Processor chips have eliminated price differences for mobile phones.

The latest forecast from Omdia shows that in 2026, global smartphone shipments will decline by 12.2% year - on - year to 1.093 billion units, but the total market value will increase by 6.1% year - on - year during the same period. Fewer units are sold, but more money is made. The average global smartphone selling price is expected to rise from $467 in 2025 to $565 in 2026, an increase of 21% or $98. Both figures are at record highs in the industry.

It is the storage cost that drives up the price. In the first quarter of 2026, the average prices of DRAM and NAND flash memory increased by more than 80% quarter - on - quarter. The surging demand for HBM from AI servers has siphoned off a large amount of the production capacity of memory manufacturers, resulting in a tightening of the supply for consumer electronics. Omdia predicts that even if the growth rate slows to single - digits in the second half of the year, the component costs will remain high, and manufacturers will have to continuously pass on the pressure to retail prices. The market structure will then diverge: low - end models will be forced to shrink due to cost increases, while high - end models will benefit in terms of market share, and the refurbished second - hand market will expand simultaneously.

In response, the Chinese government's approach is to provide demand - side subsidies. On June 18, eight departments including the Ministry of Commerce issued the "Implementation Opinions on Accelerating the Development of 'Artificial Intelligence + Consumption'", clearly stating that the policy of fiscal interest subsidies for personal consumer loans will be used to support consumers in purchasing AI phones, smart computers, AI glasses and other products. While phone prices are rising, AI phones have been listed as a national consumption strategic category.

The overall number of phones is decreasing, but the high - end segment is growing. At product launch events, phone manufacturers are no longer discussing how fast the processors are.

For the flagship phones launched in the summer of 2026, almost all of them are equipped with the same Snapdragon 8 Gen 5 processor. Xiaomi is talking about its 7000 mAh Jinshajiang battery, vivo is talking about the ceiling of folding - screen battery life, iQOO is talking about gaming heat dissipation, and moto is talking about fitting a 6000 mAh battery into a folding body. In the Android camp, the topic of SoC performance has faded, and the processor chips have become invisible in product launches, just like 4G technology.

Phone manufacturers are caught in the middle and have to find other ways out.

End - side AI competition: Trillion - parameter models are a gimmick, while power consumption and memory are the real constraints

The parameter arms race in end - side AI is very intense in terms of numbers. A report from Counterpoint indicates that in 2026, smartphones with GenAI capabilities will account for 45% of global shipments. However, analysts also point out that there is a significant gap between devices capable of running AI and devices that users actually use AI functions on in their daily lives. Although the hardware capabilities have been enhanced, user behavior has not changed accordingly.

MediaTek and vivo jointly demonstrated running a 33B - parameter large model on the Dimensity 9300. Huawei claims that its Kirin flagship can perform local inference on a sparse model with trillions of parameters. Xiaomi has publicly announced that its flagship has successfully run a trillion - parameter MoE model. The numbers 7B, 13B, 33B, and 100B keep increasing every few months. Then vivo made a reverse decision: it switched its main end - side model from 7B back to 3B. This is not because of a lack of technology, but because the 3B model only occupies 2GB of memory, has a power consumption of about 750mA, and can continuously process 128K long texts. In daily use, the 3B model can achieve similar results to the 7B model, but the phone won't get hot and the battery won't drain quickly. What users want is not a model with larger parameters, but an AI that can truly change their daily usage experience, respond at any time, not get hot, not consume much power, and actually help users in every scenario when they open their phones.

Behind this judgment is a fact that is not often mentioned in the industry: those end - side parameters in the tens or hundreds of billions are essentially based on a sparse MoE architecture. Although the total number of parameters is large, only tens of billions are actually activated during each inference. After INT4 quantization and compression, the actual computational load is similar to that of a 7B Dense model. The trillions represent the total capacity of the warehouse, not the amount used each time.

This trend means that the AI capabilities of phones are jointly determined by the LPDDR5X memory capacity, NPU computing power, and power consumption budget. The stable implementation standards almost all converge around 7B. After INT4 quantization, the 7B model requires about 4GB of memory, which is within the available range of 12 - 16GB LPDDR5X in flagship phones. MediaTek has clearly stated that the APU 790 of the Dimensity 9300 can perform inference on the 7B model at a speed of about 20 tokens/s. OPPO has deployed the 7B end - side model for more than 100 AI functions. Although Qualcomm does not disclose the number of parameters, the actual benchmark level of its AI engine is the same. Beyond this, the requirements for phone memory capacity and heat dissipation will exceed the actual limits of most flagship phones.

The meaning of this number for the chip industry has changed accordingly. In the past, the standard for evaluating NPUs was peak TOPS, and the higher the computing power, the better. However, when phone manufacturers start to actively replace large models with small models, what NPUs really need to do is to stably run a long - context inference task within a power consumption budget of 750mA, rather than chasing peak scores.

Metrics such as the space on - chip SRAM used for KV Cache, the scheduling efficiency of memory bandwidth, and the native support for INT4/FP8 low - precision formats are closer to the AI experience that users actually feel than the TOPS number.

The bottleneck in inference lies not only in the NPU computing power but also in whether the storage bandwidth can feed the model weights in a timely manner. A read speed of 10.8GB/s directly affects the model loading speed and the KV Cache refresh efficiency, which, like the NPU's TOPS number, determines the AI response speed that users feel.

Storage manufacturers have realized this. The UFS 5.0 solution launched by Samsung on June 23 has a sequential read speed of 10.8GB/s, more than twice that of the previous - generation UFS 4.1, and the overall energy efficiency has been improved by more than 40%. Samsung positions this product as "the core underlying infrastructure for end - side AI". However, mass production of UFS 5.0 will not start until the fourth quarter of this year, which means it will appear in next year's flagship phones, not at this year's product launches. Counterpoint's analysis points out that storage constraints are one of the core reasons why GenAI phones are currently priced above $400. UFS 5.0 can bring a performance leap, but the initial cost will not be low, and the pattern where high - end phones benefit first will not change in the short term.

The focus of the competition in phone AI is shifting from the device itself to the AI model layer running on the device. Counterpoint's research shows that in the high - end market, Google Gemini is becoming the core of this layer. Gemini supports Apple's rebuilt Siri, is the foundation of Samsung Galaxy AI, and also drives the AI capabilities of the overseas versions of major Chinese phone brands. OEM manufacturers are responsible for orchestrating logic, user experience, and ecosystem integration on top of the model. This is where the real competition will take place in the next stage.

The rise of a new track: Coprocessors + End - side models

The competition logic of end - side AI has changed, but one thing remains the same: there is no room for differentiation in the processor level of flagship phones. Two phones can use the same SoC, but product launch events cannot cover the same topics. Differentiation can only be found in areas that the SoC cannot reach: imaging algorithms, gaming experience, battery life scheduling. The general design of the SoC cannot naturally cover the optimal solutions for these experience - level competitions.

The choice of phone manufacturers is to design their own chips to excel in areas where the SoC performs poorly.

Since Apple used its A - series chips to create a performance gap with Android phones, "self - developed SoC" has become the ultimate aspiration in the phone industry. Many phone manufacturers have attempted to develop chips, but real - world data shows that directly developing a flagship SoC that can compete with Qualcomm and MediaTek is not a cost - effective option.

What phone manufacturers have realized later is that there is no need to replace the Snapdragon. They only need to design a small chip for areas that the Snapdragon cannot cover.

iQOO's Q2 gaming coprocessor is a typical example. It does not touch the CPU, GPU, or NPU. It only focuses on super - resolution and super - frame rate for game graphics. The Adreno GPU of the Snapdragon 8 Gen 5 can also perform this task, but it also has to handle system graphics rendering, UI composition, and other loads. Therefore, the super - resolution effect and power consumption are not optimal. The Q2 coprocessor takes this task separately and uses a dedicated chip to achieve the best results, allowing the main SoC to free up resources to maintain a stable frame rate.

Xiaomi's self - developed imaging chip follows the same logic. It does not replace the ISP of the Snapdragon. Instead, after the ISP completes basic processing, it undertakes tasks that require higher computing power and lower latency, such as computational photography, multi - frame synthesis, and long - focal - length image quality optimization. The division of labor between the two chips is more efficient and generates less heat than having a single chip handle everything.

The cost - effectiveness of this approach is much higher than that of self - developed SoCs. The functional boundaries of coprocessors are clear, and the development cycle is short. They mostly use mature 12/16/28nm processes, and the tape - out cost is only a fraction of that of advanced processes. There is no need to support a complete compiler and driver ecosystem. A gaming chip can be developed from project initiation to mass production within the replacement cycle of the SoC, which is one to two years faster than waiting for Qualcomm to update the GPU in the next - generation Snapdragon.

This trend has a two - way impact on the chip industry. The demand for dedicated chips using mature processes is increasing, which benefits the utilization rate of 12/16/28nm production lines. At the same time, Qualcomm and MediaTek are forced to adapt to this trend. For phone manufacturers' coprocessors to smoothly access the SoC's data path, more underlying interfaces need to be opened. The cooperation model has changed from "selling a single chip" to "providing a collaborative platform".

OpenAI is also planning to make phones

OpenAI plans to launch an AI - centric phone in 2028, and it will cooperate with Qualcomm and MediaTek to develop the chips. This choice is worth noting. When the world's largest AI company decides to enter the phone market, it does not attempt to self - develop the SoC but directly partners with two existing phone chip platforms. This once again shows that the SoC layer is not the key. The real goal is to seize the model layer that Gemini has already occupied.

This points in the same direction as what is happening in the phone industry: the SoC is becoming infrastructure, and the real differentiated competition is dispersed across three levels: the AI model layer, the coprocessor layer, and the application layer.

The competition in the AI model layer is about which end - side model can run longer under a 750mA power consumption, and which orchestration logic can make users actually use it; the competition in the coprocessor layer is about which can excel in specific scenarios such as game super - resolution and image processing; the competition in the application layer is about which can truly change users' usage habits with end - side AI.

The demand for phone chips is being pushed towards efficiency by AI on one hand and towards integration by cost on the other. Both directions are compressing the exclusive value of flagship SoCs and opening up new opportunities for mature processes and local players.

In the future, the price difference of phones will come from manufacturers' innovation capabilities and the implementation speed of AI, which are the real key points for breakthroughs.

This article is from the WeChat official account "Semiconductor Industry Insights" (ID: ICViews), author: Liu Qian. It is published by 36Kr with authorization.