Xpeng Merges Four Product Lines Into Two To Rein In R&D Spending

The Chinese EV maker merged its I and F lines into its existing G line, leaving only the G and D lines, according to 36Kr, which cited several unnamed industry sources. CnEVPost carried the report on Monday.

Nothing changes for buyers yet. The company still sells the same cars.

The reorganisation concerns internal product management and R&D responsibilities, not the number of models on sale, the report says. Before the merger, the G line handled large SUVs including the GX and G9L. The F line covered products including P-series sedans. The I line looked after overseas products, and the D line ran the Mona series.

After the change, Mona stays under D. Everything else falls under G. The former head of the F line now oversees product definition for the G line, with no change in rank, 36Kr reported, citing an industry source close to Xpeng. The former head of the I line has moved to overseas affairs, excluding vehicle sales in overseas markets.

Why the merger landed now

The timing is not subtle. 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. Monthly figures have improved for three straight months, but not enough to close the gap opened earlier in the year.

August deliveries came in at 39,107 units, up 3.71% year on year and 2.84% from July. That is still below the run rate the company needs to hit its own targets.

Costs are moving the other way. Xpeng's second-quarter R&D expenses rose 32.1% year on year to 2.91 billion yuan, which the company attributed mainly to spending on new models and AI-related technologies. Deliveries in the same quarter rose just 0.1% to 103,295 vehicles.

The pressure to translate investment into returns is also evident in its financial results.

The company posted a second-quarter net loss of 1.34 billion yuan ($199 million), wider than the 480 million yuan loss a year earlier, though narrower than the 1.78 billion yuan loss in the first quarter. Vehicle margin was 12.1% in the quarter, flat from the first quarter but below 14.3% a year earlier. Xpeng blamed model transitions. Services and other businesses lifted overall gross margin to 20.7% from 17.3%.

Against that backdrop, consolidating overlapping lines looks less like tidying and more like triage. The 36Kr report notes that merging the lines could focus R&D resources, lower costs and reduce internal competition between models with overlapping positioning and pricing.

The lineup keeps growing

Xpeng is not shrinking its product plan. It launched the G9L flagship five-seat SUV on 17 September, slotting it alongside the six-seat GX for different family needs. The Mona series is expanding too: deliveries of the Mona L03 began in July, and chairman and CEO He Xiaopeng said on an August earnings call that the Mona L05 SUV was scheduled to launch in China in the fourth quarter.

With those models, He said the company aims to exceed 60,000 single-month deliveries in the fourth quarter. CnEVPost notes that target is about 53% above the August level. Whether the internal merger helps or hinders that push is an open question. Fewer product lines with clearer ownership could speed decisions. It could also concentrate risk if one line misreads the market.

Xpeng has not publicly confirmed the restructuring, and the company did not respond in the CnEVPost report. The account rests on 36Kr's industry sources, so treat the management changes as reported rather than announced.

A different kind of cost problem

Xpeng's R&D squeeze is a hardware story: engineers, tooling, model transitions. The other cost pressure running through the tech industry this week is compute, and it is moving in the opposite direction for buyers.

On 23 September, developer Flavio Copes published an inference cost calculator that compares 27 models across OpenAI, Anthropic, Google, xAI, Mistral, OpenRouter and Workers AI. It takes assumptions about daily active users, calls per user, input and output tokens, and an optional prompt-cache hit rate, then produces a monthly bill.

The tool is a planning estimate, not a quote. Its own description says the figures come from published API rates and ignore batch pricing, enterprise discounts, image or tool surcharges, and any caching layer a customer builds. Still, it is a useful reminder that the dominant cost of an AI app is the model calls, and that the spread between cheap and capable models is wide enough to matter more than list-price headlines.

Read alongside the recent run of pricing news, the picture gets messier. The Register reported in mid-September that Nebius raised AI cloud prices, and a 25 September item noted an 18.3% inference price increase at the same company. Context headlines this week also point to B300 GPU pricing above $24 per hour in short-term contracts, and to research arguing agentic AI can multiply token use per task by up to 100 times.

None of those figures come from the same market as Xpeng's R&D line. They are useful only as a contrast: one company is cutting internal duplication to save money, while the compute market its AI features depend on is getting more expensive per unit of work.

Cheaper heat, at a price

Even the hardware fringes are feeling cost pressure, in their own way. On 27 September, Tom's Hardware reported that a customer ordered a custom Kubb Fanless mini PC with a pure 24-carat gold passive chassis, at a cost of around $1.7 million.

The machine is a premium version of the roughly 5-inch (120 x 120 x 120mm) Kubb Fanless PC. The gold shell pushes weight from 4.6 pounds to 28.7 pounds (2.1kg to 13kg), more than six times heavier. Fanless Tech, which first posted about the build, called it a cooling marvel and said gold's thermal conductivity is about 30% to 50% faster than aluminium. Tom's Hardware notes copper would have been far cheaper and offers even better thermal conductivity.

The specs underneath are ordinary: Intel Core Ultra 5 225H, Core Ultra X7 358H, or AMD Ryzen AI 5 340, with Wi-Fi 7 or Wi-Fi 6 depending on the chip. Standard aluminium Kubb Fanless units in graphite, blue or bronze are listed at €1,595 ($1,822) at Kubb.eu, with no stock.

It is an extreme example of the same theme: what you pay for compute hardware and its supporting design is diverging sharply from what the silicon inside can do.

Elsewhere: Brazil solar and a solo build

Cost pressure is not confined to chips. On 26 September, pv magazine reported that average PV system prices in Brazil rose 7% in the first half of 2026 for projects up to 300 kW, according to Greener's Distributed Energy Solutions study. Final system prices ranged from BRL 2.02 ($0.38) per watt for 30 kW and 50 kW installations to BRL 3.62 per watt for 2 kW systems.

Equipment costs drove the increase. The average price of PV kits for 4 kW systems rose 18.3% between January and June, from BRL 1.42 per watt to BRL 1.68 per watt. For 300 kW systems, the kit price rose just 2.0%, from BRL 1.02 to BRL 1.04 per watt.

Demand softened at the same time. New distributed generation connections fell 16% year on year, from 488,000 to 411,000, and the number of new consumer units receiving credits dropped 43%, from 951,000 to 541,000. Residential systems accounted for 65% of added capacity, up from 39% in 2019. Only 33% of integrators' sales involved financing, down eight percentage points from 2025 and the lowest share recorded in the period analysed.

On the software side, a solo developer launched ParkourNote on 28 September, a research workspace built in a month with Claude, according to its Show HN post. It imports papers, books, web pages, videos and recordings, transcribes and translates them, and exposes the library to Claude, ChatGPT or Codex through an MCP server. It runs in the browser, in Chrome's side panel and on iPhone and iPad.

The pricing angle is implicit rather than stated: the project leans on existing frontier models rather than building its own, which is one way to keep inference costs off your own balance sheet.