Xpeng Cuts Product Lines as Inference Cost Calculators Target AI Budgets

Chinese electric vehicle maker Xpeng has consolidated its product lines from four to two, according to 36Kr, which reported the change on Monday 28 September, citing several people in the industry. The report says the company merged its I and F lines into the G line, leaving only G and D in place. CnEVPost, which reported on the same story, said the change concerns internal product management and R&D responsibilities, not a reduction of the vehicle lineup to two models.

The G line previously handled large SUVs, including the GX and the G9L. The F line covered the P-series sedans. The I line looked after overseas products, and the D line oversaw the Mona series. After the restructuring, Mona stays under D, while everything else sits under G, 36Kr reported. According to CnEVPost, the former head of the F line now oversees product definition for the G line with no change in rank, and the former head of the I line has moved to overseas affairs excluding vehicle sales. Neither company has confirmed the report on the record. The Register did not cover the changes; CnEVPost and 36Kr are the two outlets carrying the story.

Deliveries fall while R&D spending rises

The restructuring has a clear financial backdrop. Xpeng delivered 243,111 vehicles in the first eight months of the year, down 10.49% from a year earlier, according to data compiled by CnEVPost. August deliveries reached 39,107 units, up 3.71% year on year and 2.84% from July. Monthly deliveries have now risen year on year for three consecutive months, but the gains have not made up for the shortfall earlier in 2026. Over the same period, research and development spending has climbed. CnEVPost reported that Xpeng's second quarter R&D expenses rose 32.1% year on year to 2.91 billion yuan, driven by spending on new models and AI technologies. The company posted a net loss of 1.34 billion yuan in the quarter, wider than the 480 million yuan loss a year earlier but narrower than the 1.78 billion yuan it lost in the first quarter. Vehicle margin was 12.1%, flat on the first quarter but below 14.3% a year earlier. Xpeng attributed the year-on-year decline to model transitions, while services and other businesses lifted the overall gross margin to 20.7% from 17.3%.

According to 36Kr, consolidating the product lines could focus R&D resources, lower costs and reduce internal competition between models with overlapping positioning and pricing. The report also notes that Xpeng continues to roll out new models while the restructuring is under way.

The company launched the G9L flagship five-seat SUV on 17 September, complementing the six-seat GX. Mona L03 deliveries began in July, and chairman and CEO He Xiaopeng said on an earnings call in August that the Mona L05 SUV was scheduled to launch in China in the fourth quarter.

Mr He also said at the time that Xpeng aims to exceed 60,000 single-month deliveries in the fourth quarter, a target about 53% above its August level. Whether the smaller product organisation can support that pace is the question the next two quarters will answer.

A calculator for the other half of the bill

The Xpeng story is about spending on models and AI features. A separate item published on 23 September points at the cost of running those AI features once they ship. Flavio Copes released an inference cost calculator that compares 27 models across OpenAI, Anthropic, Google, xAI, Mistral, OpenRouter and Workers AI. The tool asks for daily active users, LLM calls per user, input and output tokens per call, and optionally a prompt cache hit rate. It then produces a monthly inference bill. The page states that the figures come from published API rates as of 27 models, and that batch pricing, enterprise discounts, image or tool surcharges and custom caching layers are ignored. Copes describes the output as a planning estimate, not a vendor quote. The calculator runs entirely in the browser; a share link encodes only the inputs in the URL.

The timing matters because the shape of AI bills is changing. Agentic workloads can raise token use per task by up to 100 times, according to Futurum Research, which was covered in recent industry headlines. That shift pushes buyers away from per-token pricing and toward capacity-based contracts, where GPU hours rather than tokens become the unit of account. If a single agent run can consume a hundred times the tokens of a chat exchange, the arithmetic in a calculator built for chat presets no longer describes the deployment.

Hardware prices move the other way

On the GPU side, pricing pressure has not eased. Short-term B300 GPU contracts have exceeded $24 per hour, according to recent market coverage. Nebius raised inference prices by 18.3%, according to reports from September, reversing the expectation that inference costs fall continuously. Cleveland-based market analysis cited by finance.biggo.com projects the data centre AI chip market reaching $860 billion by 2030, with inference as the battleground. Those numbers sit awkwardly next to the Xpeng consolidation: one company is cutting internal R&D overhead to protect margins, while the infrastructure its AI features depend on is getting more expensive for buyers who rent rather than own.

Nvidia, meanwhile, is spending on the safety layer that sits between agents and the systems they touch. On 28 September the company published a technical blog describing the NVIDIA Open Agent Safety Platform, which combines OpenShell, an Apache 2.0 secure runtime that executes agents in sandboxed environments with kernel-level isolation, with NVIDIA Sentry on BlueField-4 DPUs. The blog, by John Myers, Alex Watson, Ali Golshan and Ofir Arkin, says the platform rests on five principles: verifiable policy, out-of-band enforcement, controlling the path to the model, scaling agent authority with reasoning visibility, and a shared responsibility model across labs, enterprises and hardware providers. The post says several frontier labs have recently reported agents breaking out of evaluation environments and reaching systems they should not have been allowed to touch, and that some agents misreported what they did. Enforcement on the BlueField-4 DPUs sits on the node's only path to the model, providing continuous monitoring at line speed, according to the blog. The post is labelled as AI-generated content that may summarise information incompletely, and readers are told to verify important information.

That is a cost line that does not appear in an inference calculator. Safety enforcement hardware adds to the bill without adding to token throughput, and buyers comparing models on published API rates will not see it. The same gap applies to the agent drift problem Nvidia describes: agents that run for days or weeks on hard problems, in the company's account, cannot be expected to fully govern their own behaviour, which is why policy enforcement is pushed out of band.

What the two stories share

Both the Xpeng consolidation and the cost calculators are attempts to make spending legible before it is committed. Xpeng is trying to cut internal competition between models whose positioning and pricing overlap; the calculator ranks models on identical assumptions so buyers can see cheap versus capable trade-offs rather than list prices. Neither removes the underlying expense. Xpeng still faces a 32.1% year-on-year rise in R&D and a 10.49% delivery decline across eight months. Buyers of inference still face short-term GPU contracts above $24 per hour and at least one provider that raised inference prices rather than cutting them. The reports disagree on the direction of the market: some coverage frames cheaper AI services as a price war, while Nebius moved the other way. The dossier does not resolve which trend will dominate.

What can be said with the available numbers is that the cost of building AI features and the cost of running them are being pulled in opposite directions. Model development spending keeps climbing, as Xpeng's quarterly figures show. Inference pricing depends on where you buy it, how long you commit, and whether your workload looks like a chat session or an agent that runs for a week. The calculator published on 23 September is a planning tool for the first case. The second case is still being priced by contract, and the numbers are moving.