AI Strategy for Enterprises: Siting Risk

Siting is now a supply‑chain risk

This week’s development — a judge allowing the NAACP’s lawsuit over alleged pollution from a data centre linked to xAI to move forward — is not an isolated headline. It’s a signal that AI capacity is exposed to local politics, permitting battles and grid constraints. Whether that particular case succeeds or fails, the direction of travel is clear: objections can slow, shrink or reshape facilities that your models depend on.

We are also seeing the financing and architecture side evolve. Reports of a major chip vendor weighing whether to co‑sign a third‑party AI campus underline how capital, energy access and community acceptance now intertwine. Experimental deployments — from novel “satellite” data‑centre concepts to public remarks about memory shortages impacting server builds — show the stack is still in flux. In short: supply is tight, public scrutiny is rising and plans can change late.

If you treat compute like a utility that is always there, you will be surprised. If you treat it like a delicate supply chain, you can harden it. That’s the mindset shift I want leadership teams to make: compute siting belongs on the same risk register as a key component supplier, a logistics bottleneck or a regulatory approval. The practical question is not “will someone object?” but “what do we do when they do, and how quickly can we move without losing the quarter?”

From ESG talk to enforceable SLOs

Most sustainability commitments sit in glossy reports, far from the clauses we enforce when a service runs hot. Bring them together. For AI workloads, I ask providers to accept service level objectives for energy, water and carbon alongside availability and latency. The goal is not to punish suppliers; it’s to align operational decisions (where, when and how we run) with the environmental constraints that increasingly determine whether a site can keep operating.

A workable approach is to express SLOs that attach to the workload and the region. For example: targets on the average emissions intensity of power used over a period, boundaries on water consumption practices in stressed basins, and commitments to publish timely, verifiable metrics. You don’t need exotic measures; you need clear definitions, data sources and remediation paths. If a provider can’t measure it today, insist on a roadmap and interim reporting.

Enforcement should resemble any other SLO: credits, make‑good capacity, and the option to migrate workloads without penalty if a site drifts out of band or becomes politically untenable. Avoid greenwashing: offsets are not the same as reduced impact at the point of use. The purpose here is realism. If a grid operator triggers curtailment on hot afternoons, we design and price for that reality. If a water permit looks fragile, we know the triggers that send us elsewhere.

Contract levers that keep you shipping

When I review AI and cloud agreements now, I look for three families of protections: curtailment, relocation and price integrity. Curtailment clauses should define planned and unplanned reductions in power or cooling capacity, the notice we receive, our priority tier, the maximum expected hours per period, and the remedies. If the provider participates in demand response, I want clarity on how that participation affects my capacity and how I’m compensated when I am asked to stand down.

Relocation rights are next. If a facility is enjoined, delayed or materially constrained by community opposition, permitting actions or grid restrictions, I want the right to move capacity to a pre‑approved list of alternate regions or providers without egress penalties, with price parity for the contracted term and with provider support to re‑platform where needed. This has to be written down before the incident, not improvised on a bad Friday.

Finally, handle legal and financial shocks. Change‑in‑law clauses should avoid turning every new environmental requirement into a pass‑through surcharge. Force majeure should be tightly defined so that foreseeable siting issues don’t excuse poor planning. I also ask for audit and reporting rights tied to the environmental SLOs, so we can evidence what we run and where, which helps with emerging disclosure regimes and, frankly, with board conversations when questions come.

Diversify regions and providers with a plan B

Capacity diversification is not a slogan; it’s a portfolio design problem. Spread training and inference across a short list of regions with different grid profiles, water stress and permitting landscapes. Mix hyperscalers with specialist AI clouds and, where it fits, on‑prem or colocation capacity near low‑risk energy. The aim is to ensure that no single local objection, weather pattern or grid event can halt your roadmap.

For the checklist leadership teams use before committing to AI, see The AI Readiness Assessment.

Pre‑approved fallback capacity is the difference between theory and practice. Reserve “warm” capacity where the business case justifies it: minimal standing instances, reserved but idle GPU slices, or contractual access to burst pools with defined SLAs. Keep a “cold” option too: a vetted provider and region with security, compliance and procurement cleared, images published and runbooks ready, even if you’re not paying to hold silicon there today.

Portability is the enabler. Standardise container images, model artefacts and data schemas. Keep feature stores, vector indices and checkpoints versioned and replicable. Use cloud‑agnostic identity and key management patterns wherever sensible. Then test. A quarterly failover exercise that actually trains or serves in the fallback region is worth more than a hundred slide decks. Capture the latency and cost deltas, fix the surprises, and update the risk register.

Do not forget data residency and sovereignty. Some workloads cannot leave a jurisdiction; others can, with controls. Make those determinations up front, tag datasets and pipelines accordingly, and map them to the diversification plan so that compliance isn’t discovered mid‑incident.

Costs: premiums, penalties and insurance

There is a cost to resilience, and pretending otherwise just hides the bill. Greener or less‑stressed regions can carry a premium. Cross‑region replication adds storage and data transfer. Warm standby capacity ties up budget. Failover tests consume real time and money. Engineers will spend cycles on portability and observability that aren’t visible in a demo but pay back in a crisis.

Price those items deliberately. I treat resilience as an insurance line on the AI P&L: a small percentage uplift justified by the value of on‑time launches and uninterrupted service. Attribute it to specific workloads so product owners can decide where to buy greater safety and where to accept more risk. Put the spend under FinOps so you can track benefits as avoided incidents or reduced exposure when a region faces curtailment.

There are offsets. Commit discounts still apply across diversified footprints if negotiated carefully. Interruptible or lower‑priority inference can ride cheaper pools if latency budgets allow. Training can be time‑shifted to lower‑cost, lower‑carbon windows if check‑pointing and job control are mature. Some providers will waive egress during forced migrations if you negotiate it now. And visibility into energy and water usage can itself reveal efficiency wins that repay the investment.

The biggest hidden cost is lock‑in that prevents movement. A slightly cheaper, more proprietary stack that you cannot move for six months is not cheaper once a permitting fight forces a relocation. Price in the option value of portability.

What can go wrong and how to govern it

The failure modes are predictable. Vague environmental SLOs that can’t be verified. Contracts that treat every curtailment as your problem. Monitoring that tells sustainability stories but not operational truth. Fallback regions that pass a security review but fail latency tests. And reputational risk when your brand is seen to benefit from facilities that neighbours feel harm them.

Where the organisation needs an independent baseline first, an AI readiness assessment scores strategy, data, architecture, people, governance and measurement.

Set governance that closes those gaps. Put siting risk on the enterprise risk register with a defined owner. Form a working group across Architecture, SRE, Sustainability, Procurement, Legal and FinOps. Define triggers (regulatory notices, community challenges, grid alerts) that escalate to an executive review. Establish a cadence for reporting energy, water and carbon alongside availability and spend. Importantly, link those metrics to decisions: where we schedule jobs, where we scale, when we move.

On supplier due diligence, look beyond glossy net‑zero slides. Ask about community engagement, local hiring and mitigation measures. Understand how the provider responds to curtailment and drought conditions. Request references for similar regions. Ensure your comms team has a plan for when — not if — a local story touches your brand, even if you don’t own the facility.

When incidents happen, move decisively. Execute the pre‑approved migration. Waive non‑critical experiments to protect customer‑facing throughput. Brief the board with a clear timeline, cost implications and the reason you’re not scrambling: because you anticipated the risk and invested in a fallback.

AI Strategy for Enterprises: a practical plan

Here is the shortlist I use with leadership teams:

– Map your critical AI workloads (training and inference), their latency and data requirements, and the regions and providers they currently rely on.

– Score siting exposure for each workload: permitting risk, grid stress, water constraints, community sentiment and regulatory trajectory. Use supplier input and independent signals.

– Define environmental SLOs for energy, water and carbon that fit your business, and negotiate them into AI and cloud contracts with measurement, reporting and remedies.

– Add curtailment, relocation and price‑integrity clauses. Pre‑approve alternate regions and at least one secondary provider per critical workload.

– Build portability: containerise, version artefacts, externalise state, and standardise identity and keys. Automate environment creation in primary and fallback targets.

– Reserve enough fallback capacity to meet your recovery‑time objective for training and your peak throughput for inference. If warm standby is too costly, secure burst rights and test them.

– Run a quarterly failover exercise that moves at least one production workload. Track cost, latency and operational friction. Update the plan and budget accordingly.

– Establish governance: assign owners, set triggers, align reporting, and brief the board on the risk and the mitigations you’ve put in place.

If you want an experienced hand to help you move from intent to contracts, runbooks and tested failovers, this is where an Enterprise AI Consultant earns their keep. A seasoned Technology Strategy Consultant can translate sustainability aims into enforceable SLOs and procurement positions, while CTO Advisory Services can align capacity diversification with product roadmaps and budget. The point is not to admire the problem; it’s to ship reliably despite it.

Compute siting will remain contested as AI demand grows. Teams that plan for it will keep their commitments; teams that don’t will relearn operational lessons the hard way. Treat siting as a supply‑chain risk, write it into your deals, diversify deliberately and rehearse your escape lanes. That is technology made practical.