2026.09.30 | TB | Internet | Didi Ride-Hailing Competition Pt.2

Analyst-led • 21 Sep 2026

Companies: BABA US, 2643 HK, DIDI.Y US, BIDU US

Expert Project Operations Manager at Beijing Bailongma Yunxing Technology Co Ltd

Eric Shen (ES) — Third Bridge Sector Analyst

Agenda

• Amap’s progress on lowering commission rates and rectifying order reselling after regulatory talks, plus commission cap enforcement intensity and changes in Alibaba’s (NYSE: BABA) transport business subsidy strategy
• Caocao’s “Dual 100,000 Plan” – cities and operational scenarios where autonomous driving replaces human drivers first, plus other operators’ robotaxi progress
• Impact of rising “dual-license” compliance rates and tightening of operating licenses on transport capacity supply, plus potential room for lifting per-order platform profits and overall industry net margins
• Market competition dynamics in ride-hailing, plus new opportunities across fleet O&M (operations and maintenance), dispatch, charging and refuelling amid autonomous driving trends

Key Insights

• Didi's net margin still has room for improvement, but the slope will continue to slow down. Commission rectification and capacity clearance do not constitute substantial pressure. Dual-pricing only accounts for 1-2% of industry orders, and comprehensive rectification has little impact on revenue per order. The capacity clearance brought by dual-certificate compliance will only manifest after 5-6 years, with a magnitude of about 10%, posing low short-term risk
• Amap's driver-side subsidy amid its commission-free season only rose from below 5% normally to below 7%, and commission-free cards are mostly purchased by drivers themselves, making the campaign cost lower than order-rushing rewards. The platform's daily average completed orders in July-August were about 11.5 million, an increase of about 10% MoM
• The per-vehicle profitability metrics publicly disclosed by robotaxi companies such as Pony.ai exclude vehicle-purchase capital costs and vehicle idling. The industry breakeven point is about 15 orders per day, while Pony.ai, WeRide and Baidu overall still average under 10, and Caocao is estimated at about 7-8 in Hangzhou

Introduction

ES: [Expert]

Expert], before we start today’s Interview, please state if you agree or disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

Expert [00:00:32]: I agree.

Questions

00:00:35] In our last Interview, you mentioned that the ride-hailing industry has a 27% commission cap. Is this commission cap currently calculated from the platform's perspective or the fulfilling party's perspective?

Expert [00:00:47]: The requirement states that it should be based on both ends—displaying how much the passenger originally paid and how much the driver received on the fulfilling driver's interface, and then calculating the ratio. Therefore, looking at it from this angle, it must be the fulfilling party. However, many instances where the platform side exceeds 27% involve "yin-yang bills" and multi-platform reselling, where the lower-level segments cannot access the actual revenue of the upper-level segments. Thus, under normal circumstances, it should be based on the fulfilling party's perspective. According to the regulatory authorities' requirements, it should involve both ends: if a passenger places an order on Platform A, and Platform A passes it down a chain to Platform B and others, until Platform C or Platform D finally fulfils it, the calculation should bridge both ends. However, these intermediate platforms only know their immediate upstream and downstream data, and cannot access figures across multiple tiers.

ES [00:02:06]: In practical terms, there may still be instances where commissions exceed the cap. Which categories usually absorb the excess portion? How is it calculated?

Expert [00:02:18]: Didi certainly has quite a few of these currently, but the exact number is truly unclear. Didi’s usual practice is to inform you that this specific order has indeed exceeded the cap, but it promises that once you complete a certain order target, it will be refunded next month. Furthermore, they promise that the monthly average should be around 25%; they have a strategy regarding this aspect, which is the first point. Secondly, the "Sunshine Initiative" rule required by the Ministry of Transport is that the passenger's actual payment over the driver's income. Of course, there is a deduction beforehand: the passenger's discount amount should not be borne by the driver. For example, if a RMB 100 fare is discounted by RMB 20, the passenger's actual payment under normal circumstances is RMB 80, which is used as the denominator, so the calculated result is highly likely not to exceed the cap. The usual practice in this scenario is to directly use the amount after the passenger's discount as the denominator, and if there are incentives, they will be added to the numerator, which lowers the commission rate to a certain extent. However, this somewhat contradicts the principle that drivers should not bear the passenger's discount portion.

Thirdly, there are "dual-pricing bills" (yin-yang bills), and every tier will likely try its best to avoid exceeding the cap. However, if it does exceed the limit in this regard, different platforms may handle it quite differently. Some do not embellish it at all; Didi uses the first method, the majority use the second method, and some do not really care at all. Of course, enforcement is quite strict now, and all parties will keep it below 27% at the level where they can be caught.

ES [00:04:44]: Is there still room for such manoeuvres at present? If the regulator merges the relevant subjects into a unified submission, will this impact the per-order revenue of all parties along the aggregation value chain?

Expert [00:05:32]: Firstly, regulatory submissions cannot capture the connection between different platforms under resale order scenarios anyway, because the reporting is dimensioned by specific platforms. Therefore, even if the data is transmitted this way, the transport commission will only see separate, individual entities. Furthermore, while the figure for those exceeding the date limit is unknown, "dual-pricing" bills account for roughly 1-2% of total orders, at most 2% at a push. Returning to the second sub-question. Of course, this slightly differs from your question, and my point is how much the revenue of all parties along the value chain would decrease if this part were completely levelled out. As I mentioned just now, everyone currently tries to add as little as possible; this is especially true for single-layer resales, where the mark-up is kept relatively low, trying best not to exceed 27%. Consequently, even if this part is fully brought in line, there might only be a margin of 1-2pp.

00:06:54] Over 100 platforms participated in Amap’s commission-free season from July to September this year. In terms of capacity-side settlement, what was the approximate magnitude of the per-order subsidy during this round of commission-free initiatives?

Expert [00:07:12]: Overall, driver-side subsidies were under 7%, compared to under 5% normally, which is an increase of about 2 percentage points. This was not limited to the summer commission-free campaign; it covers the entire summer period, roughly July to August, with school reopening in September.

ES [00:07:46]: For example, to generate an incremental volume of 10,000 orders, what is the approximate difference in unit cost between this commission-free driver-side subsidy model and the original order-rush rewards or subsidy strategies?

Expert [00:08:02]: First, this commission-free campaign is divided into two types. One is jointly launched by the aggregator platform and its subordinate platforms, which involves relatively high costs. The other is conducted independently by the CP or the aggregator platform itself, where their commission-free initiatives are actually profitable. This is because commission-free cards cannot be claimed at no cost; they all require some level of payment, and the larger commission-free cards are quite expensive. To my knowledge, some can cost up to thousands of pounds, covering a longer timeframe, potentially on a monthly basis. Furthermore, if you want to drive an incremental 10,000 orders, one approach is the commission-free model, and the other is through order-rush rewards or order-completion rewards. This commission-free model incurs lower costs because it integrates with an intelligent dispatch system.

Of course, the intensity varies by region, CP and time slot. When you buy a commission-free card, normally, if you are lucky, five out of the ten orders given to you might be relatively large. However, once you actually buy the commission-free card, under its global optimisation, you might get a few fewer orders than usual. So this commission-free feature does actually save money, but roughly how much cost can be saved? Under 1%.

ES [00:09:49]: What is the QoQ growth of Amap’s average daily order volume during this commission-free season?

Expert [00:10:01]: The average for July to August was around 11.5 million orders, and the average for May to June was 10.45 million orders, so it increased by about 10%; this is the completed order volume. The 10-odd million mentioned just now is the total order volume, including API, SaaS and taxis.

ES [00:10:48]: After the commission-free season stopped in September, how much has the order volume dropped in mid-to-late September?

Expert [00:10:57]: The summer commission-free campaign is done basically every year, just with different levels of intensity. Comparing this level of intensity with last year's, I really don't know which one was greater or smaller. However, based on last year's summer and the post-summer period, orders dropped by about 7-8%, meaning September dropped to around 93% of July-August.

ES [00:11:40]: You mentioned that commission waivers are roughly equivalent to a 2% increase in driver-side subsidies. Converted proportionally, is the elasticity of the order drop corresponding to a 1% subsidy on Amap roughly 4%?

Expert [00:12:05]: The drop in order volume after the summer peak is not solely because subsidies were reduced. Of course, that is one of the reasons, but the larger drop in volume is driven by falling demand. Scientific research relies on the controlled variable method, and this does not achieve a controlled variable since demand has also dropped. Therefore, relating 1% to the aforementioned points is not entirely accurate.

ES [00:12:45]: In the last interview, you mentioned that Alibaba’s investment in the mobility segment is declining. From the settlement statements on the capacity side, what other signals can you observe? How will follow-up investment evolve?

Expert [00:13:08]: This actually just depends on subsidies. As for other signals, besides online subsidies, it’s mainly offline activities, but I’m not very clear on that area.

00:13:34] Looking at orders within the same city and for the same distance, what is the approximate net income gap per order between Caocao’s direct-operated model and Amap's aggregation model? Over the past year, has this gap been widening or narrowing?

Expert [00:13:53]: The uniform AOV is currently around RMB 24. Caocao’s standard commission rate is 9%, though there are rebates. Based on the AOV and the normal commission rate, it calculates to roughly RMB 2.16. Any change in this variance is simply related to two variables: the AOV and the specific commission amount or channel fee. The AOV has risen slightly compared to last year, by about 0.2 to 0.3 RMB, while the commission has remained relatively unchanged. The actual commission rate should be around 8%, and less than 8%—somewhere in the high 7s—once rebates are deducted. Therefore, with the AOV rising slightly and the commission remaining stable, the variance is on a widening trend.

ES [00:14:58]: Given that orders from aggregation platforms account for more than 85%, can it be inferred that the average commission expenditure per order of automaker-backed platforms such as Caocao Mobility is slightly lower than the RMB 2.16 you just mentioned?

Expert [00:15:16]: For automaker-backed ones, meaning Caocao and T3, their actual standard commission rate is 9%, which comes to around 7.7% or 7.8% after factoring in rebates. Multiplying the commission rate by the average order value gives the commission per order, which is calculated based on roughly RMB 24.

00:15:57] The proportion of proprietary traffic for automaker-backed platforms is now relatively stable, but these platforms generally feel that the proportion of traffic from aggregated platforms is too high and hope to increase their own proprietary traffic share. At roughly what percentage of proprietary traffic will the platform's profit level show good performance? Are there relevant benchmark standards in the industry?

This

Expert [00:16:23]: This is a pseudo-proposition. Why is the proportion of aggregated order volume relatively large? Proprietary traffic cannot be built up on its own, or building it up would cost a lot more money; it is all driven by heavy subsidies. Therefore, without large-scale subsidies, proprietary traffic definitely cannot get off the ground. If we exclude the extra money spent and only look at the traffic results, it can be roughly calculated like this: once proprietary traffic is established, the average net gross margin should increase by the aforementioned average commission rate of 7-8%, weighted against how much it is losing now, and calculated roughly by proportion. That is the logic.

For example, Caocao is currently losing 10%, and its aggregated traffic accounts for 85%. How much does the aggregated traffic earn? How much does the proprietary traffic earn? Proprietary traffic certainly earns more than that part. Increase this weight to the point where it becomes profitable, and that is roughly how it is calculated. As for the profits of automaker-affiliated ones like Caocao, I am honestly not sure.

00:18:03] From 12 September 2026, Ningbo will require new cars to be registered within 30 days to obtain a transport licence, while also raising the thresholds for wheelbase and driving range. What is the approximate magnitude of the impact of such policies rolling out across more cities on fleet capacity and mobility scale?

Expert [00:18:24]: First, compared with previous policies in various cities, the only raised threshold under this policy is the 30-day window, whereas other factors like wheelbase, driving range, and base vehicle price have not changed substantially. I have a rough estimate: nationwide, brand-new vehicles procured by leasing companies account for only around 40%, with the remaining 60% being nearly-new or second-hand cars. Is this a move towards full compliance? Certainly not, because full compliance cannot possibly be achieved in a short time. Moreover, transport licences for existing ride-hailing vehicles are generally valid for eight years or 600,000 kilometres—calculated as eight years—so it will not have a clearing-out effect on existing capacity in the short term. The impact will definitely be minimal since the existing volume remains quite large.

Naturally, if the overarching pace is not full compliance in the short term, the impact remains small. The true impact may manifest 5-6 years from now; by then, all newly registered transport licences nationwide might require brand-new vehicles, and the vast majority of existing capacity will likely have expired, necessitating licence renewals, new vehicle purchases, or new licence applications. However, the vehicle supply structure will also evolve by that time due to increased demand for new vehicle procurement, representing a gradual transition. Without the impact of this policy, the difference should be at most around 10% based on current estimates. Moving from the current 40%, it will gradually bring a degree of improvement to the vehicle supply.

00:20:47] How much higher are the vehicle allocation costs under the new rules compared to before? Has the rising compliance cost had an impact on the supply side?

Expert [00:21:10]: Currently, the vast majority of orders are economy orders. Taking the BYD Qin pure electric model as an example, which is quite mainstream, the on-the-road price for a new car should be around RMB 130,000. With an eight-year depreciation period, the annual depreciation cost is roughly RMB 16,000. Commercial insurance currently on the market ranges from RMB 8,000 to RMB 12,000, with the highest likely being RMB 12,000. Since different insurance companies offer varying schemes, commercial insurance fluctuates between several thousand RMB and RMB 10,000. Additionally, regarding the acquisition of the transport certificate, many cities have hardware installation requirements, such as GPS and alarm devices. From what I understand, this is a one-off fee; though it might involve replacements if broken, the cost is very low compared to insurance and depreciation. On top of that, the remaining costs consist of daily energy replenishment, maintenance, traffic fines, and other sundries. For instance, energy expenditure across the country normally costs an average of about RMB 40 per day, according to previous statistics.

The annual costs for maintenance, repairs and traffic violations average around RMB 5,000. By this calculation, the annual cost is about RMB 45,000. With an average of 26-27 driving days per month, that comes to 310-320 days a year, so the average daily fixed cost is about RMB 140-150. With an AOV of RMB 24 and a commission rate of about 25%, each order brings in only about RMB 17-18. Dividing the daily fixed cost of over RMB 140 by RMB 18, it takes roughly 8-9 orders to break even. However, if renting a vehicle instead, it would probably take about 10 orders, because the rental company also needs to make a profit.

00:23:51] Compliance costs are currently on an upward trend. What are your thoughts on the corresponding C-end churn elasticity?

Expert [00:24:03]: The rise in compliance costs has a major prerequisite, which is the enforcement intensity against non-compliance. If enforcement remains tight at times and loose at others, compliance costs will increase, but there certainly won't be a large-scale contraction in the short term. In the long term, it will contract, but by 10% at most, several years down the road. Another factor to consider is active vs inactive drivers. Some are fully compliant, some are full-time, and some are part-time. This also has a certain impact, though it is a secondary reason.

00:25:10] If there is a certain impact on the supply side, do you think the first manifestation on the platform side would be an increase in subsidies or a decrease in commission rates? What would the approximate magnitude of the corresponding change be?

Expert [00:25:37]: Subsidies will definitely rise, because capacity is lower and drivers are in one big nationwide pool—they can drive for Amap, Didi or other platforms. When drivers have fewer choices and competition among platforms remains fierce, the normal approach is to increase subsidies and reduce commission rates in order to attract high-quality or scarce drivers to the platform. According to the general trend of policies, the commission rate will definitely decline, but it won’t have a great impact in the short term. The Sunshine Initiative has been launched for several years, and the average commission rate is still about 25%. The trend is definitely declining because the competent department aims to ensure drivers’ income, so subsidies will certainly increase slightly. However, this small increase may follow activities and budgets, but there will be a certain increase on the whole.

ES [00:26:58]: Under this trend, what do you think of the changing trend of platform profitability? Can the platform hedge against this pressure with other means?

Expert [00:27:13]: As I mentioned last time, net profit is continuously rising. The universally recognised 5% stable or perfect net profit margin you spoke of will rely more on the precision or intelligent targeted deployment of subsidies, as well as the compression of some personnel or R&D costs to increase further. Subsidies will increase, which will definitely impact the achievement of this 5% stable, perfect net profit margin; however, to summarise, it simply increases the difficulty, but the impact should not be too significant. This is because this situation is much the same for all platforms. Didi and Amap are competing for capacity, and even if B-side subsidies do not change much now, both sides are at the same starting line. Previously, one company would give RMB 10, and now the available drivers, just to give an example, have decreased from one million to 900,000, but a company still subsidises RMB 10 or RMB 5. From the drivers' perspective, choosing Didi or Amap remains the same as before, so there might be a slight impact, but it will be very limited.

00:28:47] In the medium to long term, if a robotaxi is to achieve a marginal cost lower than that of a human driver in a city, what daily order volume and per-order cost magnitude does the industry estimate would be required?

Expert [00:29:06]: The marginal cost will only manifest once a certain volume is achieved; right now, it is just okay, and the industry believes it is around 15 orders. Therefore, it is likely that from around 15 orders onwards, the marginal cost will become apparent, and the efficiency of autonomous driving will surpass that of a human. As a fundamental supplement, if the orders are too few or the scale is too small, there is no question of marginal cost.

ES [00:29:57]: Do you think the robotaxis currently deployed in Wuhan, Guangzhou, Hangzhou and other cities can achieve an average of about 15 orders per day?

Expert [00:30:11]: I heard that Baidu's figure seems to be slightly higher than this, and Pony.ai should have already reached this figure in mature areas. So this statement is basically referencing what certain executives from Pony.ai have said, and internally we also conduct some surveys similar to robotaxi, or deepen our industry knowledge. It is roughly that, but specifically it depends on the phase, because according to local policies, for instance, if there are 1,000 vehicles on a certain autonomous driving platform, it is also divided into several phases. It is roughly divided into four phases: according to our understanding, one is with a driver in the main seat, one with someone in the front passenger seat, one with someone in the back seat, and finally fully driverless.

Furthermore, when exactly it can meet the standards for carrying passengers is also phased. Therefore, if a platform in a certain city is still in the primary phase and is only permitted to carry people in partial scenarios, that is a different calculation altogether. This is because autonomous driving involves demonstration testing, demonstration operation and actual commercial operation, and there should be several phases like these, and some are purely running without being allowed to carry people, just accumulating mileage.

00:31:53] The cost items of autonomous driving companies differ significantly from those of platform operators. In the cases of Pony.ai in Guangzhou and Shenzhen, what are the primary gaps in cost items between the vehicle-level profitability metrics and the transport capacity company metrics?

Expert [00:32:12]: I will first address which scenarios and time slots are more suitable. There are essentially two scenarios: hub-to-hub transfers and highway scenarios, which naturally intersect. Why are these two scenarios currently more suitable for autonomous driving? To put it simply, unlike pure urban ride-hailing, where uncontrollable or unpredictable factors are numerous—entirely within the city.

Returning to the previous question, what items make up the gap between the vehicle-level profitability metrics of an autonomous driving company and the calculations made by a transport capacity company itself? Taking Pony.ai as an example, its unit profitability metrics only offset the operating expenses generated by vehicle operations against hardware depreciation or personnel wages. However, a true transport capacity company will look at it from a complete business operations perspective. What cost items are missing? A transport capacity company will consider the cost of purchasing vehicles, the interest on vehicle financing policies, plus the portions left vacant or idle that even policy support cannot cover.

So there are two dimensions: autonomous driving companies focus on whether revenue and depreciation can offset each other, while others have to consider more purely business factors. The examples I just mentioned, such as vehicle purchase, financing policies and idle rates, are things that autonomous driving companies such as Pony.ai generally do not pay much attention to. One is based on revenue and overall hardware depreciation, which roughly offset each other; the other is based on pure business operations, where the additional metrics or items are exactly those generated by business operations.

00:34:56] What is the approximate percentage of the actual daily average order volume of robotaxis deployed by Caocao in Binjiang District, Hangzhou, compared to human-driven cars? Will this area be rolled out faster in some cities in the future?

Expert [00:35:12]: Firstly, its true order volume is unclear, but I estimate it to be around 7-8 orders; that is about that level, as it may involve scheduling and different stages of the vehicles. However, in Hangzhou, human-driven ride-hailing cars achieve at least 15 orders or more per person, so there is nearly a twofold difference. For Pony.ai, WeRide, plus Baidu, the average across all of Baidu's cities combined is estimated not to exceed 10 orders. They can only perform well in certain specific areas, perhaps reaching over a dozen or even more than 20 orders, but to our knowledge, the overall average does not reach 10 orders.

ES [00:36:19]: Are the fleet operators of robotaxis currently autonomous driving companies such as Caocao Mobility and Pony.ai?

Expert [00:36:44]: They currently are, but next, there will emerge many collaboration models referred to within the industry as the "iron triangle" or "golden triangle", which feature separate operating parties. Right now, entities including Caocao look after all aspects, but there are also many that purely focus on technology. In the case of Pony.ai, if I recall correctly, it was the year before last or sometime around then in Hangzhou, that they found a local public transport group to handle proxy operations. That model is the "iron triangle" or "golden triangle" model, and this type of model will become increasingly common in the future. At present, the ones truly operating are indeed the autonomous driving companies themselves, recruiting safety operators and purchasing vehicles. Of course, the vehicles are certainly produced in collaboration with automakers, but it is limited to collaboration, and the holder is mostly the company itself, though not entirely itself. Didi's situation might be slightly different, but the general framework is much the same.

ES [00:38:00]: Following the "iron triangle" and "golden triangle" models you mentioned, will the fleet operator receive a fixed custody fee or a per-trip commission similar to human drivers? What is roughly the revenue-sharing proportion? Around 5%, how does that come about?

Expert [00:38:22]: Under that model, the traffic brought in by the autonomous driving company itself will surely serve as a supplement or account for a very small share; for the foreseeable future, we will still rely more on these traffic companies. Traffic companies currently take a policy commission of around 9% or 10% from the bottom layer, and in the early stage, the share allocated to drivers or actual vehicle revenue will certainly be around 25%, just like current human-driven ride-hailing services. Therefore, the actual operator, analogous to the sub-platforms under the current aggregated model, takes about 10%, and there is also a technology provider in the middle, so splitting it down the middle makes it around 5%.

ES [00:39:23]: What is the approximate fixed cost of a 1,000-vehicle robotaxi fleet in terms of station O&M [operations and maintenance] and safety officers per year?

Expert [00:39:35]: The annual cost of a vehicle, including the recruitment of safety officers, site rental, maintenance and O&M, is more than RMB 10,000. Therefore, 1,000 vehicles must be worth tens of millions of RMB, but it cannot reach RMB 20m. My answer is only a rough estimate of the overall price. The influencing factors or specific circumstances vary from one company to another, but it must be over RMB 10,000, or it may be RMB 19,000 per vehicle, and the total could be RMB 16m, RMB 15m or RMB 14m. According to our internal discussion before, it should be about RMB 16,000-17,000.

00:40:43] How many vehicles does a safety officer currently manage on average? Initially, relevant standards stipulated that it should not exceed five vehicles. Has this been adjusted now?

Expert [00:41:09]: Initially, the Ministry of Transport likely issued a document specifying three vehicles. However, in actual operation, according to our understanding of Pony.ai or WeRide, WeRide claimed to have reached 1:40 in 2025, which seems quite reasonable. If the ratio remains at 1:3 or 1:5, the significance of truly mature autonomous driving operations in the future might not be as substantial. Therefore, I personally feel that WeRide's claimed 1:40 is reasonable and close to the ratio in real-world scenarios, which is certainly different from the Ministry of Transport's 1:3 or 1:5. This is because, as far as we know, the policies of various cities are derived from the overarching documents of the Ministry of Transport, yet each city still retains a certain level of autonomy based on actual conditions.

ES [00:42:17]: Do you think this ratio of 1:40 will continue to rise, or is it relatively reasonable?

Expert [00:42:24]: A ratio of around 1:50 is likely a comparable figure foreseeable in the future. Companies like Pony.ai and WeRide both claimed last year that they aimed to reach the thousand-vehicle scale. Previously, they might have had a few hundred vehicles in actual operation, but now they should have reached the thousand-vehicle level.

00:42:51] How is the annual fleet management fee calculated based on the triangular model you mentioned, assuming a robotaxi is handed over to a third-party fleet for management? What is the variance in per-vehicle cost between a platform building its own O&M [operations and maintenance] vs outsourcing to a ground fleet? In the future, is it possible for manufacturers to adopt a self-operated management model rather than outsourcing?

Expert [00:43:23]: Currently, the cost of outsourced fleet management is less than twice as expensive as a completely self-operated model, somewhere between 1.x times, roughly RMB 20,000 to RMB 30,000 per vehicle. Furthermore, third-party fleet management companies must also generate a profit, so the likelihood of moving away from third parties to handle it entirely in-house in the future is extremely small. For autonomous driving to genuinely capture a share of the ride-hailing market, the volume of capital required is immense. Traditional human-driven ride-hailing can leverage the capital of local leasing companies or individuals, as many vehicles are purchased by individuals; however, for autonomous driving to operate commercially in the short term, the primary owners will undoubtedly be B-side entities. Therefore, under the premise of completely self-operating while scaling up volume, the probability of bypassing third parties entirely is low, unless they possess massive financial capacity. They might not focus heavily on profit in the early stages, but profit metrics cannot remain infinitely low either. Consequently, the future will still rely on the "golden triangle" model, namely, engaging third-party operational agencies and fleet managers.

ES [00:44:53]: Can the existing players in the current ride-hailing rental and vehicle service ecosystem seamlessly transition into this sector? Or will this industry experience higher concentration? My understanding is that the triangle model should shrink further compared to the original ecosystem.

Expert [00:45:14]: Traditional leasing and fleet management systems are essentially responsible for several aspects. Firstly, the management of heavy assets and documentation—the latter primarily being driving licences and vehicle licences. While driving licences belong to individuals, here it specifically refers to vehicle licences. Additionally, the existing fleet management system covers energy replenishment, maintenance, insurance claims, shift scheduling, and dispatching. These are all within the capabilities of traditional or current leasing companies and fleet management networks. However, when it comes to acting as a third-party operator for robotaxis, they lack several key capabilities. The first is the specialised maintenance of autonomous driving hardware, such as in-car equipment, lidar, or millimetre-wave/millisecond-level radar, as well as the large in-car screens. Traditional fleet management personnel cannot handle this without secondary training, which presents a relatively high barrier to entry.

Secondly, regarding the personnel previously handling scheduling and maintenance, true third-party operation of autonomous vehicles requires the role of a safety operator. In the early stages, these operators may be in the vehicle, but later they will operate remotely. Remote control of robotaxis also carries a certain technical threshold—which we understand is not excessively high, but it exists nonetheless—and current traditional leasing and fleet management systems do not possess this capability.

Furthermore, for autonomous driving, I understand that places like Beijing require dedicated ledgers, which differ from routine car rental records. They must log accident rates or intervention frequencies, something that did not exist before. There are also exclusive emergency procedures. For instance, if an autonomous vehicle gets into an accident, bumps into something, or collides with another car, what happens? The autonomous driving offices in various regions each have a set of specialised protocols. My impression is that a response must be made within five or ten minutes. This is because it affects more than just the brand awareness or recognition of a single autonomous driving company; it impacts the stability of the entire autonomous driving industry. If this is not properly managed, the future large-scale rollout of autonomous driving could run counter to the long-term development goals of the autonomous driving offices.

ES [00:48:15]: From the perspective of the vehicle service system, excluding labour costs, the O&M cost will increase compared with the original level. Is the combined cost not necessarily lower?

Expert [00:48:41]: Yes, absolutely. There will certainly be an increase, which is simply due to the maintenance of specialised intelligent driving equipment. It is impossible to turn to the autonomous driving company, as the technology provider, for every single issue. The response speed or smoothness of handling issues will definitely not be as fast as resolving minor faults ourselves. This includes emergency procedures, which are currently a mandatory requirement by the Autonomous Driving Office. For instance, traditional fleet servicing lacks a regulatory framework; as long as no major accidents or cost losses occur, it is generally considered acceptable. However, managing this area now involves a dedicated regulatory system. While fleet servicing is not regulated by the Municipal Transport Commission and is more of a corporate action, actual autonomous driving operations are governed by the Autonomous Driving Office.

ES [00:49:38]: Has Didi recently made any new progress on its robo-taxi rollout?

Expert [00:49:51]: I’m not really sure.

00:50:02] For a single vehicle in a high-intensity operating fleet, what is the approximate annual energy expenditure? How does the corresponding expenditure of a robotaxi compare with that of a human driver?

Expert [00:50:38]: It is quite different. The empty cruising rate for robotaxis is higher. For example, to complete 15 orders, a human-driven ride-hailing vehicle might have an empty cruising rate of only around 30%, whereas a robotaxi is estimated to reach over 40%. Consequently, under the premise of completing the same number of orders, it covers more mileage. Naturally, this is not limited to service mileage, so its energy expenditure in this regard is higher than that of a human-driven ride-hailing vehicle. On top of that, there is another difference: the electricity of a human-driven ride-hailing EV is primarily used to propel the vehicle, but an autonomous vehicle is equipped with more intelligent hardware that human-driven EVs lack. This hardware consumes power, so the overall energy consumption is expected to be over 10% higher than that of a human-driven ride-hailing vehicle, given equivalent service parameters, such as the number of orders.

ES [00:51:15]: After robotaxis begin to scale up, how will the related business and per-vehicle return rate of Didi’s self-operated rental vehicles be affected?

Expert [00:51:52]: I really don’t know about that.

00:52:11] Regarding Didi’s current development and potential inflexion point, do you have any additional insights to share?

Expert [00:52:14]: Major ride-hailing companies have laid out plans for autonomous driving, yet its development speed lags far behind platform expectations. The primary bottleneck or friction point stems from government policies, which currently lack absolute uniformity. For instance, Hellobike reportedly intended to launch operations in Wuhan at some point last year. Their previous fleet might have comprised a miscellany of vehicles—with no clear indication of which brand predominated—but to our knowledge, their operation in Wuhan mandated a complete fleet transition to the Dongfeng system. This is because Dongfeng is seemingly headquartered in Hubei, though whether specifically in Wuhan remains uncertain.

Local governments have superimposed numerous political performance considerations onto the advancement of autonomous driving. Why enforce the use of Dongfeng? Because Dongfeng is based there and generates tax revenue; the more Dongfeng vehicles utilised, the higher the tax revenue the Dongfeng Group contributes locally. This is demonstrably unreasonable. For autonomous driving to thrive, it simply needs to safely transport passengers, maintain a low intervention frequency, exhibit a minimal accident rate, and fully match the passenger-carrying capabilities of a human-driven ride-hailing vehicle. Restricting it to specific vehicles, models, or brands is evidently irrational.

Consequently, the future evolution of autonomous driving will remain heavily constrained by local government policies. Whilst a revised policy framework is reportedly in the pipeline, it has been anticipated for over a year, and to our knowledge, the overarching master plan has yet to materialise; we would have heard of its release otherwise. Didi might be progressing well with relatively solid profitability. Paying excessive attention to autonomous driving feels somewhat redundant at this stage. This current phase likely spans the next three to five years, during which it will not yield substantive operational optimisation and will remain largely conceptual. It may take roughly five to six years—admittedly a ballpark estimate—before volumes gradually begin to scale. This encapsulates my general assessment of Didi's pursuit of autonomous driving.

ES [00:55:05]: Didi's autonomous driving segment has developed relatively slowly. Do you approve of its influence and decision-making?

Expert [00:55:50]: Yes, indeed, its presence is not as prominent as that of Pony.ai, WeRide or Apollo Go. Personally, I speculate they might believe this sector will not yield significant returns in the short term, yet it remains a crucial layout for the future. Take Baidu, for instance; it completely bypassed ride-hailing and placed its entire bet on autonomous driving, specifically, via Apollo Go. To this day, Baidu's market volume in the ride-hailing business remains quite low within the industry.

Apart from this area, as we mentioned last time, its profit margin reached 4.6%, and my overall impression of Didi is that it should continue to grow, though the outlook is not overly optimistic, as the industry has now become relatively mature. It certainly has not reached 100% maturity, but it is like a student at school: progressing from 50 to 70 marks can happen very quickly, but moving from 95 to 100 marks is a very slow process.

ES [00:57:41]: Thanks again to [Expert] for their input today. This now concludes our meeting. Goodbye.