Why your next AI infrastructure decision might be a hardware purchase

Enterprises face a critical turning point in their AI infrastructure strategy as the operational costs of cloud-based large language models (LLMs) escalate. While cloud infrastructure provides unmatched capabilities for complex tasks, a shift toward local hardware and hybrid architectures is emerging as the solution for sustainable cost management and data sovereignty. By optimising local environments and leveraging governed platforms, organisations can balance performance with strict enterprise compliance.

The economics of cloud AI are starting to bite. Enterprises that leaned hard into cloud LLMs for their workflows, for example, summarising Jira tickets, triaging service requests, drafting tests and documentation, are now discovering that this costs a lot when done at scale.

A few chatbot queries a day doesn't really make a dent, but as teams start running continuous AI agents against large amounts of data the maths can turn on you quite quickly. AI cost management has become "the single most sought-after skill set for technology finance teams" in the past twelve months. Deloitte's 2025 AI expenditure research found enterprises using generative AI tools had seen a 35% average increase in cloud AI spend in a single year.

Similar to the infrastructure-as-a-service cloud revolution that led many companies to move workloads away from cloud providers, such as AWS, GCP and Azure, and into their own data centres, there's also a hardware story running alongside the exponential growth of cloud-based LLMs. Engineers are buying Mac Minis, spinning up Ollama or similar open-source model hosting software, and running capable open-weight models entirely on local infrastructure. It might seem like just hobbyist tinkering, but there's an undertone of re-examining where AI should actually live.

The hardware case is more straightforward than it sounds

The business case for on-premise hardware is very pragmatic. Apple's M-series unified memory architecture makes it unusually efficient for running quantised LLMs. A Mac Mini M4 Pro runs 24/7, handles models like Llama 3, Mistral, and Qwen locally, and doesn't cost much to run after the initial purchase. The break-even point against cloud API spend for teams running high-volume agentic workflows is typically less than a year. After that, the cost of local inference is essentially just the electricity.

Tools like Ollama provide an OpenAPI-compatible endpoint, making it straightforward to point existing integrations at a local model instead of a cloud one. For teams already building agents and workflows on top of APIs and MCP servers running on SaaS products, the model and where it runs can itself become swappable infrastructure. The application logic doesn't care whether the inference is running on AWS, on a box in your server room, or indeed on a box under your desk. Atlassian's own Rovo platform already supports this kind of model flexibility: Rovo runs across OpenAI's GPT, open-source models including Mistral and Llama, and third-party-hosted models like Claude and Gemini, selecting between them to optimise for latency and context. The model underneath is infrastructure; the integration layer is where the value sits.

Open-weight models are roughly three to six months behind frontier models in capability, which is a real trade-off for tasks requiring complex reasoning. But this is more than good enough for many of the high-frequency, well-scoped tasks that enterprise AI agents actually do—for example, categorising, summarising, and extracting—so running the latest greatest cloud-based model isn't essential for functionality.

Governance is the argument that closes deals

Cost gets people's attention. Governance gets sign-off.

When data is sent to a cloud LLM, it passes through a third-party API, is subject to that provider's data retention policies, and sits under the legal jurisdiction of wherever those servers happen to be. For many organisations, that's a compliance showstopper—particularly those in financial services, healthcare, legal, and the public sector.

Over 100 countries now have some form of data sovereignty or localisation law, and the requirements conflict across jurisdictions. The EU's GDPR restricts data transfers to countries without adequate protection standards. The UK's own Sovereign AI Unit, announced in 2025, signals the same direction of travel from a regulatory standpoint.

When a local LLM processes data, that data never leaves the enterprise perimeter. No API call, no third-party retention policy, no jurisdictional ambiguity. Atlassian's Rovo brings comparable governance controls to the cloud tier. Rovo strictly respects existing Jira and Confluence permission structures:

An icon of an outstretched hand with a key above it

Access restrictions

If a user cannot see a space or page, Rovo can't either.

An icon of two hands shaking each other.

Data privacy assurances

LLM providers used by Rovo do not use customer inputs or outputs to train their models.

An icon of a plane flying around the globe

Geographic pinning

Rovo supports data residency pinning to specific geographical regions to meet local data protection requirements.

Audit transparency

Rovo logs all agent and admin activity to an audit trail accessible via Atlassian Guard, giving compliance teams the visibility they need.

This matters at a practical level, too. Adaptavist's own work with financial services customers and the recent partnership with Kosli to bring automated governance into AI-assisted delivery pipelines are guided by the same principles. AI value and AI governance don't have to be in tension, but you do have to design for both from the start.

Atlassian's bet on the governed browser

Many of the previously mentioned Mac Minis are used to run all-encompassing personal productivity agents such as OpenClaw, which expose personal and corporate data to AI inference and hallucination. But we can't just write off this technique without further consideration; the genie is out of the bottle regarding the use of AI as a personal productivity assistant, and individuals and enterprises are on a path to using it safely.

In 2025, Atlassian acquired The Browser Company for $610 million and quickly set a goal to make their Dia browser the browser for knowledge work in the AI era—with AI that is contextually aware across the tools workers already use, built with enterprise security and compliance controls baked in from the start. The company's own statistics are telling: 85% of enterprise workflows happen in a web browser, but fewer than 10% of organisations use a secure browser. Where Rovo already brings AI-native search, chat, and agents within the Atlassian platform, Dia extends that governed context layer to everything else happening in the browser, across every tab, every SaaS tool, every document, and has the ability to put them all under the same enterprise controls. As Josh Miller, founder of The Browser Company, put it: "That context, plus access to your tools, is incredibly valuable for AI. Atlassian gets that."

Dia's ability to summarise and synthesise information across multiple tabs and SaaS applications simultaneously (Jira, Confluence, all your emails just for starters) operates within an enterprise-controlled environment, subject to Atlassian's admin controls, rather than pushing data through a third-party AI service. This is a direct answer to the data sovereignty problem in the browser layer: rather than a user copying content into an external AI tool for summarisation, the intelligence operates within a governed workspace, with the same admin policies that Atlassian customers already apply across Rovo, Jira, and Confluence.

How to architect your hybrid AI strategy today

A binary choice between local and cloud AI isn't the right framing, as both have genuine uses. Cloud LLMs remain the better option for complex, open-ended reasoning tasks, where the space-age capabilities of the frontier models offer a big advantage. But local inference wins on cost, compliance, and control for high-volume, well-scoped agentic workflows. Rovo and Atlassian Intelligence sit in a third category: cloud-delivered but governed by Atlassian's enterprise controls and subject to the same admin policies as the rest of the Atlassian toolchain (although Atlassian's August 2026 data collection changes are making plan-tier selection a governance decision in its own right). And when Dia is generally available, it will extend that governed layer into the browser itself.

Organisations implementing hybrid approaches report 15–30% cost savings compared to pure-cloud or pure-edge deployments. Practically, the architecture decisions we make now about which workflows use which inference layer, compound over time. Organisations that treat their AI deployment choices with the same rigour they bring to their Atlassian environment configuration will be better placed to adapt as costs, regulations, and capabilities continue to shift.

The teams best positioned to take advantage of where Atlassian is heading with Rovo and Dia are those whose Jira and Confluence environments are already well-structured: clean data, logical project hierarchies, and consistent use of fields and metadata. The richer and more coherent the context, the more useful the AI operating across it, regardless of where the AI is actually operating. We thought we'd left the days of running a secret server under the desk behind us (though it was Jenkins or some other CI tool back in the day), but maybe there's a justifiable renaissance coming.

Let’s build your AI architecture strategy together

If you're thinking through your AI architecture strategy—whether that’s modelling the cost case for local inference, designing governance frameworks for Rovo deployments, reviewing your Atlassian plan tier ahead of the August 2026 data collection changes, or preparing your environment for what Dia will enable—talk to the Adaptavist team. We can work with you across cloud, Data Center, AI strategy, and DevSecOps, engaging with both your infrastructure and tooling, without treating them as separate problems.

Speak to Adaptavist