AI strategy eventually becomes hardware strategy too
AI can look like a software subscription, but every workload runs somewhere: on a laptop, in a server room or in a data centre. The infrastructure choice affects cost, capability and what the organisation eventually has to maintain or retire.
The AI decision has a physical side
Behind an AI workflow sit processors, storage, networking, cooling, electricity and end-user devices. A cloud subscription moves most of that equipment out of the business’s office; it does not make the infrastructure disappear. For an SME, the practical question is not to calculate the footprint of every prompt. It is to make sensible architecture and purchasing decisions for work that matters.
A local model may be useful where response time, offline operation, specialist data handling or direct device control matters. A cloud service may be a better fit when the business wants to avoid specialist equipment, reach users on older devices or scale without a large upfront purchase. Neither location is automatically cheaper, more private or more sustainable in every configuration.
Before adding an AI PC or dedicated GPU to a budget, ask whether the chosen task needs it. A marketing team summarising approved documents may be well served by current devices and a managed cloud application. A specialist team processing data offline may have a reason to test local inference. Requirements—not a product label—should drive the refresh.
What does the headline e-waste estimate count?
In September 2026, Basel Action Network published The Coming AI Waste Wave, a white paper estimating that 395–617 million tonnes of AI-driven electronic equipment could be retired between 2025 and 2050. It is a modelled range, not a measured forecast of landfill. The report’s scope is broader than accelerator and server estimates: it includes multiple data-centre equipment categories and an “AI waste contagion” scenario in which AI adoption could accelerate retirement of equipment beyond the data centre.
The result depends on assumptions about infrastructure growth, equipment mass, replacement cycles and how AI adoption changes retirement of PCs and other devices. BAN’s figure should be read as one organisation’s scenario under stated assumptions—not as guaranteed waste or a claim that every retired device goes to landfill. Equipment can be maintained, repaired, refurbished, reused or recycled. Read BAN’s September 2026 white paper and its summary of the estimate.
Choose where the work runs after defining what it needs
A company searching technical manuals across a workshop might compare a hosted AI service accessed from current laptops, a small on-premise server or a local model on upgraded devices. The answer depends on sensitivity, connection quality, latency, maintenance capacity, licensing and the model performance needed. A hardware purchase before that comparison risks paying for capacity no one uses.
- What task is needed, how often will it run and what quality is sufficient?
- Does it need low latency, offline operation or local control—or is managed cloud more practical?
- Which existing computers can run the workflow? Test before replacing equipment.
- Does the full cost include purchase, energy, support, cloud use, refresh and disposal?
- Who will repair, redeploy, securely wipe or responsibly recycle the equipment?
Compare the full operating picture: hardware, electricity, cloud usage, support, security updates, integration, refresh timing and the effort required to maintain a local model. A lower monthly inference price can still lose to a machine that is underused; a hosted service can become expensive if the workflow is poorly bounded or runs constantly.
Treat useful life as part of the business case
When equipment reaches the end of its current role, consider whether it can be repaired, redeployed to a lighter workload or refurbished before replacement. If it must leave service, use a responsible recycling route and ensure business data is securely removed. These steps can protect the value of a purchase and avoid unnecessary material turnover without claiming a specific environmental saving.
For most SMEs, a proportionate plan is straightforward: test the task on available equipment, document any performance or privacy requirement that existing devices cannot meet, and buy only what closes that measured gap. Revisit the decision when the workload, model or service changes.
Good AI strategy includes the equipment behind the interface
Model selection is one part of implementation. A business also needs to understand its data, process, people, infrastructure and lifecycle. The purpose is not to turn every AI project into an environmental study; it is to avoid treating physical constraints and costs as someone else’s problem. For local, edge and cloud trade-offs, read AI is moving from the cloud onto your devices. For the commercial view, see The real cost of business AI.
At scale, AI strategy is not only software, data, people and process. It includes the infrastructure required to run the work—and what happens when that infrastructure reaches the end of its useful life.