AI Scale

Ukrainian drones struck Russia's largest data centre, a facility central to the country's artificial intelligence computing capacity. The attack, reported by Reuters, is framed as a blow to Moscow's AI ambitions, as large-scale AI development depends heavily on such data centre infrastructure. The strike shows Ukraine extending its drone campaign beyond energy and military targets to technology assets supporting Russia's war economy.

Anthropic has published a new engineering post describing how the company redesigned its source control setup. The article, written by the AI firm's engineering team, outlines the changes made to how code is managed internally as engineering practices scale. It is drawing strong attention from developers discussing the trade-offs of version control systems and what a large AI lab's internal tooling choices reveal about modern software engineering.

Krea has introduced Krea 3, an AI image model designed for editing at scale. The announcement has drawn attention among startups and technology watchers, with several hundred people engaging with the news so far. Krea positions the model as focused on large-scale image editing workflows, though detailed benchmarks or availability information have not been widely discussed yet.

Humxn Research Labs has committed $50 million to purchase data on skilled trades, farming and manufacturing, which the company says will be used to train AI systems for robotics. The move highlights growing competition for specialised, real-world training data as robotics firms seek to teach machines physical tasks that are poorly represented in existing datasets.

Microsoft is developing the Maia 300, a custom AI accelerator intended to run artificial intelligence workloads on its own silicon rather than relying fully on outside suppliers. The move fits a broader industry push by hyperscalers to cut dependence on Nvidia GPUs and lower cloud computing costs. The Maia 300 is expected to be deployed across Microsoft's Azure data centers as the company scales its AI infrastructure.

AI chip startup Tenstorrent has raised $1.5 billion to scale its AI computing systems, betting that serious competition to Nvidia requires billions in funding. Led by veteran chip designer Jim Keller, the company sells complete computing machines and licenses its processor designs so customers can build their own chips, taking a different approach from Nvidia's integrated model.

Chinese tech giants ByteDance, Alibaba and Tencent have together placed more than $35 billion worth of orders for Nvidia chips, according to a Reuters exclusive. The scale of the orders underscores how heavily China's leading internet companies are investing in artificial intelligence computing power, even as US export controls restrict what advanced semiconductor technology can be sold to Chinese buyers.

Inception Labs has introduced Mercury 3, described as a large-scale diffusion-based language model. The announcement has drawn attention among developers and AI researchers, who are weighing whether diffusion approaches can rival traditional autoregressive models like those behind mainstream chatbots at scale.

IBM Research-backed open-source project llm-d has released version 1.0, focusing on the operational side of running large language model inference in production. The release emphasizes 'day 2' operations, meaning the scaling, monitoring and reliability work needed after initial deployment. The project aims to make serving AI models on distributed infrastructure more manageable for enterprises using Kubernetes-based environments.

A new Python project, agentic-awesome-skills by developer sickn33, is drawing interest on GitHub. AAS Core acts as a local, agent-first control plane handling catalog discovery, agent-owned skill selection, stack validation and planning, backed by a library of more than 2,600 agentic skills. It ships with a command-line tool, a local MCP server, a skill catalog, plugins and a Workbench interface. The project is being discussed among developers building tooling for autonomous AI agents.

A Washington Post opinion piece argues that advances in artificial intelligence now pose biological risks serious enough to warrant a Manhattan Project-style national program for biodefense in the United States. It urges large-scale, government-led investment to detect and counter AI-enabled biological threats. The argument adds to a growing debate in Washington over how quickly AI is expanding capabilities in sensitive scientific domains.

A new research paper, ScaleNav, presents a method for pretraining local navigation models on large-scale internet video using language instructions. The approach aims to let robots or navigation agents learn from abundant online footage rather than costly purpose-collected data, conditioning movement decisions on natural-language goals. The paper is currently circulating among AI researchers, with discussion just beginning on academic preprint platforms.

llm-d, an open-source project for AI inference, has reached its 1.0 major release. The milestone signals the software is considered stable and production-ready. Details beyond the announcement, including what features shipped and who backs the project, remain sparse, but major releases of inference infrastructure typically draw attention from developers deploying large language models at scale.

Chinese AI company DeepSeek is releasing the open source code for the software infrastructure it uses to run its AI models at large scale. The move gives other developers access to the systems behind its efficient, low-cost AI operations, and is likely to fuel debate about DeepSeek's openness and its challenge to Western AI labs.

Reuters reports that ByteDance, Alibaba and Tencent have ordered around $35 billion worth of Nvidia chips, underscoring the enormous demand for advanced semiconductors among China's largest technology companies as they race to build out artificial intelligence capabilities. The scale of the orders highlights Nvidia's central role in the global AI boom and the intensifying competition in cloud and AI infrastructure between Chinese tech giants.

Nscale, an AI cloud computing provider, is in talks to sell AI computing power to Anthropic, the maker of the Claude chatbot, in a deal that could be worth about $45 billion, according to a Reuters report citing a person familiar with the matter. The potential agreement highlights the massive infrastructure spending behind frontier AI development and adds to a string of large-scale compute deals between cloud providers and AI labs.

Cisco has published the Splunk State of Observability 2026 report, focused on the hidden costs that come with deploying AI agents. The report examines the operational and financial challenges organisations face as they scale agentic AI, and how observability practices can help manage them. The full report is available through Cisco's newsroom.

New research from ActiveState suggests enterprises could spend more than 13,900 engineering hours annually on AI-assisted development. The finding points to the hidden time costs of adopting AI coding tools at scale, including maintenance and oversight work. It comes as companies weigh productivity gains against the operational burden of integrating AI into engineering workflows.

Splunk, a Cisco company, has released its State of Observability 2026 report, focused on the hidden costs of deploying AI agents. The report examines the monitoring and infrastructure challenges organisations face as they scale agentic AI, and how observability practices can help control those expenses.

Google's AI research lab DeepMind is quietly recruiting nuclear physicists, according to a report by Futurism. The hiring push suggests the company may be pursuing work that requires deep physics expertise, possibly related to energy or advanced scientific computing. Details about the roles, their scale, or DeepMind's exact aims have not been disclosed, leaving observers to speculate about the lab's ambitions beyond artificial intelligence research.

A new commentary argues NVIDIA has reached a scale where its failure would shake the entire technology sector and the broader economy, effectively making it too big to fail. The piece draws a contrast with Meta and Google, contending that despite their enormous market values, neither carries the same systemic weight because NVIDIA supplies the chips underpinning the AI boom.

llm-d has released version 1.0 of its open-source project, which is positioned as a way to run large language model inference more cheaply and reliably at production scale. The release is aimed at organisations deploying AI in production, emphasising cost efficiency and openness. Details on benchmarks or adoption partners beyond the announcement itself have not yet been widely reported.

Tencent is said to have arranged the leasing of 100,000 AI chips from Oracle, a move framed as a way to sidestep US export restrictions on advanced semiconductors to China. If confirmed, the arrangement would mark a significant workaround to American chip controls and could intensify scrutiny of cloud providers' dealings with Chinese tech firms.

South Korean bookstore chain Kyobo has turned down an order for 100,000 books that it considered suspicious, with concerns about AI copyright infringement cited. The bulk purchase reportedly raised fears that the books could be used to train or build artificial intelligence datasets without proper licensing. The move highlights growing anxiety among publishers and retailers over unauthorized large-scale copying of copyrighted works for AI development.

Amazon has released TorchNeuron, a new developer tool for running PyTorch machine learning workloads on its custom Trainium chips. The move is being framed as Amazon's most direct challenge yet to Nvidia's dominance in AI training hardware, offering developers an alternative path for building and scaling AI models in the cloud.

Analysts are pointing to forecasts that spending on data center construction could roughly double next year, driven by surging investment in artificial intelligence infrastructure. Tech companies are racing to build the computing capacity needed to train and run AI models, pushing capital expenditure to record levels and raising questions about how sustainable the boom is.

A former Google DeepMind research lead who helped create AlphaGo is launching Metis Reasoning, a London-based AI company, today. The 12-person team says it is building a compute-efficient alternative to the massive frontier models pursued by large AI labs. The launch adds to a wave of startups betting that smaller, more efficient approaches can compete with scale.