Open-Source or Commercial AI for Small Business?

Sonny HovsepianPublished 17 Nov 2026Updated 17 Nov 2026Open-source or commercial AI
Open-Source or Commercial AI for Small Business?

Key Takeaways

  • Most Australian service SMEs should start with a commercial business plan or a managed open-weight model, then test the model against a specific business task.
  • Open-weight AI can carry licence conditions, and running a downloadable model still requires hardware or hosting, security and upkeep.
  • Self-hosting deserves consideration when information must stay on a business's own machine, usage is consistently high, or an internal operator can maintain the system.
  • As at 17 November 2026, Claude's pricing page says business content receives no model training by default; local hosting still requires access controls, secure backups and maintenance.
  • Amazon Bedrock announced open-weight models with Sydney-region endpoints on 12 February 2026, but the availability announcement provides no Australian data residency promise.

I'd start most Australian service businesses on a commercial business plan or a managed open-weight service. I'd consider self-hosting after identifying a requirement that justifies owning the maintenance. For the wider workflow decision, start with Truespeak's guide to AI for small business.

What Is the Difference Between Open-Source, Open-Weight and Commercial AI?

The three terms cover what you can download, what the licence permits and who runs the service.

TermWhat the term tells youWhat to check
Open-source AIThe Open Source Initiative definition requires training-data information as well as weights.Whether the release meets the definition.
Open-weight AIYou can download the model weights and run the model under its licence.Commercial-use conditions, attribution and acceptable use.
Commercial AI serviceA vendor runs the service under commercial terms.Training policy, hosting, retention, access controls and charges.

As at 17 November 2026, Meta's Llama 3.1 licence is royalty-free but includes attribution and acceptable-use conditions. OpenAI's gpt-oss uses Apache 2.0. Mistral's documentation lists different licences for different models. Ask for the exact model and licence before approving commercial use.

Is Open-Source AI Actually Free for a Small Business to Use?

A model can have no licence fee while the system costs money to run. Your budget still needs to cover the computer or cloud instance, power, setup, security and maintenance.

Ollama's pricing page, checked on 17 November 2026, says running models on your own hardware is unlimited. Unlimited local use does not pay for the machine or the person fixing a failed update.

I'd ask for two costs in a proposal: getting the workflow working, then keeping the workflow working. Include the CRM connection, inbox access, approval steps and records of what happened.

Should a Small Business Self-Host an LLM, and What Does That Cost?

I'd shortlist self-hosting when data must stay on a particular machine, steady usage could justify dedicated capacity, or someone inside the business can operate the system. With high usage, work out the monthly running cost before assuming self-hosting is cheaper.

Hardware requirements depend on the model. As at 17 November 2026, OpenAI says its larger gpt-oss model fits on a single 80GB GPU. A downloadable model can still need substantial equipment.

Renting equipment avoids buying the GPU. The following figures were published by Hugging Face as at 17 November 2026. Currency and GST treatment were not stated; these are dollars as published, without conversion to Australian dollars.

Dedicated GPUPublished hourly costPublished monthly cost
NVIDIA T4, 14GB$0.50$360
NVIDIA A100, 80GB$2.50$1,800
NVIDIA H200, 141GB$5.00$3,600

These figures describe different hardware capacities, without establishing which model will handle your work. Dedicated capacity left running incurs charges during idle time. Confirm where an instance runs before sending client information to it.

Before committing, I'd request a paid test with a clear scope: representative documents, agreed acceptance checks, a usage estimate and a monthly operating budget.

Is Self-Hosted AI More Private Than ChatGPT or Claude Business Plans?

Local hosting gives you control over the machine. Privacy also depends on who can access the information and where connected services, logs and backups send copies.

Ollama says it does not see prompts or data when you run locally. As at 17 November 2026, Claude's business-plan comparison states no model training on your content by default. That statement addresses training; check retention and access separately.

For ChatGPT Business and other plans, Truespeak's guides cover AI training and business data and whether AI data needs to stay in Australia.

Self-hosting still leaves someone responsible for updates, permissions, backups and breaches. The small-business AI guide developed with ASD and COSBOA provides a practical security checklist.

Can I Use Open Models Hosted in Sydney?

Yes. AWS announced open-weight model support in Amazon Bedrock's Sydney region on 12 February 2026. A managed service lets you use open models without buying hardware.

The announcement describes region-specific inference endpoints. Get written residency terms for your proposed configuration before promising clients their information stays in Australia.

Check the service location each time. As at 17 November 2026, Ollama says its cloud compute runs primarily in the United States. Running Ollama locally and using Ollama cloud involve different hosting arrangements.

How Should I Choose, and What Red Flags Should I Watch For?

  1. Name one job, such as preparing an enquiry reply for approval.
  2. Identify the information required and where each copy will go.
  3. Test representative examples and record the corrections a person makes.
  4. Compare the complete operating cost, including connections and maintenance.
  5. Name the person responsible for failures, updates and approvals.

I'd question these recommendations before signing:

  • “Open source” without an exact model licence.
  • “Free” hosting that excludes equipment and upkeep.
  • “Australian hosting” without written residency terms.
  • One platform recommended for every workflow without testing alternatives.
  • A self-hosted system with nobody responsible after launch.

Ask how you would move the workflow later. Truespeak's guide to avoiding AI vendor lock-in covers that decision.

Where Does Truespeak Fit?

I'm Sonny Hovsepian, Truespeak's Sydney-based founder, and clients work directly with me. I work hands-on across Claude, GPT, Gemini, Grok and DeepSeek, and my own Company Brain runs five AI models on one shared memory. Truespeak matches the model to the job, so a client does not have to pick a platform first.

For an Australian service business that wants automation built and then run around its existing tools, with a person approving anything sensitive, Truespeak is built for exactly that.

Truespeak's Sydney AI automation service designs and builds the workflow, and managed AI operations keep it running after launch: monitoring, fixing failures, reviewing exceptions and improving rules. Start with a discovery call about the work you need done.

Frequently Asked Questions

Should my small business use open-source AI or pay for ChatGPT or Claude?

I'd start most Australian service SMEs with a commercial business plan or a managed open-weight service. Test the actual workflow, check the current contract and compare ongoing costs before considering self-hosting.

Is open-source AI free for commercial use?

Licence conditions differ by model. A royalty-free model can still carry attribution and acceptable-use requirements. Hardware, hosting, setup, security and maintenance also need a budget.

How much does a self-hosted LLM cost?

The total depends on the model, hardware, usage and maintenance. As at 17 November 2026, Hugging Face published dedicated GPU monthly figures from $360 for a T4 to $3,600 for an H200. Currency and GST treatment were not stated. Those figures exclude workflow setup and do not establish model suitability.

Does self-hosting an LLM keep client data private?

Local hosting gives a business control over the machine, but privacy also depends on access controls, connected services, logs and backups. Self-hosting still requires someone to secure and maintain the system.

Can open-weight AI models run through a Sydney cloud service?

AWS announced open-weight model support in Amazon Bedrock's Sydney region on 12 February 2026. The announcement describes regional inference endpoints without promising that all data stays in Australia. Obtain written terms for the proposed configuration.

Sources

Checked 16 Nov 2026.