Google released Gemini 4 Hydrogen, a faster model that outscores its larger Argon sibling while costing less. At the default high thinking level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index versus Argon’s 50, reaching 54 at the new xhigh setting. API pricing is $1.50 per million input tokens and $7.50 per million output tokens.
Hydrogen offers low, medium, high, and xhigh reasoning modes, plus a one-million-token input and output window. Google is rolling it out globally across Gemini, Search, AI Studio, the API, business products, and Antigravity. Its Antigravity Orchestrator switches between Hydrogen and Carbon, using 24% less quota and finishing 19% faster than a single-model workflow.
Why it matters: Hydrogen gives Google a cheaper execution model between Gemini 3.8 Flash and Carbon-level reasoning, while Antigravity teams can use its orchestrator to cut quota consumption by 24% and processing time by 19% on Terminal-Bench 4.0.
Democrats opened an inquiry into how the Trump administration handled Nvidia’s chip sales to China. Rep. Richard Neal and Sen. Ron Wyden asked Commerce Secretary Howard Lutnick to preserve records, alleging the arrangement mixed export licensing, tariffs and political influence.
The letter says President Trump agreed to let Nvidia’s H200 chips reach China after the Taiwan-made chips first entered the United States, where a tariff took a cut, before Commerce issued an export license. It asks why Singapore-based Megaspeed kept buying chips after U.S. intelligence reportedly raised concerns about executive Alice Huang, and why other Asian companies listed no physical presence at their addresses.
Why it matters: Nvidia now faces scrutiny over whether its China export approval was influenced by its relationship with Trump, while Lutnick’s Commerce Department is being asked to preserve records that could clarify the decision.
Nebius plans to spend $40 billion in 2027, using customer prepayments and loans backed by customer contracts to finance its AI data-center buildout. Nebius ended the quarter with $7.4 billion in cash.
But its planned projects face delays. At Microsoft’s Vineland, New Jersey, data center, only just over half of contracted capacity is planned for delivery by year-end 2026; the rest is scheduled for Q2 2027, beyond the original 2025-26 window. DataOne is winding down temporary generators and switching to Bloom Energy fuel cells after New Jersey regulators fined the facility $1.1 million for violating federal law.
Nebius forecast 2027 revenue of $12 billion-$13 billion and pretax break-even by Q4. Its shares fell 17%.
Why it matters: Nebius expects prepayments and contract-backed loans to cover up to 85% of 2027 capital spending, making Microsoft’s delayed, up-to-$19.4 billion contract a financing risk: missed dates can trigger termination after a grace period if no alternative capacity is offered, plus financial penalties.
AWS customers can now serve models on Trainium through SGLang’s familiar API. Amazon made SGLang-Neuron publicly available as an AWS Deep Learning Container released by RadixArk with Annapurna Labs, giving teams a common server interface and generation APIs across hardware.
The plugin runs model execution through native PyTorch, with Trainium-specific attention, mixture-of-experts, and graph-runner components. It also supports prefill-decode disaggregation: a prefill server transfers the prompt KV cache to a decode server using NIXL over Elastic Fabric Adapter.
The quickstart targets Trainium3 EC2 instances and demonstrates serving openai/gpt-oss-120b with eight-way tensor and expert parallelism. AWS provides no pricing or performance results here.
Why it matters: SGLang-Neuron makes AWS Trainium a less disruptive option for teams already built around SGLang, shifting the decision toward hardware economics rather than serving-stack rewrites. The eight-way Trainium3 example points to large-model workloads, but AWS supplies no comparative performance or cost data.
Nvidia representatives privately told Rep. Bill Foster that the company could implement geotracking for advanced AI chips—the core requirement of the Chip Security Act, which aims to deter smuggling to China. Foster said the representatives never confirmed it in writing, but responded “yeah, of course we can do this” during technical discussions.
The admission contrasts with the Semiconductor Industry Association’s March 2026 opposition to the bill as “new, untested, and potentially infeasible.” Nvidia has also declined to say whether its Fleet Intelligence system could prevent diversion, though it can indicate the country where chips operate.
Foster plans stricter legislation next year requiring licenses that GPUs would verify before connecting.
Why it matters: Nvidia’s private acknowledgment weakens the industry’s central feasibility objection and gives Foster a stronger case for geographic validation, while the current bill would locate a smuggled GPU without disabling it.
An invoice extractor can return a structurally correct object—12,500 cents for a line reading EUR 1,250.00—because neither a parser nor JSON Schema checks whether the value matches the source text.
The gate validates the contract, then compares quoted fields and amounts with the invoice, returning accepted, needs_review, rejected or shape_error. Retry only shape_error, feed back up to five field-level reasons, and stop after three application attempts. Constrained decoding prevents schema violations but cannot prove a value matches evidence.
The examples use Python 3.10.12 and Pydantic 2.14.1; set OpenAI SDK max_retries=0 if you need that three-attempt budget to survive transport retries.
Why it matters: For teams managing OpenAI API costs, max_retries=0 turns a three-attempt extraction policy into a demonstrable ceiling; otherwise SDK backoff can expand it to nine HTTP requests.
Anthropic launched Claude Meetings, putting Claude into video calls as a visible participant. In beta for Team and Enterprise plans, Claude can answer questions, generate documents and whiteboards, and log notes and action items. It responds in text by default and accesses connected tools only when granted permission.
After a meeting, Claude saves its notes, transcripts, and generated documents to a dedicated page on claude.ai. Privacy controls prevent it from joining calls with external participants by default. Google Meet is the only supported platform so far, though Anthropic says more will follow.
Organizations using HIPAA or customer-managed encryption configurations cannot use Meetings yet. The feature extends Claude’s recently announced whiteboard and data-dashboard capabilities into an all-in-one office tool.
Why it matters: Anthropic has turned Claude into a live meeting tool for Team and Enterprise customers—not just a place to store post-call notes—while organizations using HIPAA or customer-managed encryption still cannot deploy it.