Some AI numbers are now too large to read as normal technology spending.
TrendForce expects the top nine cloud service providers to spend more than $830 billion on AI data-center capital expenditure in 2026. Its estimates put Amazon Web Services above $230 billion, Microsoft at about $190 billion, Google between $180 billion and $190 billion, and Meta between $125 billion and $145 billion.
OpenAI says it has moved $122 billion of committed capital into the organization. Anthropic announced a $50 billion U.S. AI infrastructure commitment. CoreWeave's reported $5.13 billion of 2025 revenue also shows why AI infrastructure companies sit in a different category from normal software adoption.
For a Luxembourg SME or mid-market company, the practical decision is not to copy those numbers. It is to check whether the AI ambition, the budget, the owner, and the operating plan match.
If AI is only a learning topic this year, a small budget may be fine. If AI is supposed to change how sales, finance, legal, operations, customer service, or product work, the budget has to look different. The claim and the investment need to live in the same reality.
This is a benchmark, not an accounting standard
The table behind this article compares public AI investment proxies with the latest available revenue or annualized revenue estimates.
That wording matters.
For public companies, revenue is usually clean enough. Annual reports and investor releases give a clear denominator. For private AI companies, the denominator is often reported ARR or an annualized revenue estimate. That makes the ratio directional, not audited.
The investment side is also mixed. Some rows use 2026 AI data-center capex estimates. Some use infrastructure commitments. Some use funding rounds. Some use AI-relevant strategic investments.
So if the benchmark says a company is at 67% of revenue, it does not mean every dollar is one-year pure AI expense. It means the public AI-related investment proxy is very large compared with the company's annual revenue base.
That is still useful.
It tells us which companies are treating AI as normal tooling, which are treating it as infrastructure, and which are treating it as the future shape of the business.
What the numbers show
The largest cloud companies are treating AI as infrastructure.
Amazon, Microsoft, Google, Meta, and Oracle are not buying a few tools for knowledge workers. They are buying compute capacity, data centers, chips, power, networking, and platform advantage. The spending is large because the ambition is large.
The AI-native companies break normal budgeting logic.
OpenAI, Anthropic, CoreWeave, and similar companies often show ratios above 100% because they are raising or committing capital ahead of revenue. That can make sense for a company whose product depends directly on compute scale. It does not make sense as a casual benchmark for a services company, retailer, law firm, accounting firm, or logistics operator.
The semiconductor and infrastructure companies are making capacity bets.
TSMC, Samsung, Nvidia, and similar suppliers sit closer to the supply chain. Their AI investment is not only about using AI internally. It is about selling into the AI build-out.
Most Luxembourg SMEs are in a different category.
They are not trying to own the AI infrastructure layer. They are trying to make work faster, reduce process drag, improve customer response, standardize knowledge, strengthen governance, and sometimes build a focused AI workflow into the business.
That needs a different budget conversation.
Low spend can be reasonable
Spending less than 1% of revenue on AI is not automatically a problem.
For a company at the beginning of adoption, that may be the right level. It can cover leadership training, team workshops, tool testing, a light policy, and a few controlled experiments. If the goal is learning, the budget should look like learning.
The problem starts when the language is much bigger than the operating commitment.
If a company says AI is central to its future, but has no budget for training, data cleanup, workflow redesign, implementation, governance, or internal ownership, the AI claim is not operational yet.
It is still a statement.
High spend can also be a problem
The opposite mistake is spending aggressively without knowing what the money is supposed to change.
Buying licenses for everyone is not the same as adoption. Building an agent is not the same as improving a workflow. Hiring a consultant is not the same as creating internal capability.
The useful test is simple.
Can the company name the workflow, the owner, the expected change, the data involved, the review rule, and the metric?
If not, the budget may create activity without capability.
A practical budget ladder
For SME leaders, the percentage is best used as a planning band, not a rule.
Under 1% can support learning, light tooling, and early experiments.
1% to 3% can support useful adoption across selected teams.
3% to 7% starts to look like serious workflow redesign and implementation.
7% to 15% suggests a strategic AI capability build, with leadership attention and stronger governance.
Above 15% is a major business-model bet. It may be right for an AI-native company or a company rebuilding a core product around AI. For a typical SME, it needs board-level clarity and a very clear reason.
These bands are not financial advice. They are a way to stop the conversation from floating.
My opinion
My opinion: the right AI budget is not a percentage first.
It is the financial shape of a decision the company has already made.
If AI is a productivity tool, the budget should look like enablement: training, tools, support, and manager visibility.
If AI is changing workflows, the budget should include process redesign, data preparation, implementation, testing, and adoption.
If AI is becoming part of the product, the budget has to look more like product investment. That means stronger ownership, more technical depth, clearer risk review, and a longer-term funding plan.
The percentage matters because it exposes whether the company is serious about the version of AI it says it wants.
What to do this quarter
Start with the ambition, not the tool.
- Decide what AI is supposed to be in the business this year: learning topic, productivity layer, workflow redesign, product capability, or business-model bet.
- Pick three to five workflows where AI could remove real drag. Use actual workflows, not broad categories.
- Name the owner. If nobody owns the workflow, nobody owns the result.
- Set the budget band. A small learning budget is fine when the goal is learning. A serious AI claim needs a serious operating budget.
- Review the spend quarterly against usage, time saved, quality, risk, revenue impact, and team adoption.
The large-company benchmark is useful because it removes one illusion. AI strategy is not only what a company says about AI. It is what the company is willing to fund, govern, and put into the hands of people doing real work.
Practical, not theoretical.