AI Costs Are Rising: Lessons for Singapore Businesses

Amazon just gave the market a blunt reminder: AI isn’t “software spend.” It’s infrastructure spend. When Amazon told investors it expects to invest US$200 billion in 2026 capital expenditure, its shares dropped more than 11%—even though quarterly results were strong (profit US$21.2B on net sales US$213.4B). That mismatch—great revenue, shaky investor confidence—happens when people realise the AI bill is real, recurring, and heavy.

If a company with Amazon Web Services (AWS) scale can spook the market with AI and data centre costs, smaller teams in Singapore should take the hint: the winner isn’t the company that “does the most AI.” It’s the one that manages AI cost per outcome. In this instalment of the AI Business Tools Singapore series, I’ll break down what Amazon’s news means in practical terms, and how Singapore businesses can adopt AI tools without stepping into an open-ended cost spiral.

Snippet-worthy rule: AI spend is only “worth it” when you can explain it as cost per business outcome (lead, ticket resolved, hour saved), not as a vague innovation budget.

What Amazon’s plunge really signals about AI economics

The immediate story from the RSS article is simple: Amazon’s AI and infrastructure spending forecast surprised investors. Analysts were expecting around US$147B, but Amazon guided US$200B—a huge gap.

Here’s the more useful takeaway: markets are treating AI like a margin story, not a growth story. Investors aren’t asking “Will AI create revenue?” They’re asking:

• How long will capex stay elevated?
• Will AI workloads increase profit, or just energy and compute bills?
• Can supply (data centres, chips, power) scale without crushing margins?

Amazon also reported AWS revenue of US$35.6B, up 24% year-on-year. Demand is there. The problem is that AI demand tends to be bursty, compute-hungry, and expensive to serve, especially when you’re racing Microsoft and Google.

A helpful lens: capex vs. “AI tool budget”

Most SMEs and mid-market firms in Singapore don’t spend capex the way hyperscalers do, but the pattern rhymes:

• Hyperscalers spend billions on data centres and chips.
• Businesses spend on AI subscriptions, integrations, data cleanup, governance, training, and ongoing usage.

Different line items. Same trap: you start with a pilot cost, and end with a permanent run-rate.

The hidden cost drivers that hit Singapore teams first

AI costs don’t usually explode because the model is “too expensive.” They explode because the organisation didn’t design for cost control.

1) Usage-based pricing that scales faster than revenue

A lot of AI business tools (and the AI features inside CRM/helpdesk/marketing suites) price by:

• number of seats
• number of automations
• number of messages/tokens
• number of workflow runs
• volume of documents indexed

If your AI feature works well, usage goes up. That sounds great—until your monthly bill rises faster than the KPI it’s meant to improve.

Practical move: set an internal metric like:

• Cost per qualified lead (marketing)
• Cost per ticket resolved (support)
• Cost per invoice processed (finance)

If you can’t calculate it, you can’t manage it.

2) Data readiness and integration work

Many teams underestimate the non-negotiables:

• cleaning customer and product data
• setting up permissions and access control
• connecting tools (CRM, email, website forms, support desk)
• building a feedback loop so AI improves instead of drifting

This is where Singapore SMEs often burn budget: they buy a shiny tool, then spend 3 months trying to make it work with their actual workflows.

Practical move: budget for implementation explicitly. A realistic starting split I’ve seen work:

• 40% tool costs (licenses/usage)
• 40% implementation (integration + process design)
• 20% enablement (training + governance)

Not perfect for everyone, but far closer to reality than “just buy the AI add-on.”

3) Reliability, risk, and rework

If AI output causes:

• inaccurate customer replies
• wrong pricing/quotes
• non-compliant marketing claims
• poor handoff between teams

…the cost isn’t the model. It’s the rework, the brand damage, and the operational drag.

Practical move: for customer-facing AI, enforce a simple control:

• “AI drafts, humans approve” for high-risk actions
• “AI acts autonomously” only for low-risk, reversible tasks

A smarter approach: choose AI tools by “workload fit,” not hype

Amazon’s story highlights a universal truth: AI is a portfolio of workloads. For Singapore businesses, a cost-controlled AI strategy usually looks like this:

1) Start with high-frequency, low-variance tasks

These are tasks where the inputs and outputs are predictable, and your team repeats them daily.

Examples (common in Singapore SMEs):

• summarising sales calls and extracting action items
• drafting first replies for common support questions
• categorising inbound leads by intent
• generating first-draft marketing variations for A/B tests
• extracting fields from invoices or purchase orders

These deliver ROI quickly because you can measure time saved or throughput increases.

2) Avoid “big-bang” customer experience automation

Many teams try to jump straight to an AI chatbot that “handles everything.” That’s the fastest path to:

• long implementation
• high knowledge-base maintenance
• unpredictable costs as queries scale
• escalation nightmares

A better sequence:

1. Agent assist (AI helps your humans)
2. Partial automation (AI handles a narrow set of intents)
3. Broader automation only after accuracy and containment metrics are stable

3) Pick the cheapest model that meets the quality bar

This is where people get emotional. Don’t.

Use a straightforward rule:

• If a smaller/cheaper model hits your accuracy target, use it.
• Use premium models only for tasks that genuinely need them (complex reasoning, multilingual nuance, long context).

AI cost control is mostly model selection + guardrails + caching, not heroics.

Cost-management playbook you can apply this quarter

If you’re implementing AI business tools in Singapore in 2026, these steps are practical and fast.

1) Define one KPI and one “kill metric” per use case

A KPI measures success. A kill metric prevents runaway spend.

Example: AI support drafting

• KPI: Reduce average handle time by 20%
• Kill metric: Cost per resolved ticket must stay under S$X

If you can’t define both, you’re not ready to scale.

2) Put a hard cap on usage while you learn

Most tools let you control usage via:

• rate limits
• seat limits
• workflow run limits
• environment separation (sandbox vs production)

Start with a cap that forces prioritisation. Scarcity is a feature early on.

3) Measure the “human minutes saved,” not just output volume

AI often increases content volume. That’s not automatically value.

Track:

• minutes saved per employee per week
• reduction in rework (edits, corrections, escalations)
• faster cycle times (quote turnaround, onboarding time)

When you can translate improvements into dollars, stakeholder buy-in becomes easy.

4) Design your data permissions before you automate

Singapore firms frequently have regulated or sensitive contexts (finance, healthcare-adjacent services, HR data, government-linked environments). You need clear boundaries:

• what data the AI can access
• where logs are stored
• how long data is retained
• who can approve changes

Even if you’re not in a regulated sector, your customers still expect discipline.

5) Prefer tools that show cost and quality metrics in the product

This is my bias: if the tool can’t show you usage, cost drivers, and accuracy signals, you’re flying blind.

Look for features like:

• per-workflow usage reporting
• audit trails
• evaluation/testing mode
• feedback capture (thumbs up/down + reason)

What Amazon’s AWS growth tells us about where AI budgets are heading

Amazon’s AWS grew 24% year-on-year, and management said customers want AWS for “core and AI workloads” and they’re “monetising capacity as fast as we can install it.” Translation: AI demand is persistent, and supply is the bottleneck.

For Singapore businesses, that means:

• AI features won’t get “free” just because they’re common.
• Vendors will continue bundling AI into tiers and charging for usage.
• If you don’t set governance early, your AI bill will creep up quietly.

Also notable: Amazon announced significant job cuts tied to restructuring and AI focus. Whether you agree with the move or not, it reinforces that AI adoption is not only a tool decision—it’s an operating model decision.

A practical stance for Singapore leaders: be ambitious, but be strict

AI adoption is worth doing. But copy-pasting a big-tech mindset (“spend now, figure it out later”) doesn’t translate to SMEs.

Here’s what works:

• Be ambitious about outcomes (faster sales cycles, better service, leaner ops)
• Be strict about unit economics (cost per lead, cost per ticket, hours saved)
• Be realistic about implementation (data + process + people)

If Amazon can post huge numbers and still get punished for AI spend, it’s a reminder that profitability discipline matters at every scale.

Where are you seeing AI costs creep in your business today—tool subscriptions, usage charges, or the “hidden” time spent fixing outputs?

Source article: https://www.channelnewsasia.com/business/amazon-bezos-ai-shares-tech-google-5913596