This brief covers AI news from 2026-09-25 UTC.
Anthropic locks in $11.6 billion of Akamai cloud capacity, Docker takes agent sandboxes off the laptop, and Databricks buys a spreadsheet startup for its Genie agent.
Akamai signs $11.6B seven-year cloud deal with Anthropic
What happened: Akamai announced an $11.6 billion contractual commitment from Anthropic over seven years to support Anthropic’s CPU workload growth on Akamai Cloud. The deal can expand by up to $9 billion more, and Akamai issued Anthropic a warrant for up to roughly 5% of its common stock, with about 2% vesting on today’s commitment.
Why it matters: Anthropic is buying CPU capacity at a scale that signals agent and inference workloads, not just training, are driving cloud contracts developers will build against.
Docker Cloud Sandboxes run agents after your laptop closes
What happened: Docker launched Cloud Sandboxes, microVM-based isolated environments running on Docker-managed compute, with one command to move agent workloads between local and cloud. Docker says the isolation model is identical to its local sandboxes and the product is available today.
Why it matters: Long-running coding agents can now keep working for hours without tying up a developer’s machine, which changes how you schedule multi-hour refactors and test runs.
Databricks acquires spreadsheet startup Row Zero for Genie
What happened: Databricks announced it acquired Row Zero, an early-stage cloud spreadsheet startup that scales beyond one million live rows. Databricks co-founder Ali Ghodsi said the company intends to do many more acquisitions like this; terms were not disclosed.
Why it matters: If you query Databricks data through agents, the spreadsheet interface plus Genie points to natural-language analysis landing inside tools analysts already use.
Moonshot AI releases Kimi K2.6 with 1T parameters
What happened: Moonshot AI released Kimi K2.6, an open-source multimodal MoE model with 1 trillion total parameters and 32 billion activated per token, a 256, 000-token context window, and a MoonViT vision encoder. Weights are on Hugging Face under a modified MIT license.
Why it matters: An open-weight model claiming 12-hour agent sessions and 300 parallel sub-agents gives teams a self-hostable option for long-horizon coding and orchestration work.
VeriLoop E2 posts 27B model with full GGUF ladder
What happened: VeriLoop released E2, a 27B post-trained model built on Qwen3.8-27B, with scores including 76.2% on SWE-bench Pro, 88.8% on Terminal-Bench 2.1, and 98.3% on AIME 2026. The release includes a llama.cpp-oriented GGUF precision ladder from BF16 down to IQ1_M.
Why it matters: A quantized ladder down to IQ1_M means you can test the same checkpoint across laptop and server hardware without swapping model families.
Databricks ships Unity Gateway CLI for coding agents
What happened: Databricks released the Unity Gateway CLI, a command-line tool that connects coding agents such as Claude Code, Codex, and Gemini CLI to centralized Unity Gateway policies for approved models, MCP servers, skills, and spending. Admins publish a configuration and developers connect with one command.
Why it matters: Teams standardizing on multiple coding agents get one place to control which models and tools those agents can reach, instead of per-developer config.
Google to test AI chips in orbit under Project Suncatcher
What happened: Google said it will launch a prototype satellite next week for the first in-orbit test of Project Suncatcher, a research effort exploring whether space could host large-scale AI computing infrastructure.
Why it matters: It is an early experiment, but it is a concrete data point on whether off-Earth compute is a real option for future training and inference capacity.