AI cloud capex is exploding—and it affects AI tool costs and availability. Here’s how Singapore businesses can build resilient, ROI-driven AI workflows.
AI Cloud Spending Is Surging—What SG Firms Should Do
Nebius, an AI cloud provider serving customers like Microsoft and Meta, just reported US$2.1 billion in capital expenditure in a single quarter, up from US$416 million a year earlier. That’s not “normal growth.” That’s a land grab for the most scarce resource in AI right now: GPU capacity and the power-hungry data centres that run it.
For Singapore businesses following the “AI Business Tools Singapore” series, this matters for a simple reason: the tools you’re trying to use—AI for marketing, operations, and customer engagement—sit on top of this infrastructure bill. When cloud providers pour billions into GPUs, it changes pricing, availability, performance, and even the way you should design AI projects.
Here’s the practical read: this isn’t just a story about one company’s spending spree. It’s a signal that AI infrastructure is becoming a strategic constraint, and Singapore SMEs and mid-market teams need to plan around it—just like you plan around hiring, compliance, and cash flow.
AI infrastructure is the new bottleneck (not ideas)
Answer first: Most companies won’t be limited by AI “use cases” in 2026—they’ll be limited by compute access, cost, and deployment readiness.
Nebius says demand from enterprise and AI-native customers is still outpacing supply, and it’s expanding with nine new data centre sites across the US, France, Israel, and the UK. It also disclosed it has over 2 gigawatts (GW) of contracted power already and expects over 3 GW by year-end—a detail many business leaders ignore, but shouldn’t.
Why power matters: GPUs don’t just cost money to buy; they cost money to run. If power and data centre capacity become the choke point, everyone downstream feels it:
- AI features roll out slower because capacity is booked out.
- Pricing becomes more variable (especially for peak demand).
- Vendors start pushing longer commitments (“reserve now for later”).
For Singapore businesses, the implication is uncomfortable but useful: don’t assume AI compute will always be cheap and instantly available. Build plans that work even when compute is constrained.
What this changes for SG teams buying “AI tools”
A lot of “AI business tools” are essentially wrappers around cloud inference. When infrastructure tightens, you’ll see:
- Usage-based pricing surprises (especially with heavy image/video and long-context text).
- Feature gating (the “good models” require premium tiers).
- Latency variability that breaks real-time customer journeys.
If your customer engagement flows rely on instant responses (chat, WhatsApp automation, live agent assist), you need to treat performance like a requirement, not a nice-to-have.
The hidden cost (and value) of AI is GPUs + data centres
Answer first: AI doesn’t get expensive because the model is “smart.” It gets expensive because the model is computationally hungry, and someone pays for GPUs, cooling, racks, networking, and power.
Nebius reported Q4 revenue of US$227.7 million (over six times growth year-on-year), yet still posted a net loss of US$249.6 million. This is a classic infrastructure phase: spend aggressively now, monetize later. Whether Nebius hits its ambitious target of a US$7–9 billion annualized revenue run-rate by end-2026 is a market question. But the business pattern is already clear: AI cloud economics reward scale.
From a buyer’s perspective in Singapore, that creates a two-speed world:
- Big buyers negotiate capacity, pricing, and priority.
- Everyone else buys “on-demand” and absorbs volatility.
A Singapore-flavoured example: Marketing AI that scales (or breaks)
Say you run performance marketing for an e-commerce brand in Singapore and you want:
- product copy generation,
- image variant creation,
- ad creative testing at scale,
- and a multilingual customer support bot.
If you prototype using a single SaaS tool, it looks affordable. But when you operationalize it—hundreds of creatives, thousands of support sessions, daily reporting—your cost curve becomes compute-driven.
What works in practice:
- Use smaller models for high-volume tasks (classification, routing, templated copy).
- Reserve premium models for high-value moments (complaints, refunds, VIP customers).
- Batch offline jobs (image generation, product enrichment) outside peak hours.
This is the operational discipline that infrastructure constraints force on you.
What “neocloud” growth means for Singapore SMEs
Answer first: The rise of specialist AI cloud providers signals that AI workloads are different enough that general-purpose cloud isn’t always the cheapest or fastest path.
Nebius is described as a “neocloud” provider—firms that primarily sell GPU-backed compute and AI infrastructure as a service. Along with players like CoreWeave, this category has benefited from relentless enterprise AI spending.
For Singapore SMEs, the practical question isn’t “Should we use Nebius?” It’s: Should we architect our AI stack so we can switch compute providers if pricing or capacity changes?
That means:
- Avoid locking your workflows into a single proprietary setup.
- Standardize prompts, evaluation tests, and deployment interfaces.
- Keep your data layer clean so you can move where compute is available.
Vendor strategy: don’t marry your first AI tool
I’ve found that companies regret early AI decisions for one reason: they treated the first working demo as the final architecture.
A healthier approach is a simple 3-layer view:
- Business layer: your use cases (lead qualification, churn reduction, faster reporting)
- Model layer: which models you call and why
- Infrastructure layer: where compute runs (cloud, neocloud, on-prem, hybrid)
If your AI initiative only works when one vendor’s capacity is abundant, it’s fragile.
A practical playbook: build AI projects that survive compute volatility
Answer first: The winners in 2026 won’t be the firms with the most AI experiments—they’ll be the firms with AI systems designed for cost control, reliability, and measurable impact.
Here’s a field-tested checklist you can use for AI business tools in Singapore—whether you’re in retail, finance, logistics, professional services, or healthcare admin.
1) Start with a compute budget, not a model wishlist
Set guardrails before building:
- Monthly AI spend cap (separate for experimentation vs production)
- Max cost per lead, per ticket resolved, per report generated
- Target latency for customer-facing workflows
When finance asks “why did AI costs spike,” you’ll have an answer tied to business throughput.
2) Design “tiered intelligence” into every workflow
Not every task needs a premium model.
A tiered pattern looks like this:
- Tier A (cheap, fast): classify intent, extract entities, route tickets
- Tier B (balanced): draft responses, summarize calls, generate first-pass content
- Tier C (premium): handle edge cases, negotiation, regulated language, VIP support
This is the easiest way to keep customer engagement AI reliable while controlling cost.
3) Measure quality with an evaluation set (even a small one)
Most AI projects fail quietly: output looks fine, but business metrics don’t move.
Create a small evaluation set:
- 50–200 real examples (customer emails, chat transcripts, sales calls)
- “Gold standard” outcomes written by your best staff
- Scoring rubrics (accuracy, tone, compliance, resolution rate)
Then test changes like you’d test ad creatives.
4) Make data centre reality part of your risk plan
Nebius’ contracted power numbers (2 GW now, 3 GW expected) are a reminder that AI scale is physical.
So your risk plan should include:
- Backup provider options for critical workflows
- Caching and fallbacks (serve a safe template if the model is slow)
- Incident processes (who gets paged when AI responses degrade?)
If AI is customer-facing, treat it like uptime.
People also ask (and what I tell clients)
“Should Singapore businesses wait until AI gets cheaper?”
No. Costs will shift, not disappear. The smarter move is to build workflows that can tolerate pricing changes: tiered models, batching, and clear ROI metrics.
“Do we need our own GPUs in Singapore?”
Usually not at SME scale. But you do need control over how you consume compute—especially for high-volume marketing content and customer service automation.
“Is AI cloud only for tech companies?”
Not anymore. The fastest adopters I see are often non-tech teams: ops managers using AI for forecasting, marketing teams running content engines, and customer support teams using agent assist.
Where this leaves Singapore businesses in 2026
Nebius’ surge in GPU and data centre spending is a sharp signal: AI infrastructure is scaling fast because demand is real—and because companies are competing for capacity. If cloud providers are spending billions to secure GPUs and power, Singapore businesses should assume AI will remain a strategic investment area, not a one-off software purchase.
The next step is practical: treat AI business tools like a capability you operate, not a subscription you forget. Put budgets around compute, build tiered workflows, and measure outcomes the way you’d measure revenue campaigns or service levels.
If you’re rolling out AI for marketing, operations, or customer engagement this quarter, what’s the one workflow you’d redesign today so it still works when compute gets more expensive—or harder to get?