eukalyptusdesign · AI Campaign Months 2026

Local AI on equal terms.
Sovereign. Practical.

Workshops, prototypes and local AI hardware for teams that want to understand AI — instead of just renting it. In depth since 2023. Productive in the market since the end of 2024.

100 % positive client feedback GDPR / AI Act compliant Hands-on, on site
Book tech sparring View offers
Since 2009 + 2025
Cologne + Weimar
Design · Web · AI
Owner-managed
Apple Silicon
NVIDIA DGX Spark

In 2026, AI is no longer a question — the question is who you entrust your data to. Cloud AI is convenient. Local AI is sovereign. We build the second option with you. And yes: in many cases it will be cheaper.

🔒
Data stays
in-house.

GDPR and AI Act compliant not through legal construction, but through genuine on-premise inference. No data processing agreements, no CLOUD Act risk.

👻
We are done
when you are fit.

We come in, work shoulder to shoulder with your team and leave again. The goal is your independence — no lock-in.

🍎
Apple or NVIDIA
instead of rented servers.

Up to 1,200 GB/s memory bandwidth (Mac Studio M5 Ultra), manageable power draw, no data-centre noise. In many cases this makes local inference cheaper than cloud alternatives.

Who we are

Design DNA.
Tech depth.
Since 2009.

Founded in Cologne in 2009, also in Weimar since October 2025: Jessica Rudzki and Stephan Schug bring together what is rare in the AI market — decades of communication skills from design and web, and real technical depth in Python, TypeScript, local LLMs, RAG and fine-tuning.

With us, AI is not unleashed on organisations by data scientists. It is introduced by people who have spent decades designing, explaining and integrating software into real companies. That is the difference.

Generative AI
since 2023
Depth without marketing haste. Productive and revenue-generating in the market since the end of 2024.
Specialist areas
local LLMs · RAG · fine-tuning · WordPress · WooCommerce · Moodle · LearnDash · Python · TypeScript
Rooted in society
Adult-education courses, federal volunteer service, open AI cafés in Weimar and Brühl near Cologne, school-holiday camps in Erfurt, volunteer work — building AI skills beyond the tech bubble.
Campaign Months 2026

Three offers.
One promise.

Offer 01
AI workshops &
prototypes

From use case to working prototype — shoulder to shoulder with your developers.

  • Use-case identification & ideation with design-thinking tools
  • Hands-on implementation: Python / TypeScript / local LLMs / RAG pipelines
  • Iterative development in your team’s repository, transparently documented
  • Proof of concept with clear success criteria — no demo theatre
Request a PoC outline →
Offer 02
Specialised
AI training

Three tracks. Two hours to three days. Tailored to your organisation — not to our slide library.

  • Users track: prompting, tools, day-to-day work
  • Decision-makers track: use-case assessment, risk, governance
  • Developers track: architecture, evals, RAG, local models
  • Fulfilment of the AI literacy obligation (Art. 4 EU AI Act) can be documented
  • Always use-case-specific — no standard presentation
Request a training format →
Offer 03
Local AI hardware &
development

Your own Apple or NVIDIA hardware as a serious alternative to cloud GPUs.

  • Consulting, procurement, setup and onboarding from a single source
  • Unified memory — large models run entirely in memory
  • Power draw 2–9 W at idle, typically 25–55 W for single users (measured on Mac mini M4 and Mac Studio M1 Max)
  • MLX, llama.cpp, Ollama, vLLM (vllm-metal), oMLX, LM Studio, Unsloth Studio, EXO — mature toolchain, partly beta
  • Hybrid setups (DGX Spark + Apple Silicon) on request

Note on availability: Apple presented the new generation on 25 August 2026, with deliveries from 22 September 2026. Our AI boxes are based on the Mac mini M6 with 32 GB of memory (for 48 or 64 GB the M5 Pro variant that Apple continues to sell alongside is a prerequisite) or on the Mac Studio (18-core CPU / 40-core GPU) with 64 GB of memory (configurations with 256 GB of memory are available; a 512 GB configuration has been announced but could not be ordered at market launch and has not yet been priced by Apple). Delivery times are currently very volatile. Until delivery, or in case of shortages, we alternatively procure available NVIDIA DGX Spark or Apple refurbished setups. [As of: 19 October 2026]

Technical arguments →
For tech leads & CTOs

Practical.
No soft focus.

You know NVIDIA. CUDA has been the standard for almost two decades. We respect that — and still make the counter-offer. Not out of Apple conviction, but out of memory-bandwidth arithmetic.

“Your developers should understand and master AI — not need us. We come in, join in, and leave again.”
Mac mini M4 Pro next to a real apple — our first PoC hardware
Mac mini M4 Pro — our first PoC hardware. Chosen deliberately, and it convinced us straight away. From 273 GB/s memory bandwidth, 5 W at idle, usually well below 100 W under AI load — local, quiet, production-ready.
Memory bandwidth decides when answering
A language model first reads your input (prefill, compute-intensive) and then writes the answer word by word (decode, bandwidth-intensive).
DGX Spark
273 GB/s
strong at ingesting long documents
Mac Studio M5 Ultra
1,200 GB/s
strong at answering
For chat and drafting text, Apple Silicon is usually faster; with very long inputs, the DGX Spark catches up. We tell you which profile your use case has.
Memory capacity: which model fits which hardware
The question today is no longer “70B or not”. Modern models are either compact (26–31 billion parameters) or Mixture-of-Experts (MoE): per word they compute with only part of their parameters and are therefore fast, but they still need all of the memory.
MemoryExample hardwareSensible models (4-bit)
16 GBMac mini M4Gemma 4 E4B and 12B
24–32 GBMacBook Air M3, Mac mini M6Gemma 4 26B A4B (approx. 14 GB), Qwen3.8-27B (approx. 18 GB), Qwen3.6-35B-A3B (approx. 24 GB, from 32 GB)
64 GBMac Studio M1 Max, Mac mini M5 Prothe same models in higher quality (8-bit) and with long context
128 GBMac Studio M5 Max/Ultra, DGX Sparkgpt-oss-120b, Mistral Small 4, Qwen3.5-122B-A10B (approx. 65–75 GB each)
256 GBMac Studio M5 UltraMiniMax-M2.5 (229B, approx. 138 GB)
from approx. 400 GBCluster or cloudKimi K2.6, DeepSeek class
Our everyday models are Qwen3.8-27B and Gemma 4 (both Apache 2.0). We no longer recommend dense 70B models of the Llama generation: they need more memory and are slower (2.7 tokens/s on the DGX Spark, measured by LMSYS), and current 27–35B models catch up with them on many tasks.

Good to know: The Mac mini M6 tops out at 32 GB; for 48 or 64 GB you need the M5 Pro. Not every open model is freely usable commercially; we check the licence for you.

Our position: For PoCs and operation up to roughly 120B (MoE), Apple Silicon is often the more economical choice. LoRA fine-tuning up to approx. 30B also works on the Mac. Training large models remains CUDA territory. Honestly.
TCO: When does local AI pay off?
Four items count: hardware, setup, electricity, maintenance. Our prices are net.
  • Hardware: AI box “local” on an Apple basis from 2,750 €, on NVIDIA DGX Spark from 7,750 €. Setup and 2 hours of onboarding are included. For large models: Mac Studio M5 Ultra with 256 GB from 10,999 € (Apple price, incl. VAT).
  • Enablement: AI deep dive from 450 € (2 hours), from 850 € (half day) or from 1,650 € (full day) per group of up to 5 people. Other AI services: 120 € per hour or 960 € per day.
  • Electricity: depending on the device roughly 20–850 € per year (see “Power consumption”).
  • Maintenance and PoC: We offer proof of concept and maintenance individually.
Worked examples (purchase plus one training day, divided by the cloud costs saved):
  • AI box Apple: approx. 4,400 €. At 300 € cloud costs per month it has paid for itself after about 15 months.
  • AI box DGX Spark: approx. 9,400 €. At 600 € per month it is about 16 months.
The cloud costs are assumptions. We determine them in the initial conversation from your actual subscriptions and API invoices. Not included is the advantage that data that never leaves the house needs no data processing agreement.
Power consumption: What does local AI cost to run?
Local AI often runs around the clock, so idle consumption matters most.
DeviceIdleUnder AI loadCost/year*Type of figure
Mac mini M42–7 W25–45 Wapprox. 20–40 €measured, KI-Lab Weimar (1 user)
Mac Studio M1 Max9 W25–55 Wapprox. 35–50 €measured, KI-Lab Weimar (1 user)
Mac Studio M5 Ultra9 W (Apple)approx. 150–250 Wapprox. 110–170 €idle per Apple, load estimated
NVIDIA DGX Spark25–38 Wapprox. 60–100 Wapprox. 110–140 €NVIDIA, measurements by ServeTheHome, Tom's Hardware
Workstation RTX 5090approx. 70–120 Wapprox. 450–750 Wapprox. 560 €system estimated, card values from Igor's Lab
Workstation RTX PRO 6000approx. 150 Wapprox. 900 Wapprox. 850 €StorageReview (image generation)
*0.30 €/kWh, 24/7 at idle plus 2,000 hours under load. At household electricity rates (37 ct/kWh) the values are a good 20 % higher. The measurements apply to one user; with several simultaneous requests the load increases. GPU workstations also generate waste heat and fan noise, which in an office or treatment room is often the real exclusion criterion.
Software maturity 2026
  • Ollama (stable): The easiest entry. From version 0.40 it automatically uses MLX on Apple Silicon.
  • LM Studio (stable): Desktop app with a graphical interface, ideal for specialists without a command line.
  • llama.cpp (stable): The de facto standard, runs on Mac, Linux and DGX Spark.
  • MLX / mlx-lm (stable): Apple’s native framework and the fastest basis on the Mac, also for LoRA.
  • oMLX (young, active): Turns a Mac into a team server with parallel requests and an SSD cache for long contexts.
  • vLLM (stable on NVIDIA): Standard for many simultaneous users. On the Mac there is the still young plugin vllm-metal.
  • Unsloth Studio (beta): Fine-tuning via LoRA/QLoRA, mature on NVIDIA, possible on the Mac via MLX since autumn 2026.
  • EXO: Connects several devices into a cluster.
Honestly assessed: For single workstations the stack is mature. For team servers on the Mac it is usable but younger than vLLM on NVIDIA. That is why we test with your real load before recommending anything.
Hybrid setups: local where it counts, cloud where it helps
Local plus cloud: A local model (Qwen3.8-27B or Gemma 4) handles day-to-day business: emails, summaries, search in internal documents, drafts. Data on individuals, clients and patients stays in-house. Only demanding tasks without sensitive content go to a cloud model such as Claude. A routing rule determines what may leave the house.
  • GDPR: Local processing needs no data processing agreement and no third-country assessment. For the cloud share we arrange the DPA.
  • § 203 StGB (German Criminal Code): Holders of professional secrecy may pass data to external service providers only under narrow conditions. Local processing avoids the problem.
  • EU AI Act: The AI literacy obligation (Art. 4) applies to local and cloud AI. Documented routing shows which data goes where.
DGX Spark plus Mac Studio: EXO Labs measured up to a 2.8x speed-up with two DGX Spark and a Mac Studio M3 Ultra (test model Llama 3.1 8B, 8,192-token prompt). This mainly pays off with very long documents; the setup is demanding.

We do not sell a tool; we recommend the right topology, even if that is just a Mac mini.
For management & C-level

Ready to act.
Legally sound. 2026.

We bring your team into a capable, legally sound AI mode — tailored to your use cases, not off the shelf.

⚖️

AI Act & GDPR

Art. 4 EU AI Act requires measures for AI literacy among staff (since 2 February 2025; since the AI Omnibus of 27 July 2026 without a guarantee of a specific level of competence). High-risk obligations apply later (Annex III from 2 December 2027). Fines can threaten a company’s existence.

We deliver documentable training under Art. 4
🫥

Stop shadow IT

Your employees have long been using cloud AI — just not securely. Local AI solutions eliminate the GDPR risk structurally, not through bans.

Data does not leave the building — neither physically nor logically
📈

Competitiveness

41 % of German companies actively use AI (Bitkom 2026) — 59 % do not yet. We shape the start with you — local and sovereign.

Funding note: BAFA up to 80 % (eastern German states), application by 31 December 2026 · QCG
“Three days, twelve employees, one use case. After that your team is ready to act — documented, verifiable, AI-Act compliant.”
Strategy call (45 min) Request an AI competence check
How we work

Four steps.
One goal.

01
Understand 🔎

Use-case workshop, as-is analysis, hardware sizing. We listen before we build.

02
Build 🏗

PoC / prototype, local or hybrid, shoulder to shoulder with your development team. No demo theatre.

03
Enable 💪

Training for users, decision-makers, developers — documented, use-case-specific, AI-Act compliant.

04
Let go 🚀

We leave. Your team can. That was always the plan.

— Until you are fit.

Why eukalyptusdesign

Facts, not
promises.

Client feedback
100 %
Not “mostly positive”. In fact 100 % positive feedback across all completed projects and training sessions.
AI experience
since 2023
Depth in generative AI — not only since the ChatGPT hype. Productive and revenue-generating in the market since the end of 2024.
Owner-managed
since 2009
No investor logic, no scaling promises. Jessica Rudzki and Stephan Schug are personally involved in every project.
Rooted in society
adult education + volunteer service + open AI cafés
AI training for working people, federal volunteers and volunteers in civic roles — plus open AI cafés in Weimar and Brühl near Cologne and school-holiday camps in Erfurt. Building skills beyond the tech bubble.
Frequently asked questions

FAQ

Do we need our own hardware?
No — not necessarily. We can start with your existing infrastructure (cloud hybrid, existing workstations) and only recommend hardware when it makes economic and technical sense. If you want to buy your own hardware, we advise, procure and implement completely.
How does Apple Silicon compare with NVIDIA DGX Spark?
When answering with large models, Apple Silicon (Mac Studio M5 Ultra, 1,200 GB/s) is usually faster than DGX Spark (273 GB/s); for ingesting very long documents, for CUDA workflows and for training, DGX Spark is stronger. We will tell you honestly which setup fits your use case. More details in the tech-lead section of this page.
Does local AI meet the obligation under Art. 4 EU AI Act?
Since 2 February 2025, Art. 4 AI Act has required companies that use AI to take measures to promote the AI literacy of their staff. Since the AI Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026), companies do not have to guarantee a specific level of competence; documented training remains an obvious way to demonstrate the measures. Our training is use-case-specific, documentable and designed to fulfil this obligation — regardless of whether you use cloud or local AI. We are happy to advise you directly on the requirement.
What does a typical PoC cost?
That depends strongly on the use case, the desired model size and the maturity of your existing infrastructure. A first tech sparring session (30 min, free of charge) gives us the basis for a concrete offer. Typical PoC projects run for a few weeks and are calculated transparently based on effort.
Do you work remotely or on site?
Both — depending on the project phase. For workshops, hardware onboarding and shoulder-to-shoulder implementation we prefer to be on site. Concept, coding and support can run hybrid. We are easy to reach from Cologne and Weimar — across Germany.
Can the training be funded through BAFA or QCG?
Possibly. BAFA consulting funding (applications possible until 31 December 2026; continuation afterwards open) covers up to 80 % in the eastern German states and 50 % in the western ones — particularly relevant for locations around Weimar. QCG (Qualifizierungschancengesetz, a German training-funding law) may apply to employee training. We are happy to check eligibility with you.
For tech leads & development teams

Book tech
sparring

30 minutes, free of charge. Architecture, hardware sizing, stack questions — no sales pressure. An honest assessment of whether and how we can help.

Arrange tech sparring → Request an architecture sketch
For management & C-level

Arrange a
strategy call

45 minutes. Use cases, compliance status, funding options, realistic entry path. Tailored to your company.

Book a strategy call → Request an AI competence check