Anthropic Partners With SpaceX on Compute, Significantly Eases Claude Usage Limits

On May 6, Anthropic entered into a compute partnership with SpaceX. The deal leverages the full compute resources of SpaceX's Colossus 1 data center in Memphis, with more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) expected to be secured within the month.

Alongside the partnership, Anthropic implemented three changes effective the same day. First, it doubled the 5-hour rate limits for Claude Code on Pro, Max, Team, and seat-based Enterprise plans. Second, it eliminated peak-hour limit reductions for Pro and Max accounts. And third, it significantly raised the API rate limits for the Claude Opus model.

Beyond this deal, Anthropic has also announced an agreement with Amazon for up to 5 GW (including 1 GW of new capacity by the end of 2026), a 5 GW agreement with Google and Broadcom (scheduled to come online in 2027 and beyond), $30 billion in Azure capacity with Microsoft and NVIDIA, and a $50 billion US AI infrastructure investment with Fluidstack. The company already trains and runs Claude on multiple hardware platforms, including AWS Trainium, Google TPUs, and NVIDIA GPUs, and this partnership adds an even more diverse compute foundation.

Anthropic has also signaled interest in eventually working with SpaceX to develop “multi-gigawatt-scale AI compute capacity in orbit.” For enterprise customers, the company plans to expand in-region infrastructure in Asia and Europe to meet data residency requirements in regulated industries such as finance, healthcare, and government.

Google Speeds Up Gemma 4 by Up to 3x With Multi-Token Prediction

Google has released a multi-token prediction (MTP) drafter for its open model family, Gemma 4. A small auxiliary model predicts multiple tokens at once while the main model verifies them in a single batch — an approach said to accelerate text generation by up to 3x.

MTP is a “speculate-and-verify” approach to inference that substantially improves effective throughput compared with conventional token-by-token sequential generation. Within Gemma 4's open ecosystem, one notable point is that the community can now easily adopt faster inference.

Google Integrates “Expert Advice” From Reddit and Other Forums Into AI Search

Google has updated its AI search features to incorporate “expert advice” from Reddit and other web forums into search results. While this offers the benefit of surfacing community-based answers to niche questions, critics note that a design that cites forum posts as the basis for AI outputs carries challenges in terms of information reliability and potential confusion.

Anthropic Announces “Dreaming” Feature for AI Agents

At its developer conference, Anthropic announced a new feature for AI agents called “dreaming.” The feature organizes and processes the “memories” an agent has accumulated, apparently inspired by how memory consolidates in humans during sleep. WIRED has voiced concerns about the AI industry's habit of naming features after human biological processes.

Genesis AI Goes Full-Stack Robotics, Unveils GENE-26.5

Genesis AI, a robotics startup backed by Khosla Ventures, has shifted to a full-stack strategy following a $105 million seed round. Alongside unveiling GENE-26.5, its first foundation model, the company has released demo footage of multiple robot arms performing complex tasks. Its plan to integrate everything from foundation AI to hardware control in-house could accelerate the vertical integration trend in robotics.

---