AI news on October 21, 2026

1. 0100:05 UTC

   Griffin calls AI 'almost certainly' a bubble as compute demand stays strong

   4 outlets · first: investinglive.com Leadership

   Machine summary Citadel founder Ken Griffin said at a New York conference that investors are "almost certainly" in an AI bubble, pointing to the dot-com era, according to reports of his remarks. He pointed to firms that are still determining their planned spending on the technology and the way it will transform the economy. Forexlive reported that the same day SpaceX posted an AI revenue beat, Super Micro received an upgrade and Elon Musk said compute remains scarce.

   1. Per Forexlive, Griffin argued that sweeping economic change is typically bound up with bubbles, pointing to the run-up to the dot-com crash, when certain business models turned out to be fanciful.
   2. MarketWatch reported that Griffin was not telling investors to dump AI stocks, and Forexlive said his framing stopped well short of a call to exit.
   3. Forexlive reported that on the same day SpaceX delivered an AI revenue beat, Super Micro received a surprise upgrade and Elon Musk said compute remains scarce.
   4. Moneycontrol reported that Deutsche Bank chief executive Christian Sewing played down Griffin's AI bubble fears.

   Reported by investinglive.com, Seeking Alpha, moneycontrol.com and marketwatch.com.

   • SpaceX
   • Elon Musk
   • Deutsche Bank
   • Citadel
   • Ken Griffin
   • Super Micro

2. 0201:11 UTC

   Coverage recordforbes.com.au: “OpenAI opens door for payments to Australian creatives as part of AI data centre drive”

   2 outlets · first: forbes.com.au Regulation

   The outlets that reported this event are listed with links to their reports; no summary of ours passed our automated checks for it, which says nothing about the event (how they work).

   Reported by forbes.com.au and Forbes Australia.

   • OpenAI

3. 0300:00 UTC

   TraceWeave claims flat memory use for long-context on-device agents

   2 outlets · first: en.prnasia.com AI research Mostly press releases

   Machine summary TraceWeave says its Phase-TITAN post-Transformer architecture keeps memory use constant as context length grows, claiming 0.04 GB for a 1B model at a 1M-token context versus tens of gigabytes or more for a comparable Transformer key-value cache. The company reports internal RULER and BFCL benchmarks showing performance on par with same-size and larger Transformer models, and says agents using it can call tools and decide next steps without cloud calls.

   1. TraceWeave is led by co-founder and CEO Jaeyoung Park and builds what it calls Phase-TITAN, described as a post-Transformer architecture that keeps memory use constant as context length grows.
   2. TraceWeave claims that with quantization disabled and a 1M-token context, the key-value cache of a comparable Transformer takes up tens of gigabytes or more, whereas its 1B Phase-TITAN model takes 0.04 GB.
   3. According to TraceWeave, internal benchmarks it ran on RULER and BFCL indicate that Phase-TITAN performs comparably to Transformer models of equal size as well as larger ones.
   4. TraceWeave says it develops its own models rather than building on existing open-source models.
   5. The company says it completed a paid pilot in mobility and is pursuing commercialization across robots, drones and smart glasses.

   Reported by en.prnasia.com and thailand-business-news.com.

   • Naver
   • NAVER D2SF
   • Phase-TITAN
   • On-device AI
   • TraceWeave
   • RULER and BFCL

4. 0413:17 UTC

   Schneider Electric publishes modular AI factory blueprint with NVIDIA

   2 outlets · first: manilatimes.net Product launch

   Machine summary Schneider Electric announced its RapidScale Modular Data Center Reference Design 115 on 21 October 2026, a prefabricated blueprint for high-density AI infrastructure that integrates the NVIDIA Vera Rubin platform with liquid cooling and smart temperature control. The company said the design bundles power, cooling, IT, software and lifecycle services into a repeatable delivery model. Schneider Electric estimates global data center capacity will reach 300 GW by 2030, up 170 GW, with AI workloads making up about 65% of demand by then.

   1. Schneider Electric announced the RapidScale Modular Data Center Reference Design 115 on 21 October 2026, delivered as a Berlin-issued GlobeNewswire release.
   2. The reference design combines the NVIDIA Vera Rubin platform with liquid cooling and smart temperature controls, according to the company.
   3. The blueprint integrates power, liquid cooling, IT, software and lifecycle services into what Schneider Electric describes as a scalable, industrialized delivery model.
   4. Schneider Electric estimates global data center capacity will grow by 170 GW by 2030, reaching 300 GW total, about 2.3 times the 2025 installed base.
   5. The release cites skilled-labor shortages, grid-capacity pressure and regulatory compliance as challenges facing developers using traditional stick-built construction.

   Reported by manilatimes.net and markets.businessinsider.com.

   • NVIDIA
   • Schneider Electric
   • GlobeNewsWire
   • NVIDIA Vera Rubin
   • RapidScale Modular Data Center Reference Design 115

5. 0502:40 UTC

   HPE Spaceborne Computer-4 to run language models on the Moon

   2 outlets · first: Interesting Engineering AI research

   Machine summary According to HPE's October 20, 2026 announcement reported by Interesting Engineering, the HPE Spaceborne Computer-4 module scheduled to land near the lunar south pole in late 2026 will run large language models using retrieval-augmented generation alongside high-performance computing benchmarks. The experiments also cover image analysis and processing of LiDAR data on a lightweight module pairing a low-power CPU and GPU in a fanless design with Kioxia flash storage. HPE has not disclosed processing speeds or expected benchmark results, the reports state.

   1. HPE announced HPE Spaceborne Computer-4 on October 20, 2026, describing it as the first high-performance computing benchmarks to be run on the Moon's surface, according to the company's announcement as reported by Interesting Engineering
   2. The system is scheduled to fly aboard Astrolab's FLIP rover in late 2026 under NASA's Commercial Lunar Payload Services (CLPS) initiative, with the spacecraft expected to land near the lunar south pole
   3. Spaceborne Computer-4 is a lightweight, self-contained module with a low-power CPU and GPU, a fanless design, and solid-state flash drives using memory from Kioxia, the reports state
   4. Planned workloads include HPC benchmarks, large language models using retrieval-augmented generation, image analysis, LiDAR data processing and laser-based mapping
   5. HPE has not disclosed processing speeds or expected benchmark results for the lunar tests, according to Interesting Engineering

   Reported by Interesting Engineering and hothardware.com.

   • NASA
   • Kioxia
   • LiDAR
   • FLIP rover
   • Astrolab
   • Commercial Lunar Payload Services

Each summary and its key facts are a machine summary of the reports linked under it; the outlets, times and counts come from those reports' own records. How a rundown is built.