Urgent.News

600+ sources. One page. See who else covered it.

Editions

AI

[Today’s Signal] $3.5 Trillion Is Pouring Into AI, but Wall Street Is Betting Against It

Artificial intelligence is producing two seemingly contradictory signals. Trillions of dollars are flowing into data centers, GPUs and power infrastructure, while some of Wall Street’s most sophisticated investors are increasing bets against companies building that capacity. CoreWeave captures the t

Artificial intelligence is fueling two contrasting trends: a massive influx of capital into data centers, graphics processing units (GPUs), and power infrastructure, while some Wall Street investors are betting against companies driving that expansion. CoreWeave exemplifies this paradox: its revenue backlog shot up to $104.2 billion in Q2, yet the Nvidia-backed AI cloud provider was among the most-shorted stocks in July according to Hazeltree research.

This divergence doesn't necessarily signal Wall Street turning against AI, but rather investors distinguishing between firms that can convert infrastructure spending into lasting cash flow and those whose capital demands outpace returns. According to Morgan Stanley, global data center construction spending is forecasted to reach $2.9 trillion between 2025-2028, totalling roughly $3.5 trillion when considering chips, power, and related infrastructure.

The AI narrative is evolving from conventional software investment cycles to a capital-intensive industrial buildout measured in trillions of dollars. CoreWeave's Q2 revenue surged 112% to $2.58 billion, with a backlog of $104.2 billion. However, expansion comes at a steep cost. The company's Q2 interest expense hit around $640 million, and it raised its 2026 capital expenditure outlook to $39 billion.

Investors are therefore faced with the twin challenge of demand and financing costs. Hedging firms increased CoreWeave, Nebius, and Super Micro Computer short positions in July while reducing long exposure in Nvidia, despite their overall Nvidia position remaining net long. The story is more complex than a blanket rejection of AI. Investors are now selective about where AI spending can yield sustainable returns.

The industry's financing hurdle is set to rise. Morgan Stanley estimates $1.4 trillion of data center construction could be funded through hyperscaler cash flows, $200 billion via corporate debt, and $150 billion via securitized credit. This leaves an $800 billion window for private credit and asset-based financing. Power is another constraint, with Morgan Stanley projecting U.S. data center demand to hit 74 GW by 2028, with available power potentially falling short by 49 GW.

Thus, AI infrastructure success hinges on access to capital, GPUs, electricity costs that preserve returns, and the ability to convert contracts into free cash flow. This reshapes how the AI boom is judged. Companies with the most GPUs or largest data centers may not create the most value. Key metrics like utilization rates, power costs, financing expenses, and converting long-term contracts into free cash flow will matter increasingly.

In South Korea, continued AI infrastructure investment supports demand for SK hynix and Samsung Electronics' high-bandwidth memory and power equipment, as well as opportunities for semiconductor suppliers. However, as global investors shift from rewarding capacity expansion to demanding returns, customers will exert greater pressure on the cost and efficiency of the entire AI supply chain.

CoreWeave's $104.2 billion backlog and rising short positions reflect this transition. Wall Street is betting not that AI will vanish, but that not every AI beneficiary will generate adequate returns on its capital investment.

Written by urgent.news from Korea IT Times's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.

Read the original at koreaittimes.com →

More in AI

The Browsing LLM's Frustrating Limitations

直接在瀏覽器跑 LLM,既兼顧隱私又不用複雜的GPU設定,太完美了吧? 從 WebLLM 到 Transformers.js,前端社群有一群反骨仔吹起一股「邊緣 LLM」的熱潮,可是,當真正將模型落地到使用者的瀏覽器時,第一個面對的考驗就是,WebGPU 真的有比 WASM 快嗎? 這陣子實測的結論比想像中更加戲劇化, 500M 以下的微型模型,WASM 反而快了 12%,但是對於 3B…

  • Browser-based LLMs face performance limits
  • WASM outperforms WebGL/WebGPU for models <500 million parameters
  • Chrome crashes with models >3 billion parameters due to lack of physical head

The Agentic Economy Needs a Market for Work

Most AI agents still live inside a chat window. They can write code, search for information, call an API, or prepare a document, but they usually stop when the task leaves the boundaries of their own…

  • Agents can hire other agents or pay humans for tasks beyond their capabilities
  • Agentic economy enables marketplace for task collaboration and payment
  • Smart contracts ensure immutable financial transactions and rules

Learn Provider-Agnostic Model Routing by Building a Tiny LLM Switchboard

Every few weeks a new model drops and my study group chat fills up with screenshots: "this one is cheaper," "this one is better at code," "switch now." I can never verify any of it quickly, because my…

  • Python script enables provider-agnostic routing between multiple LLM providers
  • Switchboard routes tasks to different backends based on task type (summarize, code, chat)
  • Script adds single entry to adapt to new models, maintains evaluation criteria