Google prices Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens across its one-million-token context window. It says the model is available through AI Studio and the Gemini API, among other products.
The company reports 57% on Terminal-Bench 4.1, a fixed set of terminal tasks for comparing models, versus 18.2% for its older Gemini 3.8 Flash. These are Google’s reported results.
Antigravity’s Orchestrator switches between Hydrogen and the larger Carbon during a session. Google says Hydrogen is rolling out to paid Pro and Ultra users there, while Orchestrator is initially for Ultra subscribers.
Caveat: Google’s two launch pages disagree on the Terminal-Bench version used for the Orchestrator comparison, so its quoted speed and quota savings need clarification.
(HN 370 points · 244 comments · Google launch and Antigravity announcement)
On a Xiaomi TV box with 1.9 GB of RAM, the author reports Qwen3.5-2B generating 7.7 tokens per second in chat. About 1 GB of the device’s memory is normally usable by the process.
The TV-box notes describe loading parts of larger models from flash as needed. The ready-made model files require the project’s patched llama.cpp rather than the stock program.
On a first-generation iPad Air, the author reports 23.7–25.8 tokens per second for Bonsai-1.7B in a fixed three-turn test. The iPad setup requires a jailbreak, and its larger models use short contexts to fit.
Caveat: These are the author’s device-specific measurements, and the iPad’s simple three-question check is too small to establish general answer quality.
Cornell’s report describes a November 9 Nature Energy study using electricity-price records, transmission plans and utility filings across four major US grids.
For 2025, Cornell reports $8.25 billion in infrastructure costs linked to data-center growth, of which data centers paid $3.12 billion. It says the remaining $5.13 billion fell largely on residents.
The paper’s abstract says rural communities saw energy-burden increases 76% larger than urban communities. It also reports a median 47% of regulated tariff costs allocated to residential users, a separate measure from Cornell’s infrastructure total.
Caveat: The full paper is subscription-only, so the methods behind these estimates could not be checked in full here; they cover data centers generally, not only AI workloads.
(HN 94 points · 85 comments · Cornell report and paper abstract)
Nvidia’s board announcement says Jensen Huang resigned as president, CEO and board member effective immediately. Kress remains chief financial officer while serving as interim CEO.
The board named Harvey C. Jones its first chair and says it will consider both internal and external candidates for a permanent CEO.
The Verge reports the resignation followed reports of two extramarital relationships, including one involving an alleged Chinese spy. That allegation is not a finding of guilt.
Caveat: The board says it expects the product roadmap and customer commitments to continue, but its permanent leadership decision remains open.
(Techmeme · The Verge · Nvidia board announcement)
Anthropic IPO site — the offering timeline and roadshow began before this cycle; the site provides financial information but no newly verified developer-facing change.