New Horizon No. 195 / 2026-07-14 · Berlin
Radical Transparency

Receipts

The autonomous company, audited by itself.

This company claims to run itself. Claims are cheap, so this page shows the books instead: what the pipelines actually shipped, what it costs to keep them running, what they have earned, and where they broke. Every output count on this page is computed live, on each request, from the same files the pipelines write when they publish — there is no marketing copy between you and the data.


What the Machine Shipped

353D models listed
56Blog posts
24Reels produced
74Digest issues
93Days running 24/7
13Newsletter subscribers
4Inbound leads
0Leads converted

Yes, thirteen subscribers. This page would be easier to write with three zeros appended — it would also be fiction. The point of an audit is that you can check it:

→ The models on CGTrader → The reels on Instagram → The blog → The digest archive


The Part Nobody Publishes

REVENUE — ALL TIME as of 2026-07-13
  • €0 REV CGTrader asset sales — payouts to date
  • €0 REV Consulting engagements closed
  • €0 REV Other income
  • €0 Σ Total revenue since day one
RUNNING COSTS — MONTHLY ~ = estimate, being finalised
  • ~€185 COST AI coding & operations subscription (Claude Max)
  • ~€30 COST 3D generation credits (Meshy)
  • ~€10 COST Cloud LLM for the site chat (Ollama Cloud)
  • ~€8 COST Hosting + domain (Strato)
  • ~€35 COST Electricity — GPU workstation + NAS, 24/7
  • €0 COST Analytics (Umami, hobby tier)
  • €0 COST Instagram / Meta API
  • ~€268 Σ per month, every month
€0Earned to date
~€819Spent to date
~−€819Net, honestly

The machines work. The business, so far, does not — the products are live, the funnels exist, and the revenue line is a flat zero. That is what the first ninety-odd days of an autonomous company actually look like, and publishing it is the experiment: an operation this cheap only needs a handful of paying customers to flip the sign. Hardware (one GPU workstation, one NAS) was owned before the company started and is excluded above.


Where It Broke

An autonomous company is mostly a machine for discovering new failure modes. These are real incidents from our own logs — kept here because they are more informative than anything that worked.

SELECTED INCIDENTS
  • 2026-05 VIDEO The yogurt incident. Our first video model rendered every liquid — ocean waves, pouring coffee, rain — with the physics of yogurt. No metric caught it; a human did. The engine was swapped for a two-stage pipeline and liquids got their own routing rule.
  • 2026-06 QA The judge that failed everything. We tightened a vision-model QA prompt with phrases like “reads as AI blob” — and the judge began parroting them back as verdicts, failing 100% of 3D assets for days while humans rated most of them fine. Lesson: never put negative vocabulary in a judge prompt; require a named defect in a named part.
  • 2026-07 DEPLOY The approval gate that deleted itself. A stale local config, pushed by a routine sync, silently removed the human sign-off step from the reel pipeline — two videos went out unreviewed before metadata analysis caught it. The gate is back, and config mirrors are now part of the deploy contract.
  • 2026-07 LEADS The sales team was off for ten days. The same class of sync accident disabled the inbound-lead processor on July 3rd. Discovered July 13th during a traffic audit. The inbox was empty, so nothing was lost — but “the watcher now watches the watcher” is a sentence we get to write from experience.

What It Runs On

Control Plane
Windmill
  • Self-hosted on a NAS
  • Every pipeline is a versioned flow
  • Schedules, retries, audit trail
Generation
Local first
  • One GPU workstation: video, images, LLMs
  • Cloud only where it pays (3D meshes, coding agent)
  • Apache-licensed models for anything commercial
Oversight
Traced
  • Self-hosted tracing + offline LLM judges
  • Vision-model QA on every asset
  • A watchdog that files alerts, not excuses

Human involvement, stated exactly: one approval tap per reel before it posts (accept or reject, no edits), and corrections when an incident like the ones above surfaces. Product generation, writing, QA, publishing, lead replies and this page's numbers run without anyone in the loop.


How These Numbers Are Made

Output counts are file counts: the 3D tally comes from the gallery manifest the publishing pipeline writes, blog and digest counts from the article files on this server, the reel count from the reel index, subscribers from the length of the subscriber list (the list itself never leaves the server). The euro ledger is a config file we update by hand and mark with ~ where a figure is still an estimate. Nothing is fetched from an analytics service, and nothing is rounded up.

→ Want a machine like this for your business?


The AI news that matters — in your inbox by 07:30 CET. Free, no spam.