Anthropic’s listing metrics dismantle the bubble hypothesis: a dollar of computing generates two dollars of revenue
Listen to this article
Read by Anchor
The talk of an imminent AI bubble burst has become commonplace among observers, yet a review of financial statements and capital obligations tells a completely different story. While skeptics claim that valuations of the mega-cap companies that have crossed the trillion-dollar threshold are on the brink of a market-shaking collapse, an analysis of spending structure shows that the massive cash flows represent foundational costs for building data centres rather than recurring operating expenses. Once these computing facilities are completed, capital outlays drop sharply, opening the way for operating profitability that demonstrates the inference metrics’ ability to cover their financial commitments.
Anthropic’s filing shows that every dollar spent on computing infrastructure generates $2.08 of revenue, a ratio that clearly indicates that running models yields lucrative returns once the initial setup costs are excluded.The company’s revenues have exceeded the combined cost of computing and infrastructure investmentsits revenue jumped from $41 million in 2023 to $386 million in 2024, then reached $4.585 billion in 2025, and hit $16.3 billion in the first half of 2026 alone. This upward curve demonstrates that expanding software consumption is effectively underpinned by firms’ ability to curb immediate costs by suspending free-tier accounts or slowing costly research when needed.
Cumulative liabilities and spending in the AI sector amount to roughly $5 trillion, of which $1.9 trillion has already been spent and $3 trillion remain as announced commitments. Although settling obligations of this magnitude represents a long-term burden, spreading the amount over twenty years equals about $250 billion per year for the entire industry, a figure comparable to Alphabet’s 2019 revenue of approximately $162 billion (about $213 billion after inflation adjustment). A substantial portion of these commitments was issued as convertible bonds, and Anthropic recorded $72 billion of book-losses in favour of lenders in the first half of 2026 as its valuation rose to roughly $1 trillion, meaning that sufficient equity exists to satisfy creditor claims without draining cash reserves.
Unlike the dot-com bubble, which featured unprofitable firms with price-to-earnings multiples in the hundreds, major players such as Alphabet, Meta and Nvidia maintain moderate profitability multiples in the twenties.Revenue growth is driven by actual product consumption rather than server rentals aloneIn Alphabet’s results, Google Services revenue grew by $37.8 billion at a 12 percent rate, while cloud computing revenue rose by $15.5 billion at 36 percent, showing that absolute service growth outpaces infrastructure growth. The same pattern appears at Meta, which does not rent data centres to third parties but still posted an annual revenue increase of about $30 billion for two consecutive years, powered by embedding the technology in its core products.
Fears of losing a solid competitive moat are also receding; despite the leaked Google memo in 2023 that acknowledged the absence of a permanent defensive moat for closed-source models versus open-source alternatives, market reality shows that typical users do not migrate to self-hosting models because of complexity and cost, preferring easier-to-use interfaces.Institutional lock-in and long-term contracts bind companies to the same providerlarge-scale systems are tied to pre-purchase agreements for specified volumes of compute tokens, making a switch to a competing platform costly and procedurally complex.
This financial reality directly influences technology decision-makers in the Gulf, Egypt and the broader Arab region. Rather than postponing investment while waiting for a presumed collapse in global model pricing, tech leaders at banks and regional telecoms need to recognise that compute inference is sold with stable commercial margins and that international providers aim to lock in customers through long-term supply contracts. Accurately calculate the cost of consuming compute tokens before signing, avoid committing to volume agreements that tie up your operating budget for years, and focus on retaining data sovereignty and ensuring software-stack flexibility to lower migration costs if pricing terms change.
These numbers do not eliminate risk for the sector; opposition to data-centre energy consumption, potential regulatory hurdles and reliance on complex financing mechanisms remain challenges that could slow growth if expansion stalls. Yet the assessment of AI-sector sustainability is now linked to the operational ability to service debt and expand revenue-generating applications, not to bubble-theory narratives or rapid-collapse scenarios.