DeepSeek has open-sourced DeepInfer, the engine it runs its models on, and for now it serves only DeepSeek-V4.1-Flash. If your GPU sits under a desk, this engine is not for you: it targets Nvidia SM100 data-centre GPUs and Huawei Ascend 950 chips. Also: Advantech and DEKRA, the breaches Spain’s AEPD counted and Intel’s D-series GPU.
DeepInfer ships with one model
Illustration generated with AI on our own machine.
What happened. DeepSeek open-sourced DeepInfer, the software it uses to run its AI models in production, and announced it on X on Tuesday, The Next Web reports. DeepInfer supports only DeepSeek-V4.1-Flash and runs on Nvidia’s SM100 data-centre GPUs and Huawei’s Ascend 950 chips. DeepSeek’s own benchmark figures used four SM100 GPUs for processing prompts and 64 for generating output. Bloomberg’s Saritha Rai reported there is no independent benchmarking of those gains yet.
What we think. We wait for an independent measurement before we judge the speed; you, on a single GPU, stay on your current engine. A benchmark with 64 data-centre GPUs generating says nothing about your card. The repository passed 8,000 stars by Tuesday afternoon, and stars are not a benchmark.
The good bit. DeepSeek says DeepInfer was first built mainly on vLLM, so the open code shows what they changed.
What you can do. If you already run DeepSeek models locally, weigh any gain against what you have in our DeepSeek-R1 local review.
Advantech and DEKRA put conformity in writing
Illustration generated with AI on our own machine.
What happened. Advantech signed a Memorandum of Understanding with DEKRA, Industrial Cyber reports. DEKRA independently assessed Advantech’s AFE-E630 industrial edge system, which received an Attestation of Conformity. According to DEKRA, its security functions meet IEC 62443-4-2 SL2 requirements.
What we think. We ask for the assessment report, not the press release; you, if you buy connected industrial equipment, ask for the attestation of that exact model. Advantech’s Miller Chang says compliance is becoming a critical consideration for customers. An attestation covers the AFE-E630, not the network you plug it into.
The good bit. DEKRA CEO Stan Zurkiewicz says companies need verifiable evidence of safety and cybersecurity in manufacturing, energy and transport, and a third-party assessment is exactly that.
What you can do. List the connected equipment on your site with its certification next to it; the EU AI Act conformity checklist helps with the AI part.
Spain’s AEPD counts 163% more breaches
Illustration generated with AI on our own machine.
What happened. The AEPD registered more than 4,500 data breach notifications between January and August 2026, a 163% increase on the same period last year, HuffPost Spain reports. In March it detected an incident at a single processor that caused a breach across several controllers. Researchers found more than 250,000 stolen Spanish ID cards in an incident affecting Chekin.
What we think. We look at suppliers first, because one processor exposed several controllers; you, if a software supplier holds your customers’ data, ask them this week where and how. In July the AEPD saw incidents at hotel booking software providers. The hotel that signed the contract still answers for its guests’ data.
The good bit. The AEPD describes the patterns it detects, so you know which kind of supplier to check first.
What you can do. Ask every supplier that processes personal data for its latest incident log, and read why GDPR and local deployment fit together.
Intel bets on memory
Illustration generated with AI on our own machine.
What happened. Intel revealed the full specifications of its GPU D-series, previously known as Crescent Island, at the Open Compute Project Global Summit, Shattered reports. It packs up to 480GB of LPDDR5X memory onto a single air-cooled PCIe card. XenoSpectrum’s breakdown shows the 480GB applies only to partner-built cards; Intel’s own card has 160GB.
What we think. We separate the headline figure from the card Intel actually sells; you, if you are planning a server, budget for 160GB, not 480GB. Intel says a single card can host models of up to approximately 300 billion parameters. Capacity is not speed, and Intel does not build the 480GB card.
The good bit. It is a standard air-cooled PCIe card, designed for the server chassis you already have.
What you can do. Before waiting for the D-series, work out how much memory your models need with the VRAM guide for local LLMs.
In one line
DeepInfer is open and real, but until someone measures it outside DeepSeek’s own servers, its speed is a promise.
Work with us
If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.