Big Tech Is Building Its Own AI Chips. The Target Is Inference, Not Training.
Microsoft's Maia 200, Meta's MTIA rollout, Google's Ironwood TPU, and OpenAI committing 2 gigawatts of Amazon Trainium. The inference market — billions of queries per day — has strong economics for custom silicon. Training stays NVIDIA's; inference is the contested ground.
Read full story →AI Is an Energy Problem. It Might Also Be the Energy Solution.
TechCrunch argues the best AI investment in 2026 might be an energy tech company. AI demand is driving capital into renewables, storage, and grid infrastructure at a pace climate policy alone couldn't achieve. Scientific American finds AI could cut 5.4 billion tons of annual emissions — more than data centres add — if deployed at scale.
Read full story →Doctors Think AI Has a Place in Healthcare. They Don't Think It's as a Chatbot.
OpenAI is rolling out to Cedars-Sinai, Stanford, UCSF and others. A clinical study across 40,000 patient visits found a 16% reduction in diagnostic errors with AI support. DAX Copilot is running at 500 institutions. But physicians are clear: they want AI handling documentation and flagging missed diagnoses — not acting as a patient-facing chatbot.
Read full story →AI Should Have Eliminated Millions of Jobs By Now. It Hasn't. Here's What the Data Says.
MIT finds AI can technically replace 11.7% of the US workforce. Yale Budget Lab finds the occupational distribution hasn't shifted since ChatGPT launched. Anthropic's own research confirms limited empirical displacement. The gap between exposure and impact is the most important number in the AI jobs debate — and what explains it has real implications for what comes next.
Read full story →The Left Wants to Stop Building Data Centers. The White House Has Other Plans.
Sanders and AOC's data center ban bill and the Trump administration's AI framework landed in the same week, and both reveal the same problem: Washington's regulatory conversation is running years behind the capital already committed. The real story is not their disagreement but the vacuum they are both trying to fill.
Read full story →NVIDIA Bets on Both Sides: A 10-Gigawatt OpenAI Deal and an Open-Model Family in the Same Week
NVIDIA signed the largest compute contract in AI history with OpenAI while simultaneously releasing Nemotron 3, a family of open models anyone can run without paying for proprietary APIs. The company makes money from compute regardless of which way the market goes, and this week's moves show exactly how that strategy works in practice.
Read full story →OpenAI Disbanded Its Alignment Team. Then Funded Someone Else to Do the Work.
OpenAI shut down its internal Mission Alignment team in February, then committed $7.5 million to independent external alignment research. Google DeepMind took the opposite approach, deepening a joint safety partnership with the UK's AI Security Institute. The gap between those two strategies reveals the hardest unsolved question in AI governance: who is actually accountable when something goes wrong.
Read full story →Google DeepMind Is Putting Its Robotics Models Into Real Machines
Google DeepMind's partnership with industrial robot maker Agile Robots is an attempt to move Gemini Robotics from demo to deployment. Real-world factory data is the scarcest resource in physical AI, and whoever gets their models into production robots first builds a data advantage that is very hard to close. This week's announcement is a land-grab as much as a product deal.
Read full story →OpenAI's GPT-5.4 Arrives With a Million-Token Window and a Thinking Mode for Everyone
GPT-5.4 sets new records on computer-use benchmarks and ships with a one-million-token context window in the API. Three simultaneous tiers cover everything from enterprise Pro users to casual Free users — revealing that "GPT-5.4" is now less a model than a platform, and raising the question of how meaningfully these variants relate to each other.
Read full story →The White House AI Framework Is Out. Whether It Is a Policy or a Placeholder Is Another Question.
The Trump administration's six-pronged AI legislative outline aims to preempt state laws and set a single federal baseline. CNBC called it a significant move; Bloomberg's opinion desk called it anything but a plan. The gap between those two reads reveals exactly where the framework is strongest and where it is thinnest.
Read full story →NVIDIA's Rubin Platform Arrives, and It Brings a Surprising Passenger: the CPU
NVIDIA's Rubin cuts inference costs by 10x and training GPU requirements by 4x over Blackwell. But the real surprise at GTC was Jensen Huang's argument that agentic AI workloads are bringing the CPU back to the centre of AI infrastructure, and that Rubin's six-chip co-design is the first platform built with that transition in mind.
Read full story →Meta's Four-Chip Gamble and the Custom Silicon Arms Race
Meta plans to roll out four generations of its own AI inference chips by the end of 2027 — a pace that treats hardware development like a software release schedule. The bigger picture: Bloomberg Intelligence projects the AI accelerator market will hit $604 billion by 2033, and the companies building their own silicon will have a structural cost and capability advantage over those that don't.
Read full story →The Data Center as Power Plant: AI Factories Enter the Energy Grid
NVIDIA and Emerald AI are partnering with NextEra, Vistra, and four other major energy companies to build "flexible AI factories" that act as grid assets, not just grid consumers. The concept is real but the claims are optimistic, and what makes this announcement interesting is less what it promises than what it reveals about where the AI industry thinks its energy reputation problem lies.
Read full story →Speech Joins the Open-Source Wave
Mistral and Cohere both released open-source voice AI models on the same day: a nine-language TTS system and a lightweight transcription model built for consumer hardware. Individually modest, together they signal that speech is following the same open-source trajectory that already transformed text and vision AI.
Read full story →Europe Moves on Deepfakes While Washington Debates Frameworks
The EU parliament's civil liberties committee approved a ban on AI-generated non-consensual sexual imagery this week, while Anthropic signed the EU Code of Practice. Together they show the EU's layered regulatory strategy in action: hard bans at the extremes, voluntary accountability frameworks at the frontier, and binding deadlines arriving in August.
Read full story →The Model Welfare Debate Is the Rights Debate in Disguise
Anthropic is asking whether its AI might have interests worth protecting. Microsoft says the question is dangerous. Both are missing what the debate is actually about — it is the question of what kind of relationship humans will have with artificial minds, and whether we build the frameworks for coexistence now or wait until it is economically catastrophic to acknowledge them.
Read opinion →