GPU capacity

Kimi

2026-07-19 14:55:00

Kimi halts new subscriptions as Kimi K3 demand pushes GPU capacity near its limit

Kimi.ai said on July 19 that it has temporarily stopped accepting new subscriptions after demand for Kimi K3 surged over the past 48 hours, pushing available GPU capacity close to its current limit. The company said the move is meant to protect the experience of existing subscribers, with compute resources now being prioritized for current members. Kimi added that services for users who already hold subscriptions will not be affected. The company also said it is speeding up capacity expansion and plans to reopen subscription slots in batches. At the same time, Kimi is restructuring its membership system into two separate products: “Kimi Membership,” which will cover Web, App, and Work scenarios, and “Kimi Code Membership,” which will target code workflows. According to the announcement, the split is designed to match compute resources more precisely across different usage needs.

2360

Meta

2026-07-06 03:29:56

Meta Weighs Meta Compute to Rent Out GPU Capacity and Host Third-Party AI Models

Meta is reportedly considering a new business line, Meta Compute, to commercialize its vast AI infrastructure by renting GPU capacity to external customers, hosting third-party models such as Claude, and offering model-as-a-service products. Bloomberg and SemiAnalysis say the company has not slowed its data center expansion: in the first six months of this year alone, Meta signed more than 5GW of cloud and colocation capacity, while total compute-related deals since early 2024 have approached 10GW. The infrastructure would continue to support Meta’s own model training, including Muse Spark and the next-generation Watermelon model, while also being deployed for ad recommendation systems. SemiAnalysis argues Meta could also emulate the neocloud model by leasing a portion of its compute externally at high margins. In parallel, Meta is said to be in late-stage talks with Anthropic for private-instance access to Claude, a move that could help Meta build a platform similar to Amazon Bedrock, Microsoft Foundry, or Google Vertex. The shift comes as Meta’s AI spending accelerates, with 2026 capital expenditure guidance raised to $125 billion-$145 billion. Investors appeared to welcome the strategy: Meta shares rose nearly 9% on the news, while AI cloud names including CoreWeave and Nebius came under pressure.

1320