Climate alarm in Texas; Meta and Alibaba ship open models

Five developments from the past 48 hours arrive together, mapping both the cost and the spread of the current AI expansion: a single data-centre complex with emissions on the scale of a small country, two open-weight model releases aimed at developer desktops, a new browser purpose-built for AI agents, and an enterprise acquisition that widens one company's reach into daily work software.

Amazon's Proposed Texas Data Centre Could Become the Largest US Climate Polluter

Analysis of regulatory filings shows that a planned Amazon data-centre complex in Texas would emit an estimated 33 million metric tons of CO₂ annually — a figure that would place it among the largest single-site greenhouse-gas emitters in the United States, exceeding the output of most coal-fired power plants. The projections are drawn from permit applications and energy contracts filed with state regulators; Amazon has not confirmed the site's specifications publicly. Environmental groups have called for federal review before construction approvals are granted, arguing that projects of this scale require national-level environmental assessment, not state-level permitting alone. The case puts a precise number on the structural tension at the heart of AI infrastructure: computing capacity is expanding far faster than the grid can add renewable generation.

OpenAI Acquires AI-Powered Presentation Startup NextSlide

OpenAI has acquired NextSlide, a startup that built an AI system for generating and editing presentation decks. Financial terms were not disclosed. The deal brings a production team and a working product into OpenAI's portfolio at a moment when the company is competing with Microsoft Office, Google Workspace, and Canva for daily-use productivity software. It follows OpenAI's earlier moves into voice interfaces, persistent memory, and code generation — a pattern suggesting the company is assembling a suite of workplace tools rather than remaining solely a model provider. NextSlide's product is expected to be integrated into OpenAI's consumer and enterprise platforms; no timeline has been announced.

Cloudflare Launches Kitesurf, a Browser Built for AI Agents

Cloudflare has unveiled Kitesurf, a web browser built for AI agents rather than human users. Unlike conventional browsers, Kitesurf is engineered for the tasks autonomous systems need to perform at scale: logging into services, navigating multi-step forms, parsing structured data, and executing transactions without human oversight. Cloudflare positions the tool as infrastructure — plumbing for agent developers rather than a consumer product. The announcement reflects an emerging category of software designed specifically for AI that browses and acts on the web autonomously, with meaningful implications for how websites detect and manage non-human traffic.

Meta Releases Muse Glimmer, a 30-Billion-Parameter Model for Consumer Hardware

Meta has published Muse Glimmer, a 30-billion-parameter language model designed to run on a single consumer machine — a laptop or high-end desktop — without cloud access. The weights are available on Hugging Face under an open licence. Muse Glimmer is optimised for agentic tasks: planning, multi-step reasoning, and tool use in environments where data privacy or offline operation matters. Twelve months ago, a 30-billion-parameter model required a rack of data-centre GPUs; the fact that it now runs on consumer hardware is a tangible indicator of how rapidly model compression and quantisation have advanced. Meta's open-weights strategy is designed to embed its models in the ecosystem of local AI tools before competing closed providers can establish dominance at this size range.

Alibaba Releases Qwen 3.8-Max, a Coding-Focused Open-Weights Model

Alibaba has published Qwen 3.8-Max, a model aimed at software development tasks: code generation, debugging, documentation, and technical reasoning. The weights are available for self-hosted deployment, letting teams run the model on their own infrastructure without routing queries through Alibaba's cloud. Qwen 3.8-Max enters a competitive field that includes Meta's Code Llama, Google's Gemini Code, and the models underlying GitHub Copilot — but its fully open-weights release, combined with strong benchmark performance claims, positions it as a credible option for organisations that prefer not to depend on a single Western cloud provider. The release is a further sign that Chinese AI laboratories have moved from catch-up mode to active competition on specialised developer benchmarks.