美国AI云服务商CoreWeave在2026年10月密集发布技术产品并宣布全球扩张,但其高杠杆商业模式引发市场剧烈震荡。公司推出Forge平台及Nvidia Vera CPU支持,并在北达科他州和印度推进数据中心建设;然而,其债券收益率飙升逾70%,股价单周跌近四分之一,导致英伟达市值蒸发3650亿美元。尽管高盛等机构认为市场过度反应债务风险并上调评级至买入,但Janus Henderson等投资者指出其净债务约300亿美元且租赁承诺巨大,融资成本上升可能抑制未来芯片需求。目前争议焦点在于:CoreWeave能否通过GPU资产抵押或大客户预付款缓解流动性压力,还是将因资本开支高峰陷入再融资困境。
Shreyan Basu Ray, author for the Weights & Biases blog under CoreWeave, explains that agentic workflow costs are driven by input tokens rather than output tokens—each loop iteration resends all accumulated context, resulting in a long-tailed cost distribution. The article identifies model-tier orchestration (expensive models for planning, cheap models for execution) as the strongest architecture-level cost lever among published evidence. τ-bench's pass^k consistency data indicates that single-attempt success rates severely underestimate the true cost calculated per successful task.
CoreWeave blogger Min Jun explains why the serial, multi-hop traffic pattern of agentic AI—where latency is the sum of every hop rather than the worst one—forces network design choices in topology, load balancing, and resilience strategies that differ fundamentally from those required for training.
CoreWeave self-reported 441k tokens/sec system throughput for DeepSeek-R1 inference in the inaugural MLPerf Endpoints benchmark, emphasizing production-environment validation.
CoreWeave blog authors Tara Madhyastha and Nisha Nadkarni explain why latency and availability in production inference services can degrade continuously without triggering any alerts: the combination of GPU resource contention, batch configuration drift under load, and autoscaling response lag causes p99 and TTFT to rise silently even when averages appear normal. The article presents three diagnostic signals—divergence between p99 and p50, rising TTFT under stable load, and the gap between goodput and throughput—and argues that reliability should be designed as an architectural property rather than a monitoring issue.
John Lloyd, global head of multi-sector and corporate credit at Janus Henderson, holds CoreWeave 9.75% bonds due in 2031 through his fund and views the more-than-70% yield jump since April as an overreaction. Lloyd argues CoreWeave can navigate its capex peak through GPU-collateralized financing, larger prepayments from customers like Microsoft and Meta, and a near-term maturity-free window, though he acknowledges the bond recovery will not be a straight line.
Nvidia shares fell 6.6% as rising borrowing costs for leveraged data center operators like CoreWeave raise doubts about the sustainability of AI infrastructure spending.
Investors are buying credit default swaps to hedge against a potential downturn in the AI boom, driving up protection costs for heavily indebted hyperscalers like CoreWeave and Oracle.
A CoreWeave blog post explains the engineering practice of distinguishing between two AI agent modes in physical infrastructure monitoring: offline construction of deterministic detectors and online open-ended reasoning. The article details how this architectural split addresses non-determinism, controls costs, and retains human control over critical decisions.
Applied Digital announced an additional 75 MW of ready-for-service capacity at its Polaris Forge 1 campus, bringing total operational load to 250 MW, over 60% of CoreWeave's contracted volume.
The cloud vendor wires inference and reinforcement learning into one loop; hot-loading lowers iteration friction, but customer results are still unproven.