AI Demand Forecasting Lessons from Onsemi’s Miss

On 10 Feb 2026, Reuters reported that chipmaker Onsemi missed quarterly revenue estimates—posting US$1.53B vs US$1.54B expected—and guided a first-quarter midpoint below consensus. Shares dropped about 6% after-hours. The headline sounds like a Wall Street story, but it’s really an operations story: when demand signals get noisy and inventory piles up, revenue misses are often the last domino to fall.

For Singapore businesses—especially manufacturers, distributors, and retailers—this matters because the same pattern shows up in smaller ways every day: over-ordering when lead times feel scary, under-shipping when demand unexpectedly spikes, and forecasting based on “last year plus a bit.” The reality? Volatility is the default now, and your spreadsheet forecast won’t keep up.

This post in our AI Business Tools Singapore series uses Onsemi’s situation as a case study. Not to criticise one company, but to pull out practical lessons: how to spot an inventory glut early, how AI forecasting changes the game, and what a realistic rollout looks like for Singapore teams that need results this quarter—not a two-year transformation programme.

What Onsemi’s revenue miss really tells operators

A revenue miss is usually described as a sales problem. Operationally, it’s often a demand-supply alignment problem.

In Onsemi’s case, the report pointed to a persistent inventory glut as customers worked through excess chip stockpiles ordered during earlier supply chain disruption. The article also notes pressure in silicon carbide (SiC), tied to slower-than-expected EV growth and increased competition from Chinese companies.

Here’s the operator’s translation:

• Ordering behaviour shifted (customers are consuming inventory rather than buying new supply).
• End-market growth slowed (EV demand didn’t absorb capacity as forecast).
• Competitive supply increased (price and share pressure, particularly in SiC).

Those three forces hit revenue even if your product is good and your sales team is doing fine.

Inventory glut is a forecasting failure—just delayed

Inventory gluts don’t happen overnight. They build up quietly:

1. Lead times expand → buyers panic and over-order.
2. Supply normalises → deliveries arrive.
3. Demand cools or shifts → stock sits.
4. Buyers pause purchasing → suppliers see a sudden revenue drop.

A lot of companies only react at step 4 because that’s when finance feels it. The better time to act is step 2–3, when you can still throttle production, rebalance procurement, and adjust pricing or channel strategy.

Snippet-worthy truth: Inventory is demand forecasting error stored in a warehouse.

Why traditional forecasting breaks in 2026 (and what AI does differently)

Traditional forecasting usually relies on one of two approaches:

• Time-series extrapolation (trend + seasonality)
• Sales input plus “gut feel” overrides

These can work in stable markets. But Onsemi’s story is a reminder that many markets aren’t stable:

• Customer ordering is distorted by past shortages.
• Macro policy shifts (for example, changes to clean energy incentives) ripple into demand.
• Competitors can flood supply, especially when capacity expands regionally.

AI forecasting is not magic—it’s better signal processing

AI demand forecasting systems don’t “predict the future.” They combine more signals, update more frequently, and quantify uncertainty.

A practical AI forecasting stack for most Singapore businesses does three things well:

1. Multi-signal forecasting: Uses shipments, quotes, POS (if available), web demand, promotions, lead times, macro indicators, and even competitor price signals.
2. Probabilistic forecasts: Gives ranges (P10/P50/P90) instead of one fake-precise number.
3. Exception-based planning: Flags where humans should intervene (big deltas, unusual patterns) instead of asking planners to touch every SKU.

When demand becomes volatile, the win isn’t “perfect forecasts.” The win is faster detection of turning points—like customers moving into destocking mode.

The KPI that matters: time-to-detect

Most teams track forecast accuracy. I care just as much about:

• Time-to-detect a demand regime change (weeks, not quarters)
• Time-to-replan (days, not monthly S&OP only)

If you detect destocking early, you can avoid the classic spiral: build too much → discount to clear → margin falls → CFO freezes spend → you can’t invest in better planning.

A Singapore playbook: using AI to prevent “surprise” revenue drops

Singapore companies are exposed to global swings—electronics, precision engineering, pharma supply chains, trading, and regional distribution. Even service businesses feel it through staffing, utilisation, and procurement.

Here’s a grounded playbook I’ve seen work with AI business tools in Singapore, without turning it into an IT science project.

1) Start with “where inventory hides mistakes”

Answer first: Pick one product family where stockouts are painful and overstocks are expensive. That’s your pilot.

Good pilot candidates:

• High-value components with long lead times
• SKUs with volatile demand (promotions, project-based orders)
• Items impacted by external cycles (construction, automotive, consumer electronics)

Define three numbers upfront:

• Current days of inventory on hand (DOH)
• Write-downs/obsolescence risk
• Service level or fill rate

AI is only useful if it changes one of those.

2) Move from one-number forecasts to ranges

Answer first: Forecast ranges help procurement and production make safer decisions.

Instead of “Forecast = 10,000 units,” you want:

• P50 (most likely) = 10,000
• P90 (high demand) = 12,500
• P10 (low demand) = 7,500

Then set policies:

• Build to P50 for standard capacity
• Secure materials to P70/P80 if lead times are long
• Trigger review if actual demand tracks below P25 for 2–3 weeks (destocking signal)

This is how you keep flexibility without hoarding inventory.

3) Add an “inventory glut early warning” dashboard

Answer first: If you can see glut signals in week 2, you won’t be shocked in quarter-end earnings.

Minimum viable signals:

• Customer order cadence: order frequency and average order size
• Backlog movement: cancellations, push-outs, requested delivery shifts
• Channel inventory proxy: distributor sell-through vs sell-in
• Lead-time compression: suppliers suddenly offering faster delivery
• Price pressure: more discount requests, competitor undercutting

You don’t need perfect data. You need a consistent, weekly view.

4) Automate the boring part: exceptions and scenarios

Answer first: Planners should manage exceptions, not copy-paste sheets.

Use AI analytics to generate:

• Exception list: SKUs with forecast error above a threshold
• Scenario plans: “If demand drops 15%, what happens to cash tied in inventory?”
• Recommended actions: pull forward promotions, pause PO releases, redeploy stock across channels

This turns planning into decision-making instead of administration.

“Can AI help avoid revenue misses like Onsemi’s?” (A practical answer)

Yes—if you define the job correctly.

AI won’t prevent macro headwinds (like slower EV sales). But AI can reduce the operational lag between market reality and company actions.

Here are concrete outcomes AI forecasting and supply chain analytics typically improve:

• Lower inventory without hurting service levels (because you stop buffering blindly)
• Faster response to destocking (because you monitor leading indicators)
• Cleaner revenue planning (because finance sees ranges and scenarios)
• Better capital allocation (because you can quantify where demand is structurally weakening)

The key is to treat AI as part of a loop:

1. Sense (signals)
2. Predict (ranges)
3. Decide (policies)
4. Act (procurement/production/pricing)
5. Learn (feedback)

Most companies get stuck at step 2—pretty forecasts that don’t change behaviour.

People Also Ask: What data do I need to start AI demand forecasting in Singapore?

You can start with surprisingly little:

• 12–24 months of sales orders (weekly granularity is fine)
• Current inventory and open POs
• Basic product hierarchy (category/family)
• Customer segments or channels

Then iterate. Add POS data, marketing calendars, lead times, and pricing later.

People Also Ask: Does AI forecasting replace S&OP?

No. It upgrades S&OP.

S&OP is still where trade-offs get decided. AI makes S&OP less emotional because:

• assumptions are explicit
• uncertainty is quantified
• scenarios are quick to generate

What to do next (especially if you’re planning for 2026 volatility)

If the Onsemi story resonates, don’t file it under “semiconductor news.” Use it as a prompt to stress-test your own planning.

Run this 30-minute internal check:

1. Where do we see customer destocking first? (Which segment/channel?)
2. What’s our time-to-detect? (Weekly signals or month-end surprises?)
3. Do we forecast a single number or a range?
4. What decision changes if we’re wrong by 15%? (Inventory, staffing, cash)

If you want a practical starting point, begin with one product line and build an AI-driven early warning view. Once leaders see that “turning point detection” is possible, scaling becomes much easier.

The bigger question for Singapore operators in 2026: When the next demand swing hits, will your team see it in week 2—or explain it in week 12?

Source article (case study): https://www.channelnewsasia.com/business/chipmaker-onsemi-misses-quarterly-revenue-estimates-shares-fall-5918906