Introduction and Research Questions
China’s online food delivery (OFD) sector is one of the most consequential areas of urban infrastructure to emerge in the past decade, affecting how tens of millions of residents eat on a daily basis. The OFD ecosystem connects consumers, food service merchants, drivers, and food courier networks to allow food to be delivered directly to homes and workplaces. Today, China represents the world’s largest OFD market, generating hundreds of millions of daily orders and supporting an industry worth hundreds of billions of yuan.
Fueling the continued expansion of OFD is the rapid growth of delivery oriented food service models, particularly ghost kitchens (GKs). GKs are delivery only facilities built by major brands and investors that operate without a public storefront. These facilities only prepare food for handoff to couriers fulfilling online orders, which allows them to eliminate expenses such as dining tables, customer service staff, and paying more expensive rent for locations with high foot traffic. Alongside GKs, hybrid restaurants combine traditional dine in service with OFD capabilities by maintaining physical dining spaces and employing waiters and hosts, while also designing kitchen layouts, assembly lines, and staff workflows to manage delivery orders from third party food couriers.
At the same time, underlying demand for OFD continues to grow. Industry forecasts project China’s OFD market to expand at roughly 9 to 15% annually through the early 2030s, while the country’s GK segment is expected to grow by more than 25% annually. The Guangdong Province, where the city of Shenzhen is located, is also expected to lead national growth. As demand increases, however, growth in delivery capacity may not occur evenly across space.
This essay reports on a summer research project conducted in partnership with the Department of Urban Planning and Design at the University of Hong Kong. The project examined two questions. First, what spatial patterns characterize the location of GKs in Shenzhen, based on PhD candidate Weipeng Deng’s LLM generated dataset of GK locations, and how do these locations relate to traditional restaurants, land use, and underlying food demand? Second, does the existing GK network provide sufficient delivery accessibility to meet current demand, and where might additional GK capacity most effectively address spatial gaps as demand continues to grow? Together, these questions aim to identify where GKs are located, assess whether their distribution aligns with demand, and recommend where expanding the network could improve food delivery accessibility while being economically viable for the industry.
Background
This project uses the Two Step Floating Catchment Area (2SFCA) method, a spatial accessibility measure rooted in public health geography, first proposed by Radke and Mu (2000) and later formalized by Luo and Wang (2003) to address physician shortage areas in Illinois. The method proceeds in two steps: first, for every supply location (originally a clinic, but here a restaurant or kitchen), it defines a catchment containing all demand points within a fixed travel distance or time, and computes a supply-to-demand ratio within that catchment. Second, for every demand location, it sums the supply-to-demand ratios of all supply points whose catchments reach it, producing a single accessibility score per demand point. The 2SFCA family has since been applied well beyond healthcare to other sectors where spatial mismatch between supply and demand applies, such as grocery access and banking deserts, making it a good fit for modeling OFD accessibility.
This project’s supply layer drew on a restaurant classification developed by the supervising research group, which used an in-house large language model pipeline (LLM4GKID) to identify GK locations across Shenzhen based on business listing signals. Land use was drawn from EULUC China 2.0, a nationally consistent essential urban land use classification product, and the road network came from OpenStreetMap. Together, these sources allowed the project to ask not just where GKs are, but what kind of land they tend to occupy relative to traditional restaurants and customer demand.
Data and Results
The empirical work proceeded in four steps: market composition analysis, baseline accessibility modeling, spatial clustering analysis, and scenario based stress testing. All map analysis is done in reference to the map of Shenzhen, China.
Market Composition and Land Use Overlap
First, three restaurant supply layers (Traditional Dine In, Hybrid, and GKs) were derived using an LLM-based extraction pipeline developed by the Department of Urban Planning and Design at the University of Hong Kong (HKU) to produce geolocated datasets. They were then counted within each 500 meter grid cell and mapped against Essential Urban Land Use Categories (EULUC), a land use classification system for Chinese cities that uses 10 meter satellite imagery to categorize areas into residential, commercial, industrial, and other urban land types, to identify locational trends.
Across the study area, total dining supply reached 142,669 establishments: 85,509 traditional dine in restaurants (59.9 percent), 44,869 hybrid establishments (31.4 percent), and 12,291 ghost kitchens (8.6 percent). Ghost kitchens are present in only 32 percent of grid cells, compared with 61 percent for traditional restaurants, and where they do occur, GK density averages 2.07 facilities per cell but reaches as high as 89 in a single grid cell, indicating a market structure defined by a modest number of intensely GK saturated clusters rather than uniform citywide dispersion.
A EULUC land use analysis shows that ghost kitchens and traditional restaurants largely overlap in their dominant locations: both concentrate heavily in residential zones (53.4 percent of traditional restaurants versus 51.1 percent of ghost kitchens) and commercial zones (24.96 percent versus 23.86 percent). However, GKs are consistently overrepresented, relative to their overall market share, in educational, industrial, and transportation land use categories. For example, GKs account for a larger relative share of restaurants sited in transportation stations and industrial zones than traditional dine in restaurants do. This pattern is consistent with them engaging opportunistically in lower rent areas rather than a deliberate strategy to chase underserved demand. Moreover, GKs display tighter, more isolated spatial clustering in peripheral grid cells than traditional restaurants.
Demand Estimation
Second, demand was estimated independently of any single delivery platform’s proprietary order data, drawing instead on companion travel and consumption survey data collected in Shenzhen by the HKU Department of Urban Planning and Design. A fixed 500 meter grid was built across Shenzhen in a projected coordinate system, and the city’s roughly 18 million residents were distributed across cells in proportion to population density. Two citywide average behavioral frequencies, drawn from the data, were applied uniformly to each cell's estimated population: 2.48 weekly delivery occasions and 3.71 weekly dine out occasions per person. The resulting dataset estimates roughly 44.6 million delivery meals and 66.8 million dine out meals per week citywide (a combined demand of about 111.3 million meals). This formed the demand side of the accessibility model.
Third, using this demand estimate alongside a Two Step Floating Catchment Area (2SFCA) model on the same 500 meter grid, the project built a baseline map of delivery accessibility across Shenzhen’s Built Up Area (BUA), areas identified by the HKU Department of Urban Planning and Design as particularly relevant for study. A 2SFCA accessibility index (Ai) was computed for every grid cell using total restaurant supply as the numerator and the constructed weekly demand estimate as the denominator, within a delivery relevant catchment radius. The baseline catchment was set to 3.0 kilometers, equivalent to roughly a 25 to 30 minute delivery window for motorcycle delivery drivers. This time frame was chosen because delivery courier platform data report that order completion probability drops off sharply as courier travel time increases. This produced a continuous Online Accessibility Index surface across the BUA.
Baseline Accessibility (2SFCA)
Under the 3.0 kilometer baseline catchment, 40.0 percent of BUA grid cells (1,749 cells) were classified as high dual deficit zones. These are areas where both online and offline supply fall short of estimated local demand. Only 0.02 percent, or a single grid cell, showed a critical GK reach. High density peripheral zones exhibited the most severe baseline deficits, revealing a mismatch between Shenzhen’s central restaurant clusters and its growing suburban population (Figure 2).
Figure 2. Baseline Online Accessibility Index (Ai) across Shenzhen’s BUA under a 3.0 km catchment (left), and the resulting recommended restaurant development strategy by grid cell (right).
Spatial Clustering (LISA + Gi)*
To test whether restaurant supply followed a spatially random or systematically clustered pattern, two Local Moran’s I (LISA) were computed on total restaurant supply using a 1,016 meter distance threshold to identify statistically significant High-High and Low-Low clusters. A High-High cell is a cell with a high value (in this case high supply of a specific restaurant type) surrounded by neighboring cells that also have high values, or a cluster of areas with rich supply next to each other. Oppositely, a Low-Low cluster would have areas with poor supply next to each other.
Figure 3. Local Moran’s I (LISA) clusters of total restaurant supply, FDR adjusted (q ≤ 0.05).
The Local Moran’s I (LISA) analysis of total restaurant supply confirmed that this pattern is not spatially random. Of the significant clusters identified, 805 grid cells formed High-High clusters while 1,392 cells formed Low-Low clusters, concentrated toward the northern and northeastern periphery of the BUA. Spatial outliers were rare: only 13 cells were classified as Low-High, and 2,716 cells showed no statistically significant clustering.
Figure 4. Getis-Ord Gi* hot and cold spot classification of total restaurant supply, FDR adjusted, with 90/95/99% confidence tiers.
The Getis-Ord Gi* hot-spot analysis corroborated the LISA results while adding confidence tiered detail: 557 cells were classified as 99 percent confidence hot spots and a further 261 and 189 cells as 95 percent and 90 percent confidence hot spots respectively, again concentrated centrally and to the south. On the cold spot side, 384 cells reached 99 percent confidence, concentrated overwhelmingly in the northeastern periphery, with 786 and 412 additional cells at the 95 percent and 90 percent confidence levels.
Together, the two tests show that in Shenzhen, most of the investment in restaurants and GKs is toward already well served central districts rather than the underserved periphery.
Scenario Stress Testing
The baseline model was stress tested under four scenarios that change catchment radius, demand level, or supply configuration: (1) a current market mobility shock, in which the catchment radius was compressed to model traffic congestion or severe weather; (2) a future demand growth scenario simulating a 20 percent increase in online demand over two years, consistent with sector CAGR projections; (3) a targeted GK siting intervention layered onto the deficit map from the baseline scenario, or Figure 2; and (4) a scenario simulating the conversion of 20 percent of randomly selected traditional dine in restaurants into hybrid ones.
(1)
Shrinking the effective delivery catchment from 3.0 to 2.0 kilometers is a proxy for traffic congestion, courier shortages, or severe weather such as the typhoons that periodically affect Shenzhen.
Figure 5. Spatial access under a 2.0 km delivery catchment shrinkage, modeling mobility shocks such as congestion or severe weather.
(2)
Next is a demand growth scenario simulating a 20 percent increase in online orders over a two year horizon (a trajectory broadly consistent with sector wide CAGR forecasts in the 9 to 15 percent range).
Figure 6. Newly emerged delivery gaps under a simulated 20% online demand growth scenario over a two-year horizon.
This identified 646 newly vulnerable grid cells, or 14.8 percent of the entire BUA, that shifted from sufficient to deficit status. These new gaps form concentric vulnerability rings around Shenzhen's central commercial hubs, indicating that order volume is projected to outpace fixed supply capacity fastest at the immediate edges of the current core, rather than in the already deficient far periphery (Figure 6).
(3)
This is the result of building ghost kitchens where Figure 2 identified deficits to restore food accessibility.
Figure 7. Spatial recovery under targeted GK siting in identified deficit zones.
This figure visibly closed high deficit rings around the urban core without requiring costly new storefront retail. This suggests that precision in where new delivery capacity is added matters most.
(4)
This converts 20 percent of existing traditional dine in restaurants into hybrid operations.
Figure 8. Spatial access under a scenario converting 20% of offline dine in supply into hybrid F&B capacity.
This produced only a thin band of newly restored accessibility, concentrated near locations that were already close to sufficient. Thus, solution (3) is likely more economically viable.
Investment Prioritization
An investment priority tiering was then produced by layering the Scenario 3 growth projection against the Ai baseline, classifying cells into a Tier 1 “Critical” category with zero GK reach or a severe deficit, a Tier 2 “High Priority” category (Ai < 0.5), and a Tier 3 “Moderate Priority” category (Ai < 1.0).
Figure 9. Composite Level 2B investment urgency tiers, combining baseline deficits with projected demand growth.
Tier 2 High Priority cells (Ai < 0.5) dominate the map, forming thick, ring shaped bands around the built up perimeter and along the transitional zones between Shenzhen's major supply clusters. A small number of Tier 1 Critical cells appear as isolated points embedded within these Tier 2 bands, most likely corresponding to locations with nearly zero GK reach despite meaningful local demand, while only a handful of cells fall into the intermediate Tier 3 category.
Facility Deployment Strategy
A facility deployment strategy was produced to recommend where operators should build different facility types based on each cell’s underlying supply-demand gap and existing balance between dine in and delivery infrastructure. Multi Brand Cloud Kitchen Hubs are large scale facilities housing multiple brands, recommended for areas with the most severe deficits. Standard GKs are delivery only facilities that serve food and beverages from a single brand, recommended for areas with moderate deficits in food accessibility that are primarily O2O related. Delivery First Hybrids combine GK capacity with a small pickup counter, recommended for areas with some dine in demand. Equal Dual Service Hybrids balance dine in and delivery capacity evenly for areas where both channels contribute meaningfully to demand.
Figure 10. Recommended facility deployment strategy by grid cell based on supply-demand deficits and the existing balance of dine in and delivery infrastructure.
Discussion
Several limitations should temper how these results are used. The demand layer applies uniform, survey derived per capita delivery and dine out frequencies to every grid cell, rather than allowing for seasonal, demographic, income, or weather related variation in consumption. It is a coarse exploratory estimate. The restaurant classification relies on an in house large language model pipeline that may misclassify ambiguous hybrid cases. The 2SFCA implementation used here applies a single fixed catchment radius per scenario unlike enhanced models such as E2SFCA or kernel density 2SFCA; adopting these enhancements could better differentiate accessibility within each catchment. Finally, the scenario shocks are illustrative stress tests calibrated to plausible ranges drawn from sector literature, not calibrated predictions of any specific future event.
Future work could extend this framework to other Chinese Tier 1 cities and integrate real time or forecasted traffic and weather data to develop dynamic resilience monitoring. Moreover, future work could examine the up and coming drone based delivery market in China and how it will impact food delivery access and the GK industry.
Conclusion
This project set out to determine whether Shenzhen’s rapidly growing online food delivery network distributes accessibility evenly across the city, and whether that network could withstand plausible future shocks. The evidence shows that this is not the case for both: ghost kitchens largely mirror the geography of traditional dine in restaurants rather than correcting it, and roughly 40 percent of Shenzhen’s BUA already sits in a state of dual OFD deficit before any additional stress is applied. Under a realistic demand-growth scenario, nearly 15 percent of the city’s BUA is projected to newly join that deficit, and even routine mobility disruptions are sufficient to push some suburban zones into deficit. Framing food delivery infrastructure investment through a spatial accessibility frame gives operators and planners the evidence base needed to shift from reactionary market competition toward targeted and resilient network expansion.