Test mode: payments run through the Stripe sandbox. Nothing is charged, reservations do not count in the counter.

Skip to content
llmcrowd
DE Deutsche VersionReserve

How llmcrowd works.

Everything a managing director wants to know before reserving: what we buy, where your money goes, what happens when, and what you can realistically expect. Calculated openly, with sources.

Planned figures as of 10 November 2026. Technical terms are explained where they first appear.

The concept in plain words

Crowd pool

Several companies share one server's compute. Whoever uses little frees capacity for others. That is why there are no hard quotas and no price per token.

Own servers in the country

The servers run in the country where you use llmcrowd: in Switzerland for Swiss customers, in Germany for German ones. No US cloud provider processes your prompts.

Open-weight model

«Open-weight» means the model's weights are freely available and we run it ourselves. Today that is Kimi K3.1 Pro by Moonshot AI (China) with 1 million tokens of context. There is no connection to its maker; switching models is possible by contract.[13]

Own harness, own database

The «harness» is the software around the model: chat, parallel agents, connectors to your systems, permissions and audit log. Your knowledge sits in your own database, separate from other companies.

What exactly is bought

One server per tranche. These figures come from our planning; we confirm prices and delivery times in writing with the dealer before ordering.

8 × B300
NVIDIA GPUs with 288 GB each, 2,304 GB GPU memory in total. The model needs at least 1,874 GB.[12]
1.5–2 TB
RAM, plus fast NVMe SSDs to cache long documents
~14 kW
Power draw per server. Power and cooling are the bottleneck, not space.
DE · CH
Data centre (colocation) in the country of use, with a provider without ties to a US group
~625,000
Purchase including installation, once per server (planned)
~8,000
Operation per month: power, cooling, data centre, support (planned)

Amounts in the country currency (CHF or EUR), set equal as in the concept paper.

Where your money goes

Per seat and year with annual payment, in a server's first year of operation. The calculation in full:

  1. 400 seats × 2,400 = 960,000 per year and server
  2. Server including installation ~625,000 + operation ~96,000 = ~721,000 in the first year
  3. Per seat: 625,000 ÷ 400 = ~1,563 hardware, 96,000 ÷ 400 = 240 operation, the rest goes to team, development, support and reserve

2,400 per seat and year

~1,563Hardware share
~240Operation and power
~597Team, development, support, reserve

From the second year the hardware is paid; most of the money then goes to operation, development, spare parts and the next servers. Monthly payers pay 25% more (250 instead of 200 per month). Planned figures before payment fees and taxes.

The process, step by step

And if a tranche does not fill? If server 1 has not started by 30 June 2027, we refund every reservation in full. If your tranche's server is not operational within 4 months of the payment becoming due, you may withdraw and get everything back.

Once tranche 1 is full, tranche 2 (server 2) starts with the next 400 seats, and so on up to 3,000 seats per country.

  1. 01

    Reservation

    99 per seat, from 10 seats, via Stripe. Your payment method is saved for later.

  2. 02

    The tranche fills

    The counter on the home page shows live how many of the 400 seats are reserved.

  3. 03

    Tranche full

    We announce the first payment by email.

  4. 04

    Payment due

    Annual 2,400 or the first monthly instalment of 250 per seat, each less the 99 reservation.

  5. 05

    Server order

    We order the tranche's server from the dealer.

  6. 06

    Delivery and installation

    The delivery time is open; we get it confirmed in writing before ordering and share it.

  7. 07

    Guided start

    We set up your environment, connect the first systems and measure quality and speed together.

  8. 08

    Operation

    Your 12-month term runs. Monthly payers pay instalments 2 to 12 from then on.

What you can expect

Honestly, with planned figures. What we cannot guarantee is here too.

1 million tokens of context

About 1,500 pages of text in one question. Whole case files, contracts or codebases at once.[13]

Response speed

In chat about 100 tokens per second (faster than you read), background agents 30 to 50 tokens per second. Planned figures to be confirmed in the pilot.

Parallel agents

Several agents per seat work at the same time. The admin sets the limit per person; the pool keeps it fair.

Long prompts take time

Reading 340,000 tokens at once takes about 9 to 15 seconds, a full million more like 30 to 45 seconds. Repeated questions on the same file are faster thanks to caching.[12],[13]

Fair-use pool

Whoever uses little frees capacity. When the pool is full, every seat gets at least its share; chat stays fast, background agents wait briefly.

Availability with one server

Until server 2 runs there is no redundancy: during maintenance or a failure the service stops. We announce maintenance windows. From server 2 we run redundantly.

400 seats per server

That is our cautious plan. How many seats a server really carries we measure in the pilot and adjust the tranches before it gets tight.

One country, its own servers

Every country gets its own servers and its own tranches. Your data is processed on servers in your country. We start in Switzerland and Germany; Austria, Italy, Spain, Portugal and Poland follow in the order of waitlist entries.

See countries and counters

Data protection and law, in short

  • Data processing agreement under Art. 28 GDPR with a list of all sub-processors
  • Processing and storage in the country of use (CH or DE)
  • Deletion concept with fixed periods, export at any time
  • Audit log of every question, file access and agent action
Read the reservation terms

Sources

  1. [1] Anthropic, Claude Pricing (Team & Enterprise), retrieved 10 November 2026
  2. [2] Anthropic, API Pricing (Claude Opus 5.5), retrieved 10 November 2026
  3. [3] OpenAI, ChatGPT Pricing (Business & Enterprise), retrieved 10 November 2026
  4. [4] OpenAI Help Center, Managing billing and seats in ChatGPT Business, retrieved 10 November 2026
  5. [5] Microsoft, Microsoft 365 Copilot Pricing, retrieved 10 November 2026
  6. [6] Google, Google Workspace Pricing, retrieved 10 November 2026
  7. [7] OpenAI, API Pricing (GPT-6 Astra), retrieved 10 November 2026
  8. [8] Google, Gemini API Pricing (Gemini 4 Argon), retrieved 10 November 2026
  9. [9] 18 U.S.C. § 2713 (CLOUD Act), Cornell Law School LII, retrieved 10 November 2026
  10. [10] Startup Fortune, Kimi K3.1 Pro: Weights und Benchmarks, retrieved 9 November 2026
  11. [11] MarkTechPost, Kimi K3.1 Pro Open-Weight-Modell, retrieved 9 November 2026
  12. [12] ComputingForGeeks, Kimi K3 lokal betreiben: Hardware und KV-Cache, retrieved 9 November 2026
  13. [13] vLLM, Day-0-Support für Kimi K3.1 Pro, retrieved 9 November 2026

Convinced? Then secure your seats.

Reserve seatsCheck the value