Live Feed
Meta Locks 1GW+ MTIA Deal With Broadcom Through 2029, Targets Multi-Gigawatt Buildout
Meta and Broadcom have extended their custom silicon partnership through 2029, with an initial commitment exceeding one gigawatt of MTIA chips — the first AI accelerator on a 2nm process. Broadcom CEO Hock Tan is leaving Meta's board as the commercial relationship expands beyond what typical governance standards allow.
Microsoft Plans to Sell Software Seats to AI Agents — Treating Each Agent as a Paid Employee
Microsoft executive Rajesh Jha has outlined a pricing model that charges AI agents for enterprise software seats the same way companies pay for human users. As AI reduces human headcount, each agent would get its own identity, login, and tool access — turning the AI agent into a billable digital worker.
Microsoft Cuts Image Generation Costs 41% With MAI-Image-2-Efficient — $19.50/M Output Tokens
Microsoft launched MAI-Image-2-Efficient on April 14 with image output pricing dropping from $33 to $19.50 per million tokens, 22% speed improvement, and a claimed 40% latency edge over Google Gemini image models. No waitlist.
SPAN XFRA Puts Enterprise NVIDIA Blackwell GPUs in Homes to Solve AI's Power Gap
With data center interconnection queues running 7–10 years, SPAN is distributing liquid-cooled NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs into residential and small commercial spaces — converting stranded grid capacity into enterprise AI compute. NVIDIA is a launch partner.
Sub-32B Open Weights Hit GPT-5 Intelligence: Qwen3.5 27B Scores 42, Gemma 4 31B Scores 39
Two recently released open-weights families now match mid-tier GPT-5 on the Artificial Analysis Intelligence Index — but a 35-point hallucination gap still separates them from the frontier. The ceiling for small open models just moved, not collapsed.
Trump's AI Framework Is Creating the Patchwork It Was Designed to Prevent
Six weeks after the White House released its national AI policy framework, independent analysis finds the absence of binding rules is accelerating state-level regulation, not suppressing it. Companies cannot rely on federal pre-emption until Congress passes actual legislation.
EU AI Omnibus Trilogues Begin — Formal Talks to Revise AI Act Obligations Open in Brussels
The European Commission's Digital Omnibus AI package has entered trilogue negotiations between the Parliament, Council, and Commission. The talks aim to ease AI Act compliance burdens before full enforcement hits in August 2026 — but with 16 weeks left, most companies are already running late.
India's Motion Farms: Human Labor Is Still the Cheapest Way to Train Humanoid Robots
Humanoid labs can't bootstrap physical intelligence from text. India's factory workers, fitted with head-mounted cameras, are producing the supervised motion data robots need — and the economics reveal the core bottleneck in physical AI.
OpenAI Acqui-Hires Hiro Finance — ChatGPT's Next Move Is Your Money
OpenAI absorbs Hiro Finance and its ~10-person team, adding an AI personal financial planning capability as the company expands ChatGPT from conversational assistant into consumer life infrastructure. Hiro shuts down April 20; all user data is deleted by May 13.
OpenRouter Ships Its Own 100B Model — Elephant Alpha Is Free on Day One
OpenRouter, which routes API traffic to 200+ third-party models, has released its own 100B-parameter model at $0 per token. The move signals a strategic shift from pure aggregator to model producer — and raises questions about what Elephant's free pricing is actually optimising for.
US Federal Agencies Are Using AI to Write Regulations — DOT Names Google Gemini Its Rulemaking Tool
The Department of Transportation is using Google Gemini to draft proposed federal rules, calling it the "point of the spear" for AI in government. DOGE separately proposed using an LLM to rescind half of all federal regulations. AI has entered the rulemaking chain — and the legal implications are unresolved.
CoreWeave Closes $8.5B Investment-Grade GPU Loan — First A3-Rated AI Infrastructure Debt
CoreWeave secured an $8.5 billion delayed draw term loan on March 31, rated A3 by Moody's — the first investment-grade financing backed by GPU clusters and customer contracts. The deal, anchored by a $19B Meta backlog, marks the formal arrival of AI compute as an institutional asset class.
Musk Sets June Deadline for Grok to Match Claude Opus 4.6 — After Admitting xAI Was "Not Built Right"
Elon Musk said Grok will be close to Claude Opus 4.6 by May and match it by June. The admission follows a March acknowledgement that xAI is undergoing a full architectural rebuild — and a public concession that it currently trails OpenAI, Anthropic, and Google by roughly seven months.
New York RAISE Act Is Finalised: 72-Hour Incident Window, $1M Penalties, Effective January 2027
New York signed the final version of its frontier AI safety law on March 27, giving large developers nine months to comply. The 72-hour critical incident disclosure window is stricter than California's 15-day equivalent — creating a new compliance bar for any model operating in the state.
The 36x Reasoning Tax: DeepSeek R1 vs o3-pro in the 2026 API Economy
DeepSeek R1 charges $0.55/M input tokens. OpenAI's o3-pro charges $20/M. The gap is 36x on input and 36x on output — while R1 posts within 2 points of o3's graduate-level science score. This is the cost calculus defining production reasoning in 2026.