XPeng Leads, Tesla Localizes as EV AI Costs Rise

China’s EV industry delivered a striking mix of expansion, localization, and supply-chain pressure this week. On June 24-25, XPeng confirmed production at its new Malaysia plant, Tesla China’s infotainment system was reported to be integrating ByteDance’s Doubao large model, and new industry data showed automotive-grade memory chip prices have surged by about 180% over the past three months. Together, these developments show how Chinese EV makers are scaling globally while the software-defined car race increasingly depends on AI capabilities, localized digital ecosystems, and a fragile semiconductor supply chain.

XPeng accelerates overseas production with Malaysia plant

XPeng has officially started production at its Malaysia plant in Melaka, marking its third localized production base globally. The first XPeng G6 units have already rolled off the line at the EPMB factory, with volume ramp-up now under way.

This matters because XPeng is no longer simply exporting vehicles from China. It is gradually building a regional manufacturing footprint designed to reduce tariffs, improve logistics, and support local market expansion.

XPeng’s international manufacturing footprint

XPeng’s recent operating metrics also suggest it is scaling from a stronger financial base than in prior overseas pushes:

• May deliveries: 32,158 vehicles
• Month-on-month growth: 4%
• Q1 gross margin: 20.6%, up 5 percentage points year on year
• R&D spending in Q1: RMB 2.91 billion, up 46.8% year on year
• Cash on hand: RMB 42.09 billion
• Q2 delivery guidance: 100,000-106,000 vehicles
• Q2 revenue guidance: RMB 19.6-20.8 billion

The bigger story is that XPeng is trying to become one of the few Chinese EV startups able to industrialize globally while still investing aggressively in software and intelligent driving. That combination is hard to sustain without margin improvement, and XPeng’s latest numbers suggest it now has more room to do so.

Tesla China turns to local AI with Doubao integration

One of the most revealing software developments came from Tesla China. According to reports shared from ByteDance’s Volcano Engine event and screenshots from Tesla’s internal test software version 2026.14.11, Tesla’s China-market vehicle system is preparing to integrate the Doubao AI assistant.

In practical terms, the move appears aimed at improving Tesla’s localization in mainland China, especially in natural-language voice interaction.

What the reported Tesla-Doubao setup could do

• Wake-word control via “Hey, Tesla”
• Vehicle functions such as:
  • air conditioning
  • windows
  • navigation
• A new Doubao AI assistant entry in the app list
• A split-AI structure in China, with:
  • Doubao handling vehicle control and voice interaction
  • DeepSeek reportedly supporting conversational Q&A and more advanced localized AI tasks

For Tesla, this is less about adding a novelty chatbot and more about closing a local product gap. Chinese consumers increasingly expect their smart EVs to function like mobile AI terminals, with fluid voice control, contextual understanding, and app-like service integration. Domestic brands such as NIO, XPeng, Li Auto, and Huawei-backed competitors have spent years training buyers to expect exactly that.

Tesla’s response is notable because it suggests the company is becoming more pragmatic in China: rather than relying purely on its global software stack, it is leaning into local AI ecosystems where needed.

The hidden bottleneck: automotive memory chip prices jump 180%

While the EV industry talks endlessly about autonomous driving, in-cabin AI, and on-device large models, all of that increasingly depends on memory. That is why the latest supply-chain warning may be the most important story of the week.

According to CCTV Finance, overall prices for automotive-grade memory chips have risen by around 180% in the past three months.

The core reason is simple: AI servers are absorbing most of the new capacity.

Samsung, SK hynix, and Micron control more than 90% of the global DRAM market, and they are prioritizing higher-margin products such as HBM and server-grade DDR5 for data-center customers. The automotive sector, by contrast, accounts for only about 3% of global DRAM demand, leaving it at the back of the queue.

Why cars are getting squeezed

• AI customers sign long-term agreements and pay premium prices
• HBM consumes far more wafer capacity than traditional DRAM
• Carmakers still rely heavily on LPDDR4/LPDDR4X for cockpit and ADAS systems
• New capacity is expected to favor HBM and server memory first
• Meaningful supply relief may not arrive until late 2027 or 2028

Cost pressure is rising fast

At the same time, vehicle memory demand is climbing rapidly:

This is a structural problem, not a short-term fluctuation. As large language models move into the cockpit and ADAS stacks become more compute-intensive, memory becomes both a cost item and a production-risk item.

How Chinese automakers are responding

The industry response is beginning to take shape around three main strategies.

1. Long-term supply agreements

Some large automakers have locked in pricing and volume through long-term contracts, creating short-term protection from spot market volatility. But that only delays the problem until renewal, especially if future supply is already reserved for AI and data-center clients.

2. Smarter E/E architecture

The more durable solution may be architectural rather than purely procurement-based.

A key industry theme is cockpit-driving integration: instead of separate chips and separate memory pools for infotainment and ADAS, automakers are increasingly looking at unified SoC architectures with shared memory.

Benefits include:

• lower DRAM usage
• less redundant data copying
• lower bill of materials
• better system efficiency
• reduced exposure to LPDDR4/4X supply constraints

Qualcomm has also argued that integrating cockpit and driving functions into a single system can materially reduce total system cost, including memory requirements.

3. Domestic substitution and inventory buffers

Chinese memory suppliers are making progress, with names such as CXMT, GigaDevice, and Puya increasingly part of the discussion. But near term, domestic alternatives still cannot fully replace imported high-end automotive memory at scale.

That means most carmakers will need a blended strategy:

• secure long-term contracts where possible
• redesign software and model structure to run in smaller memory footprints
• shift toward newer memory standards such as LPDDR5 over time
• build local backup supply where feasible

Battery reality check: CATL says solid-state is not ready for mass scale

In another important reality check for the EV market, CATL chairman Robin Zeng (Zeng Yuqun) said at the Summer Davos forum in Dalian that the current solid-state battery technology route is only at Level 4, while Level 9 is needed for true mass production.

That is a significant statement in a market where solid-state batteries are often treated as the next inevitable breakthrough.

CATL’s key message on solid-state batteries

• Current technology maturity: Level 4
• Mass-production readiness target: Level 9
• CATL’s 2027 goal: roughly Level 7-8, enabling small-batch production
• Before 2030, million-unit vehicle deployment is unlikely

Zeng’s point was not just technical. He emphasized that a viable battery technology must also clear three commercial tests:

• sufficient supply capability
• product reliability and safety
• market acceptance at workable cost

This is especially relevant for Chinese EV brands that must balance innovation messaging with product economics. For the next several years, lithium iron phosphate, high-nickel chemistries, fast charging, pack integration, and cost engineering are still likely to matter more commercially than headline-grabbing solid-state claims.

AI infrastructure is becoming an automotive issue too

Even though OpenAI’s newly unveiled Jalapeño AI chip and Nvidia CEO Jensen Huang’s latest remarks were not automotive announcements, they matter for the EV sector because they reinforce a critical trend: cars are now competing with AI data centers for compute-related components, investment, and engineering attention.

Some of the headline figures are striking:

• OpenAI says Jalapeño could cut inference cost by around 50%
• The chip reportedly went from design to tape-out in 9 months
• OpenAI plans gigawatt-scale data-center deployment by the end of 2026
• Nvidia reported annual revenue of $216 billion, up 65%
• Nvidia data-center revenue reached $194 billion, up 68%

For the auto industry, the takeaway is clear: AI is no longer adjacent to mobility. It is reshaping semiconductor allocation, software stack design, and the economics of intelligent vehicles.

Why This Matters

China’s EV sector is entering a new phase where competitiveness no longer depends only on battery cost or vehicle design.

The next battleground includes:

• global manufacturing localization, as seen with XPeng in Malaysia and Europe
• localized in-car AI ecosystems, as seen in Tesla China’s Doubao move
• supply-chain resilience, especially for memory and compute components
• realistic technology roadmaps, as highlighted by CATL’s caution on solid-state batteries

In other words, the winners in Chinese EVs will be the companies that can industrialize globally, localize digitally, and redesign their hardware-software architecture around an AI-constrained supply chain.

Looking ahead

Over the next 12 to 18 months, expect three trends to intensify.

First, more Chinese EV makers will push local assembly abroad to protect margins and expand market access. Second, in-car AI assistants will become a mainstream product differentiator in China, especially as brands compete on voice interaction, agent-like functions, and localized services. Third, automakers will increasingly treat memory efficiency as a strategic engineering priority, not just a component sourcing issue.

That may be the most important lesson from this week’s news: in the era of software-defined vehicles, the smartest EV is not just the one with the most AI features. It is the one built on the most resilient industrial stack.