# AUKORA: illustrative from-scratch rebuild budgets

Prepared 5 October 2026. These are author-created planning scenarios, not measured project expenditures, vendor quotes, replacement-cost appraisals or a valuation of AUKORA. The person-year ranges are judgment calls based on the functional scopes below; they have not been validated against project payrolls or an independent estimating exercise.

## The question and counterfactual

What might a company budget to develop comparable scoped layers if it could not reuse Cordis, DeepSeek Harness, an existing trained model, OpenShell or Nostr implementations? It may still use ordinary operating systems, languages, compilers, databases, cloud hardware and established cryptographic libraries. This is not a cost to reinvent computing, invent the underlying science, acquire users or reconstruct an entire ecosystem.

Assume conventional staffed engineering. AI assistance and access to prior research are not forbidden, but unusually effective automation could substantially reduce required staff effort. The scenario is not a validated claim about how much each team would need today.

A fully loaded software-engineering year is assumed to cost $250,000, including salary, benefits and ordinary overhead. Specialist model research/engineering is assumed at $350,000 per person-year. These are explicit US-style commercial planning assumptions, not salary surveys, global wage floors or supplier quotes. One person-year can be spread over a larger team and a shorter interval; it is effort, not elapsed time.

## Five separate scopes and calculations

| Layer | Scope and assumed effort | Arithmetic | Illustrative USD range |
|---|---|---|---|
| Cordis | Comparable plugin/service lifecycle framework, replacement, testing, docs; 2–6 engineering-years | 2–6 × $250,000 | $500,000–$1,500,000 |
| DeepSeek Harness | Agent loop, provider adapters, file/tool/editor interface, core plugins, tests and packaging; 8–24 engineering-years | 8–24 × $250,000 | $2,000,000–$6,000,000 |
| V4.1-Flash-class model development | 20–100 research/engineering-years, training/experiment compute budget $5–50m, data/annotation/evaluation services $3–15m | $7–35m + $5–50m + $3–15m | $15,000,000–$100,000,000 |
| OpenShell | Comparable execution/control plane on existing OS primitives, policy/credential interfaces, test suite and hardening; 12–48 engineering-years plus $1–3m external security review/test infrastructure | $3–12m + $1–3m | $4,000,000–$15,000,000 |
| Nostr | Comparable signed-event specification, reference relay, basic clients, interoperability tests; 2–12 engineering-years | 2–12 × $250,000 | $500,000–$3,000,000 |
| Total of scoped scenarios | Not a probabilistic confidence interval | Low endpoints $22m; high endpoints $125.5m | About $22–126 million |

The Harness line excludes the Cordis framework, model training and OpenShell implementation. The model line excludes building the general-purpose Harness and OpenShell, but includes model-specific training and evaluation work. External data/evaluation allowances are for non-payroll procurement and services, not a second copy of staff costs. OpenShell's external-review allowance excludes the engineers already counted. Nostr's network adoption, reputation, existing clients, worldwide relay fleet and community cannot be purchased by this estimate. General multi-year operations after the initial project are excluded.

The five component endpoints can be added for this deliberately partitioned scenario, but shared teams and integration can create efficiencies or extra costs. Summing endpoints does not establish a likelihood for the total. No contingency for unlimited research failure is priced. Reaching a comparable quality level is not guaranteed by any budget.

## Why use DeepSeek V4.1-Flash as the reference?

The official model card and report describe a multimodal model trained from scratch on 45T tokens, with separate post-training and data/environment work. That establishes the scope being targeted, not this estimate's dollar total. The $15–100m line is our hypothetical company program, not DeepSeek's reported cost, API price, hardware purchase or fine-tuning cost. The sources reviewed did not provide an audited complete company-build cost for V4.1-Flash. No V3 or Llama estimate is used in the table.

The model line has particularly low confidence: staff expertise, data access, research failure, pretraining efficiency, posttraining and evaluation workloads are not known for our hypothetical company. A cheaper successful reproduction or a far more expensive failed attempt is possible. Scope and assumed expenditure are not a promise of benchmark equivalence.

## The approximately $10,000 prototype comparison

Peter supplied the rough account of months of AI-assisted work and approximately $10,000 in direct spending. It has not been independently audited and excludes unpaid founder labor and the upstream work inherited or accessed. AUKORA did not train the reference model, create the upstream projects or finish all their integrated safety properties for that amount. This compares funding an original stack with assembling an unfinished prototype on shared foundations; it is not a like-for-like savings calculation.

The cost to operate a finished AUKORA deployment, qualify its boundaries, support users and maintain it is not estimated here. Models, code and protocols have their own license terms, and access to open materials does not make computation or hosting free.

## Sources for functional scope, not for our budget ranges

- Cordis: https://github.com/cordiverse/cordis
- DeepSeek Harness: https://github.com/deepseek-ai/deepseek-harness
- DeepSeek V4.1-Flash model card: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- DeepSeek V4.1-Flash report: https://arxiv.org/html/2609.19969v1
- OpenShell: https://github.com/NVIDIA/OpenShell
- Nostr: https://github.com/nostr-protocol/nostr
- AUKORA current implementation/limits: https://github.com/aumara-xyz/aukora-sovereign/blob/main/docs/GOLDEN-BOUNDARY-PAPER.md#17-what-exists-today

These projects do not endorse this estimate. Public architecture descriptions support the named functions; all dollar ranges, team sizes, non-labor budgets and calculations are our own disclosed assumptions.

## Revision, 10 October 2026: the model row is now a range from small to large

The homepage row now reads "Open-source models" and shows $1–100m. The upper part is the V4.1-Flash-class line above ($15–100m). The lower end is a separate, lower-confidence scenario for training a small open model of a few billion parameters: 2–6 research/engineering-years at $350,000 ($0.7–2.1m) plus $0.3–3m of compute and data, rounded to about $1–5m. With that row the low endpoints sum to about $8m and the high endpoints to about $126m. These remain author-created planning scenarios, not measured costs.