Google Capped Meta's Gemini Access Over AI Compute Shortage, Revealing Industry Bottleneck

Google limited Meta's access to Gemini AI computing in March 2026 due to capacity constraints, forcing Meta to accelerate its own Muse Spark model as Google Cloud's order backlog nearly doubled to $460 billion. Amazon is seeking cheaper AI alternatives as Anthropic shifts to token-based pricing, while China's Meituan open-sourced a 1.6 trillion parameter model trained entirely on domestic chips—underscoring the global AI compute and cost crunch. Meanwhile, OpenAI silently rolled GPT-5.6 to some Codex users despite promising gated access, and DeepSeek unveiled DSpark, an inference framework that delivers responses up to 85% faster.

Written by AI from the sources it cites; it can be wrong — follow any citation to confirm.

8 sources2 min read

Google limited Meta's Gemini access as AI compute demand outstripped supply Google capped Meta's access to Gemini AI models in March 2026 because it lacked enough computing capacity to meet Meta's demand, according to the Financial Times. Google Cloud's order backlog nearly doubled to $460 billion in Q1 2026, forcing Meta to accelerate development of its own Muse Spark model and ration internal AI usage. 1 2

Amazon seeks cheaper AI alternatives as Anthropic shifts to token-based pricing Amazon is exploring OpenAI and other models after a renegotiated contract will shift Anthropic billing to per-token pricing next year, potentially raising costs sharply. Some Amazon engineers are already distilling Claude models internally to build smaller, cheaper versions, and the company has committed large investments in both Anthropic and OpenAI. 3 4

Meituan open-sources LongCat-2.0, a 1.6 trillion parameter coding model trained on Chinese chips Chinese delivery giant Meituan released LongCat-2.0, a Mixture-of-Experts model with 1.6 trillion parameters and a 1-million-token context window, trained entirely on a 50,000-card domestic compute cluster. It scores 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, showcasing China's ability to train cutting-edge AI without restricted Nvidia hardware. 5

California partners with Anthropic to deploy Claude across state agencies at half price Governor Gavin Newsom announced a partnership giving California's public sector access to Anthropic's Claude at a 50% discount, along with free workforce training and developer support. It marks the first wide-scale AI deployment in a US state's government, intended for document analysis, drafting, summarization, and customer service. 6

OpenAI silently rolled GPT-5.6 to some Codex users despite promising gated access Developers discovered that OpenAI had deployed GPT-5.6 to some Codex users shortly after announcing the model was restricted to government-vetted partners. A hidden system prompt parameter called the Juice value exposed the swap, raising transparency questions about how AI vendors communicate model changes to paying customers. 7

DeepSeek unveils DSpark, an inference framework delivering responses up to 85% faster Chinese AI startup DeepSeek introduced DSpark, a speculative decoding framework that increases inference speed by up to 85% through parallel generation and adaptive verification. The release targets the industry's high inference costs, enabling more users per GPU at a time when AI compute demand continues to outpace supply. 8