Nebius's headline $40 billion is contract backlog, not spending; its actual 2026 capex guidance is $20 billion to $25 billion, funded by prepayments and new debt as delivery delays test the model.
A Commerce Department rule let licensed Chinese firms buy Nvidia's H200 chips after a vetting check, but the New York Times found one licensee's Beijing office standing vacant.
A "Conversational Apps" flag found in the iOS 27.3 beta hints that Apple's smart home hub could soon host chatbots like ChatGPT or Claude, though the feature isn't wired up yet.
Fireworks VP Rob Ferguson argues that routing queries across a portfolio of models, not locking into one provider, cuts AI bills by half or more. He says caching gaps and untracked data value are hiding the real cost of running a single frontier model.
Amazon's new Alexa text-to-shop feature is rolling out by invitation only, with full access promised to millions of US customers in the coming weeks. It arrives weeks after Amazon blocked Meta's Muse agent from checking out on Amazon.com, and amid an unresolved court fight over whether Amazon can block rival shopping agents at all.
OpenAI says it will acquire ten-person startup Pierre Computer to fold its git-hosting tool into Codex, but the deal has no disclosed price, no closing date, and is still subject to customary conditions.
Ai2's new benchmark finds frontier models drop to 13 percent on the hardest state-tracking tasks, and it quietly checks a 2024 theory paper's prediction about which architectures should do better.
MIT CSAIL had GPT-6 Astra edit just 0.016% of Qwen3.5-0.8B's weights by hand and saw math scores jump sixfold, with no gradient-descent training involved.
Anthropic's revised S-1 filing puts total compute commitments near $895 billion, pushing past $1 trillion with additional needed capacity, while Q3 revenue growth slowed to 46 percent. The filing warns the 1,206 percent year-over-year growth rate it posted through September won't repeat.
An internal Amazon memo on AI use, written by Distinguished Engineer Terk Terkowitz and sent to staff in September, says skipping AI tools can be "a finding, not a failing." Nothing in the document changes how promotion or performance reviews score output, so the stated policy and the actual incentive point in different directions.
Apple CEO John Ternus told Wired he won't "ship AI for AI's sake," a line issued days before his first earnings call, backed by a January deal licensing Google's Gemini to run the new Siri.
Google's orbital TPU prototype, Suncatcher, beamed down a Gemma 4-generated haiku, a near-free bit of inference riding on a satellite that costs millions to launch and faces real radiation limits.
Runware is building a factory to crank out shipping-container AI data centers it says can power on in 48 hours, but grid interconnection delays, not assembly speed, are what keep most projects waiting years.
