Articles / Explore / AI Business Value
If an AI service became unavailable tomorrow, what would your team actually lose? A convenient writing assistant is one answer. The ability to complete routine work is another. That difference changes what an AI investment is worth.
On September 14, 2026, The Guardian reported Christine Lagarde’s call for Europe to develop more of its own AI capability. Speaking in Vienna that day, the European Central Bank president argued that dependence on foreign technology could expose many sectors to changes in access. Her concern was not simply who builds the cleverest chatbot; it was who can keep using the technology.
The news: adoption and access belong in the same discussion
The ECB’s published speech sets out three ambitions: more European computing capacity, models suitable for ordinary tasks running on European infrastructure, and continued access to the technological frontier. Lagarde presented these as complementary investments, not a claim that Europe can make every component itself.
The speech also cites a scenario in which rapid AI adoption raises the level of total factor productivity by up to 4% over a decade. That is a conditional economy-wide estimate—not 4% annual growth, not a measured company result, and not a return that belongs in every AI business case. Source: ECB speech, including footnote 9.
Our interpretation: enterprise buyers should separate the usefulness of today’s service from the durability of access to it. A compelling demonstration answers the first question. It says much less about the second.
What “building AI” costs beyond a subscription
Infrastructure makes this debate concrete. In a May 14, 2026 cost model, researchers Amelia Michael and Ben Cottier at Epoch AI estimate roughly $38 billion in upfront capital for a hypothetical one-gigawatt US AI data center. They convert capital expenditure into annual costs using asset lifespans and financing assumptions, then add operating expenses.
Their resulting total is about $8.5 billion per year. Servers account for approximately $5.0 billion. In the published annualized table, the precise displayed amounts are $5.021 billion for servers and $8.514 billion overall; subtracting servers from the reported total leaves $3.493 billion for other costs. Figure 2 shows that comparison.
This is a stylized estimate for US hyperscaler infrastructure, not a bill for an actual facility or a European procurement quote. Hardware, location and financing assumptions matter. It also should not be scaled down mechanically to price a small company’s AI project. Epoch AI explains its assumptions and limitations.
The business distinction is important: using a hosted service buys access to infrastructure, while owning infrastructure creates responsibilities for keeping it useful. A lower monthly service bill and a lower total cost are different claims. Neither can be inferred just from a model’s benchmark score.
An ordinary workplace example: the proposal desk
Illustrative scenario—not a reported customer case. Imagine a small equipment supplier whose sales-support team uses an AI service to draft proposals from approved product descriptions. People still check every proposal. The service does not decide prices or send offers.
On a normal morning, the team values faster first drafts. If access disappears, the consequence depends on where the working knowledge lives. Suppose approved descriptions, templates and completed proposals remain in the company’s own shared workspace. Staff can continue manually, though more slowly.
Now change one assumption: suppose the latest product explanations and reusable instructions exist only inside that service. The interruption also becomes a reconstruction problem. The team must locate and check the material before it can resume dependable work. The drafting task has not changed, but the exposure has.
A second AI subscription would not automatically solve this example. Someone would still need to check that the replacement handles the same documents and produces acceptable drafts. Conversely, buying servers would not by itself recreate missing working knowledge. The useful investment might be a maintained set of approved inputs and a tested manual route, not another model.
There is no invented outage probability or savings percentage here. The example illustrates why the value of an alternative depends on the work it preserves. A team that can tolerate a delay has a different case from one whose customer commitments depend on immediate turnaround.
What the announcement does—and does not—settle
Lagarde’s speech is a policy argument, not evidence that domestic hosting guarantees continuity or that overseas suppliers will cut access. It also notes that future open-model releases and licensing terms are not guaranteed. The original speech supports taking dependency seriously; it does not identify the cheapest architecture for your company.
For business readers, the interesting shift is in the investment question: alongside “What can this system do?”, ask “What capability are we securing?” That makes room for useful differences between buying a service, keeping an alternative ready and owning infrastructure. It does not make one choice universally best.
Three takeaways
1. Access is part of value. Time saved today matters, but so does the ability to keep essential work moving.
2. Ownership has a full cost. The infrastructure estimate shows why a building or electricity bill alone cannot represent the whole expense.
3. National ambition is not a company ROI forecast. Evaluate your own workload and alternatives rather than importing a macroeconomic percentage.
One proportionate next step: ask the owner of one important AI-supported activity to describe what the team would do during a one-day interruption. Record what remains available and what must be rebuilt. Start with the answer, not a new purchase.
So what would your team lose tomorrow? The answer tells you whether you are buying convenience, supporting a capability, or creating a dependency that deserves a better alternative. That is the practical business question behind Europe’s latest AI investment debate.
Sources and related reading
The Guardian: Europe must build own AI or risk getting cut off by US or China — September 14, 2026; accessible editorial reporting. The report credits a Reuters contribution, so those two reports are not treated here as independent evidence.
Christine Lagarde, ECB: A new age of capital: growth, sovereignty and AI — speech delivered and published September 14, 2026. Primary evidence for the stated policy position and conditional productivity scenario.
Amelia Michael and Ben Cottier, Epoch AI: Servers account for 60% of the total cost of ownership of a one-gigawatt AI data center — May 14, 2026. Background cost model, not a new announcement. All sources accessed September 15, 2026.
Continue with our analysis of AI activity versus business value, or explore the AI Business Value collection.
Leave a Reply