Quick Hits — August 31, 2026
An afternoon sweep: Europe's biggest AI supercomputer deal lands at €387.8M on AMD silicon, two Chinese provinces push tokens to metered-utility status, and a small devblog argues the cleanest agent memory is the one you stop scoring.
Europe locks in LUMI-AI — a €387.8M, AMD-MI430X supercomputer for Finland. The EuroHPC Joint Undertaking signed a procurement contract with state-owned French integrator Bull to build LUMI-AI at a new CSC data centre in Kajaani, with deployment set for the second half of 2027. The system pairs AMD's next-generation Instinct MI430X GPUs with 256-core EPYC processors on Bull's BullSequana XH3500 platform, and is billed as delivering ten times the AI capacity of the existing LUMI machine. IBM handles storage, Nokia handles networking, and the rack draws fully renewable power with waste heat fed into Kajaani's district grid. It is the sixth AI Factory contract EuroHPC has signed this year — the message is that sovereign AI infrastructure is now a scheduled procurement program, not a one-off flagship.
China Telecom Hunan launches a "Token factory" — and shows the unit economics. At a Changsha launch on Monday, Hunan Telecom unveiled the "Zhiyun Hunan" platform alongside a 384-super-node domestic compute cluster and the company's TeleAgent personal/enterprise/on-prem product. The pitch is the metered-utility framing: pay-as-you-go, package, committed-volume and tiered pricing for AI services, with metering separated from billing. Hunan Telecom cited a third-party benchmark showing TeleAgent completed complex tasks in 21.1 minutes versus 43.6 minutes for the OpenClaw comparison product, while consuming roughly 38 percent as many tokens. That is the first hard token-efficiency number a Chinese telecom has published against an open benchmark — even if the test was vendor-supplied, the framing signals how seriously state-affiliated carriers are now treating tokens as the new voice minutes.
Sichuan's first "Token factory" goes live in Ya'an — with Southeast Asia as the first customer. The Tianjiang–Baize Token factory, a joint project between Tianfu Jiangxi Laboratory and the Ya'an economic development zone, came online this week with a September target of roughly 3 trillion tokens per month and a year-end goal of 10 trillion. Output is sold on a "no downgrade, no routing" basis — clients get the full-capability model on a dedicated concurrency queue, with cryptographic traceability from production to consumption. The immediate commercial hook is the export side: Ya'an Telecom has shipped the first batch of Chinese tokens to Southeast Asia and Belt-and-Road customers under China Telecom's TokenHub program. Tokens are starting to leave China the way mobile-phone minutes did a decade ago.
A local Rust agent dev deletes his "forgetting curve" — and writes the most useful agent-memory post of the month. Jan Paul Dahlke's blog post about his Eris agent walks through a real lesson: he built a three-tier scoring ladder for his Markdown vault with TTLs, decay constants and mention boosts, then realised it never measured importance — it measured string overlap with a slug he wrote himself. The cleaner design is two states (explicit commit, then rank at read time over plain text with embeddings he already had), and a Markdown vault as the source of truth. It is the rare post that is honest about deleting code instead of polishing it. Read it if you have ever shipped a "memory subsystem" that nobody — including the agent — ever reads the output of.
Sources: LUMI-AI contract (Fair EDIH) · AMD on LUMI-AI (AMD Newsroom) · Hunan Telecom Token factory (Voc.com.cn) · Hunan Telecom Token factory (Chinanews Hunan) · Sichuan Ya'an Token factory (Sichuan Online) · Ya'an Token exports to SEA (Sina Sichuan) · Eris agent memory retrospective (Eris System Dev Blog) · Eris agent repository (GitHub)
