Journal บทความ

Why we built AI that stays in the building

The most useful thing an AI assistant can do at work is help with the work itself: the draft contract, the customer spreadsheet, the recording of yesterday's board meeting. That is also the material most organizations are least comfortable sending to a third party. For a lot of teams, the result is an awkward compromise. People either avoid AI for anything sensitive, or they use it anyway and hope nobody asks.

We started Brainwave because we think there is a simpler answer: put the AI on your own network, on hardware you own, and let everyone in the office use it.

Privacy you can explain in one sentence

With a cloud service, the honest description of where your data goes is long. It involves the provider's data centers, their retention policy, their subprocessors and whatever their terms say this year. With Brainwave the description is short: prompts, files and outputs are processed on a machine in your office. Nothing has to leave the local network for the AI to work.

That matters in practice. It is far easier to get a legal team, a client or an auditor comfortable with “it stays on our hardware” than with a stack of vendor documents.

A cost you can plan for

Most AI services charge per person, per token or per generated image. That works for light use, but it makes budgeting hard and quietly discourages the heavy users who get the most value. Teams start rationing, or someone has to watch a usage dashboard.

Brainwave is a one-time hardware purchase with no subscription and no per-seat or per-token fees. Once it is on your network, the only limit is what the hardware can do. A long transcription job or a batch of image generations does not show up on next month's invoice.

Speed that comes from distance

A request to a cloud model travels across the internet, waits in a shared queue and comes back. A request to Brainwave travels across your office network. For everyday chat that difference is noticeable; for things like real-time translation in a meeting, or a workflow that makes hundreds of small calls in a row, it adds up. Local AI also keeps working when your internet connection does not.

Control over what runs

When AI is a service, the provider decides which model you get, when it changes and what it will or will not do. On your own hardware, you choose. Brainwave runs open models, and an administrator can decide which ones are available to which people. If a model works well for your contracts team, it stays that way until you decide to change it.

The CLI and local database connections mean you can point your own tools and data at it, too, without building an integration with an outside service.

What local AI is not

We want to be straightforward about the trade-offs. The very largest frontier models still run in large data centers, and a single machine in your office will not match them on every task. What we have found is that most day-to-day work, such as drafting, summarizing, searching internal documents, transcribing and translating, is well served by capable open models running locally. For those tasks, keeping data in the building costs very little and gives you a lot.

Brainwave is currently in pre-launch. You can reserve a unit today with a fully refundable US$500 deposit.

What “local AI” means for your data

“Local” is easy to say and easy to misunderstand. If you are the person in IT or compliance who has to sign off on a new AI tool, you need a clearer picture than a slogan. This is how we think about it, and the questions we suggest asking of any on-premise AI product, including ours.

What happens on the device

On Brainwave, the AI work itself happens on the machine. When someone on your network sends a chat message, uploads a document to a shared workspace, transcribes a meeting or generates an image, the model that does the work runs on Brainwave's own GPU. The inputs, the outputs and the files connected through the local database are stored on Brainwave's own storage. None of this needs an internet connection, and none of it is sent to us.

In practical terms, the machine behaves like any other server in your office. It sits behind your firewall and your existing network rules apply to it.

What the administrator controls

Because the data lives on hardware you own, the decisions about it are yours as well. Through the Admin Dashboard, an administrator manages who has an account, what each person or group can access, and which models and features are available. Shared workspaces can be limited to the people who need them, so a finance team's documents do not appear in a marketing workspace.

Retention works the same way. How long conversations, transcripts and generated files are kept is something your organization decides, in line with your own policies, rather than something set by a vendor.

What connects outward, and only if you choose

Some things are useful to fetch from the internet, such as software updates and new open models. We treat these as things you turn on, not things that happen silently. If your policy is that the machine should never reach the internet, it can run fully offline; you will simply need to plan how updates reach it. If you connect your own external tools through the CLI or an automation workflow, those connections are ones you configure, and data goes only where you send it.

Practical guidance for IT and compliance teams

• Map the data flows. Record Brainwave in your systems inventory as an internal server, and note which internal data sources you connect to it.
• Apply least privilege. Set up groups and workspaces before you invite everyone, and give access to sensitive sources only to the people who need it.
• Decide on retention up front. Agree how long chats and transcripts should be kept, and configure that before rollout.
• Review outbound settings. Decide whether updates and model downloads are allowed, and from which network segment.
• Treat it as physical infrastructure. Place the machine somewhere access is controlled, as you would any server holding business data.

Data protection laws such as Thailand's PDPA still apply to personal data processed on your own hardware. Running AI locally does not remove those obligations, but it does make them simpler to meet: the data stays within your organization, under your controls, with no additional third-party processor to assess for the AI itself. Your own legal advisers should confirm how this applies to your situation.

For more on how we handle the limited information we do collect, such as waitlist sign-ups, see our Privacy Policy.