Topic
Accelerators, model releases, inference economics and on-device machine learning.
Google’s AI search features, ChatGPT search, Copilot and Perplexity turn web results into answers. The convenience is real. So are the gaps between a convincing summary and evidence you can rely on.
Open-weight text generators publish their trained parameters so anyone can run them on their own hardware. That single decision changes cost, privacy, customization and risk — and it is not the same thing as open source.
AI that surpasses human experts across many fields could reshape science, work and security. But impressive benchmarks are not proof of superintelligence, and predictions about its arrival are not established facts.
The price of a useful AI response can fall while the bill for building the infrastructure rises. That tension is reshaping software budgets, semiconductor demand, electricity investment and the expectations built into stocks and credit markets.
AI can connect suspicious activity across accounts, devices and cloud services, and help analysts investigate faster. But a convincing explanation is not proof of an attack. Here is how detection actually works, where the evidence stops, and what a responsible deployment looks like.
Machine-learning forecast systems trained on decades of atmospheric data have moved from research demos into the workflows of national weather services, outperforming traditional models on some measures while leaving harder questions — extreme events, uncertainty, physical interpretability — unresolved.
Agent-initiated payments need something browsers never had: verifiable proof that a human authorized the transaction. Multiple protocols are racing to provide it — and merchants get to choose which wins.
The most economically important technique in AI isn't a new architecture — it's the quiet practice of making small models behave like large ones, which is how frontier capability gets onto your phone.
The Robot didn't arrive as a machine; it arrived as a model. Vision-language-action systems trained on human demonstration are replacing hand-coded control — and the economics, not the demos, will decide what happens next.
The frontier gets the headlines, but the deployed reality of corporate AI is increasingly open-weight models running on infrastructure the company controls. The reasons are cost, data custody and predictability — not ideology.
Browsers that browse for you are the breakout AI product category of the year. They are also a prompt-injection delivery system pointed at your logged-in sessions.
Publishers spent two years suing AI companies over training data. The quieter story is the deal flow: licensing agreements are turning archives into recurring revenue and drawing a line between the licensed web and the scraped one.
Software agents now book travel, move money and touch production systems on behalf of users. The identity infrastructure built for humans and service accounts was never designed for this.
The price of generating a token has fallen by orders of magnitude in three years — yet enterprise AI spending keeps rising. The economics of cheap inference are stranger than the headline suggests.
The gap between agent demos and agent deployments turns out to be made of permissions, audit trails and error budgets — and the companies succeeding are the ones that narrowed the task before trusting the model.
Two years after RAG was the hot architecture, grounding models in retrieved documents is now assumed infrastructure — and the hard problems have migrated to evaluation, freshness and permissions.
Hyperscale operators are committing to tens of gigawatts of new capacity, but transmission bottlenecks, transformer shortages and multi-year interconnection queues mean the electricity cannot arrive fast enough. Here is where the strain is showing.
While frontier models dominate keynotes, businesses are discovering that models small enough to run on a laptop handle the bulk of real work — cheaper, faster, private and often more accurate on narrow tasks than their giant siblings.
The gap between open-weight and proprietary models has narrowed enough that many companies now run capable models on their own infrastructure — and the reasons are about control and cost more than benchmark scores.
Oracle told investors its AI infrastructure ran at nearly full capacity last quarter, backed by a bulging cloud revenue backlog, offering a rare hard data point on just how supply-constrained Nvidia GPU capacity remains.
Google's latest Android feature drop adds Guided Vision, a visual-description tool for blind and low-vision users, plus a Gemini-powered memory feature for tracking misplaced items, extending the Gemini Intelligence push it started in May.
Every AI stack diagram tells you how intelligence gets built. Almost none of them tell you where the money stays. That is the whole difference — and it matters if you are pitching, buying or funding AI in 2026.
Spend classification, contract review and supplier scoring are moving from quarterly exercises to continuous processes. Here is what actually changed, and what buyers should ask for.
Anand Arivukkarasu's framework argues that intelligence behaves like a supply chain — and that the money ends up at the bottlenecks, not at the chatbot everyone can see. Here is the map, in plain English.
Google's seventh-generation TPU, sold commercially as TPU7x, is designed around inference-heavy 'thinking' models rather than training, and Google says it is also markedly more energy efficient than its predecessor.
Advanced packaging and stacked DRAM capacity now gate how many accelerators actually ship.
Quantisation and unified memory have moved useful models from the data centre to the laptop.
