IT News Week 37/2026: EUR 261 of own AI capacity against EUR 7,640 in API fees, Article 4 of the AI Act without a fine, and what the slot counter really costs

This week the in-house stack takes centre stage for the second time, with a bill instead of benchmarks: a single production day at the gateway made the shape of this traffic visible, and that shape is the whole reason for the calculation. The second piece is duty rather than pleasure: Article 4 of the AI Act, applicable to every company since 2 February 2025, with no size threshold and no fine of its own. Both pieces follow the rule this newsletter runs on: numbers measured in the building, not taken from the press release.

Deep dive: EUR 261 against EUR 7,640, a month of agent traffic

The traffic has a shape

One production day at the gateway, all lanes: 15,914 requests, 603 million prompt tokens, of which 546 million came from the prefix cache, a hit rate of 90.6 per cent. Genuine prefill work stayed at 56.6 million, answer tokens 5.4 million. Input to output sits at 111 to 1, on average 37,904 tokens in and 341 out. A chat workload would be closer to 5 to 1. Over the month that adds up to 18.3 billion tokens.

Against that: EUR 76 of electricity and EUR 185 of depreciation for the cards, EUR 261 a month together. The post states what that covers, and it is the GPU box alone, with the gateway host, the embedding card, storage, backups and the operating work left out. Spread over 18.3 billion tokens that is 1.4 cents per million, 0.4 cents of it power. The same token mix at Claude Sonnet 5 list pricing costs EUR 7,640 a month if the provider recognises our repetitions as cleanly as our prefix cache does, and up to 133 times as much if it does not. A list price is not a measurement, it is claude.com/pricing from 10 September multiplied by our own token mix. Full breakdown, assumptions and lanes in the complete post.

What the calculation does not support

The API side is not a measurement but list price multiplied by our own token mix, and no head-to-head run against Claude Sonnet 5 was ever carried out, because there is no API key in the inventory. The cache-friendly variant also assumes the provider spots the repetitions as reliably as our automatic prefix cache does, which on their side needs explicit breakpoints and comes with a short cache lifetime. The sentence that survives is smaller than the headline: with this shape of traffic, long repeated context and little output, the marginal price per task is close to zero. That is precisely why we run tasks that would never amortise per task. The bottleneck today is not the KV pool but the slot counter, and next month's homework is testing MTP properly and reporting prefill and decode separately.

Assessment: The 29 to 133 times is not an argument against the cloud, it is an argument against per-task billing. Rare, short requests stay cheaper on an API and save you the operating work that is not part of the EUR 261 in the first place. Run agents with long context and the decision is not about token price but about utilisation, which sat below 30 per cent this month.

---

Deep dive: Article 4 of the AI Act, the duty without a fine

No offence, four routes

Article 4 requires providers and deployers of AI systems to take measures to support AI literacy of their staff, applicable since 2 February 2025, with no size or industry threshold. A company using ChatGPT, Copilot or Claude is covered. There is no fine for missing literacy as such, Article 4 is not in the sanction catalogue of Article 99. The Digital Omnibus, Regulation (EU) 2026/1744, widened that catalogue in July to cover Article 25(2) and (4), and left Article 4 where it was. The duty becomes costly along four other routes: information requests from market surveillance, an assessment of staff capability once a liability case is open, works-council consent duties under § 96a ArbVG, and the general duty of care under § 1313a ABGB. The full article with citations: Article 4 of the AI Act: what the AI-literacy duty asks of a company.

Austria: same regulation, different table of authorities

Austrian implementing legislation for the AI Act does not exist, and no market surveillance authority has been named. The RTR runs an AI service desk as an information hub, that is § 20c KOG and § 194a TKG, not market surveillance with sanctioning power. Waiting for an authority to send a letter means waiting for the wrong one. In a liability case it is an expert who asks whether the staff were capable for the specific task, and that question lands on the table automatically.

Assessment: Literacy is provable with three sheets, not with a training certificate: one line per person and measure, one line per AI system in use, one log entry per published piece of AI-assisted content. For a thirty-person company the effort is a few half days. Whoever holds those three sheets is also prepared for the questions actually asked in Austria.

Digest: Other news that mattered

Models and inference

• DeepSeek V4.1-Flash: 552 billion parameters, one eighth of the KV cache: model card published on 10 September, MIT licence, one million token context. The card names 552 billion backbone parameters while the shipped weight files add up to 763 billion. The architectural trick is called CED, twenty encoder and twenty decoder layers, 8 active parameters per token in prefill and 16 in decode. DeepSeek puts the persistent KV footprint at roughly an eighth of V4-Flash, which is their own measurement without an independent benchmark. A single workstation still cannot hold the model; what matters on multi-GPU boxes with an input-heavy workload is the KV footprint, because that is what decides concurrency.
• vLLM v0.29.0: Model Runner V2 becomes the default, with breakages: 594 commits from 277 contributors, released 9 September. Model Runner V2 is now the runtime for every model, V1 sunsets at v0.32. Ten legacy architectures were removed, python -m vllm.entrypoints.openai.api_server is deprecated in favour of vllm serve. New and directly relevant for us: two admission control thresholds, --max-num-queued-reqs and --max-num-queued-tokens, plus FlashInfer all-reduce as default, Mamba prefix caching with 9 to 25 per cent better time to first token, and expanded NVFP4 and MXFP4 paths. Review start commands and model lists before upgrading.
• No release from the usual suspects: Qwen, Google with Gemma, Mistral and Meta shipped nothing this week, the Hugging Face organisation profiles still show Flash-Next from 24 August as the latest Qwen build. A negative finding worth stating, because weekly planning otherwise runs towards an announcement that does not arrive.

Security

• Windows patch day: around 1,000 fixes, two of them exploited: CVE-2026-81963 is a link following flaw in the Windows update stack, CVE-2026-85880 a heap overflow over ALPC, both privilege escalation, both added to the KEV catalogue on 8 September. The patch count varies by counting method between 964 and 974, Krebs writes "nearly 1,000" with 113 rated critical. Rather than compare numbers, compare inventory.
• GitLab CE/EE: unauthenticated file read: CVE-2026-85706 is a path traversal, all releases from 18.7 affected, fixed in 19.1.8, 19.2.6 and 19.3.2. The 10.0 score comes from GitLab itself as CNA, while NVD still listed the record as "Received" and unrated when we checked. Added to the KEV on 11 September, which means exploitation is confirmed. Self-hosted GitLab gets patched today, not at month end.
• MikroTik RouterOS: two entries on one day: CVE-2026-86060 and CVE-2026-67277 sit in the SSH login handler, a username starting with a forbidden character lets you switch the policy mask and escalate. The MikroTik security page was unreachable when we checked, so the list of affected builds stays open. In a mid-sized network, routers are usually the devices nobody has looked at for years.
• The KEV is a lagging indicator: on 9 September CISA added CVE-2026-19490 for Citrix NetScaler, disclosed on 19 August, and CVE-2025-25249 for several Fortinet products, an entry from last year. Bind your patch window to KEV additions and you patch weeks late on average.
• Chrome: CVE-2026-87491: out-of-bounds write in V8, sandbox escape through a crafted page, fixed in 153.0.8010.36, in the KEV since 9 September.
• A reassuring negative finding for our own stable: no KEV entry and no security advisory we could read this week for Postfix, nginx, containerd, Kubernetes, Proxmox, vLLM, Ollama, PostgreSQL or Samba. Next.js 16.3.5 from 11 September is bug fixes only.

Regulation

• Cyber Resilience Act: reporting duty live since 11 September: manufacturers must report actively exploited vulnerabilities and severe incidents, an early warning within 24 hours, the full notification within 72 hours, a final report 14 days after the corrective measure is available or one month after the first notification. Reports go through the ENISA Single Reporting Platform, which started operating on the effective date, and on to the CSIRT of your main establishment. Open source software stewards are not caught until 11 December 2027. On the side: the German version of the Commission page is machine translated and says "Ratingagentur" where it means the Cyber Resilience Act, so quote the English text.
• Austria: no solid primary source on the network security act registration duty surfaced this week, the authority portal stays empty, so we leave it here and report once the register itself says something.

Hardware

• RTX Spark: NVIDIA silicon on Windows on Arm from October: shown at IFA, two configurations with N1X and up to 128 GB of unified memory, Lenovo attached its name to the Yoga Pro 9n. NVIDIA named no prices on the product page, and core counts are only circulating through secondary reporting, because the page itself would not parse fully. For inference the rule from last week's digest holds: memory bandwidth decides, and it is still unconfirmed in this class.
• AMD Ryzen AI Halo: USD 3,999 against USD 4,699 for the DGX Spark: AMD's own benchmark page puts its 128 GB Strix Halo machine against the same model, measured in May, USD 700 cheaper. The DGX Spark went from 3,999 to 4,699 dollars over the summer, with the memory market as the stated cause. Vendor-measured against vendor-measured, so read the comparison with care.

---

Sources for this edition

1. Own measurement: LiteLLM gateway statistics, vLLM metadata and power readings from 07 to 13 September 2026, written up in EUR 261 instead of EUR 7,640 in API fees.
2. CISA Known Exploited Vulnerabilities catalogue, version 11 September 2026, as JSON in the original, retrieved 13 September. Every CVE number and addition date in this edition comes from that file.
3. National Vulnerability Database for the individual records CVE-2026-85706, CVE-2026-86060, CVE-2026-87491 and CVE-2025-25249.
4. Project releases: vLLM v0.29.0, Next.js 16.3.5 and the model card DeepSeek V4.1-Flash.
5. Legal texts: Regulation (EU) 2024/1689 Articles 4 and 99, Regulation (EU) 2026/1744, plus the Commission CRA reporting obligations and the ENISA Single Reporting Platform, retrieved 13 September.

What we could not substantiate this week: the Austrian registration practice under the network security act, because the authority portal stays empty; the affected RouterOS builds, because the vendor security page returned 404 on retrieval; and the RTX Spark core counts, which we only had through secondary reporting.

From the blog

• EUR 261 of own AI capacity instead of EUR 7,640 in API fees — a month of agent traffic over LiteLLM and vLLM on four RTX 5090, measured at the gateway.
• Article 4 of the AI Act: what the AI-literacy duty asks of a company — why there is no fine, no Austrian authority, and which three sheets carry the proof.

---

Compiled on 14 September 2026. Sources: own gateway and power measurements, Hugging Face, GitHub (vLLM, Next.js), CISA KEV, NVD, Krebs on Security, ENISA, European Commission, NVIDIA, AMD, Anthropic pricing.