AI investment has reached a strange stage.
Some companies are worried that they are already too late.
Others are spending money because they do not want to appear behind their competitors.
Management teams are approving AI licences.
Departments are testing chatbots.
Developers are connecting applications to language models.
Cloud usage is increasing.
But one question is still difficult for many organisations to answer:
What are we getting back?
The latest results from Microsoft and Amazon show why that question matters.
Both companies are spending extraordinary amounts on data centres, processors, memory, networking and power. They are not investing because AI sounds interesting. They are investing because customers are buying more cloud and AI services.
Microsoft reported quarterly revenue of $90 billion, an increase of 18% from the previous year. Operating income reached $40.6 billion, while net income was $35.8 billion. The company attributed the results to continued strength in its cloud and AI businesses.
Amazon reported that Amazon Web Services sales increased 37% year over year to $42.2 billion during the second quarter. Amazon’s total operating income increased to $27.5 billion from $19.2 billion during the same period a year earlier.
These are very large numbers.
But the more useful lesson for ordinary businesses is simple.
Technology spending becomes easier to justify when it produces visible demand, revenue, productivity or operational improvement.
Why Microsoft’s Results Received So Much Attention
Companies have been spending heavily on AI infrastructure for several years.
Investors have increasingly asked whether those investments will generate enough revenue to justify the cost.
Microsoft’s latest results provided some reassurance.
Reuters reported that Microsoft expects Azure cloud revenue growth of approximately 45% in the coming quarter, above analyst expectations. The company also maintained its infrastructure-spending plans rather than announcing another major increase.
Microsoft’s shares rose more than 15%, adding close to $450 billion to the company’s market value in a single trading day. The market reaction reflected growing confidence that Microsoft’s spending on AI and cloud capacity is being matched by customer demand.
This does not mean every Microsoft AI investment will succeed.
It means the company was able to show a stronger connection between infrastructure spending and business growth.
That connection is what every company should look for.
Amazon Is Spending More Because Demand Is Still Ahead of Capacity
Amazon’s situation is slightly different.
Its AWS cloud business grew 37%, reaching $42.2 billion in quarterly revenue. Reuters described this as AWS’s strongest growth in more than four years.
Amazon also increased its expected 2026 capital spending from approximately $200 billion to $220 billion.
Part of the additional cost is connected to rising memory-chip prices. Amazon also says demand for AI computing remains greater than the infrastructure capacity it can currently provide. Much of its available cloud capacity for 2027 is reportedly already committed, while customers are also planning workloads for 2028.
This tells us something important about AI infrastructure.
The cost is not only the AI model.
It includes:
- Data-centre buildings
- Processors and GPUs
- Memory
- Storage
- Networking
- Power
- Cooling
- Backup systems
- Engineers
- Security
- Software platforms
When usage grows, all these layers may need to grow with it.
The Lesson Is Not to Spend Like Big Tech
A normal company should not look at Microsoft and Amazon and conclude that it needs to spend aggressively on AI.
These companies operate global cloud platforms.
Their customers pay them for computing capacity.
Infrastructure is part of what they sell.
Most businesses are different.
A retailer sells products.
A manufacturer produces goods.
A professional service company sells expertise.
A logistics company moves cargo.
A hosting provider delivers infrastructure and managed services.
The purpose of AI investment should therefore be linked to the company’s actual business.
The first question should not be:
“Which AI tool should we buy?”
It should be:
“What business problem are we trying to solve?”
Begin With the Problem, Not the Product
A weak AI project often begins with a tool.
Management sees a demonstration and decides that the company should use it.
The team then searches for somewhere to apply it.
A stronger project begins with a problem.
For example:
- Customer enquiries take too long to classify.
- Sales staff spend too much time preparing standard proposals.
- Employees cannot find information across internal documents.
- Support teams repeatedly answer the same questions.
- Finance spends too much time extracting data from invoices.
- Management reports take several days to prepare.
- Security alerts are not prioritised quickly enough.
Once the problem is clear, the business can evaluate whether AI is actually the right solution.
Sometimes the answer may be AI.
Sometimes the better answer may be:
- A simpler workflow
- Better staff training
- A system integration
- A reporting dashboard
- Improved documentation
- A normal automation rule
- Additional computing capacity
AI should not be used to hide a broken process.
Define the Expected Outcome Before the Trial
An AI trial should begin with a measurable expectation.
For example:
“Reduce the average time required to classify a customer ticket from six minutes to two minutes.”
“Reduce the time needed to prepare the first draft of a quotation by 40%.”
“Allow employees to find approved policy information within one minute.”
“Reduce the number of repetitive support questions handled manually by 25%.”
These targets do not need to be perfect.
They give the trial a clear purpose.
Without a target, almost any demonstration can appear successful.
The AI produces an answer.
The team is impressed.
The company purchases more licences.
Several months later, nobody can explain whether the business is actually working better.
Measure the Cost Per Useful Result
Looking only at the monthly AI bill is not enough.
A company may spend RM10,000 per month on an AI service and receive strong value.
Another company may spend RM1,000 and receive almost nothing useful.
The better measurement is the cost per outcome.
The FinOps Foundation describes unit economics as a way to connect technology spending with business-relevant value or demand. Depending on the organisation, this may include cost per customer, transaction, workload, team or service delivered.
For AI projects, practical measurements may include:
- Cost per customer enquiry resolved
- Cost per document processed
- Cost per report generated
- Cost per sales opportunity supported
- Cost per successful automation
- Cost per software feature delivered
- Cost per employee using the system actively
- Cost per hour of manual work saved
This gives management a clearer view than simply reporting token consumption or total cloud spending.
Do Not Confuse Usage With Value
High usage does not automatically mean the AI system is successful.
Employees may generate many prompts because the tool is difficult to use.
An AI agent may make several model calls to complete one simple task.
A chatbot may receive many questions but fail to resolve them accurately.
A coding assistant may generate a large amount of code that still requires substantial review.
Usage tells the company that the tool is being used.
It does not tell the company whether the output is valuable.
A useful review should combine:
- Usage
- Accuracy
- Completion rate
- Staff time saved
- Customer impact
- Error rate
- Security incidents
- Total cost
The system may be popular but expensive.
It may be accurate but too slow.
It may save time but introduce unacceptable risk.
The decision should be based on the full result.
Include the Hidden Costs
AI vendors usually present clear subscription or API pricing.
The wider implementation cost may be less obvious.
A serious business deployment may also require:
- Data cleaning
- System integration
- Cloud storage
- Additional processing
- Cybersecurity review
- Access control
- Monitoring
- Backup
- Staff training
- Policy development
- Technical support
- Output verification
- Legal or compliance review
These costs should be included in the business case.
For example, an AI tool may save 20 hours of administrative work each month.
But if the IT team spends 30 hours maintaining its integration, the overall result may not be positive.
The project should measure the complete operating effort, not only the invoice from the AI provider.
AI Costs Can Be Difficult to Forecast
Traditional software is often priced per user per month.
AI services may use several different pricing methods.
These can include:
- Subscription licences
- Tokens processed
- API requests
- GPU hours
- Images generated
- Documents processed
- Storage
- Data transfer
- Dedicated capacity
Usage may also be unpredictable during the early stages of a project.
The FinOps Foundation notes that AI pilots can be especially difficult to budget because consumption volumes may be hard to predict before real usage begins. It recommends connecting usage and cost information with business-value measurements rather than managing the bill in isolation.
This is why an AI trial should have a budget limit.
The company should know:
- Who is allowed to use the service
- What the monthly limit is
- Who receives spending alerts
- What happens if usage suddenly increases
- Whether individual departments are accountable
- How non-production usage is controlled
Unlimited experimentation can create an unexpected invoice.
Separate the Trial From Production
A trial environment should not automatically become a production system.
During the trial, the company is trying to answer questions.
Does the tool work?
Is it accurate enough?
Will employees use it?
Does it save time?
Is the cost reasonable?
A production environment needs stronger controls.
These may include:
- Managed user accounts
- Multi-factor authentication
- Access restrictions
- Data-retention rules
- Monitoring
- Backup
- Support ownership
- Incident response
- Availability requirements
- Documented recovery steps
A small trial may depend on one employee and one spreadsheet.
A production service should not.
Before scaling, the company should decide who owns the system and who is responsible when it fails.
Choose the Right Model for the Job
The most advanced AI model is not always the best choice.
A large model may provide stronger reasoning.
It may also be slower and more expensive.
A smaller model may be sufficient for:
- Classification
- Data extraction
- Standard summaries
- Sentiment analysis
- Simple question answering
- Template-based drafting
More demanding tasks may require a stronger model.
The correct approach is to test several options against the actual workload.
Do not compare models only through general benchmarks.
Measure them using the company’s own documents, questions and quality requirements.
The cheapest model is not useful if the output is wrong.
The most expensive model is wasteful if a smaller one performs equally well.
Avoid Using AI Where a Normal Rule Is Enough
Not every automation needs artificial intelligence.
A normal rule may be more reliable and cheaper.
For example:
- Rename a file according to a fixed pattern.
- Route an invoice according to supplier name.
- Send a reminder three days before expiry.
- Reject an upload above a certain size.
- Generate a report from structured database fields.
These tasks may be handled through existing software, scripts or workflow tools.
AI is more useful when the input is unstructured, variable or difficult to interpret using fixed rules.
Examples may include:
- Understanding natural-language enquiries
- Summarising long documents
- Extracting information from different layouts
- Comparing complex text
- Assisting with draft responses
- Searching across unstructured knowledge
Choosing the simplest reliable solution usually produces better long-term results.
Security Must Be Included in the Cost
AI projects often require access to business data.
That creates security responsibilities.
The company should know:
- What data is being processed
- Where the data is stored
- Whether it is used to train the provider’s model
- Who can access prompts and responses
- Whether logs are retained
- Whether confidential information is filtered
- How access is removed
- What happens during a breach
Security work may add cost.
It is still necessary.
A low-cost AI tool can become expensive if it leads to data leakage, incorrect customer communication or a compliance problem.
The project budget should therefore include security review and ongoing monitoring where appropriate.
Human Review Still Has a Cost
Many AI systems require human review.
That is not necessarily a problem.
AI may still save time even when a person checks the result.
But the checking time should be measured.
For example:
- AI prepares a quotation draft in two minutes.
- A sales executive spends ten minutes checking it.
- The previous manual process took twenty-five minutes.
The actual saving is thirteen minutes, not twenty-three minutes.
This is still useful.
The measurement simply needs to be honest.
For high-risk work such as legal, financial, technical or customer-facing output, human review may remain essential.
Set a Clear Decision Date
Trials have a habit of continuing indefinitely.
The company continues paying.
Employees continue testing.
Nobody decides whether the project should expand, change or stop.
A better trial has a defined review date.
At that point, management should choose one of four outcomes:
Scale
The project meets its quality, security and financial targets.
Expand it carefully.
Improve
The idea is useful, but the workflow, model or integration needs more work.
Run a revised trial.
Limit
The tool is useful only for selected teams or tasks.
Keep the scope controlled.
Stop
The project does not deliver enough value.
Cancel it before more money and time are spent.
Stopping an unsuccessful project is not failure.
Continuing without evidence is worse.
What Management Should Ask Before Approving More AI Spending
Management does not need to understand every technical detail.
It should ask practical questions.
- What problem are we solving?
- Who owns the outcome?
- How is success measured?
- What is the expected monthly cost?
- What costs are excluded from the vendor quotation?
- Which data will the tool access?
- How will output be checked?
- What happens if usage doubles?
- Can a smaller model perform the task?
- Can a normal workflow solve the problem?
- What is the cost per useful result?
- When will we decide whether to scale or stop?
Clear answers show that the project is being managed as a business investment.
Vague answers such as “everyone is using AI now” should not be enough.
Cloud Spending Requires Shared Responsibility
AI and cloud spending should not be managed only by the IT team.
IT understands infrastructure and security.
Finance understands budgeting and cost.
Department managers understand workflow and business value.
The users understand whether the tool is genuinely helpful.
A strong AI programme brings these groups together.
The FinOps approach encourages cooperation between engineering, finance and business teams so technology spending can be connected with value and accountability.
Without shared responsibility, each group may see only one part of the problem.
IT sees usage.
Finance sees the bill.
Management sees the promised benefit.
Users see the daily workflow.
All four perspectives are needed.
Do Not Assume Cloud Capacity Is Unlimited
Microsoft and Amazon are investing enormous amounts because demand for AI infrastructure continues to grow.
Amazon says demand remains greater than the capacity it can provide, even after raising its planned spending to approximately $220 billion.
This does not mean ordinary businesses will suddenly be unable to access cloud services.
It does show that cloud capacity depends on physical infrastructure.
Processors, memory, power, cooling and data-centre space cannot be created instantly.
Businesses planning large or time-sensitive AI workloads should therefore confirm:
- Whether the required capacity is available
- Which cloud region will be used
- Whether reserved capacity is necessary
- What happens during capacity shortages
- Whether another region or provider is available
- How data can be moved
A project should not assume that any amount of computing power will always be available immediately.
When Self-Hosted AI May Make Sense
Most businesses will begin with a cloud or software-based AI service.
This is usually easier.
The provider handles the underlying model and infrastructure.
Self-hosted AI may become worth considering when the company has:
- High, predictable usage
- Strong privacy requirements
- Suitable technical staff
- Existing GPU infrastructure
- Low-latency requirements
- A need for greater model control
- A clear long-term workload
Self-hosting also introduces responsibility.
The company must manage:
- Hardware
- Power and cooling
- Model deployment
- Security
- Updates
- Monitoring
- Backup
- Capacity planning
- Technical support
The correct comparison is not cloud subscription versus hardware price.
It is total cost of ownership for both options.
The Bigger Lesson
Microsoft and Amazon are not being rewarded simply because they spent billions on AI.
They are being rewarded because cloud demand and revenue are growing alongside the spending.
Microsoft’s results showed strong cloud and AI performance, while Amazon’s AWS business delivered its fastest growth in more than four years.
That is the real lesson for businesses.
Spending is not the strategy.
Buying licences is not the outcome.
Running more prompts is not the return.
The investment becomes meaningful when it improves something the company cares about.
Revenue.
Customer service.
Productivity.
Risk reduction.
Decision-making.
Delivery time.
Quality.
Closing Thoughts
Microsoft and Amazon have shown that large AI and cloud investments can produce strong results when real customer demand follows.
Microsoft reported $90 billion in quarterly revenue, while Amazon Web Services grew 37% to $42.2 billion. Amazon is now preparing to spend approximately $220 billion during 2026 as it expands infrastructure to meet continued demand.
Most businesses operate at a completely different scale.
But the financial discipline should be the same.
Start with a real problem.
Define the expected result.
Run a controlled trial.
Measure total cost.
Track the cost per useful outcome.
Include security, integration and human review.
Decide whether to scale, improve, limit or stop.
AI does not create value merely because it is advanced.
It creates value when it helps the business produce a better result at an acceptable cost.
At Net Onboard, we help businesses design and manage practical cloud environments through cloud hosting, dedicated servers, cybersecurity, backup and business continuity services.
When an AI or digital project requires additional infrastructure, the environment should be sized according to the actual workload, security requirement, expected growth and business value—not industry excitement alone.