Nextcloud shipped Hub 8 on Wednesday, and the release notes read the way every collaboration release reads this year: an assistant, a chat interface over your own content, meeting summaries, suggested replies. The interesting part is not the feature list. It is the question the release quietly puts to every organisation running Nextcloud on its own infrastructure. Until now, self-hosting was a storage decision. From this week it is also an inference decision, and those are governed by different economics.
The real question is not whether the features are good enough. It is whether you self-hosted for control or for cost — because those two answers have just diverged
Organisations that chose Nextcloud to keep data on their own hardware get an answer they will like. Organisations that chose it to avoid per-seat pricing are about to discover what a graphics card costs.
What is actually in the release
Assistant 2.0 sits across the suite as a single entry point rather than as separate features bolted into each app. Context Chat lets users ask questions against their own files rather than against a general model. Talk gains call summaries, Mail gains suggested replies, and the whole set carries the ethical rating labels Nextcloud introduced to indicate how open the underlying model, training data and code actually are. Alongside it, the company has signalled an AI-as-a-Service arrangement with a European hosting provider during this quarter, which matters more than any single feature because it defines the middle option.
| Mode | Question to document |
|---|---|
| Local inference | Who owns hardware, runtime, capacity and patching? |
| Contracted hosted inference | Where is processing performed and under what terms? |
| External API | Which content classes and endpoints are approved? |
Qualitative summary of this article's source text, not a measured outcome or performance estimate.
Three deployment modes, and the trade is honest
Fully local. The model runs on hardware you own. Nothing leaves. You pay for a graphics card, the electricity, and somebody to keep the runtime alive. Capability is whatever open models currently deliver, which is respectable and not state of the art. AI-as-a-Service from a European provider. No hardware, no runtime to maintain, and the processing happens in a defined jurisdiction under a contract. Data does leave your server, which is the whole point of the distinction and the thing to be clear about internally. External commercial API. Cheapest per unit of work, strongest capability, weakest posture. Entirely defensible for some content and indefensible for other content, which is why the decision should be made per content class rather than once for the platform. The ethical rating labels are more useful than they first appear, because they force the organisation to state which trade it took and to write it down somewhere a procurement questionnaire can reach.
The arithmetic nobody runs before buying hardware
A card capable of serving a useful model to a working team is a real capital item, and the operating question is utilisation. Assistant traffic is bursty and concentrated in a few hours; a card sitting idle for most of the working day is an expensive way to summarise thirty documents. Before committing to hardware, estimate concurrent active use rather than total user count. Most organisations of two hundred people discover that peak concurrency is in the single digits, which changes the comparison against a hosted arrangement considerably — in either direction, but at least it is a comparison rather than an assumption.
Context Chat will disappoint before it delights
Retrieval over your own file store inherits everything about that file store. Shared folders contain drafts, superseded versions, three copies of the same policy and a folder called Final that is not. A retrieval system cannot tell which of four similar documents is current, because nothing in the file store says so. The model is rarely the problem. Scope the index to a small set of curated locations for the first month, see what people actually ask, and expand from there. Pointing it at the whole share on day one produces confident answers drawn from a 2019 draft, and one of those is enough to end the internal goodwill.
What self-hosting genuinely buys in this release
Three things, and they are worth naming because they are easy to lose sight of. The ability to not enable a feature, permanently, rather than to opt out of it each quarter. Version control over when behaviour changes, rather than discovering on a Tuesday that summaries now read differently. And a defensible answer on a customer security questionnaire that does not depend on a third party's sub-processor list. None of those are capability advantages. All three are governance advantages, and for a certain kind of organisation they are the entire reason the platform is there.
Practical Guidance for Deploy Nextcloud Hub 8
- Upgrade in staging first and check every third-party app for compatibility.
- Do not enable the AI features on day one — separate the upgrade from the launch.
- Estimate peak concurrent use before pricing hardware.
- Choose an inference mode per content class, not once for the platform.
- Scope Context Chat to curated folders for the first month.
- Record which model and mode each feature uses for questionnaires.
- Test summaries on your own recordings before announcing anything.
- Name the person who patches the inference stack — in writing.
The Regional Angle
The first thing to check is whether your own tender commitments survive the upgrade. Public-sector, semi-government and regulated contracts across the Gulf increasingly carry explicit data residency clauses, and a self-hosted platform is frequently how organisations satisfy them — the files are on a server in the country, the clause is met, the auditor is content. Enabling an assistant changes that in a way no storage diagram captures: the file stays where it was, but its contents are transmitted to wherever inference happens. A European hosted option is a European guarantee, not a local one, and a commitment to keep data in-country is not satisfied by keeping it in Frankfurt. Read your own contractual wording before enabling anything, check where the endpoint physically runs rather than where the vendor is headquartered, and remember that in-region graphics capacity remains scarce enough this year that fully local processing may be the only literal way to comply. The second is language, and it is the most common cause of quiet disappointment in regional deployments. Summarisation, transcription and suggested replies degrade noticeably on Arabic, and they degrade further on the code-switched speech that is normal in a Gulf office — a call conducted in Arabic with English technical terms, or a message thread that alternates by sentence. Open models available for self-hosting are generally weaker here than the large commercial services, so the deployment mode that is best for residency is also the one most exposed on language. Test with your own recordings and your own message threads before announcing the feature internally, and be willing to enable summaries selectively by team or channel rather than universally. Setting the expectation by language at launch costs a paragraph. Recovering from a summary that mistranslates a commitment costs considerably more. The third is operational capacity, which is where self-hosting decisions are usually made too optimistically. A great many regional organisations run Nextcloud precisely because it is undemanding: a server, a backup, an annual upgrade, and a partner on call for the awkward bits. A model runtime is a different kind of surface — driver versions, memory limits, queue behaviour under load, and failure modes that look like the application being slow rather than the application being broken. If nobody in the organisation can be named as responsible for patching that stack, the hosted option is the honest answer, and choosing it deliberately is far better than running a local model that quietly stops being updated in September.
The objection worth taking seriously
The strongest objection is capability. Self-hosted assistants run a generation or more behind the commercial services, and everyone in the building has the better one on their phone. Spending on hardware and operational effort to deliver a visibly weaker experience invites the obvious outcome: staff paste the document into the tool that works, and the organisation has bought an expensive way to feel compliant while the actual content walks out through a browser tab. That is not a hypothetical, and anyone who has audited real usage knows it is the default outcome rather than the worst case. The response is to stop treating this as a comparison between two assistants. It is a routing decision about content. A meaningful proportion of what people summarise and draft is entirely unremarkable — internal notes, scheduling, first drafts of ordinary correspondence — and there is no reason to hold that back from a capable commercial service under a reasonable contract. A much smaller proportion is genuinely sensitive: client material under confidentiality, anything touching a residency clause, personnel matters. That is the content the local option exists for, and it is small enough that a modest capability gap is tolerable. The failure mode is not choosing the weaker tool; it is providing nothing adequate for the ordinary work and assuming a policy will hold. Hub 8 is most useful to organisations that make that split explicit, enable the local path for the narrow category that needs it, and stop pretending the rest is a security problem.
Common Questions
Do we need a graphics card to use any of this?
Not necessarily. Some features run acceptably on processors, and the hosted service option removes the question entirely. The heavier tasks are where hardware starts to matter.
Can we enable the assistant for some users only?
Yes, and you should at first. Group-scoped enablement lets you find the awkward cases with twenty people rather than two hundred.
Will Context Chat respect our existing permissions?
It is designed to, and you should still test it adversarially with a deliberately low-privileged account before wide release. Assume nothing about retrieval and access until you have tried to break it.
What should we expect over the next twelve months?
Expect the hosted European inference option to arrive this quarter and to become the default choice for organisations that wanted the governance without the hardware. Expect open model quality to keep closing the gap, which makes local deployment steadily more attractive rather than less. Expect competing suites to ship similar assistant features during the year, which will turn the conversation from whether to have one into where it runs. And expect procurement questionnaires to start asking that question explicitly, at which point having written the answer down will look like foresight.
Deploy Nextcloud Hub 8 — we upgrade cleanly, route each content class to the right inference mode, and make sure the assistant does not quietly break a residency clause you already signed.