CoreWeave revela el cuello de botella de la IA: los límites físicos frenan la entrega de hardware

La carrera por la supremacía en inteligencia artificial se topa con un muro de hormigón. No es la falta de algoritmos brillantes o de capital de riesgo—es la cruda realidad de la física.

El cuello de botella es físico, no digital

CoreWeave, un jugador clave en la infraestructura en la nube para IA, señala lo obvio que muchos en Silicon Valley prefieren ignorar. Fabricar y entregar los chips especializados que alimentan modelos como GPT no es cuestión de simplemente presionar 'replicar' en una fábrica. Hay límites en la cadena de suministro, en la capacidad de fabricación de semiconductores y en las leyes fundamentales de la termodinámica que gestionan la disipación del calor.

La promesa versus la realidad del silicio

Mientras los folletos de inversión pintan un futuro de crecimiento exponencial, la industria choca con la linealidad—y a veces la escasez—de la producción física. Cada avance en densidad de transistores se enfrenta a nuevos desafíos de rendimiento y potencia. Es la paradoja de la IA: el software más etéreo depende del hardware más tangible y complejo de producir.

Un recordatorio para los mercados

Para los entusiastas de las criptomonedas y la tecnología descentralizada, esta situación resuena profundamente. Así como la red Bitcoin encontró su límite en el consumo energético y la capacidad de hash, la IA encuentra el suyo en la disponibilidad de GPUs y la infraestructura de refrigeración. La descentralización, en teoría, podría ofrecer caminos alternativos—pero incluso las redes distribuidas necesitan hardware para funcionar.

Al final, la física siempre cobra su factura. Puedes tokenizar una promesa, pero no puedes minar un chip H100 de la nada—algo que los CFOs de las startups de IA están descubriendo ahora que el capital barato se ha evaporado y los inversores exigen resultados reales, no solo roadmap.

CoreWeave believes physical limits is what is slowing AI hardware delivery

“The primary constraint is a physical bottleneck associated with getting the most performant compute into the hands of the most cutting edge players,” Michael said.

Michael added that this pressure reaches far beyond cloud firms and chip suppliers. Michael shared details from a recent conversation with a mining company executive, who he did not name, where the discussion focused on materials needed to build infrastructure. He said the mining executive explained that the strain goes two levels deeper into the supply chain, reaching raw metals and copper used to support AI systems.

Michael said the mining executive made a direct request for cooperation across industries to increase output. “We need to work together as a group,” the executive told him. Michael said similar comments from leaders inside AI companies are often criticized as proof of a circular economy. “If I say that in the AI space, I get accused of being in a circular economy,” he said. “So that’s all I’ll say on the circular economy is, like, you do that by working together.”

Concerns from critics have also focused on risk tied to debt and customer exposure. Some warn that if CoreWeave fails to refinance obligations or loses a major client, lenders could release large volumes of used GPUs into secondary markets, hurting prices and causing disruption. Michael addressed that scenario by pointing to demand trends instead. He said the growth he sees is fast and aggressive, not fragile.

CoreWeave faces stock swings as demand from big tech stays intense

Michael said CoreWeave sits at the center of rising demand because of its focus on parallel computing, which supports modern AI workloads. He said requests from large technology firms continue without pause.

Since its IPO, CoreWeave’s stock has been wobbly, but Michael only pointed out that shares now trade near $90, compared with an IPO price of $40.

Michael also addressed past reliance on Microsoft, which once accounted for 85% of company revenue, saying that exposure has dropped after active diversification. He claims that no single customer now makes up more than 30% of the backlog, urging investors to ignore short delays that draw heavy reactions, such as a data center opening pushed back by a week, which he said triggered “bedlam” from short term observers.

Michael described the current phase as a macro super cycle driven by a move from sequential to parallel computing that opens access to far greater compute capacity.

Michael said the challenges slowing delivery remain tied to policy limits, physical infrastructure, and energy access. “The reasons that you have challenges in delivering that compute is because of policy, because of physical infrastructure, because of energy,” he said. “You do that by working together.”

Claim your free seat in an exclusive crypto trading community - limited to 1,000 members.