Amazon is reportedly planning to sell up to $8 billion worth of Nvidia AI GPUs to investors, a move that would help the company offload inventory and free up cash. The chips, used in data centers for AI workloads, come as Amazon faces stiff competition from Microsoft and Google in the cloud computing space. The potential deal, if it goes through, would mark a significant moment in the tech industry's ongoing GPU supply crunch.
Meanwhile, the broader AI investment picture in the U.S. is coming under scrutiny. A Goldman Sachs report projects AI investment will reach 1.9 percent of GDP in 2026, though most of that spending flows to data center construction and GPU sales rather than actual AI deployment. The analysis draws a parallel to the railroad boom of the late 1800s as the last comparable industrial buildout, and notes that tech's contribution to nominal GDP has remained flat for two years.
On the practical side, Ford Pro and Rockwell Automation are putting AI to work in the field. An AI copilot built by Rockwell now guides dealership technicians through complex engine removals, cutting training time from nine months down to three. Machine downtime at one plant dropped 33 percent, and a BCG report suggests AI could boost field-service productivity by 20 to 30 percent and profit per technician by up to 80 percent.
In China, Zhou Yuxiang is chasing an industrial AI breakthrough through his company Black Lake Technologies. He hired food delivery riders to sell AI factory-management software to manufacturers. The company says it has been profitable since 2025 with revenue growing over 60 percent annually. Zhou argues that AI has so far only replaced individual tasks rather than restructuring any entire industry.
Research into how well AI actually performs is also advancing. Datalab launched OmniExtractBench, a benchmark designed to reduce bias and opacity in existing evaluation methods. It uses 620 documents and scores systems by comparing their JSON output against a gold standard, checking for errors like misreads and fabricated data. In testing, Datalab's accurate mode led with 93.85 percent accuracy.
Another study, ReLiveGym, tests long-lived AI agents over simulated weeks of real-world news, market, and social media streams. Across eight language models, the research found that action timing and feedback use matter significantly for how well agents perform over extended periods. Separately, ScholarCatalyst tested whether AI can find inspiring research papers and found that agentic search performed no better than basic embedding retrieval, with even a Claude Fable 5.1 agent reaching only modest recall scores.
Meta is opening up its hardware ecosystem, releasing free SDKs for building Muse AI gadgets. The October 2, 2026 launch includes an ESP32 Device SDK and a Linux Device SDK under an Apache 2.0 license, plus a free USB-C Muse Home Link device for U.S. Muse subscribers. Builders can use kits like a Raspberry Pi 5 or ESP32 board to create custom projects that connect to displays, sensors, and home devices.
On the creator side, PewDiePie unveiled the Ajax AI model, designed to run on home PCs and described as uncensored. The creator claims OpenAI banned him twice over the model distillation techniques used to build the product. The launch adds another player to the growing field of personal AI models.
Key Takeaways
- Amazon plans to sell $8 billion of Nvidia AI GPUs to investors, aiming to reduce inventory and free up cash amid competition with Microsoft and Google.
- Goldman Sachs projects U.S. AI investment will hit 1.9 percent of GDP in 2026, though most spending goes to data centers and GPUs rather than real AI use.
- Ford Pro and Rockwell Automation's AI copilot cut technician training from nine months to three and reduced machine downtime by 33 percent at one plant.
- Black Lake Technologies, founded by Zhou Yuxiang, claims profitability since 2025 with revenue growing over 60 percent yearly selling AI factory-management software in China.
- Datalab's OmniExtractBench uses 620 documents and a gold-standard scoring system, with Datalab's accurate mode achieving 93.85 percent accuracy in testing.
- ReLiveGym tests long-lived AI agents over simulated weeks, finding that action timing and feedback use are critical to performance across eight language models.
- ScholarCatalyst benchmark reveals agentic search performs no better than basic embedding retrieval when AI tries to find inspiring research papers.
- Meta open-sourced firmware and SDKs for Muse AI hardware gadgets under Apache 2.0 license, including an ESP32 SDK, Linux SDK, and a free USB-C Muse Home Link device.
- PewDiePie launched the uncensored Ajax AI model for home PCs, claiming OpenAI banned him twice over model distillation used to build it.
- A BCG report estimates AI could lift field-service productivity by 20 to 30 percent and profit per technician by up to 80 percent.
Amazon plans to sell $8bn of Nvidia AI chips
Amazon wants to sell $8 billion worth of Nvidia GPUs to investors. The GPUs are used in data centers for AI work. This sale would help Amazon reduce inventory and free up cash. The company is investing heavily in AI while facing competition from Microsoft and Google in cloud computing.
Amazon considers selling $8bn of Nvidia AI chips
Amazon is thinking about selling about $8 billion worth of advanced Nvidia chips for its cloud computing business. The company has been investing heavily in AI and machine learning. The report did not give details on the potential deal. This would be a major development in the tech industry if it happens.
Datalab launches OmniExtractBench to fix benchmark problems
Datalab released OmniExtractBench, a structured extraction benchmark built to fix bias and opacity in existing benchmarks. It uses 620 documents from four sources and scores systems by comparing their JSON output to a gold file. The scorer checks six types of verdicts per value, including matched, misread, and fabricated. Datalab's accurate mode led with 93.85 accuracy in testing.
ReLiveGym tests long-lived AI agents over weeks
Researchers created ReLiveGym, a diagnostic environment for testing long-lived AI agents. The agents act over simulated weeks of real-world news, market, and social media streams. Across eight base language models, the study looked at how model choice and harness design affect performance. Results show that action timing and feedback use are important for long-lived agent design.
Zhou Yuxiang chases China's industrial AI breakthrough
Zhou Yuxiang founded Black Lake Technologies and hired food delivery riders to sell AI to factories. His company makes factory-management software with AI agents that can schedule production on their own. Black Lake says it has been profitable since 2025 with revenue growing over 60% a year. Zhou believes AI has only replaced tasks so far and has not yet restructured any industry.
Ford and Rockwell Use AI to Speed Up New Mechanic Training
Ford Pro is using AI to guide dealership technicians through complex engine removal work. Rockwell Automation built an AI copilot that cut technician training from nine months to three. The copilot draws on decades of engineer expertise and plant data. Machine downtime at one plant fell 33 percent. A BCG report says AI could lift field-service productivity 20 to 30 percent and profit per technician by up to 80 percent.
Meta Releases Free SDKs for DIY Muse AI Hardware Gadgets
Meta open-sourced firmware and device SDKs for building Muse AI hardware gadgets on October 2, 2026. The release includes an ESP32 Device SDK and a Linux Device SDK under an Apache 2.0 license. Meta also offered the Muse Home Link, a free USB-C device for U.S. Muse subscribers. The gadgets let Muse connect to displays, buttons, sensors, and home devices. Builders can use kits like a Raspberry Pi 5 or an ESP32 board to create their own projects.
ScholarCatalyst Tests How Well AI Finds Inspiring Research Papers
ScholarCatalyst is a new benchmark that tests whether AI can find prior research papers that inspire new work. Researchers had 184 lead authors label 207 papers that helped their projects. The study found that agentic search performed no better than basic embedding retrieval. Even a Claude Fable 5.1 agent reached only modest recall scores. The results show that AI still needs better training to match a scientist's intuition for finding key research.
Report Examines How AI Investment Is Shaping the U.S. Economy
A report notes that AI investment in the U.S. is projected to hit 1.9 percent of GDP in 2026 according to Goldman Sachs. Most of this spending goes to data center construction and GPU sales rather than actual AI use. The analysis says tech's contribution to nominal GDP has been flat for two years. The article also discusses how government software price adjustments may undercount real costs. The railroad boom of the late 1800s was the last comparable industry buildout.
PewDiePie Launches Uncensored Ajax AI Model for Home PCs
PewDiePie unveiled the Ajax AI model designed to run on home PCs. The creator says OpenAI banned him twice over model distillation used to build the product. The model is described as uncensored. The article reports on the launch of this new AI product built for personal computer use.
Sources
- Amazon seeks to offload $8bn of Nvidia chips to investors
- Market Chatter: Amazon Mulls Offloading Nvidia's AI Chips Worth $8 Billion
- Datalab Introduces OmniExtractBench to Fix Bias and Opacity in Extraction Benchmarks
- ReLiveGym: Evaluating Long-Lived Agents over Weeks of Replayed Reality
- Zhou Yuxiang hired food delivery riders to sell AI to factories. Now he’s chasing China’s industrial AI breakthrough
- Ford and Rockwell Fast-Track Rookie Technicians With AI
- Meta Open-Sources Muse Gadget SDKs for DIY AI Hardware Devices
- ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research
- Premium: How Has AI Changed The Economy?
- PewDiePie unveils ‘uncensored’ Ajax AI model built to run on home PCs — creator says OpenAI banned him twice over model distillation used to build his product
Comments
Please log in to post a comment.