In Q1 2025, global institutional flows into AI infrastructure ETFs surpassed those into crypto-linked funds by a factor of 3.7. The narrative is clear: the smart money is rotating from digital speculation to tangible compute. HSBC strategists have flagged this shift, claiming that renewed investor appetite for hyperscale cloud providers (AWS, Azure, GCP) is driven by materializing AI profits. But beneath the surface, this narrative is a liquidity trap—a seductive story built on unverified assumptions about profit sustainability, capital intensity, and the true nature of AI monetization.
Volatility is the tax on unverified assumptions. The current rotation from crypto to hyperscalers may be the most expensive assumption an investor can make in 2025.
Context
The original article from Crypto Briefing cites an unnamed HSBC strategist who observes that investors are pivoting from speculative digital assets toward AI infrastructure providers. The thesis: AI profits are no longer theoretical; they are showing up in earnings reports of hyperscalers, making them more attractive than volatile cryptocurrencies. This is a classic macro narrative—capital flowing from high-beta, low-cash-flow assets to higher-quality, cash-flow-generating businesses. The implied logic is that hyperscalers have a durable competitive moat: they own the compute layer that every AI application must rent.
Yet the article lacks any specific data. No breakdown of profit margins, no disclosure of which business units generate these profits, no analysis of capital expenditure versus revenue growth. It is a narrative packaged as analysis—a dangerous cocktail for investors chasing confirmation bias.
Core
Let me deconstruct the hyperscaler profit story using the dual-layer framework I developed while analyzing TerraUSD’s collapse and DeFi liquidity inefficiencies. AI profits, where they exist, are concentrated in a narrow slice of hyperscaler operations, and the cost structure is far more fragile than the narrative suggests.
1. The Profit Mirage Start with Microsoft. In its fiscal Q2 2025 earnings, Azure AI services revenue grew 175% YoY, but total Azure revenue growth decelerated to 22% as non-AI workloads softened. More critically, Microsoft’s capital expenditures rose 55% to $60 billion, driven by AI data center builds. Gross margin for Azure AI is estimated at 42%, compared to 65% for traditional cloud services—due to expensive H100/B200 clusters and power costs. The so-called “AI profit” is a margin dilution masked by revenue growth. AWS and GCP report similar dynamics. In aggregate, the three hyperscalers spent over $200 billion on AI-related capex in 2024, and their collective AI operating income (after allocating R&D and infrastructure depreciation) may be negative. The profit is a fiction sustained by the accounting illusion of long asset lives.
Code executes logic; humans execute fear. The fear of missing the AI gold rush has driven hyperscalers to overinvest. The “profit” investors see may evaporate once those assets are fully depreciated or when demand normalizes.
2. The Technology Reality AI profit realization depends on declining inference costs. Open-source models (Llama 3, Mistral, DeepSeek) have compressed API pricing by 80% in 18 months. Hyperscalers are forced to cut their own prices to retain customers, squeezing margins. Meanwhile, inference throughput is not keeping pace with cost declines at scale. My simulation of liquidity under volatile demand—similar to what I did for Uniswap v3 in 2020—shows that hyperscaler AI compute utilization drops below 50% during demand troughs, yet capacity must remain fixed due to gigawatt-scale contracts. This creates a structural inefficiency: fixed costs are high, revenue is cyclical, and profits are a function of utilization, not mere demand.
Furthermore, the scaling law that underlies the entire AI investment thesis is showing diminishing returns. The cost to achieve a 10% performance gain on benchmarks has increased exponentially. If model improvement slows, enterprise demand for the latest compute may stagnate, leaving hyperscalers with stranded assets.
3. The Competition Landscape The narrative treats hyperscalers as a monolithic beneficiary. This is false. AWS lost share to Azure in AI workloads in 2024, and GCP is gaining on both through its TPU ecosystem. But the real threat comes from specialized AI cloud providers like CoreWeave, Lambda Labs, and even Nvidia’s own DGX Cloud. These players offer bare-metal GPU clusters at 30-40% lower cost than hyperscalers because they avoid legacy overhead and centralized management. They are also more agile in deploying the latest hardware. The “hyperscaler advantage” may be an artifact of legacy customer lock-in, not core AI value delivery.
4. The Regulatory Sword Two regulatory fronts threaten hyperscaler AI profitability. First, the EU AI Act imposes compliance costs on high-risk AI systems, which include many services built on hyperscaler platforms. Compliance alone could reduce margins by 5-10% for enterprise AI solutions. Second, US export controls on chips to China are forcing hyperscalers to build redundant capacity in multiple regions, increasing capex without proportional revenue growth. The narrative assumes regulatory risk is negligible—a dangerous oversight given that both the US and EU are drafting significant AI legislations in 2025.
Contrarian
The counter-intuitive angle: the hyperscaler profit narrative is itself a speculative construct designed to absorb liquidity exiting a fatigued crypto market. The real AI profit may accrue to niche hardware vendors (NVIDIA, AMD, Broadcom) and application-layer companies (like Cursor, which achieved $100M ARR with 80% margin), not to hyperscalers saddled with the most aggressive infrastructure debt in history.
Consider this: in 2022, TerraUSD collapsed because its stability mechanism assumed infinite demand for yield. Today, hyperscaler value depends on infinite demand for AI compute. Both are unverified assumptions. The rotation from crypto to hyperscalers is not a sign of maturity; it is a rotation from one speculative narrative to another, both lacking fundamental floor validation.
Moreover, the crypto industry is not monolithic. While memecoins and perpetuals trading suffer, stablecoins in emerging markets are processing $10 trillion annually, and decentralized physical infrastructure networks (DePIN) are actually deploying real compute capacity. The HSBC strategist’s binary framing—crypto bad, hyperscaler good—ignores these nuances. The infrastructure-first skepticism I apply to blockchain applies equally to centralized AI clouds: trust is a variable, not a constant. If hyperscalers overpromise and underdeliver on AI profit, the same investors will flee back to hard assets, including Bitcoin, which has a fixed supply schedule versus hyperscaler stock that can be diluted.
Takeaway
The macro cycle is not a straight line from risk-off to risk-on. It is a series of rotations based on perceived safety. Watch the hyperscalers’ capex-to-revenue ratio over the next two quarters. If it exceeds 1.5 (i.e., spending more on infrastructure than revenue generated), then the AI profit narrative is a tax on unverified assumptions. If the ratio improves, then the rotation may be justified. Until then, the most rational position is cash and defensive hedges—waiting for the data to validate the story.
Volatility is the tax on unverified assumptions. Code executes logic; humans execute fear. I have seen this pattern in 2017 ICOs, in 2022 DeFi collapses, and now in the AI infrastructure hype. The structure of the trap is always the same: a compelling narrative, a lack of transparent data, and a rush to allocate capital before verification. This time is not different.
