Food waste in Singapore and where AI can help.

The short version.

Singapore produced about 790,000 tonnes of food waste in 2025, roughly 11% of all its waste, and recycled 18% of it. The rest was incinerated.³¹ The tonnage has stayed in the same band for a decade, and the recycling rate has been 18% or 19% every year since 2019.² Policy is moving: since March 2024 new large malls, hotels and food factories must segregate, treat and report their food waste, and existing premises will follow when the food waste line at Tuas Nexus comes online, now expected progressively from 2027.⁴¹³

This case looks at three questions a programme with universities would ask. Can AI help predict and prevent food waste? Can collection and recovery be improved? Can the work create measurable value on a model that pays for itself? The evidence is strongest for the first, weakest for the third, and the second is held back by one problem that keeps recurring in Singapore: contaminated food waste streams.⁴⁰³⁷³⁵

Universities here are active, but mostly on turning food waste into products: okara, spent grain, coffee grounds, peels and insect protein.⁵⁸²³ Published work is thin on forecasting for hawkers and caterers, on measuring household and hawker waste, on the logistics of collecting segregated food waste and on a cost and carbon model for Singapore. Those gaps are where a new programme would add most. The new NEA and NTU research centre, TREASURES, funded to 2030, is the most direct way in.⁴⁸

Status as of 12 October 2026: the latest official figures are for 2025, released in June 2026. The Zero Waste Masterplan review is due in 2027, and no date has been set for existing premises to segregate food waste. Every time-sensitive claim on this page was checked on that date.

Food waste in Singapore has not moved in a decade.

NEA's series shows food waste between about 744,000 and 817,000 tonnes a year since 2015, apart from a COVID-19 dip to 665,000 tonnes in 2020.² The recycled share rose from 13% in 2015 to 18% in 2019 and has stayed there. Most of what is recycled is clean, single-type waste from manufacturers, such as spent grain, okara and bread, which goes mainly to animal feed.⁶

Two features of the numbers matter for anyone designing a programme. First, the government puts commercial and industrial sources at about 40% of food waste and has not published a split for the rest, which includes households, hawker centres and smaller food outlets.⁵ NEA's 2017 household study found food was about half of what households threw away, and more than half of that was avoidable.²⁶ Second, food that is rescued or redistributed is not counted at all. The Minister said so in Parliament in July 2026.³ Prevention, the outcome that matters most, therefore does not show up in the headline statistic.

The wider targets give the urgency. The Zero Waste Masterplan of 2019 aims to cut waste sent to landfill per person per day by 30% by 2030, against 2018.⁸ In February 2025 the figure was "about the same" as in 2018.⁹ Semakau, the only landfill, is more than half full and is expected to run out of space by 2035.¹⁰ A review of the masterplan was announced in June 2026, due in 2027.⁷

The rules now reach large new premises.

Under the Resource Sustainability Act, new buildings whose plans were submitted from 2021 have had to segregate, treat and report food waste since 8 March 2024. The thresholds are malls and hotels with more than 3,000 sqm of food and function space, food factories with more than 750 sqm of operations, and large multi-user food factories.¹¹⁴ NEA estimated in 2020 that about 360 premises would be in scope once existing buildings were included.¹²

Existing buildings wait on Tuas Nexus, where segregated food waste will be pulped and digested with used-water sludge, a method NEA trialled at Ulu Pandan from 2016 with about 40% more biogas than digesting the two apart.¹⁵ The food waste line is designed for 400 tonnes a day.¹³ Its date has moved several times.

Households are outside the rules. The ministry said in October 2025 that it has no plan to require households to segregate food waste, because trial participation could not be sustained.¹⁶ In July 2026 it said that Korean-style charging by weight, using RFID bins, is hard in high-rise blocks that share a refuse chute.¹⁷ Hawker centres sit between the two: tray return has been mandatory since 2021, which gives a natural point to separate food waste, but no published results from hawker-centre pilots were found.⁶⁸⁵⁵

Where AI fits, problem by problem.

Predicting and preventing food waste.

This is where the evidence is best. The most useful independent result comes from the Pacific Coast Food Waste Commitment, which ran AI demand-forecasting pilots in more than 1,300 stores of two large US grocers from 2020 to 2022. Food waste fell by 14.8% per store on average, with results visible after about eight weeks. The main barrier it reported was organisational: merchandising, finance and operations teams had conflicting goals.⁴⁰

Kitchen tracking, in which a camera over the bin identifies and weighs what is thrown away, has the strongest company results. IKEA's stores cut production food waste by 54% against 2017 over about four years, using Winnow and training 20,000 staff; the figure excludes customers' plate waste.³⁹ Lumitics, the Singapore company now owned by Winnow, reports an average 30% cut in a hotel's first six months.³¹ SATS reports that its food waste intensity fell from 4.1% to 2.3% between FY2021 and FY2023 after it introduced AI tracking in its kitchens.³⁴

Two limits apply in Singapore. Vision systems are tested mostly on hotel buffets and production kitchens. A 2025 university study found they did worse on stews and fragmented food after eating, which describes most of a hawker centre's or school canteen's waste.⁴⁷ And prediction only pays if someone changes a decision because of it. In each of these cases the gain came from kitchen and store staff acting on the numbers, not from the model alone.³⁹⁴⁰

There is a second kind of prevention with Singapore research behind it. NTU's food safety hub, FRESH, reported in April 2026 an AI model that predicts the growth of bacteria such as Salmonella under real storage conditions, so that food which is still safe is not thrown away on a fixed date. NTU says supermarket chains are exploring it.⁵⁰

Collection and recovery.

Singapore's record on recovery points to one constraint above the rest: what goes into the bin. IUT Global, once the country's largest food waste digestion plant, folded in 2011. It was built for 800 tonnes a day, collected 120 to 130, and up to 40% of what it received was contaminated with plastics.³⁷ Changi Airport has 11 digesters across its terminals, yet only about 300 tonnes a year passed through them in 2019. Its own problem statement to innovators gave the reason: cleaners and food outlet staff cannot tell digestible from non-digestible waste.³⁵ For recyclables, the blue-bin contamination rate "remained at about 40%" in February 2026.¹⁹

That makes contamination detection the most valuable collection use case here, ahead of route optimisation. The pilots so far have not published results. FairPrice and A*STAR ran AI and IoT sorting bins for recyclables at three sites from 2022.³⁸ 800 Super's fill-level sensor trial was delayed by procurement and manpower problems.²⁰ The minister cited smart recycling boxes that track habits and reduce contamination at the 2025 launch of Cora Environment, the former SembWaste.²¹ Route optimisation has a Singapore research paper from SMU, but for general municipal waste rather than segregated food waste.⁵³

Matching surplus food with people who can use it works at useful scale. The Food Bank Singapore redistributed about 612 tonnes in 2024.²⁷ Yindii, a surplus-food app, reported more than 300,000 meals rescued in Singapore in 2025.²⁹ FairPrice sold 1,095 tonnes of imperfect produce and redistributed about 75 tonnes through volunteers in 2024.²⁸ The Good Samaritan Food Donation Act, in force since January 2025, protects donors who meet its safety conditions.²⁵ Too Good To Go, often assumed to be here, is not; its first Asian market is Japan.³⁰

Measurable value on a lasting model.

Behavioural measures have the best controlled evidence, and most need no AI. In a field experiment in Nordic hotel buffets, plates 3 cm smaller cut food waste by 19.5%, and a sign inviting guests to come back for more cut it by 20.5%.⁴⁵ A Singapore version exists on a small scale: a 2021 trial of a rice-portion menu card in four NUS canteens, run by students with the ministry as mentor.⁵⁴

The business case has been studied across many sites. Champions 12.3 found a return of about 7 to 1 for hotels (42 sites) and for restaurants (114 sites), with first-year cuts of 21% and 26%. The often-quoted 14 to 1 is the cross-sector average from 2017, not the hotel figure.⁴²⁴³⁴⁴ These sites chose to take part, so the figures show what is possible, not what is typical.

Carbon is the weak leg of the value case in Singapore. Residual food waste is incinerated rather than landfilled, so carbon methods that credit avoided landfill methane are unlikely to apply. The carbon tax, S$45 a tonne from 2026, falls on about 50 large emitters, not on a hotel's or a hawker's bin.²⁴ On my reading, the value case rests on food cost, disposal cost, compliance with the segregation rules and, for the state, landfill life. No published model puts those together for Singapore.

Who is doing what in Singapore.

The activity is real and spread across government, industry, universities and volunteers. It is spread thinly: few of these efforts share data, and fewer publish results.

The collection market is changing hands as well. The three public waste collectors are 800 Super, ALBA and Cora Environment, which took over SembWaste's sectors after Sembcorp sold the business in March 2025.¹⁸⁶⁹ Winnow bought Lumitics in July 2026, and Olio cut back its Singapore team in 2023.³¹⁷⁰ A programme that relies on one vendor or app should plan for that vendor changing.

What Singapore universities are working on.

A government list of food waste research at universities and research institutes runs to dozens of projects, almost all on valorisation: turning a waste stream into feed, food ingredients, fertiliser or materials.⁵⁸ NEA's latest valorisation awards, three projects worth S$5.3 million combined, follow the same pattern.²³ The projects that use AI or address the three problem statements more directly are fewer.

Where university work is thin.

1. Forecasting for hawkers, caterers and canteens. I found no university project on demand forecasting for the parts of the food trade outside supermarkets and hotels. The only academic AI on prevention is NTU's spoilage model and an SMU student dashboard.⁵⁰⁵²
2. Measurement of household and hawker waste. There is no repeated household survey or diary panel like those at Wageningen, Leeds or Ohio State. The last NEA household study dates from 2017.²⁶⁶⁴⁶⁷
3. Vision models for local food. Mixed rice, noodles in gravy and shared dishes are harder to classify than buffet trays, and no published Singapore data set of food waste images was found.⁴⁷
4. Logistics for segregated food waste. As existing premises join the scheme, someone has to collect clean food waste and move it to Tuas Nexus or an on-site digester. No optimisation work on that flow was found.⁵³¹³
5. A cost and carbon model for Singapore. A*STAR's audit method and TREASURES' waste-flow node are the only anchors. Neither has published a food waste cost-per-tonne or carbon model.⁵⁹⁴⁹

The crowded area is valorisation chemistry. A new funder adds least by paying for more of it, unless the project solves the feedstock problem that sank IUT Global.

Routes into the universities.

TREASURES is the most important of these. It is run by NEA and NTU with S$35 million from January 2026 to March 2030, and its first node, Waste Flow and Analytics, maps and models waste streams.⁴⁸⁴⁹ Its launch material does not mention food waste. A food waste data workstream would fit the node and would give a programme a home that outlasts a single grant.

What universities elsewhere have done.

Questions to put to the universities.

These turn the three problem statements into questions a university team can answer in a term or a grant cycle. Each fills one of the gaps above.

1. Baseline. How much food waste do households, hawker centres and small food outlets produce, by type, and how much is avoidable? A repeated survey and diary panel, on the Leeds and Ohio State model.⁶⁵⁶⁷
2. Forecasting. Can a simple demand model cut overproduction at a central kitchen, caterer or chain of hawker stalls, measured against a control group, as the US retail pilots were?⁴⁰
3. Local vision data. Can a shared, open set of images of Singapore food waste raise classification accuracy on mixed dishes enough to be useful?⁴⁷
4. Contamination at source. Can a camera or sensor at a digester or bin catch contamination before it reaches the machine, at Changi-like sites and the first existing buildings to join the scheme?³⁵³⁷
5. Collection design. What collection routes and frequencies would move segregated food waste from existing premises to Tuas Nexus at least cost?¹³
6. Nudges at the tray return. Do portion choices and signage cut hawker-centre waste as they did in hotel buffets and NUS canteens?⁴⁵⁵⁴
7. The value model. What is the cost per tonne avoided for each measure, for the operator and for the state, counting food cost, disposal, compliance and landfill life? Published openly, on the ReFED model.⁶⁶

What leaders can take from it.

1. Measure first. The national statistic cannot see prevention or rescued food. A programme that wants to show results needs its own baseline before it starts.
2. Fix the input before the machine. Contamination has defeated more food waste projects in Singapore than technology has. Put detection and staff training ahead of new treatment capacity.
3. Back prediction where someone will act on it. Forecasting and tracking work when a kitchen or store changes what it orders or cooks. Choose partners who will change the decision.
4. Ask universities for the thin areas. Forecasting outside retail, household and hawker measurement, local vision data, collection logistics and the value model. Valorisation is well served.
5. Use the existing institutions. TREASURES, NEA's Closing the Resource Loop calls and SMU-X give funding, a host and a pace. A standalone initiative will struggle to outlast its first grant.
6. Plan for vendor change. The market has seen a bankruptcy, a sale and a retreat in five years. Own the data, and make sure it can move.
7. Count value in Singapore terms. Food cost, disposal cost, compliance and landfill life. Carbon credits are not an established source of income here.

For government. Counting rescued and prevented food would make the main measure reflect the outcome policy wants. The 2027 masterplan review is the natural moment to do it, and a university baseline would give that count a method.³⁷

A checklist for a food waste AI pilot.

To test a pilot's value claim against these points, record it in the AI value ledger.

What is often misreported.

• Food waste is not at a record high. 2025 (790,000 tonnes) was below 2021 (817,000).²
• Food waste segregation is not yet required of all large premises. Only new buildings are bound; existing ones have no date.⁴
• There was no mandatory food waste reporting from 2021. Reporting started with segregation in March 2024.¹¹
• Singapore is not adopting RFID bins for households. The ministry judged it hard with shared chutes.¹⁷
• IKEA did not cut all food waste by half. The 54% is production waste.³⁹
• The 14 to 1 return is a cross-sector average. Hotels and restaurants were about 7 to 1.⁴⁴⁴²
• Hilton Singapore Orchard's 30% is a target, not a result.³³
• Too Good To Go does not operate in Singapore.³⁰
• The non-domestic recycling target is 81%, not 80%.⁸

Dates to re-check.

Full timeline.

Common questions about the case.