Samsung co-leads a €200m-plus round for a Dutch inference-chip startup
Euclyd, founded in 2024 at the High Tech Campus in Eindhoven, announced on 15 September a Series A of “more than €200 million” co-led by Samsung, Somerset Capital Partners, the EQT-managed Scaleup Europe Fund and Innovation Industries. Peter Wennink, ASML’s former president and CEO, joins as chairman of the board. Worth noting how the number travelled: the company states only a euro floor and discloses no valuation, while DCD reported the round at $231m and CNBC at $230m. The money behind it is the part an EU buyer should read twice — participation from Denmark’s state Export and Investment Fund, imec.xpand, the Brabant Development Agency and Quadri, while one of the co-leads, the EQT-managed Scaleup Europe Fund, is described in Euclyd’s own release as bringing “together public and private capital.”
Nothing is shipping. The roadmap is two products, craftwerk and the craftwerk station, which the company calls “the world’s first agentic AI silicon” and “the world’s lowest-power exascale AI factory” — its own language, with no independent measurement behind either phrase and no silicon in customers’ hands. CNBC reports first systems rolling out in 2028. The money goes to engineering headcount and the silicon roadmap, ahead of what the release calls commercial deployment across “enterprise, sovereign, and hyperscale AI markets.” Treat it as a supply-side option on cost-per-token that does not exist yet rather than as a procurement choice. What it does tell you is that European public money is now betting on the inference layer specifically, not on model labs.
The same open-weight model costs 11× more depending on who serves it
Artificial Analysis added five independent provider measurements for DeepSeek V4.1 Flash across 14 and 15 September — Fireworks, Parasail and Databricks on the 14th, Novita and Inco on the 15th — bringing it to seven measured providers, after Baseten on 11 September and DeepSeek’s own API on the 10th. One model, one set of weights, and the spread is enormous: cost per task runs from $0.19 to $2.18, roughly 11×, while output speed runs from 96 to 545 tokens per second. Artificial Analysis puts the speed gap at 470% and the blended-price gap at 4.9×. These are its own measurements at a 10,000-input-token workload, taken over a trailing window, so the figures move.
The useful detail is that the two price metrics disagree about who is cheap. On blended price per million tokens the lowest is $0.08; on cost per task that same provider ranks fourth, at $0.34 — because how much a provider charges per token and how many reasoning tokens its configuration emits are separate variables, and only the second one shows up on your invoice. Served context also differs for identical weights, from 524k tokens on one provider to 1.05M on another. If your procurement compares models on a published price list, it is measuring the wrong thing; the serving decision is worth more than the model decision here.
A 10M-parameter probe turns 27 other vendors’ models into self-monitoring ones
inclusionAI published 27 SingProbe checkpoints on 14 September, extending to third-party base models a technique it had only shipped for its own two Ling models thirteen days earlier. The covered set is broad — gpt-oss-20b and 120b, three Gemma 4 variants, two Llama releases, GLM-5.2 and 5.3, MiniMax, a DeepSeek Flash checkpoint, Step-3.7-Flash, a Hunyuan checkpoint and fourteen Qwen models. Rather than running a separate guard model alongside the request, SingProbe taps the base model’s own hidden states mid-generation (layers 20, 41 and 62 on the 27B variant) and scores query intent, response safety and hallucination risk per token from roughly 10M parameters.
The vendor’s own card claims decode overhead below 0.5% and a 0.03% false-positive rate on benign responses, with streaming-safety AUC ahead of an 8B guard baseline — and, to its credit, query-intent F1 of 0.8683 slightly behind the 0.8714 baseline it prints beside it. None of it is independently replicated. Two things stop this being a production dependency today: it needs trust_remote_code, and the serving integration exists only in personal forks of vLLM and SGLang rather than upstream. The licence is also asserted as Apache-2.0 in the card metadata with no LICENSE file in the repository. But a per-token audit signal at sub-1% overhead is a different economic proposition from an 8B model in front of every call, and that is worth a pilot.
Quick Hits
- AI Act × GDPR guidelines reach an EDPB plenary as a discussion item. The 123rd plenary meets remotely on 17 September and lists “Guidelines on the interplay between the AI Act and EU data protection laws” under agenda items for discussion — not, unlike the fining guidelines and the updated DSA/GDPR guidelines on the same agenda, for discussion in view of adoption. A draft exists and is simply not public: a state-of-play on it was already an information item at the 116th plenary on 10 March. So nothing binding lands this week, but the file is moving, and it is the file that decides how controllership sits between model provider and deployer. Contracts signed before it lands should keep that allocation renegotiable.
- Scotland’s parliament votes today on pausing data centres above 50MW. The Scottish Greens are using their Holyrood opposition debate to force a vote on halting new hyperscale proposals until a national strategy and a definition of “green data centre” exist, DCD reported on 15 September. The claim that 24 planned projects would draw 1.5 times Scotland’s peak national electricity demand is the Green Party’s own, and DCD attributes it to them without corroboration. The SNP’s national council has backed a moratorium, while DCD reports Prime Minister Andy Burnham telling parliament last week that his government would not be banning data centres. Either way, announced pipeline capacity in Scotland now carries a planning risk that has to be priced.
- An Italian device-security firm raised $270m with the EIB in its cap table. Exein, which secures embedded and robotic systems, says it raised the round at a $1.7bn valuation on 15 September led by Headline, with the European Investment Bank Group and KfW Capital alongside Goldman Sachs and Sofina — and claims the title of Europe’s most valuable cybersecurity scaleup, which is its own characterisation. Every growth figure quoted — 4× ARR, a thirtyfold valuation increase in two years, two billion connected devices — is the company’s own and unaudited. Its foundation models are slated for Q1 2027, which is to say they do not exist.
- Tencent shipped sparse-attention weights with no licence whatsoever. The Hunyuan team published Simple Attention Sparsification on 14 September: tiny learned gates (33–42M parameters, ≤81 MiB) that pick which KV blocks a frozen Qwen3-4B/8B/14B attends to, with a default 2,048-token sparse decode budget. The repository carries no licence tag and no LICENSE file, and the same is true of its GitHub counterpart. With no grant, default copyright applies — this is not deployable commercially as published, whatever the KV-cache economics look like. Serving it also needs an out-of-tree SGLang fork maintained by the authors.