Ask What the AI Investment Actually Builds

The monitor estimates that global foreign direct investment rose 37% in the first half of 2026 against the average half-year level of 2025—not against the first half of 2025. Its project series use different observation windows and annualised comparisons. These are broad international-investment measures, not a census of usable AI capacity. ¹

There is an important counterweight to the decline in developing-economy data-centre greenfield announcements. Telecommunications project-finance value, driven mainly by data centres, increased. UNCTAD notes that large data centres are increasingly financed through project structures rather than investment shouldered by individual multinationals. A falling measure in one category therefore cannot establish a general collapse in building. ¹

For a board, the practical consequence is simple: do not combine a flow of money, an announced project and an operating asset into one growth story. Nor should an apparently weaker category automatically trigger retreat. First ask whether the underlying activity moved, the funding structure changed, the comparison period shifted or the project itself changed.

This discipline matters in both directions. Optimism can turn capital announcements into assumed delivery. Pessimism can turn a narrower statistical decline into a presumed shortage. An investment committee should challenge either shortcut with the same question: which observation supports the decision we are about to make?

Imagine a company preparing an AI-enabled customer service for three markets. Its supplier announces a financing partnership for a new data centre. The company then treats that announcement as confirmation that its launch can expand. This is a hypothetical planning mistake, not a finding about any named supplier.

Several links remain unproven. The financing may concern an existing asset, a future expansion or a pool of projects. The relevant site may not serve the planned market. Even a completed facility may not offer the required model, data-handling arrangement or customer allocation. The buyer needs evidence about its own service, not merely confidence in the supplier's total spending.

Start with a single project identity. Record its location, sponsor, financing route and stated purpose. Then distinguish announced support, committed finance, construction, commissioning and customer availability. These are proposed management labels, not new accounting rules. Their value is that finance and operations can point to the same thing when they discuss progress.

Keep a separate field for what the project would enable for your organisation. A capacity reservation, a service launch and a regional availability statement answer different questions. Ask the supplier which one it can substantiate and what remains conditional. Do not ask a public announcement to carry contractual promises it never made.

The International Energy Agency supplies a separate operational warning. Its October review says connection queues include duplicate or speculative projects; even projects that proceed bring buildings, servers and customer workloads online gradually. This is a review of electricity-system integration, not a list of which suppliers will miss delivery. It reinforces the need to check a project's actual stage instead of treating requested power as working capacity. ⁶

A global company should apply a consistent evidence standard without assuming identical delivery conditions. Its teams may face different electricity connections, network routes, data rules, languages and procurement requirements. Those differences belong in the investment case before a common launch date is agreed, not in a list of exceptions after the budget is approved.

In the Middle East and North Africa, the World Economic Forum's October reports frame AI opportunity alongside energy, water and grid constraints. The reports were developed with Bain and SRMG respectively; they are forward-looking analysis, not guarantees that a particular site will be connected on time. A regional opportunity estimate should not become an individual project's delivery assumption. ²

For Africa, UNDP and the African Development Bank describe a country-by-country route under the AI 10 Billion initiative: identify productivity opportunities, prepare projects for financing and convene investment partners. The 12 October statement describes intended work and expansion, not completed investment or realised employment. Its useful distinction is between an attractive opportunity and a project ready to finance. ³

For North American and European buyers, the review should identify which local service an international investment actually supports. For Asia-Pacific operations, ask whether the proposed location and product meet the intended workload and access requirements. For Latin America and the Caribbean, require the same project-level evidence rather than importing another region's delivery assumptions. These are diligence questions, not claims that every market in a region shares one constraint.

We have not established comparable country-level measures of commissioned AI capacity in this scan. That gap should remain visible. A board can authorise different rollout sequences where the evidence differs without declaring one region inherently ready and another inherently unready.

Once a service is available, someone still has to make it useful. A new preprint reports 21 interviews with 27 workers across 21 small and medium-sized businesses in Kenya and Nigeria, conducted from July to October 2023. Participants described checking, adapting and repairing AI-assisted drafts before sharing them with clients. These are situated accounts, not a representative survey or a measure of today's tools. ⁴

That research does not prove an infrastructure investment will fail. It exposes a separate budget question: who will make the output fit the work? A plan that funds access but omits local review, language support and accountable release may have a complete technology purchase and an incomplete service.

The week's research radar also shows why capacity cannot be judged by hardware alone. FusionEval, a preprint using expert models derived from Qwen3 and Llama, finds trade-offs between preserving reasoning and tool use when combining models. Its eight-benchmark evaluation is not a business outcome, but it cautions against assuming one apparent efficiency gain preserves every capability. ⁵

Keep those two questions separate in the approval: can the organisation obtain the service, and can that service perform the intended work acceptably? Resolving one should allow the next test to proceed. It should not silently mark the other as complete.

Ask the finance lead and the operating sponsor to choose one material AI dependency already in the plan. Avoid a company-wide inventory exercise as the first response. Select the dependency whose delay or changed terms would most affect a near-term commitment, and identify the decision date that makes the review useful.

Finance should establish what is being funded and whether the evidence describes an intention, an agreement or money already available to the project. Operations should establish what must be delivered, where and by when. Procurement should connect those facts to the organisation's actual access and remedies. Record unknowns plainly rather than filling them with the supplier's broader growth story.

Have local business and technical owners test the service assumptions. Does the planned offering support the relevant language, permitted data use and required response time? Who will review difficult cases? What would happen if the preferred route arrives later than expected? A fallback can be a narrower launch or continued existing work; it need not be a second expensive platform.

Then separate expenditure that can sensibly proceed now from expenditure that depends on an unresolved condition. Staff preparation or a controlled test may remain worthwhile while a larger rollout waits. Conversely, a verified operating service may justify progress despite gloomy aggregate investment news. The decision should follow the evidence about this project, not the mood of the headline.

At the next review, show what changed in the record. Did a financing condition clear? Did the service become available to the intended customer? Did a local test reveal a missing capability? Name the person who can change the rollout decision. This keeps a monitoring exercise connected to action rather than producing another status report.

The opportunity is to make expansion more discriminating, not slower by default. Better evidence can justify moving ahead where a general risk narrative would discourage it. It can also reveal a dependency early enough to change the plan without disrupting customers or wasting the next round of spending.

Build a shared project record. Buy the access and support the intended work needs. Test availability and usefulness separately. Watch financing and delivery changes that affect your dependency. Ignore claims that one global investment number settles all five questions. The board should leave this week's discussion knowing what its chosen AI investment enables—and what must still happen before the business can rely on it.