital at a pace and scale that has no historical precedent, and the numbers are no longer moving in a direction that makes the bull case easier to defend.
The individual company commitments tell the story plainly. For 2026, Google parent Alphabet has guided to $185 billion in capital expenditure, up from $52 billion as recently as 2024. Amazon has committed $200 billion, Microsoft $190 billion, and Meta between $125 and $145 billion. Combined, the four largest hyperscalers alone are committing approximately $700 to $725 billion in a single year, nearly double what they spent in 2025, which was itself nearly double 2024. Bank of America projects this combined figure will exceed $1 trillion in 2027.
To appreciate how quickly this has escalated, consider that combined CapEx across these companies was approximately $162 billion in 2022. By 2025 it had reached $448 billion. By 2026 it will reach $725 billion. That is a compound annual growth rate of approximately 65%, sustained over four consecutive years, on a base that was already enormous.
Three aspects of this spending deserve particular attention beyond the headline figures.
First, a significant portion of the increase is being driven not by new capacity decisions but by input cost inflation. Microsoft disclosed that approximately $25 billion of its $190 billion 2026 budget is attributable to higher component pricing, particularly memory chips. This means the industry is paying more for the same or similar capacity, compressing the return on each dollar deployed before a single revenue line has been written.
Second, the visible CapEx figures substantially understate the true financial commitment. Moody’s reported in early 2026 that hyperscalers have approximately $662 billion in data center lease commitments that have been signed but not yet commenced. These obligations sit off balance sheet under GAAP’s lease commencement standard, meaning they do not appear in the capital expenditure figures that analysts and investors typically scrutinise. This off-balance-sheet liability is larger than the combined on-balance-sheet debt of the same companies. The bull case defense that this buildout is conservatively self-funded from operating cash flows becomes significantly harder to sustain when the full picture of committed obligations is accounted for.
Third, the spending is accelerating even as free cash flow is deteriorating. At Amazon, capital expenditure has now reached $151 billion in the last twelve months, exceeding its operating cash flow entirely and pushing free cash flow into negative territory. Across the hyperscalers more broadly, the direction of travel is consistent: CapEx is growing faster than the cash it generates, a dynamic explored in detail in Section 3.4.
The competitive logic driving this spending remains the same prisoner’s dilemma our original article described. Alphabet’s CEO has stated that “the risk of underinvesting is dramatically greater than the risk of overinvesting.” That logic is internally coherent for any individual company. The problem is that when every major player in an industry applies the same logic simultaneously, the collective result is an investment cycle that is structurally incapable of self-correction until the losses become impossible to ignore.
One addition to the AI landscape since our December 2025 article deserves its own mention: the Stargate project. Announced in January 2025 as a joint venture between OpenAI, SoftBank, Oracle and MGX, Stargate represents a $500 billion commitment to AI infrastructure over four years, with the first data center already operational in Abilene, Texas. Epoch AI estimates the project will exceed 9 gigawatts of capacity by 2029. Stargate is being financed largely through debt rather than equity, and by mid-2026 OpenAI had already quietly shifted strategy toward renting compute from hyperscalers rather than building, raising early questions about the execution of the original vision. Stargate is, in many ways, the purest expression of everything this article examines: unprecedented capital commitment, debt-financed, built ahead of proven demand, on a timeline that assumes the revenue will arrive before the bills do.
The critical question remains the same as it was in our original article. It is simply more urgent now:
Can these investments generate sufficient and timely returns to justify their cost?
2.0 The Monetization Gap: Assessing the Disconnect Between Investment and Returns
Validating a capital expenditure cycle of this magnitude requires the eventual emergence of tangible, profitable, and scalable revenue streams. A critical evaluation of the current AI landscape reveals a significant and widening gap between the revenue required to justify these investments and the observed reality of AI monetization. This section quantifies that gap with current data, examines the core challenges in user and enterprise adoption, and introduces two new dimensions of the problem that have emerged since our original analysis: the hidden cost of token consumption, and the accelerating erosion of pricing power at the model layer.
2.1 The Math Behind the Gap
The revenue required to justify the current buildout is not a matter of interpretation. It is arithmetic. And three independent analyses, using different methodologies and different starting assumptions, arrive at the same conclusion.
J.P. Morgan estimates the industry needs to generate approximately $650 billion in new annual revenue to achieve a modest 10% return on the infrastructure being built. To put that figure in human terms, it is equivalent to $34.72 per month from every current iPhone user, or $180 from every Netflix subscriber, in perpetuity. The Bain projection, cited by the Wall Street Journal, is even more demanding: $2 trillion in annual AI revenue by 2030.
Marathon Asset Management, approaching the question from a capital cycle perspective, arrives at an even larger number: Morgan Stanley estimates cumulative data center investment of $3 trillion between 2025 and 2028, excluding energy costs entirely. McKinsey estimates that approximately 60% of data center investment goes into chips and hardware, implying roughly $1.8 trillion in hardware alone over that period.
If those assets generate no economic profit after 5.5 years, the hardware investment alone requires net cash flow of over $500 billion in 2028 just to cover the cost of capital on the equipment. If data center operators need to generate 20% free cash flow margins to justify their current share prices, the implied revenue requirement rises to $2.5 trillion. And if their enterprise customers in turn need similar margins on their own AI deployments, consumers and businesses would need to pay close to $3 trillion for AI services in the near term. That figure is equivalent to approximately 10% of current US GDP and 5% of global labor costs based on OECD data. Critically, it covers only the cost of the equipment and excludes energy infrastructure, R&D spending, and the broader operational costs of running these systems.
We can also build this calculation from the ground up using current company-level data. The four largest hyperscalers are committing approximately $710 billion in capital expenditure in 2026 alone. At a 10% hurdle rate, that single year of spending requires approximately $71 billion in new annual AI-attributable profit just to break even on 2026 investments. That figure does not include the return required on the $448 billion spent in 2025, the $246 billion spent in 2024, or the $162 billion spent in 2023. The cumulative capital deployed from 2023 through 2026 is approximately $1.6 trillion, implying a requirement of roughly $160 billion in new annual profit at a 10% return threshold.
To appreciate what those numbers mean in practice, consider the following. Over the last twelve months, Microsoft generated $120.7 billion in total net operating profit after tax across its entire business. Alphabet generated $113.8 billion. Meta generated $70 billion. Amazon generated $67.6 billion. These are not AI profits. These are the total profits of four of the most valuable and operationally efficient companies in the history of capitalism, across every product and service they offer. Combined, they generated approximately $372 billion in NOPAT. The $160 billion in new annual profit required simply to achieve a 10% return on the cumulative AI CapEx represents 43% of that combined total, to be generated by AI alone, on top of existing operations.
And 10% is not a number that justifies the valuations these companies currently trade at. To sustain the return on capital that their current market multiples imply, the threshold is closer to 20%. At that level, the requirement doubles: $320 billion in new annual AI-attributable profit, approaching the entire combined NOPAT of all four companies from their existing businesses. The math does not suggest these investments are merely risky. It suggests they are, under any realistic near-term revenue scenario, nearly impossible to justify on a return basis.
Against that requirement, what does the revenue picture actually look like? OpenAI, the most visible AI application in the world, is generating approximately $25 billion in annualized revenue as of early 2026. Anthropic is generating approximately $47 billion in annualized revenue. Enterprise GenAI market spending reached an estimated $37 billion across all providers in 2025. Microsoft’s AI business is running at a $37 billion annual revenue run rate as of fiscal Q3 FY2026, ending March 2026, though this figure includes broader cloud AI services rather than pure AI product revenue.
The bull case here deserves honest treatment. Cloud revenue broadly attributable to AI demand is large and growing fast. If one credits all incremental cloud growth to AI, the addressable revenue base is considerably larger than the pure AI product figures suggest. Google Cloud grew 63% year-on-year in Q1 2026, AWS 28%, and Azure 40%. Under the generous assumption that all of this incremental growth is AI-driven (it’s not), AI-attributable revenue across the major hyperscalers could be argued to reach $50 to $150 billion annually. Dividing the J.P. Morgan $650 billion threshold by the high end of that range gives a monetization gap of approximately 4x. Dividing it by the low end gives approximately 13x. The honest answer is somewhere between those two figures, depending on how much of the cloud growth one is willing to attribute to AI specifically rather than to baseline cloud adoption that would have occurred regardless.
The gap is closing. It is not closing fast enough. To illustrate the growth rate required: closing a gap from roughly $100 billion in AI-attributable revenue today to $650 billion requires compound annual revenue growth of approximately 45% per year, sustained for four consecutive years across an entire industry. Maybe it happens, maybe it doesn’t. It’s a tall bar nonetheless. For context, cloud computing, the last infrastructure buildout of comparable scale, took fifteen years to generate its first $100 billion in annual revenue.
2.2 Low User Conversion
While AI applications like ChatGPT have achieved massive user adoption, conversion to paying subscribers remains stubbornly low. Multiple analyses estimate that only 3% to 5% of all users are currently paying for premium AI services, indicating a broad reluctance among consumers to pay for services widely perceived as free utilities. OpenAI now reports approximately 900 million weekly active users and roughly 50 million paying subscribers, implying a conversion rate of approximately 5.5%, consistent with earlier projections but not materially above them despite years of product development and iteration.
The structural challenge here is not simply that conversion is low today. It is that the economics of the free tier are becoming increasingly difficult to sustain as inference costs scale. The more users engage without paying, the greater the operational loss per user, particularly as agentic AI workflows consume exponentially more compute per session than conversational AI.
2.3 Elusive Enterprise ROI
The narrative of AI transforming corporate productivity has not yet translated into widespread, measurable financial gains. A 2025 MIT NANDA study examining over 300 enterprise AI initiatives found that 95% delivered zero measurable profit and loss impact. This is consistent with earlier McKinsey findings and represents an updated and more comprehensive data point on the same underlying problem.
The granularity of the enterprise revenue problem becomes even clearer when looking beyond broad cloud figures. UBS estimates that total third-party AI product revenue across all listed software companies currently stands at approximately $2.5 billion, of which Microsoft alone accounts for more than 80%. That figure strips away hyperscaler-to-hyperscaler revenue and internal deployments, leaving only genuine external enterprise software AI revenue. It is a small number, and its concentration in a single company suggests that the enterprise adoption story, while real in pockets, has not yet diffused into the broad-based commercial demand that the infrastructure buildout requires.
The micro versus macro paradox from our original analysis persists. Microsoft-sponsored IDC research continues to report impressive firm-level ROI figures of 3.7 to 10.3 times initial spend for successful projects. Yet macroeconomic models from MIT predict AI will increase total factor productivity by less than 0.66% over the next decade. The resolution of this paradox matters enormously for the monetization thesis: if AI productivity gains remain confined to isolated deployments rather than diffusing across the broader economy, the aggregate revenue pool available to justify the CapEx will remain far smaller than the infrastructure buildout assumes.
2.4 The Token Cost Paradox: When Adoption Becomes the Problem
A new and significant dimension of the monetization challenge has emerged since our original analysis, one that cuts to the heart of the gap from an unexpected direction. The assumption embedded in most AI revenue forecasts is that more adoption equals more revenue. The reality is proving more complicated.
The shift from conversational AI to agentic workflows has driven an explosion in token consumption. Agentic applications consume between 10 and 100 times more tokens per session than simple conversational queries. According to Yipit Data, total token consumption by paying users has grown from near zero in late 2024 to approximately 24,000 billion tokens per month by May 2026, a roughly 19-fold increase in eighteen months. At the same time, the price per million tokens has fallen from approximately $1.90 in late 2024 to around $0.65 by May 2026, a decline of roughly 65%.
The bulls cite this dynamic, known as the Jevons Paradox, as proof that the market is working: cheaper AI drives more consumption, which drives more revenue. The Goldman Sachs projection that token consumption by AI agents will multiply 24 times by 2030, reaching 120 quadrillion tokens per month, is the most cited expression of this optimism.
The paradox, however, cuts both ways. Volume is growing, but it is growing at collapsing prices, and it is generating cost shock at the enterprise level that is actively constraining the revenue growth the industry depends on. The most prominent example is Microsoft, which canceled the majority of its internal Claude Code licenses in early 2026 because the scale of employee adoption made costs unmanageable at the enterprise budget level. Uber consumed its entire 2026 AI budget within four months. These are not edge cases. They represent a structural problem: even at current token prices, the volume of consumption generated by agentic AI is producing budget shock at enterprise finance departments, forcing usage caps and rollbacks that directly limit the revenue growth projections of AI providers.
Average monthly enterprise AI spend jumped from $63,000 in 2024 to $85,500 in 2025, a 36% increase, with the share of companies spending over $100,000 per month more than doubling in the same period. If Goldman Sachs is right that token consumption grows 24 times by 2030, and if enterprise budgets continue to hit ceilings at current consumption levels, the industry faces a fundamental question: who pays for that volume, and at what price?
The paradox is precise. The more enterprises adopt AI at scale, the more they hit cost ceilings that force them to pull back, cap usage, or demand lower prices. Lower prices compress provider margins. Compressed margins widen the gap between revenue and the return required on the capital deployed. Adoption, in this environment, is simultaneously the goal and the problem.
2.5 The Commoditization Trap: Eroding Pricing Power at the Model Layer
The token cost paradox is compounded by a structural shift in the competitive dynamics of the model layer itself. AI is commoditizing, and it is doing so at a historically unprecedented rate.
The price of AI API access has fallen approximately 99.7% since GPT-4's launch in March 2023. Tasks that cost $30 per million tokens then cost $0.10 today, performed by models that match or exceed GPT-4's original capabilities. The competitive moat that once justified premium pricing has been almost entirely eroded in just over two years. Gartner has formally identified this dynamic, publishing analysis titled “Navigating the Commoditization Trap" as Token Costs Fall by Over 90% Through 2030.” The competitive moat that once justified premium pricing is eroding faster than most forecasts anticipated.
The switching cost data makes this concrete. According to Ramp spending data, approximately 79% of OpenAI enterprise customers also pay for Anthropic. Enterprises are not choosing between providers; they are hedging across them, which is the opposite of a loyal, locked-in customer base.
Consumer behavior tells the same story. ChatGPT’s share of AI referral traffic fell from 89.2% in Q4 2025 to 81.4% in Q1 2026, while Gemini nearly tripled its share in the same period. BrightEdge’s analysis is unambiguous: “Loyalty is weak, and model quality moves behavior. Users are not locked into one LLM, and they will shift quickly when a model improves or when another one feels more useful.”
The deeper structural issue is that value is migrating away from the model layer entirely. Gartner notes that value will accrue to platforms that can orchestrate workloads across a diverse portfolio of models, not to the model providers themselves. The winners in the application and distribution layer are Microsoft 365, Google Workspace, and Salesforce, all of which have default access to hundreds of millions of workflows regardless of which underlying model performs best this month. The model providers, by contrast, are competing on a treadmill: each improvement is rapidly matched, pricing power continues to erode, and the revenue base needed to justify the infrastructure investment becomes harder to defend with each passing product cycle.
Together, the token cost paradox and the commoditization trap squeeze the monetization gap from both ends simultaneously. Adoption generates runaway costs that force enterprise pullbacks, while the ability to sustain pricing sufficient to cover those costs is structurally weakening. The revenue projections that underpin the bull case for AI CapEx assume neither of these dynamics at the scale we are now observing.
3.0 Financial Engineering and Structural Fragility
In late 2025, a quiet but significant accounting story briefly captured the attention of financial analysts: the depreciation schedules that hyperscalers were applying to their AI infrastructure were materially out of step with the economic reality of the assets being depreciated. The story made noise for a few weeks and then largely disappeared from the conversation. It should not have. The underlying dynamic has not resolved. It has worsened. And it sits alongside two other structural fragilities, circular financing and an accelerating reliance on leverage, that similarly generated concern when they first surfaced and have since been absorbed into the background noise of an industry that moves too fast for inconvenient questions to linger.
3.1 The Obsolescence Gap: Accounting Lifespans vs. Economic Reality
Hyperscalers have converged on a five to six year useful life assumption for their server assets. This accounting standard is directly at odds with the technological reality of the AI chip market, where Nvidia’s twelve to eighteen month product cycle renders prior GPU generations economically obsolete for primary workloads within two years of a successor’s release. The true economic lifespan of a high-performance AI chip is closer to two to three years, long before its value is fully written off the books.
The motive for extending depreciation schedules is straightforward. By spreading the cost of AI servers over a longer period, companies reduce annual depreciation expenses and inflate reported operating income. A single one-year extension in Amazon’s server depreciation schedule added approximately $3.2 billion to its annual operating income. Scaled across the industry, the effect is considerably larger. Michael Burry has estimated that the industry is understating depreciation by approximately $176 billion between 2026 and 2028, enough to inflate combined reported earnings by roughly 20%, with Oracle and Meta among the most exposed, their profits potentially overstated by 27% and 21% respectively by 2028. Amazon has already acknowledged the direction of travel, shortening its server useful life from six years to five in 2025, citing “an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”
This story was discussed when Burry’s analysis surfaced and has since been largely forgotten. It deserves to be revisited, because the problem has materially worsened since the estimate was made. As agentic AI workflows have driven token consumption to approximately 19 times its late 2024 level, data center hardware is being utilized far more intensively than the depreciation models assumed when five and six year useful lives were originally set. The assets are being consumed faster and rendered obsolete faster simultaneously.
Epoch AI’s modeling captures the financial consequence precisely: shortening IT equipment lifespan assumptions from five years to three years raises the annualized total cost of ownership of a one gigawatt data center from $8.5 billion to $12 billion, a 41% increase. Applied across the 30 gigawatts of AI data center capacity that existed globally at the end of 2025, and the far greater capacity being built through 2027, the aggregate understatement of true asset depletion costs is substantially larger than even Burry’s $176 billion estimate implies. The depreciation accounting mirage, far from resolving itself, is growing.
3.2 The GPU Value Cascade as a Counter-Argument
The primary counter-argument employed by hyperscalers to justify longer depreciation periods is the GPU Value Cascade. This concept asserts that GPUs are durable economic assets whose usefulness extends significantly beyond their initial purpose, transitioning progressively from foundational model training in years one and two, to high-end inference in years three and four, to batch processing and analytics in years five and six. CoreWeave cites 95% resale values for older chip generations like the A100 and H100 as evidence that the market itself endorses extended useful lives.
The cascade is real and should not be dismissed. The question is whether the rate of value decline has accelerated beyond what five and six year schedules can absorb. Given Nvidia’s accelerating product cycle, the explosion in token consumption driving hardware utilization far beyond original assumptions, and Amazon’s own decision to shorten its schedule in acknowledgment of exactly this dynamic, the weight of evidence suggests the cascade argument, while structurally sound, is being applied to a depreciation timeline that no longer reflects economic reality.
3.3 Circular Financing and the Return of Vendor Financing
Several months ago, the circular financing structure underpinning the AI ecosystem received considerable attention. Analysts drew comparisons to the vendor financing that amplified the 1990s telecom boom. The conversation has since faded. It should not have, because the structure has not only persisted but expanded significantly in scale.
The mechanics are by now well documented but worth restating clearly. Nvidia has made strategic equity investments in companies like OpenAI. OpenAI purchases both Nvidia chips and compute capacity from CoreWeave, another Nvidia-backed company. Hyperscalers including Microsoft, Amazon, and Google are investing billions in AI startups like OpenAI and Anthropic, with a standard condition of those investments being that the startups commit to running their models on their investors’ cloud infrastructure. The capital flows in a closed loop, and each revolution of that loop generates reported revenue for someone in the chain.
For example, Nvidia’s $100 billion investment commitment in OpenAI alone represents approximately 39% of its current annual revenue, in a relationship where OpenAI is simultaneously one of Nvidia’s largest customers.
The scale of individual transactions within this loop has grown dramatically since that comparison was first made. OpenAI raised $122 billion in its most recent funding round at an $852 billion valuation, with Nvidia contributing $30 billion. A company that is simultaneously one of Nvidia’s largest customers and one of its largest investment recipients is not an independent demand signal. It is a financial structure.
The most striking recent illustration of how deeply embedded this circularity has become is Anthropic. On May 28, 2026, Anthropic closed a $65 billion Series H funding round at a $965 billion post-money valuation, confidentially filing for an IPO the same week. The headline valuation places Anthropic in the top tier of the S&P 500 by market capitalization, ahead of companies like Walmart, JPMorgan, and Visa, each of which has decades of earnings history and hundreds of billions in annual revenue. Anthropic has approximately $47 billion in annualized revenue and has delayed its cash-flow positive target to 2028.
The circular financing embedded in that round is explicit. Of the $65 billion raised, $15 billion consists of previously committed hyperscaler capital, including $5 billion from Amazon. Separately, as a condition of Amazon’s broader $25 billion investment in Anthropic, Anthropic has committed to spending $100 billion on AWS infrastructure. Google has made similar arrangements. A meaningful portion of the capital Anthropic raises is pre-committed back to the same infrastructure providers that funded the raise, meaning the net new independent capital is considerably less than the headline figure suggests.
This is not a red flag unique to Anthropic. It is the defining financial architecture of the entire AI ecosystem. But Anthropic's near-trillion-dollar valuation, applied to a company that has not yet reached cash flow positivity and whose capital raises are substantially recycled back into the infrastructure of its own investors, is the clearest single illustration of how far the current valuation environment has detached from conventional financial logic.
3.4 Increasing Reliance on Leverage
The original article noted early signs of a shift from equity to debt financing in the AI infrastructure buildout. That shift has since accelerated materially, and it has done so against a backdrop of deteriorating free cash flow that makes the trajectory increasingly difficult to defend.
The visible CapEx figures substantially understate the true financial commitment. As noted in Section 1, Moody’s reported in early 2026 that hyperscalers have approximately $662 billion in data center lease commitments signed but not yet commenced, sitting off balance sheet under GAAP’s lease commencement standard. This figure is larger than the combined on-balance-sheet debt of the same companies and represents a shadow liability whose economic reality is no less binding for its accounting treatment. The argument that this buildout is conservatively self-funded from operating cash flows does not survive contact with that number.
The on-balance-sheet debt and equity picture is deteriorating in parallel. Oracle issued an $18 billion bond. CoreWeave secured a $2.6 billion loan and a $1.75 billion bond package. OpenAI and Oracle reportedly engaged in a $100 billion vendor financing arrangement. And on June 1st 2026, Alphabet announced it is raising $80 billion in equity offerings to help fund its AI infrastructure spending plans, comprising $30 billion in underwritten public offerings, a $40 billion at-the-market sale, and a $10 billion private placement with Berkshire Hathaway.
The Berkshire participation deserves honest acknowledgment. Berkshire Hathaway, under new CEO Greg Abel, is one of the most disciplined and patient capital allocators in investment history, and its willingness to commit $10 billion via private placement is not a signal to dismiss. It is a genuine expression of long-term conviction in Alphabet’s franchise from an investor that does not chase momentum. That conviction, however, does not change the structural arithmetic. Alphabet’s 2026 CapEx guidance of $185 billion exceeds its entire 2025 operating cash flow. The $80 billion equity raise is not the financing decision of a company comfortably funding its buildout from operations. It is the financing decision of a company that has looked at its capital commitments against its cash generation and concluded it needs outside capital to bridge the gap. Whatever one thinks of the long-term opportunity, that gap is precisely what this article has been quantifying.
The charts accompanying this section show why debt markets are being tapped with increasing frequency. These are four of the greatest cash-generating businesses in corporate history. Amazon’s operating cash flow has grown from $18.4 billion in 2017 to $148.5 billion in the last twelve months, a 708% increase. Alphabet’s has grown from $37.1 billion to $174.4 billion. Meta’s from $24.2 billion to $124 billion. Microsoft’s from $39.5 billion to $170.1 billion. The operational performance is, by any historical standard, extraordinary.
But capital expenditure has grown faster than operating cash flow at every one of these companies, and the gap is widening. Amazon’s CapEx has grown from $12 billion in 2017 to $151 billion in the last twelve months, now exceeding its operating cash flow entirely and pushing free cash flow to negative $2.5 billion. Alphabet’s CapEx has grown from $13.2 billion to $109.9 billion, compressing free cash flow from a peak of $72.8 billion in 2024 to $64.4 billion today despite dramatically higher revenues. Microsoft’s CapEx has grown from $8.1 billion to $97.2 billion, with free cash flow compressing from a peak of $89 billion in 2022 to $71.6 billion as of June 2025 and continuing lower in the most recent period. Meta is the most resilient of the four, with free cash flow of $48.3 billion in the last twelve months, though even here CapEx has grown more than tenfold from $6.7 billion in 2017 and free cash flow has compressed from its 2024 peak of $54.1 billion.
The pattern is consistent across all four companies: operating cash generation has been exceptional, but CapEx growth has been so dramatic that it is consuming an ever-larger share of that cash. Free cash flow is either flat, compressing, or in Amazon’s case, already negative. And 2026 CapEx guidance is materially higher still across all four, meaning the pressure on free cash flow will intensify before it eases. This is not a story of companies in distress. It is a story of capital deployment running structurally ahead of the cash it generates, with the gap being funded by debt markets rather than the AI revenue that was supposed to justify the spending in the first place.
The late-cycle financing dynamic at work involves funding shorter-lived assets, AI chips with a two to three-year economic life, with long-duration financial obligations and increasing leverage. The mismatch between asset life and liability duration is a fragility that tends not to announce itself gradually.
4.0 The Energy Constraint: From Narrative to Numbers
The energy demands of the AI infrastructure buildout have received considerable attention in the press, largely framed as an environmental or grid reliability concern. That framing, while valid, understates the more immediate financial problem. Energy is not simply a constraint on AI’s growth. At the scale now being built, it is a permanent and rapidly growing operating cost that must be covered before a single dollar of profit is possible, and the numbers involved are large enough to materially affect the return calculations in Section 2.
4.1 The Scale of Power Demand
AI data centers are among the most energy-intensive facilities ever built. A single modern AI data center consumes as much electricity as approximately 100,000 average households. At a broader scale, data centers in Virginia, the largest data center market in the world, already account for approximately 26% of the state’s total electricity consumption, a figure that is rising every year as new capacity comes online.
Globally, AI data center power capacity reached approximately 30 gigawatts at the end of 2025. The buildout currently underway is targeting a dramatic expansion of that base. Stargate alone is targeting approximately 10 gigawatts of capacity by 2029. The broader US buildout across all hyperscalers is targeting between 20 and 30 gigawatts of additional capacity through 2028. By conservative estimates, total global AI data center power capacity will reach between 50 and 100 gigawatts before the end of the decade.
4.2 Translating Power Into Cost
The financial consequence of this power demand becomes clear when expressed in operating cost terms rather than megawatt figures. Epoch AI’s modeling of a one gigawatt data center provides the most rigorous publicly available cost structure. Under a five year IT equipment lifespan assumption, the annualized total cost of ownership of a 1 gigawatt facility is approximately $8.5 billion per year, of which approximately $0.9 billion represents annual operating expenses including energy. Under a more realistic 3 year lifespan assumption, that annualized total cost rises to $12 billion, a 41% increase, as discussed in Section 3.1.
Applying the $0.9 billion annual operating cost figure to the current 30 gigawatt installed base implies approximately $27 billion per year in energy and operational costs simply to keep existing capacity running, before any capital recovery. As capacity scales toward the 50 to 100 gigawatt range through 2028 and 2029, annual operating costs reach between $45 and $90 billion per year. These are permanent, recurring costs that grow with the buildout and must be covered in full before the revenue generated by that infrastructure contributes anything toward recovering the capital invested in building it.
To put those figures in the context of Section 2’s monetization analysis: even under the most generous assumption that AI-attributable revenue reaches $150 billion annually by 2027, between $45 and $90 billion of that would be consumed by energy and operational costs alone, before depreciation, financing costs, or any return on the $1.6 trillion in cumulative CapEx is considered. The energy cost is not a footnote to the return calculation. It is a first charge against every dollar of revenue the infrastructure generates.
4.3 The Grid Reality
The power demands of the planned buildout are beginning to run into physical constraints that no amount of capital can quickly resolve. Electricity grids are not built in months. Permitting, transmission infrastructure, and generation capacity additions operate on timelines measured in years, often five to ten years from approval to operation in the United States.
The consequence is a growing gap between the power that AI data centers require and the power that grids can reliably deliver on the timelines the buildout assumes. In several major data center markets, including parts of Virginia, Texas, and the Pacific Northwest, grid operators have already flagged capacity constraints that are affecting new data center connections. Some hyperscalers have responded by pursuing dedicated power agreements with nuclear operators, investing in small modular reactors, and signing long-term renewable energy contracts. These are rational responses to a real constraint, but they add further capital commitments and operational complexity to an ecosystem that is already financially stretched.
4.4 The Efficiency Counter-Argument
The bull case on energy rests on efficiency gains, and it deserves honest treatment. The cost of AI inference per unit of output has fallen dramatically and continues to fall. Newer GPU architectures deliver significantly more compute per watt than their predecessors, and software-level optimizations are compressing the token cost of any given task. The DeepSeek R1 model, released in January 2025, demonstrated that near-frontier reasoning performance could be achieved at a fraction of the training cost previously assumed, suggesting that the energy intensity of AI per unit of useful output may fall faster than the buildout implies.
The Jevons Paradox, however, applies here as directly as it does to token pricing. Historical precedent across every major efficiency-improving technology suggests that lower cost per unit of output drives total consumption higher, not lower. Aviation fuel efficiency per passenger mile has improved dramatically over fifty years; total aviation fuel consumption has risen every decade. The same dynamic is already visible in AI: as inference costs fall, token consumption is rising faster, not slower. If the Goldman Sachs projection of a 24 times increase in token consumption by 2030 is even approximately correct, efficiency gains will be more than offset by volume growth, and the energy cost trajectory will continue upward regardless of improvements in performance per watt.
5.0 “Valuations Are Not that High”. Yet.
Three companies representing the frontier of the AI infrastructure buildout are planning to go public in the second half of 2026: SpaceX in June, OpenAI targeting September, and Anthropic targeting October. Combined, they are seeking valuations of approximately $3.5 to $3.7 trillion. These listings will serve as the first genuine public market stress test of the assumptions embedded in this article.
The IPOs matter for three distinct reasons.
First, they will force disclosure of true cost structures that private markets have long deferred. OpenAI is preparing to ask public investors to value the company above $1 trillion while projecting a $14 billion operating loss in 2026, with no path to profitability before 2029 or 2030. Its computing expenditure alone is projected to reach $121 billion in 2028, with a projected loss of $74 billion in that year. To contextualise the scale of this burn: Amazon accumulated approximately $3 billion in cumulative losses over its first six years before turning its first annual profit in 2003. OpenAI is on course to accumulate hundreds of billions in losses before reaching positive cash flow, a difference in scale of approximately 100 times. HSBC analysts have estimated a $207 billion funding shortfall against OpenAI’s stated growth plans. Public markets, with their quarterly earnings cadence and institutional scrutiny, are a considerably less forgiving environment for that kind of structural gap than the sovereign wealth funds and private equity firms that have funded the AI buildout to date.
Anthropic’s IPO story is materially different, and honesty requires acknowledging it. The company disclosed that it projects $10.9 billion in revenue for Q2 2026, more than doubling Q1’s $4.8 billion, and expects its first ever operating profit of $559 million for that period, a milestone that arrived two full years ahead of what Anthropic had told investors only the previous summer. By 2028, Anthropic projects $17 billion in positive cash flow on $70 billion in revenue, with gross margins approaching 77%. The divergence from OpenAI traces to client mix: approximately 85% of Anthropic’s revenue comes from enterprise and developer customers, compared to OpenAI’s roughly 85% dependence on consumer ChatGPT subscriptions, of which approximately 95% of users pay nothing. Enterprise customers generate three to five times more revenue per token, their query patterns are cheaper to serve, and their contracts are stickier. The important qualification is that Anthropic’s Q2 operating profit figure excludes stock-based compensation, which at a company that has raised $134 billion in private capital could be significant enough to erase the margin on a GAAP basis.
Second, the filings will expose the true scale of circular financing. Anthropic’s Series H round of $65 billion includes $15 billion previously committed by Amazon and other hyperscalers. Separately, as a condition of Amazon’s broader compute commitment, Anthropic has agreed to consume up to 5 gigawatts of AWS capacity, while also sourcing 5 gigawatts of TPU chips from Google and Broadcom, and signing GPU capacity agreements with SpaceX’s Colossus data centers. When these arrangements become line items in a public prospectus, the narrative that these companies are generating independent organic revenue becomes considerably harder to sustain.
Third, the listings will test whether public markets will price these companies at valuations that fundamental analysis cannot support. The math is straightforward. The numbers are not.
SpaceX is targeting a valuation of $1.75 to $1.8 trillion against 2025 revenue of $18.7 billion. That implies a price-to-revenue multiple of approximately 95 times. As for a price-to-earnings multiple, there is no earnings to speak of: the company posted a GAAP net loss of $4.94 billion in 2025, having been profitable the prior year before absorbing xAI’s losses. The AI segment alone lost $6.4 billion from operations on $3.2 billion in revenue. Goldman Sachs estimates that sustaining a $1.75 trillion valuation through 2030 would require annual revenues exceeding $100 billion, implying a compound annual growth rate above 40% from an $18.7 billion base, sustained for four consecutive years, while simultaneously closing a multi-billion dollar annual operating loss. Morningstar, the first major Wall Street institution to publish a formal valuation, assigned SpaceX a fair value of $780 billion, less than half the IPO target.
OpenAI is targeting a valuation above $1 trillion against $25 billion in annualized revenue, implying a price-to-revenue multiple of approximately 40 times. A price-to-earnings multiple is… not available. The company is projecting a $14 billion operating loss in 2026 at a non-GAAP adjusted operating margin of negative 122%, meaning it loses $1.22 for every dollar it earns. By the way, this is before stock-based compensation, which is excluded from the non-GAAP figure. Gross margins are constrained at approximately 33% by inference compute costs.
While OpenAI’s revenue growth has been remarkable, there are early signs of deceleration precisely as it approaches the public markets: ChatGPT’s weekly active users stalled at approximately 905 million on average in Q1 2026, well short of the 1 billion target OpenAI had set for 2025, and the company has already lost meaningful referral share to Gemini and Claude. Portfolio managers at Gabelli Funds have noted publicly that OpenAI’s growth appears to have slowed in late 2025 into early 2026 as it ceded share to Anthropic and Google. To reach $1 trillion in valuation at a more “conventional” (being generous) 10 times revenue multiple, OpenAI would need to reach $100 billion in annual revenue. Its own internal projections put $280 billion in revenue by 2030, requiring an 83% compound annual growth rate from today’s base, in a market where its pricing power is collapsing and its largest consumer base is growing more slowly than projected.
Anthropic presents the most credible near-term financial case of the three, and that too deserves honest treatment. At $47 billion in annualized revenue and approaching its first operating profit of $559 million in Q2 2026, it is at least heading in the right direction. By 2028, it projects $70 billion in revenue and $17 billion in positive cash flow, with gross margins approaching 77%. At those 2028 figures, the implied cash flow multiple on today’s $965 billion valuation would be approximately 57 times. A 57x 2028 cash flow multiple is still a demanding multiple, and it remains entirely dependent on growth rates staying extremely high well into the decade. Also, the Q2 operating profit figure excludes stock-based compensation, and $15 billion of its latest funding round is pre-committed back to hyperscaler cloud spend, meaning the true independent capital raised is considerably less than the headline suggests. Anthropic is the best house on this particular street. But at 57 times projected 2028 cash flow, it is still a very expensive street.
Taken together, the three companies are asking public markets to assign a combined valuation of approximately $3.5 trillion to businesses generating a combined $90.7B billion in annualized revenue, two of which are unprofitable and one of which is marginally profitable on a non-GAAP basis before stock compensation. That is a blended price-to-revenue multiple of approximately 38.5x times, against an S&P 500 that as a whole trades at approximately 3 times revenue. The premium embedded in these valuations is a direct expression of the growth assumptions this article has questioned throughout: that AI revenue will compound at rates between 40% and 83% annually for the rest of this decade, in a market where pricing power is collapsing, enterprise budgets are already hitting ceilings, and the return on capital deployed in the infrastructure buildout remains far below what conventional investment thresholds require.
6.0 Conclusion: When the Numbers Don’t Add Up
Each section of this article has examined a different dimension of the same underlying problem. This conclusion does what the original article did not: it puts all of those dimensions together into a single, cumulative picture. The purpose is not to alarm. It is to be precise.
Start with the capital committed. Between 2023 and 2026, the four largest hyperscalers will have deployed approximately $1.6 trillion in AI infrastructure. By 2027, annual spending alone will exceed $1 trillion. Add Stargate’s $500 billion commitment, Oracle’s standalone buildout, and secondary players, and total committed infrastructure spend through 2029 is plausibly $3 to $4 trillion. Add the $662 billion in off-balance-sheet lease commitments that Moody’s identified sitting outside the figures analysts typically examine, and the true economic liability of the current buildout is larger still.
The return required on that capital is not ambiguous. At a 10% hurdle rate, the cumulative $1.6 trillion deployed through 2026 requires approximately $160 billion in new annual AI-attributable profit. At the 20% return on capital that current market valuations imply, that figure doubles to $320 billion. For context, the combined total net operating profit after tax of Microsoft, Alphabet, Meta and Amazon across their entire businesses in the last twelve months was approximately $372 billion. The AI infrastructure alone, at 20% returns, would need to generate profit approaching that combined total, on top of everything these businesses already do.
The revenue reality sits far below that requirement. Three independent analyses converge on a requirement of between $650 billion and $3 trillion in annual AI revenue to justify the buildout, depending on the methodology and assumptions used. Current AI-attributable revenue, even under generous assumptions that credit all incremental cloud growth to AI, is between $50 and $150 billion annually. The gap is between 4x and 20x, depending on where in that range one anchors. The gap is closing. The rate at which it is closing does not match the rate at which capital is being deployed.
Now add the costs that reported earnings do not fully capture. Depreciation schedules set at five to six years on assets with a two to three year economic life are understating true asset depletion by an estimated $176 billion between 2026 and 2028, a figure that is growing as token consumption accelerates hardware utilization beyond the assumptions those schedules were built on. Energy and operational costs on the current 30 gigawatt installed base run approximately $27 billion per year and will reach between $45 and $90 billion annually as capacity scales toward 2029. These are first charges against revenue, not optional costs, and they are not fully reflected in the return calculations that justify current valuations.
Underlying all of this are structural fragilities that have received attention in waves and then been largely forgotten. Free cash flow across the four hyperscalers, despite extraordinary operating cash generation, is being consumed by CapEx at a rate that is pushing Amazon into negative territory and compressing Alphabet, Microsoft and Meta simultaneously. Alphabet has now confirmed the dynamic explicitly, announcing on June 1st 2026 an $80 billion equity raise to fund its AI infrastructure commitments. The circular financing structures have not resolved. They have expanded, with OpenAI's $122 billion raise, Anthropic's $65 billion Series H at a near-trillion-dollar valuation, and Nvidia's deepening investment in its own largest customers all illustrating a financial architecture in which a meaningful portion of reported revenue is recycled capital rather than independent organic demand.
The clearest test of whether any of this is sustainable is now imminent: SpaceX, OpenAI and Anthropic are targeting a combined $3.5 trillion in public market valuations in the second half of 2026, against a combined $90.7 billion in annualized revenue, two of which are unprofitable. When S-1 filings replace private market enthusiasm with quarterly earnings scrutiny, the assumptions this article has questioned will face their first genuinely public reckoning.
Put it all together and the picture is this. The industry has committed between $3 and $4 trillion to infrastructure whose annualized return requirement, at valuations current markets are assigning, is approximately $600 to $800 billion in new annual profit. Current AI-attributable revenue is roughly $50 to $150 billion, before energy costs, before realistic depreciation, and before stripping out the circular financing that inflates the headline figures. The compound annual growth rate required to close that gap by 2030 is approximately 45%, sustained across an entire industry, in an environment where API prices for frontier-equivalent models have declined from $30 per million tokens in 2023 to $0.10 today, and enterprise budget ceilings are already being hit at current consumption levels.
We are optimistic about AI as a technology. The productivity potential is real. Some of the revenue growth is real and accelerating. Berkshire Hathaway’s decision to commit $10 billion to Alphabet’s infrastructure raise is not a signal to ignore, and the cloud revenue growth numbers from Q1 2026 are genuinely strong. The bull case is not without foundation.
But a bull case with foundation is not the same as a bull case whose numbers work. The defining characteristic of every major capital misallocation cycle in history has not been that the underlying technology failed. It has been that the capital deployed ran so far ahead of the revenue it could plausibly generate, on any reasonable timeline, that the investment cycle became self-referential, sustained by momentum and circular financing rather than returns, until it could not be. The railroad boom of the 1870s, the electrification bubble of the 1920s, the fiber optic overbuild of the late 1990s: in each case the technology was real, the long-term economic value was real, and the capital cycle was still a bubble.
The math is not a prediction. Markets can remain disconnected from arithmetic for longer than anyone expects. But the arithmetic is the arithmetic. And right now, it does not add up.
This publication is for informational and educational purposes only and does not constitute investment, financial, legal, or tax advice. Investing involves risk, including the possible loss of principal. You should consult with a qualified professional before making any investment decisions. The author may hold positions in the securities discussed.
Some interesting points but lots of potential pushback...
Everyone expects capex to be refunded short term. That's not how investments or ROI works. It'll take years, decades. Google doesn't care about 2027 it cares about 2035.
Second, most assumptions are just that, assumptions. All of those made three years ago ended up being wrong. Those made last year were wrong.
Why would I pay more attention to bearish views on CapEx today when the last ones were damn wrong?
There's clear data. And unclear opinions.
One mistake you are making re: spending and revenue assumptions is based on timelines. Sure, projected spend vs current revenue looks like a huge gap. If you were to actually think like a business, however, you would inverse this: current revenue based on the historical spend.
I.E. current revenue and GW monetization is based on investments made in the past. Back out spending 2-3 years ago and look at revenue being generated today based on those investments. Do that math, and then you will understand why there is so much spending.