AI Infrastructure Costs: What Singapore Firms Miss

A lot of AI budgets look tidy on paper—until electricity and infrastructure show up.

This week, Reuters reported (via Channel NewsAsia) that Anthropic is promising to pay for grid upgrade costs tied to connecting its data centres, so those expenses don’t get pushed into consumer power bills. It’s a very US-centric story, but the signal is global: AI isn’t only a software decision anymore. It’s a power-and-capacity decision.

For Singapore businesses adopting AI business tools—whether that’s customer support automation, marketing content workflows, sales enablement, or analytics—the same dynamics apply in a quieter way. When the infrastructure layer gets expensive (energy, compute, GPUs, cloud capacity), the “cost per AI outcome” rises. That changes what you build, what you buy, and how fast you scale.

One-liner to remember: If your AI plan doesn’t include compute and energy assumptions, it isn’t a plan—it’s a wishlist.

Why AI infrastructure costs suddenly matter to Singapore companies

Answer first: AI infrastructure costs matter because they flow into the price and reliability of the AI services you actually use—cloud GPUs, model APIs, data platforms, and even your internal IT capacity.

Even if you never run your own servers, you’re still “buying” infrastructure indirectly:

• When your vendor raises prices for an AI feature because inference costs increased.
• When your cloud bill spikes due to heavier usage (more prompts, more files, more users).
• When performance throttles because shared GPU capacity is constrained.

The Anthropic announcement is basically a public admission of something most teams feel but don’t quantify: AI demand is pushing power grids and data centre capacity to their limits, and someone has to pay.

What the Anthropic move tells the market

Answer first: It signals that AI providers expect infrastructure-related costs to become politically and commercially sensitive—and they’re trying to control the narrative before regulators and communities do.

From the article:

• Anthropic says it will cover grid upgrade costs needed to connect its data centres by increasing its own monthly electricity charges, instead of passing them to consumers.
• It also says it will bring new power generation and add grid capacity to meet its data centre demand (not just buy offsets or claim existing capacity).
• Where new generation isn’t online yet, it will work with utilities and experts to estimate and offset demand-driven price effects.

Microsoft has taken a similar stance recently, committing to pay utility rates that reflect its usage and to work with utilities to expand supply.

For Singapore readers, the point isn’t who’s “nicer.” The point is: the infrastructure bill is real, and it’s growing.

The hidden expenses of AI adoption (beyond software)

Answer first: The biggest hidden costs are usage-based compute, data readiness, governance, and the operational overhead of making AI reliable.

In the “AI Business Tools Singapore” series, I keep seeing the same budgeting mistake: teams compare tools by licence fee and ignore the total cost curve.

Here are the cost buckets that tend to surprise companies after a pilot succeeds.

1) Compute cost: prompts, tokens, and “AI sprawl”

Answer first: If adoption goes well, usage explodes—so costs do too.

The first month of a chatbot might cost a few hundred dollars. Then marketing adds it to the website, support integrates it into ticketing, sales uses it for proposal drafts, HR uses it for onboarding Q&A, and suddenly you’re paying for:

• Higher volume of requests
• Larger context windows (more documents attached)
• More expensive models for better accuracy
• More “retry” traffic due to flaky outputs

Practical control: Put budgets on workflows, not just teams. A single workflow (e.g., “summarise calls + generate follow-ups”) should have an expected volume and unit cost.

2) Data costs: cleaning, indexing, permissions

Answer first: AI doesn’t run on “data.” It runs on usable data with the right access rules.

Common work needed for business AI tools:

• Standardising product names, customer records, or knowledge base content
• Creating retrieval indexes for internal documents
• Setting role-based access controls so AI doesn’t leak sensitive info
• Maintaining source-of-truth content to reduce hallucinations

If your AI tool is connected to SharePoint/Google Drive/CRM, you’ll also pay in time: permission audits, duplication cleanup, and taxonomy fixes.

3) Reliability and safety: monitoring and human fallback

Answer first: Real deployments need monitoring, evaluation, and escalation paths.

A pilot can be “good enough.” Production cannot. You’ll need:

• Response quality checks (sampling + scoring)
• Guardrails (blocked topics, PII handling)
• Human escalation (handover to staff when confidence is low)
• Incident routines (what happens when the model is wrong?)

This is operational labour, not a one-time setup.

4) Energy and sustainability pressure (yes, even for SMBs)

Answer first: Even if you don’t pay electricity directly, you’ll face it through vendor pricing and sustainability reporting expectations.

Large clients increasingly ask suppliers about sustainability practices. If your AI usage is heavy, procurement teams may ask:

• Where is processing done?
• Is energy matched with new generation or just offsets?
• What’s the policy on model selection and usage efficiency?

Anthropic’s story lands because it’s proactive: it treats energy impact as a cost-and-trust problem.

How infrastructure costs should change your AI implementation plan

Answer first: Build for efficiency first—choose the smallest model that meets the need, reduce unnecessary calls, and design workflows that produce measurable business outcomes.

If you’re rolling out AI business tools in Singapore, here’s a practical stance: optimise the workflow before you optimise the model.

Start with “ROI per workflow,” not “AI everywhere”

I’ve found that companies get better results when they pick 2–3 high-volume workflows and make them boringly efficient.

Good candidates:

• Customer service: draft replies, summarise tickets, suggest knowledge base articles
• Sales: meeting summaries, follow-up emails, account research, proposal outlines
• Marketing ops: repurpose long-form content into ads, landing pages, and email sequences
• Finance/admin: document extraction, invoice classification, policy Q&A

Then track:

• Time saved per task (minutes)
• Volume per month
• Error rate / escalation rate
• Unit cost per completed output

That last one—unit cost—is where infrastructure and compute show up.

Design patterns that cut AI costs without hurting quality

Answer first: You reduce cost by reducing calls, shrinking context, caching outputs, and routing tasks to cheaper models.

Concrete tactics:

1. Model routing: Use cheaper/faster models for routine drafts; reserve premium models for high-stakes or complex cases.
2. Prompt minimisation: Strip templates down; avoid pasting entire documents if a targeted excerpt will do.
3. RAG hygiene: Retrieval-augmented generation works best when your documents are well-structured and deduplicated.
4. Caching: If 100 users ask the same policy question, you shouldn’t pay 100 times.
5. Batching: For back-office tasks (tagging, extraction), batch jobs are often cheaper than real-time calls.

These are not “engineering-only” ideas. They’re business levers.

Decide early: buy, build, or blend

Answer first: Most Singapore companies should blend—buy AI tools for common workflows, and build light custom layers only where differentiation matters.

• Buy when the workflow is standard (meeting notes, email drafting, basic chat).
• Build when your advantage is in proprietary data/process (pricing logic, compliance workflow, specialised knowledge).
• Blend when you need tight integrations but don’t want to own model operations.

Infrastructure costs push you toward buying—until you hit usage at scale. Then hybrid designs often win because you can control unit economics.

What Singapore leaders should ask vendors (and internal teams)

Answer first: Ask questions that reveal unit economics, scalability, and energy/compute assumptions—not just features.

Use this list in your next AI tool demo or renewal discussion:

1. What’s the cost per 1,000 tasks for our expected usage (with realistic attachment sizes)?
2. Which model(s) are used by default, and can we route by task type?
3. How do you handle caching for repeated queries?
4. What happens when confidence is low—is there an escalation path?
5. How do you measure hallucinations and retrieval accuracy over time?
6. Where does processing occur (region), and what’s your availability/SLA?
7. Can we export logs and evaluation data if we switch vendors?

If a vendor can’t answer #1 clearly, you’re signing up for surprise invoices.

Bringing it back to the Anthropic story

Answer first: Anthropic paying grid upgrade costs is a public version of the same principle every business needs: who bears the infrastructure burden of AI growth?

When AI demand grows, costs land somewhere—on consumers, on vendors, or on the businesses adopting AI tools. Anthropic is choosing to shoulder some of that burden (and likely bake it into their own economics). Microsoft is signalling similar intent. These moves are about trust, community acceptance, and long-term capacity.

For Singapore companies, the most useful takeaway is straightforward: treat AI like a utility-backed operating expense, not a one-off software purchase. Budget for variable usage, design for efficiency, and pick workflows where the savings are measurable.

If you’re mapping your 2026 AI roadmap, here’s the forward-looking question worth ending on: When your AI usage doubles (because it will), will your unit costs stay predictable—or will they spike at the exact moment the business starts relying on the system?

Source referenced: https://www.channelnewsasia.com/business/anthropic-shoulder-some-costs-data-center-expansions-threaten-raise-power-bills-5924596