The usual state
Scattered
A tool per team
A developer's inference server here, a notebook there, a container somebody started in March. Nothing knows about anything else, and no single question about the whole estate has an answer.
Why Diffuse Enterprise
The deployment that passes the audit
Most organisations already run a language model somewhere. Almost none of them could survive being asked to prove who used it, on what, and under whose authority.
The consequence
Unprovable
No trail to produce
When the auditor asks who ran what against which weights on the fourteenth, the honest answer is that nobody recorded it, because nothing in that stack was built to.
The outcome
Failed
A finding, not a warning
For regulated and critical infrastructure this is not a remark in an annexe. It is the control that was not implemented, and the project stops until it is.
What goes wrong
The dev-tool deployment, examined
The tools that make a model easy to try are not the tools that make one defensible to run. Each of these is ordinary, and each is a finding.
Access
An open port is an open door
The usual inference server binds a port and serves whoever reaches it. Open it beyond the workstation and anyone on the network can send requests: no token, no account, no identity, no limit. The access control is the firewall rule, and firewall rules are edited by people in a hurry.
Transport
Plaintext across the estate
Prompts, documents and completions cross the network unencrypted because the default was http and nobody changed it. In an environment where transport encryption is a stated control, that single default is the difference between a deployment and an incident.
Identity
Nobody is anybody
There is no notion of who made a request. Not a certificate, not an account, not a key. Attribution is impossible after the fact, which means every question that begins "who" cannot be answered at all.
Audit
Nothing is written down
No record of who deployed which weights, who revoked what, who was refused and why. A log of crashes is not an audit trail, and reconstructing one from three teams' shell histories is not evidence.
Governance
No centre, so no control
Each team runs its own. Nobody can say how many models are running, on what hardware, under which licences, with which data. A control you cannot enumerate is a control you do not have.
Reach
Available to the few
Because it is somebody's machine under somebody's desk, it serves that team. The rest of the organisation either does without or starts a second one, and now there are two of everything to not audit.
Reasons for Diffuse
What we built instead
Every line above has a counterpart that was designed rather than retrofitted. This is the whole argument for this product.
Access
An identity, or nothing
Every connection between machines is mutually authenticated TLS 1.3. There is no plaintext transport in any configuration, including development defaults, and no flag that turns it off. A machine's identity comes from its certificate, never from what it claims in a request.
Applications
Keys that expire and narrow
Application access is an API key, hashed at rest, shown once, with an expiry and a scope. A reporting service gets a key that reaches one model and stops working on a date you chose.
Audit
Who did what, and what was refused
An append-only trail of every administrative action, with the actor resolved from their certificate or session. Refusals are recorded as deliberately as successes, because a trail that holds only what worked cannot answer whether somebody tried.
Confidentiality
Customer content never reaches it
Not prompts, not completions, not the rows of a training corpus, not a document's text. Asserted by tests that send a sentinel through a real request and then search every column and both service logs for it.
Governance
One centre that decides
One coordinator holds the identities, the entitlement, the placement and the trail. Enumerating what runs where is one command, and roles separate the people who deploy from the people who read the trail.
Reach
Available to the organisation
One endpoint, OpenAI-compatible, that every team points at. Existing clients work unchanged, and access is granted and withdrawn centrally rather than by who knows which port.
None of this makes an audit pass by itself. What it does is make the answers exist: the trail is there to produce, the identities are there to enumerate, and the transport was never in question. That is the difference between preparing for an audit and rebuilding for one.
Where it runs
Your hardware, wherever it is
On-premises is where this started and it is not the limit. What the product requires is that the machines are yours to control, not that they sit in your building.
Option one
The fleet you already own
Workstations and lab machines, idle after six. Capacity you already pay for becomes capacity that serves.
Option two
Your own cloud account
GPU instances in your account and your network. You pay for the machine rather than for tokens, and nothing is metered by anybody else.
Option three
One box on a desk
An AMD Ryzen AI Halo or an NVIDIA DGX Spark: 128 GB of unified memory in a compact, quiet case. One machine holds the model whole, for the whole company. Splitting is the fallback, never the requirement.
Next step
Ask us the hard question
Tell us what your auditor asked last time. If this product does not answer it, we would rather say so now than after a pilot.