Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

SK pushes Ulsan AI data center toward 900 MW
09/14/26, 06:31 AM

SK Group’s Ulsan AI data center has expanded to nearly 900 megawatts, up from a 100-megawatt launch plan, according to Yonhap’s report from Ulsan Forum 2026. The AWS-linked project is estimated at 7 trillion won ($5.2 billion) and targeted for operation in the second half of 2027; SK Telecom’s infrastructure plan puts the broader Korean buildout at 15 gigawatts, with 5 gigawatts staged from 2029. Builders should file this as a capacity-and-tenant bet, not live supply: AI’s grid-cost exposure is decided by power, contracts, and commissioning—not announced megawatts.
Wire

China proposes five-part BRICS AI plan
09/13/26, 10:45 PM

China’s Xi Jinping proposed a five-initiative BRICS cooperation program on September 13, putting AI, trade, digital industry, smart manufacturing, and talent development on the same industrial-policy agenda. The official summit announcement says China’s AI initiative will support cooperation on developing and applying large language models, while Xi’s full speech calls for open-source development and an AI-empowered new-industrialization initiative. Builders targeting Global South deployments should file this as a market signal—not an operating standard yet—and watch whether open-weight compatibility, shared R&D, and regional supply-chain partners become procurement requirements, much as Mistral’s sovereign-AI economics show that local control can be part of the product.
Wire

ByteFuture raises $55M for local AI
09/13/26, 03:28 PM

ByteFuture announced it completed a $55 million financing in June, lifting total funding above $100 million, to scale Olares One and its open-source Olares OS, according to the syndicated financing release. ByteFuture’s product description frames Olares as hybrid infrastructure for local and cloud AI inference; the release says Olares One drew 816 backers and more than $2.3 million in 2026 Kickstarter pledges. For builders weighing private hardware against hosted inference, this is a distribution bet to benchmark—not proof that a personal workstation wins—alongside sovereign-AI costs in production.
Wire

Thrive raises $2B for workflow AI
09/13/26, 02:30 PM

Thrive Holdings announced more than $2 billion in new capital at a $12 billion valuation, taking its total capital raised above $3 billion since inception, according to its fundraise announcement. The company says it has deployed AI products across more than 70 businesses it owns and operates, with those products serving tens of thousands of customers; new capital will expand its accounting and IT platforms and fund a technical-and-regulatory platform for critical infrastructure. The bet is distribution, not another model: builders should watch whether embedding AI inside repeatable operating workflows produces durable deployment economics, the same proof gap surfaced in enterprise AI ventures’ push from capital to adoption.
Wire

Black Lake's AI coach wins 7 of 10 senior votes
09/13/26, 11:47 AM

Black Lake’s AI sales coach won the preference of seven out of 10 of the company’s most senior salespeople in a comparison against its top human sales leader. Fortune’s account says the Shanghai startup fed recorded sales conversations and location data into an agent, then used it as a coach. Operators should treat this as a test of judgment transfer, not proof of replacement: Nvidia’s agent-control-plane thesis matters only if teams can audit recommendations and protect the recordings that feed them.
Wire

AI agents make 2.7M messages a power problem
09/13/26, 11:44 AM

OpenAI announced a swarm of more than 10,000 agents that sent 2.7 million messages to solve a longstanding math problem. WIRED’s power analysis says a run like that likely consumes tens of millions of dollars in processing power, and notes an independent estimate putting a heavy Claude session above the energy needed to run two refrigerators. Builders should budget by agent fan-out and run length—not user prompt count—because Google’s €13B nuclear bet shows infrastructure is already being bought around workloads that have not arrived.
Wire

Hyundai feeds 7M vehicles into Atria AI
09/13/26, 11:41 AM

Hyundai Motor Group has put its Data Flywheel into full operation, connecting an annual sales base of 7 million vehicles to its Atria AI autonomous-driving program. In its September 12 announcement, Hyundai says 42dot is standardizing sensors across Hyundai, Kia, 42dot, and Motional, while a real-world Level 4 pilot with South Korea’s transport ministry is due in Gwangju by year-end. Builders should read this as a data-and-validation bet rather than a self-driving launch: Zoox’s commercial-exemption ceiling shows why deployment permissions can matter as much as model quality.
Wire

OpenAI hires Git AI's founders for Codex
09/13/26, 06:25 AM

OpenAI is hiring Git AI’s 2 founders, Aidan Cunniffe and Sasha Varlamov, for its Codex team to help businesses measure coding-agent performance and return on investment, according to Git AI’s announcement and The New Stack’s report. Git AI says its open-source telemetry runs on hundreds of thousands of developer machines and tracks which agent, model, and prompts produced code. Builders should keep independent attribution and cost logs: the case for pricing agents by completed work gets harder to audit when model vendors own the scorecard.
Wire

Meta's agent swarm beats 100 engineers on narrow tasks
09/13/26, 04:28 AM

Meta Chief AI Officer Alexandr Wang says an internal AI-agent swarm outperformed 100 human engineers on specific, well-defined tasks, according to Crypto Briefing’s account of Wang’s Startup School 2026 remarks. Wang described a deliberately plain stack—Markdown-file memory, cron scheduling, evaluation loops, and continuous feedback—so the comparison is a company claim, not an independently reproduced benchmark. Builders should read it alongside research showing multi-agent coordination can waste output tokens: test narrow, measurable loops before treating agent swarms as general engineering capacity.
Wire

Agnes opens a 33B multimodal preview
09/13/26, 03:27 AM

Agnes AI released a 33B-parameter open-weight multimodal preview with a 262,144-token context window under Apache 2.0, according to its Hugging Face model card. Its hybrid decoder uses 54 delta-rule recurrent layers and 18 global-attention layers, so only 18 layers carry a context-growing KV cache; the bf16 checkpoint is about 66 GB and fits on one H200 or H100. Builders should treat this as a local long-context option to benchmark—not a production shortcut: Agnes says its separate API checkpoint has a 1M-token context, so validate the preview against open-weight adoption’s demand-side denominator before committing hardware.