Three data points arrived within 24 hours of each other last week, and together they describe the same thing: the spending on AI infrastructure is not slowing down. It is accelerating. TSMC posted record first-quarter revenue. CoreWeave, already flush from a $21 billion Meta commitment, signed a multibillion-dollar follow-on deal with Anthropic. And Bloomberg ran a headline that would have seemed surreal two years ago: "TSMC's Sales Beat Estimates After War Fails to Dent AI Demand."
Start with TSMC. The Taiwanese chipmaker reported Q1 2026 revenue of NT$1.13 trillion (about US$35.6 billion), a 35% year-on-year jump and ahead of analyst forecasts. March alone was up 45%. The full quarterly earnings report is due April 16, but the revenue number tells the structural story: despite a global trade environment full of tariff risk, supply chain pressure, and active military conflict in nearby waters, demand for advanced semiconductors from AI customers has kept the top line on a steep upward curve. Analyst Sravan Kundojjala from SemiAnalysis said the company would "easily exceed its 30% annual growth target." TSMC has also reportedly raised prices on its most advanced nodes, which has become a significant earnings tailwind in itself.
The same week, CoreWeave announced a multibillion-dollar capacity agreement with Anthropic. CoreWeave, which went public earlier this year and carries substantial debt from its rapid infrastructure build-out, had just disclosed a $21 billion expansion of its Meta relationship the day before. CEO Michael Intrator described the Anthropic deal as a "phased infrastructure roll-out" with expansion potential, and noted that Anthropic's annual revenue run rate had crossed $30 billion, up from $9 billion at the end of 2025. That growth rate, roughly tripling in roughly four months, explains why Anthropic needs more compute capacity than it currently has, and why CoreWeave was in a position to negotiate terms.
The two stories reinforce each other in a way that matters for understanding where the industry is. CoreWeave's business model depends on leasing GPU capacity to AI labs that need more compute than they can build themselves, which means its health is a direct proxy for how aggressively those labs are scaling. When Meta commits $21 billion and Anthropic follows the next day, it suggests that both companies see near-term demand that justifies locking in capacity at current prices rather than waiting for newer, cheaper hardware generations. That is a strong signal about expected workloads, not a hedge.
The TSMC result adds context. The chips CoreWeave is operating mostly come from TSMC fabs; the demand that drives CoreWeave's revenues ultimately flows back to TSMC's order books. When both ends of that chain are posting record numbers simultaneously, it is harder to argue the current build-out is speculative. The investment is meeting real usage. Anthropic's $30 billion run rate is not a valuation; it is revenue. Somewhere, models are running on compute that is being billed for at a rate that is generating that number.
What is less clear is the demand floor. The infrastructure investment being made now assumes continued growth in AI workloads over a multi-year horizon. The CoreWeave deals are long-term capacity commitments, not month-to-month rentals. If the adoption curve for enterprise AI flattens, or if a next-generation architecture arrives that dramatically reduces compute requirements per unit of output (as happened with DeepSeek's efficiency improvements in early 2025), the companies that locked in multi-year capacity agreements could find themselves overextended. CoreWeave's CEO acknowledged this dynamic, saying simply that "scaling is expensive."
For now, the bet is clearly on continued growth. TSMC is raising prices, CoreWeave is signing billion-dollar deals, and Anthropic just tripled its revenue in under six months. Whether the infrastructure being built today serves the applications of 2027 or becomes the next cautionary tale about overcapacity depends on questions that no one at any of these companies can answer yet. What they can do is point to the numbers, and the numbers this week were large.