OpenAI’s Jalapeño Chip Rewrites Inference Economics as Google and Claude Move Deeper Into Work

OpenAI Turns Inference Hardware Into a Product Lever

OpenAI published the first measured results for Jalapeño, its custom inference accelerator, and framed the chip as the beginning of a multigenerational platform rather than a laboratory prototype. In testing across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5, the company says the system delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than its comparison systems. OpenAI plans to begin deploying the hardware inside its own compute infrastructure by year-end, while continuing to buy accelerators from Nvidia and other partners.

The most interesting part of OpenAI’s detailed Jalapeño benchmark report is the workload definition. Agent tasks amplify latency because each step waits on the previous one, so performance per user and per kilowatt can matter more than a chip’s isolated peak. OpenAI says AI also shortened chip development, helping the team reach tapeout in nine months and optimize selected kernels beyond existing human-written implementations.

“Jalapeño can serve more AI work per unit of power while also returning responses more quickly.” — OpenAI

Google Packages Legal AI Around Permissions, Not Prompts

Google Cloud launched Gemini Enterprise for Legal in preview, positioning it as a governed operating layer for law firms and corporate legal departments. The product combines reusable legal skills, connectors into systems such as iManage, NetDocuments, Microsoft 365, Everlaw, and CourtListener, plus agents for contract review, research, regulatory scanning, discovery, redaction, and drafting. Google says inherited document permissions and ethical walls remain in force, while customer data and playbooks are not used to train its foundation models.

The architecture matters more than the list of tasks. Google Cloud’s launch brief argues that foundational intelligence alone is insufficient in a profession defined by confidentiality, changing law, matter-level access, and traceable citations. The platform’s commercial test will be whether governance survives real integrations and whether practitioners can inspect the path from source material to completed work without reconstructing it manually.

“General-purpose AI, however capable, does not meet that standard on its own.” — Thomas Kurian, CEO of Google Cloud

Claude Connects Memory Across Conversation and Execution

Anthropic has unified Claude’s memory across chat and Cowork, removing the need to rebrief the execution environment after developing a project in conversation. The company is exposing retained topics so users can read, edit, or delete them, and memory updates as a conversation develops instead of waiting for a closing summary. According to TechCrunch’s report on the Claude memory update, the feature is enabled by default for Free, Pro, and Max plans across web, desktop, and mobile.

Anthropic says Claude will avoid storing categories such as health data, political beliefs, religion, race, and gender identity by default, with an opt-in for sensitive topics and notifications when those memories are saved. Some information, including government-issued identifiers and Social Security numbers, is excluded entirely. That policy turns memory from a convenience feature into a data-governance surface: what the assistant knows now changes what an action-taking agent can do without another explicit briefing.

Ledger Moves Agent Authorization Into Hardware

Ledger outlined a 2026 security stack for AI agents built around a hardware root of trust. Its roadmap includes device-backed agent identity, command-line tools and skills, human-approved intents, hardware-enforced spending or interaction policies, and proof that a unique person stands behind an agent. The company’s Device Management Kit is available now, and Ledger points to a MoonPay integration where an agent can propose a transaction while private keys remain on the device and a person confirms the action physically.

The threat model in Ledger’s agent-security roadmap is blunt: useful agents need credentials, files, communications, and money, while prompt injection can arrive through a webpage, document, or tool response. Software-only controls share a failure domain with the agent they are supposed to constrain. Hardware does not solve malicious planning or poor policy design, but it can create a separate signing boundary for consequential actions.

Universities Challenge the Scale-First Research Economy

USC Viterbi interim dean Gaurav Sukhatme used a conversation about AI education to make a broader research argument: universities should not try to beat industry at accumulating models, data, and compute. Their advantage is the freedom to question the architecture itself. In USC Viterbi’s account of the interview, Sukhatme asks whether a radically different design could reproduce the function of a large model with one-tenth, one-hundredth, or one-thousandth of the data.

That question lands on the same day OpenAI is promoting a tightly integrated stack. Industry’s path is to optimize the dominant paradigm across silicon, software, models, and products; academia’s opening is to find a paradigm with different resource assumptions. Sukhatme also argues that frontier research can move directly into classrooms and that ethical implications belong inside the technical curriculum rather than in a separate elective.

“What if you could do exactly what a large model does, but with an entirely new architecture that needed only a tenth of the data, or a hundredth or a thousandth?” — Gaurav Sukhatme, USC Viterbi

Scientific Publishing Confronts AI Writing at Scale

A preprint analyzed biomedical literature and estimated that nearly nine in ten papers published in December 2025 in a major database showed signs of AI-assisted writing. Nature’s report on the study emphasizes that the work has not been peer reviewed and that its detection method is more sensitive than earlier approaches. The authors estimated higher use in introductions and discussions than in methods and results, while outside researchers cautioned against generalizing the figure across all scientific literature.

The distinction between assistance and integrity failure is essential. Editing language is not equivalent to inventing evidence, particularly for researchers writing outside their first language. Yet model-shaped framing can propagate bias, and AI assistance inside results sections creates a sharper risk because fabricated details can masquerade as findings. The operational problem is provenance: readers need to know where AI contributed, which claims trace to data, and which human accepted responsibility.

“Whatever bias the LLM may have will just suddenly permeate the literature.” — Dmitry Kobak, quoted by Nature