The complete archive
Independent analysis of the incentives, infrastructure and decisions shaping technology.
Explore the publication
Showing 18 articles.
The semiconductor supplier is becoming a financier of the computing capacity its customers need. The Broadcom–Anthropic arrangement shows how AI infrastructure contracts now combine chips, leases and extraordinary capital commitments.
CoreWeave has signed enormous future AI computing commitments. Its own financial disclosures reveal the expensive chain between a customer's contract, an energized data center and actual recognized revenue.
Foxconn is moving from familiar consumer-device assembly toward complete AI computing systems. Revenue is accelerating, but physical integration and contract-manufacturing margins reveal a more complicated opportunity.
Microsoft Execution Containers put file, network and process permissions outside the agent's own instructions. That is a practical answer to a growing problem: autonomous software needs boundaries it cannot rewrite.
OpenAI's first custom inference chip makes more sense as a strategy for controlling latency, energy costs, and the product experience than as a simple attempt to replace Nvidia.
Two striking annualized-revenue figures sparked questions about OpenAI's growth. The more revealing issue is how cloud-partner sales, different measurement boundaries and investor comparisons can change a private AI company's reported run rate.
A record-sized bond sale helped refinance an enormous AI investment. The important question is not just how much SoftBank believes in OpenAI, but how it funds that belief without sacrificing financial flexibility.
AI infrastructure can have hundreds of suppliers but still depend on a small group of ultimate buyers. Nscale, CoreWeave and Core Scientific filings show how customer concentration travels through the cloud supply chain.
When grid capacity arrives more slowly than AI computers, developers consider turbines and on-site generation. Projects in Mississippi, Texas and Dublin show why that workaround creates new engineering, financial and public questions.
Nvidia designs the accelerators driving AI, but its chips depend on manufacturing tools, advanced foundries and packaging that it does not own. ASML illustrates where the real bottlenecks sit.
A reported multibillion-dollar Oracle cloud lease shows how export restrictions, computing demand and data-center geography can reshape the AI supply chain without moving ownership of the hardware.
The electric bill is only one layer of the AI infrastructure build-out. Grid upgrades, capacity, stranded-cost risk and financing can land on very different balance sheets.
Gemini does not need to win every benchmark for Google to gain leverage. Search, Chrome, Android, Workspace and hardware give each model improvement many places to become habitual.
Humanoid robots are making real progress in factories, but a new industry-wide sales count shows how far the category still is from the mass market implied by its demos and forecasts.
AMD's trillion-dollar milestone is not simply a bet on one faster GPU. It is a bet that the company can become a credible second full-stack supplier for the AI data center.
Apple is willing to use outside model technology. The layer it appears determined to own is the one around the model: personal context, permissions, system actions and the privacy boundary.
Muse is more than another chatbot. Meta is testing what happens when an AI sits between a user's intent and the apps, services, credentials and transactions needed to carry it out.
The deal is not just about models. NVIDIA is buying a position inside the workflow where millions of developers discover, evaluate, customize, and deploy open AI — while promising not to close that workflow around its own hardware.
