CoreWeave closed a $3.7 billion private placement of 2.875% convertible senior notes due 2033 on Tuesday, expanding the deal 23% from its initial $3.0 billion target. The AI cloud infrastructure operator will deploy proceeds into datacenter expansion through 2027, principally to house Nvidia H100 and next-generation Blackwell GPU clusters for enterprise inference workloads.
The notes carry a 2.875% coupon and mature January 2033. CoreWeave has not disclosed conversion terms, but the structure follows the company's October $650 million convert at 4.875%, which priced at a 35% conversion premium to its then-private valuation. The upsizing occurred without reopening the books, indicating anchor orders from sovereign wealth and insurance allocators who absorbed the full greenshoe. Goldman Sachs and Morgan Stanley led the syndicate.
The raise matters because CoreWeave is building inventory for a market that doesn't yet exist at scale. Enterprise inference — running trained AI models in production rather than training new ones — requires different infrastructure than the H100 training clusters that drove 2023 revenues. CoreWeave is betting $3.7 billion that by 2026, Fortune 500 companies will prefer renting inference capacity by the hour rather than owning it, and that margin compression in training workloads will push hyperscalers to cede inference share to specialists. The company's October S-1 filing showed $1.2 billion trailing revenue but a $491 million net loss, with gross margins at 41% versus the 60-70% range AWS maintains on EC2 compute. The convertible structure defers equity dilution while CoreWeave races to prove unit economics before the 2033 maturity.
The timing is tight. Nvidia's Blackwell chips ship in volume starting Q2 2025, and CoreWeave has reserved an undisclosed allocation. The company must deploy this capital into power contracts, datacenter shell construction, and cooling infrastructure before those chips arrive, because enterprise customers will not commit to multi-year inference contracts without proof of capacity. Meanwhile, Microsoft is building its own inference infrastructure for OpenAI's models, and Google is routing Gemini inference through internal TPU clusters, narrowing the available customer base for third-party providers.
Operators should track CoreWeave's Q2 2025 capacity announcements and whether the company discloses anchor inference customers before Blackwell volume shipments. The debt structure assumes CoreWeave reaches $4-5 billion annual revenue by 2027 to service the convert and maintain datacenter utilization above 75%, the threshold where fixed power costs stop eroding margins. If enterprise inference adoption lags six quarters, the 2033 maturity becomes a refinancing event, not a conversion event.
The 23% upsize without marketing tells you sovereign allocators are treating AI infrastructure as a national interest trade, not a venture bet.