The Intake
The Intake — Monday, September 15, 2026
On the substrate
Euclyd raises €200M to build non-GPU agentic AI inference silicon
EUCLYD is an Eindhoven-based silicon startup. The company closed a €200M-plus Series A on September 15, 2026. Samsung led the round. Founders Bernardo Kastrup and Atul Sinha are Silicon Hive veterans; Peter Wennink — former ASML president and CEO — chairs the board. EQT's Scaleup Europe Fund, Somerset Capital Partners, and Innovation Industries co-invested. EIFO, imec.xpand, Brabant Development Agency, and Quadri also participated.
EUCLYD's core product is craftwerk — what the company calls agentic AI silicon. The design is a programmable ASIC built on processor-memory co-design. System-level optimization is part of the architecture. EUCLYD's stated claim is a lower per-token power profile than GPU-based inference. CWS 32 is a rack-level deployment built on craftwerk. EUCLYD positions it as an exascale-density option.
Neither craftwerk nor CWS 32 is currently available — this is a funded roadmap at Series A, not a shipping product. No near-term practitioner implication.
Shanghai AI Lab releases Atria Dawn Preview — 744B open-source agentic model, MIT-licensed, without announcement
If you've been running agentic workloads on proprietary APIs partly because there wasn't an MIT-licensed alternative at frontier-range benchmark levels, September 11 changed what's available.
Shanghai Artificial Intelligence Laboratory published Atria Dawn Preview checkpoints on HuggingFace on September 11 and 12, 2026. Both full-precision and FP8 variants are available. No blog post, pricing announcement, or press release accompanied the release. The MIT license applies to the published weights.
The model is a 744B-parameter mixture-of-experts built on the GLM-5.2 base. The technical report (arXiv 2609.15818) describes the training method as a Verifiable Experience Pipeline — tool-mediated interactions connected to executable environments with outcomes verified externally rather than by human preference labels. Author-reported scores across 16 tasks include DeepSearchQA 96.0, BrowseComp 92.5, and CyberGym 86.5. MLE-bench Lite 86.2, Terminal-Bench 2.1 78.3, and BFCL v4 77.0 were also reported. The context window is 256K tokens. Self-hosting runs on SGLang v0.5.13 or vLLM v0.23.0. A free preview API is live at api.atria-asi.ai; it accepts OpenAI-compatible Completions, Messages, and Responses requests.
If you're evaluating open-weight alternatives to paid frontier APIs, the MIT license and published checkpoints make this a concrete option to add to an evaluation sweep — bearing in mind the benchmark scores are author-reported and not yet independently verified.
---
For operators
Atria Dawn Preview is available via free API, no waitlist — evaluation and cost benchmarking decision
If you want to benchmark a 744B MIT-licensed model against your current paid-API stack, api.atria-asi.ai is live as of September 15 with no stated waitlist and no stated usage cost.
No data handling terms were published alongside the model release.
Two decisions are in scope: whether to include Atria Dawn Preview in an evaluation sweep alongside your current paid-API options, and whether to route development workloads through the free preview while it's available. If your development environment handles customer data or other material you'd treat as sensitive, the unpublished data handling terms are the boundary to confirm before routing anything through that API.
---