AI Infrastructure

Paris – September 20, 2026 -- A newly published book, "Dialogue Between a Thinker and AI," has become the test object for a comparative experiment feeding identical text to five major artificial intelligence systems: OpenAI/Astra, Perplexity, DeepSeek, Google Gemini and xAI/Grok. The French and English print editions are now available, but the project's core output is a public meta-reading exercise designed to expose how differently large language models interpret the same source material.

Five AI systems receive an identical corpus for independent analysis

Santa Clara – September 20, 2026 -- Dnotitia Inc. has received back the first ASIC samples of its Vector Data Processing Unit (VDPU) from fabrication, with chip-level characterization now underway and silicon-based server evaluations scheduled for the fourth quarter of 2026. The company disclosed the milestone at AI Infra Summit 2026, held Sept. 15-17 at the Santa Clara Convention Center, where it showcased a server-scale VDPU architecture built for vector retrieval workloads in retrieval-augmented generation and agentic AI systems.

San Antonio – September 20, 2026 -- Frost & Sullivan has named Lenovo a 2026 Global Technology Innovation Leader in AI-powered sports event technology, citing the company's role as FIFA's unified technology partner for the FIFA World Cup 2026TM.

Frost & Sullivan cites FIFA AI Pro's reach across all 48 World Cup teams

The analytics platform, co-developed by Lenovo and FIFA, gave all 48 teams competing at the tournament the ability to query petabytes of competition data in plain language and receive insights within seconds. Alaa Saayed, VP of Digital Content Services at Frost & Sullivan, said the tool changes the competitive logic of the sport by democratizing access to advanced data analysis.

Las Vegas – September 19, 2026 -- Hyperscale Data, Inc. (NYSE American: GPUS) declared monthly cash dividends of $0.2708333 per share on its 13.00% Series D Cumulative Redeemable Perpetual Preferred Stock and $0.20833 per share on its 10.00% Series E Cumulative Redeemable Perpetual Preferred Stock.

Board sets record date of September 30, 2026 for dual preferred dividends

The record date for both dividend payments is September 30, 2026, with cash distributions scheduled for October 13, 2026. The payouts correspond to the 13.00% and 10.00% coupon rates set at issuance for the Series D and Series E preferred shares, respectively.

Stockholm – September 19, 2026 -- atNorth, the Nordic high-density colocation and built-to-suit data center provider, confirmed that Fredrik Jansson is stepping down from his role as Chief Strategy and Marketing & Communications Officer for health reasons and will transition to Senior Advisor.

Jansson exits operational duties but retains strategic ties to atNorth

Jansson will no longer hold day-to-day executive responsibilities but will continue advising the executive team and Board during a transition period. CEO Eyjólfur Magnús Kristinsson credited Jansson's strategic guidance and analytical rigor for strengthening key company decisions during his tenure.

Seoul – September 19, 2026 -- SK hynix Inc. has launched a new corporate venture capital brand, SK hynix Ventures, to expand strategic investments across the global AI ecosystem, the company announced at the inaugural 'SK hynix Ventures Day' held in Silicon Valley.

CEO Kwak Noh-Jung leads executive delegation at Silicon Valley launch event

CEO Kwak Noh-Jung and senior SK hynix executives met with representatives from global venture capital firms and leading startups to outline investment strategies targeting AI data centers and optical-based systems. The event brought together startups from the U.S. AI computing and data center sectors to discuss memory, computing and system software innovations tied to advanced AI models.

San Francisco – – September 19, 2026 -- Arcjet has launched agent runtime security, a new product designed to give engineering and security teams visibility and control over AI agents operating inside production systems. The platform targets agents that read and write to databases, process refunds, call APIs, and take autonomous actions across multiple systems on behalf of users.

Arcjet builds its product around three functions: observe, enforce, and audit

The observe layer lets teams ingest agent activity without changing application code, using existing OpenTelemetry tooling or, for Claude users, the Claude Compliance API. Arcjet connects actions across sessions into a single workflow view, capturing prompts, tool call parameters, session metadata, identity, and security decisions, and maintains an inventory of agents and applications running in a given environment.

Pasadena, Calif. – September 19, 2026 -- PrismML has released Bonsai 2 27B, a ternary-compressed version of a 27-billion-parameter multimodal model that runs at just 5.9 GB while retaining 98.2% of its full-precision counterpart's benchmark performance.

Ternary compression shrinks memory footprint by more than 9x

Built on Qwen3.8 27B, Bonsai 2 27B reduces the model's memory requirement to 5.9 GB, a reduction of over nine times compared with the full-precision original. The model contains 27.8 billion parameters and was trained using Google v5 TPUs.

Fremont – September 19, 2026 -- Delta Electronics announced a system-level infrastructure approach to accelerate deployment of AI factories built on the NVIDIA DSX AI Factory Platform, integrating power delivery from onsite energy sources down to the chip level to raise token output per megawatt.

Delta's 800 VDC architecture cuts conversion stages and hits 98% power efficiency

The company's 800 VDC power architecture reduces conversion stages between facility power and compute, delivering up to 98% power conversion efficiency while responding to GPU load transients. High-density In-Row systems can deliver up to 800 kW in a single power rack. At the board level, Delta's 800 VDC DC-DC distribution technology steps down directly to 50 VDC or 12 VDC with peak efficiency reaching 98.5%.

Beijing – September 19, 2026 -- WiMi Hologram Cloud Inc. (NASDAQ: WiMi) has developed a reinforcement learning-based framework to automate the design of quantum encoding circuits, aiming to cut the resource costs and inefficiencies that have limited quantum machine learning (QML) model performance.

WiMi replaces manual circuit design with automated search architecture

The new scheme uses reinforcement learning to generate task-specific encoding circuit architectures automatically, moving away from the heuristic, manually designed approaches that have dominated QML development. WiMi states the framework targets three persistent industry problems: poor task adaptability, low search efficiency, and insufficient handling of multiple design objectives simultaneously.