8news

Tech • AI • Robotics

VIDEO
ENFR
TodayShortsTop StoriesYour topicFor youTopicsAll videosYT channelsArchivesSearchFavorites

OpenAI Jalapeno claims 104x efficiency as Nvidia eyes Hugging Face

AIFriday, August 28, 2026· 11 videos

Briefing

Audio player
0:00 / 0:00

OpenAI touts Jalapeno efficiency leap

OpenAI published benchmark claims for its custom Jalapeno inference ASIC, developed with Broadcom, arguing it outperforms Nvidia GB200/GB300 systems on power efficiency and latency. In public InferenceX tests, the company said Jalapeno delivered 1.5x to 1.9x more work per watt at peak throughput. The most eye-catching figure was 104.3x throughput per kilowatt on DeepSeek R1, with similarly large gains on GPT OSS 120B and Kimi K2.5 under a matched decoding-speed setup. OpenAI stressed the comparison was normalized by power, not chip count, making the result more about inference economics than raw absolute speed.

Nvidia nears Hugging Face deal

Nvidia is reportedly close to acquiring Hugging Face for roughly $12.9 billion to $13 billion, a move that would extend its influence from accelerators into the open-model distribution layer. Hugging Face has become a central repository for about 3 million models, 1 million datasets and 1 million Spaces applications. The prospect raises concentration concerns because Nvidia already dominates critical training and inference hardware while also shipping its own model families such as Nemotron. It also sharpens questions around European tech sovereignty, given Hugging Face's French roots and strategic role in open-source ecosystems.

OpenAI revives AGI by 2026

OpenAI executives said the company could reach its internal definition of AGI before the end of 2026, putting a concrete timeline on a historically slippery concept. Sam Altman said the target is a highly autonomous system that outperforms humans at most economically valuable work, while Mark Chen put progress at roughly 80%. The internal program most associated with that claim is Astra, which Jakub Pachocki said can already function as an automated AI research intern inside OpenAI's codebase. The company has not yet released outside validation, a technical report, or independent benchmarks to substantiate the milestone.

Claude and ChatGPT target offices

Anthropic and OpenAI both expanded workplace automation features, underscoring a race to become the default interface for office execution rather than chat alone. Claude Co-work gained a built-in browser in the desktop app, letting the assistant navigate websites, click, type and use selectively imported signed-in sessions. ChatGPT Work added trigger-based automations and deeper multi-app flows across Gmail, Slack, Calendar and Google Drive. Together, the releases push AI products toward authenticated, cross-tool task completion with the user retained as final approver.

ChatGPT Work becomes document operator

ChatGPT Work is being positioned as a production layer for routine white-collar operations, especially where information is fragmented across enterprise apps. With Google Drive connectivity, it can turn long meeting transcripts into structured Google Docs and matching Google Slides, preserving formatting and flagging missing owners or deadlines instead of inventing them. Other examples include pulling context from Gmail and Slack to complete coordination tasks such as event confirmations. The broader pitch is that value comes less from prompting and more from workflows, permissions, templates, memory and guardrails.

Musk's OpenAI lawsuit tests governance

Elon Musk's lawsuit against OpenAI is emerging as a consequential governance fight over whether an organization founded around public-benefit rhetoric can evolve into a profit engine. Musk, an early backer, is reportedly seeking $130 billion in damages tied to OpenAI's structural shift. Beyond the courtroom, the dispute mirrors a wider industry argument about who captures value from frontier systems and how mission commitments should constrain commercialization. It lands as AI tools are already reshaping translation, legal support, coding assistance and other repeatable cognitive work.

Hangzhou deploys traffic robot patrols

Hangzhou has put autonomous T2 traffic robots from Subcon Information on active street duty, marking a visible expansion of robotics into public-service enforcement. The wheeled unit uses cameras, radar and onboard computing to spot e-bike riders without helmets, unauthorized passengers, stop-line crossings and jaywalking. Rather than issue fines, it delivers spoken warnings and reproduces standard traffic-police gestures with seven-jointed arms synchronized to signal cycles. Subcon says 15 robots have been deployed since May, generating more than 170,000 warnings.

Open models fuel sovereignty push

Concern over the future ownership of Hugging Face has intensified a wider push for open and locally deployable AI across Europe. The argument is that open models let developers run systems on their own machines, inside private infrastructure or even in robots, rather than depending entirely on US cloud platforms. That vision was reinforced by attention on Reachy Mini, a robot from Pollen Robotics, the French company acquired by Hugging Face, as a practical embodiment of programmable, open software. The editorial divide is increasingly strategic: control of models, tools and deployment venues may matter as much as model quality itself.

Videos covered

Previous briefings · AI