Omdia's latest forecast indicates that global smartphone shipments will decline by 12.2% year-on-year to 1.093 billion units in 2026, yet the total market value will increase by 6.1% year-on-year during the same period. Lower volume, higher revenue. The global average selling price of smartphones is expected to rise from $467 in 2025 to $565 in 2026, a 21% increase or $98, both setting new industry records.

Rising prices are driven by storage costs. In the first quarter of 2026, the average prices of DRAM and NAND flash memory rose by over 80% month-on-month. The surge in demand for HBM from AI servers has absorbed a significant portion of memory manufacturers' capacity, tightening supply for consumer electronics. Omdia predicts that even if price increases slow to single digits in the second half of the year, component costs will remain high, forcing manufacturers to continue passing pressure onto retail prices. The market structure is accordingly differentiation (diversifying): low-end models are being scaled back due to rising costs, while high-end models are gaining market share, and the refurbished and second-hand market is expanding simultaneously.

In response, the Chinese government is focusing on 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,'" explicitly supporting consumers in purchasing AI phones, smart computers, AI glasses, and other products through fiscal interest subsidies for personal consumer loans. As phone prices rise, AI phones have been designated as a national consumption strategic category.

Overall phone prices are declining, while high-end models are rising. At product launches, phone manufacturers no longer discuss how fast their processors are.

Flagship phones launching in the summer of 2026 almost all feature the same Snapdragon 8 Gen 5 processor. Xiaomi emphasizes its 7000 mAh Jinsha River battery, vivo highlights its foldable screen endurance ceiling, iQOO focuses on gaming cooling, and moto discusses fitting a 6000 mAh battery into a foldable body. In the Android camp, SoC performance has faded from the spotlight, with processor chips becoming as invisible in launches as 4G once was.

Trapped in the middle, phone manufacturers are forced to innovate elsewhere.

01 On-Device AI Intensifies: Hundred-Billion Parameters Are a Gimmick; Power Consumption and Memory Are the Real Constraints

The parameter arms race in on-device AI is numerically lively. Counterpoint reports that 45% of global smartphone shipments in 2026 will feature GenAI capabilities, yet analysts note a significant gap between devices capable of running AI and those where users actually engage with AI features daily. Hardware capabilities have surged, but user behavior has not followed suit.

MediaTek and vivo jointly demonstrated running a 33B-parameter large model on the Dimensity 9300; Huawei claimed its Kirin flagship could locally infer sparse models with hundred-billion parameters; Xiaomi proudly announced its flagship ran a hundred-billion MoE model. 7B, 13B, 33B, 100B—the numbers double every few months. Then vivo made a counterintuitive decision: switching its primary on-device model from 7B back to 3B. Not due to technical limitations, but because the 3B model occupies only 2GB of memory, consumes about 750mA of power, and can continuously run 128K long texts. In daily use, the 3B model performs similarly to the 7B model, but without overheating or draining the battery. Users don't want larger-parameter models; they want an AI that genuinely transforms daily experiences—responsive, cool-running, and power-efficient, helping in every scenario when the phone is unlocked.

Behind this judgment lies an often-overlooked industry fact: those hundred-billion or thousand-billion on-device parameters are essentially sparse MoE architectures—the total parameter count is large, but only a few billion are activated per inference. After INT4 quantization compression, the actual computational load is similar to a 7B Dense model. Thousand-billion parameters represent warehouse capacity, not daily usage.

This trend means a phone's AI capabilities are jointly determined by LPDDR5X memory capacity, NPU compute power, and power budget, with stable deployments converging around the 7B mark. A 7B model, after INT4 quantization, requires about 4GB of memory, falling within the 12-16GB LPDDR5X range of flagship phones. MediaTek confirmed that the Dimensity 9300's APU 790 can infer 7B models at ~20 tokens/s, while OPPO deployed 7B on-device models for over 100 AI functions. Qualcomm, though not disclosing parameter counts, targets the same practical scale with its AI engine. Beyond this, memory capacity and cooling requirements exceed most flagship phones' practical limits.

This shifts the chip industry's evaluation criteria for NPUs. Previously, peak TOPS (tera-operations per second) was the standard—higher compute power was better. But as OEMs proactively replace large models with smaller ones, NPUs must now stably run long-context inference tasks within a 750mA power budget, rather than chasing peak benchmark scores.

On-chip SRAM space for KV Cache, memory bandwidth scheduling efficiency, and native support for INT4/FP8 low-precision formats now matter more to users' perceived AI experience than TOPS figures.

Inference bottlenecks stem not just from NPU compute power but also from storage bandwidth's ability to feed model weights promptly. A 10.8GB/s read speed directly impacts model loading speed and KV Cache refresh efficiency, determining perceived AI responsiveness alongside NPU TOPS figures.

Storage vendors have recognized this. Samsung's UFS 5.0 solution, announced on June 23, offers a sequential read speed of 10.8GB/s—more than double the previous UFS 4.1—with over 40% better energy efficiency. Samsung positions this product as the "core underlying infrastructure for on-device AI." However, UFS 5.0 mass production won't begin until Q4 this year, meaning it will appear in next year's flagships, not this year's launches. Counterpoint notes that storage constraints are a key reason GenAI phones are currently locked above the $400 price point. UFS 5.0 will enable performance leaps, but initial costs will remain high, so high-end models will benefit first—a pattern unlikely to change soon.

The focus of phone AI competition is shifting from devices themselves to the AI model layers running on them. Counterpoint highlights that in the premium market, Google Gemini is becoming the core of this layer, powering Apple's rebuilt Siri, Samsung's Galaxy AI, and the AI capabilities of major Chinese phone brands' overseas versions. OEMs then handle orchestration logic, user experience, and ecosystem integration atop these models—where the next phase of competition will truly unfold.

02 Emerging Battleground: Coprocessors + On-Device Models

The competitive logic of on-device AI has changed, but one thing remains: flagship phones no longer differentiate at the processor level. Two phones can use the same SoC, but their launches can't focus on the same aspects. Differentiation must come from areas SoCs don't optimize: imaging algorithms, gaming experiences, battery scheduling. SoCs' general-purpose designs inherently can't deliver optimal solutions for these experiential layers.

OEMs' solution: build a dedicated chip to excel where SoCs fall short.

Since Apple's A-series chips created a performance gap with Android, "self-developed SoCs" have become the ultimate aspiration in the phone industry. Many manufacturers have attempted chip development, but data shows that creating a flagship SoC to compete with Qualcomm and MediaTek is not cost-effective.

Manufacturers now realize: there's no need to replace Snapdragon; they just need a small chip for tasks Snapdragon doesn't handle well.

iQOO's Q2 gaming chip exemplifies this approach. It doesn't touch the CPU, GPU, or NPU—only handling game visual upscaling and frame interpolation. While Snapdragon 8 Gen 5's Adreno GPU could technically do this, it must also handle system graphics rendering, UI composition, and other workloads, compromising upscaling quality and power efficiency. The Q2 offloads this task to a dedicated chip, optimizing it while freeing the main SoC to maintain stable frame rates.

Xiaomi's self-developed imaging chip follows the same logic: it doesn't replace Snapdragon's ISP but takes over computationally intensive tasks like computational photography, multi-frame synthesis, and telephoto image optimization after the ISP completes basic processing. Dividing tasks between two chips proves more efficient and generates less heat than a single chip doing everything.

This approach is far more cost-effective than self-developing SoCs. Coprocessors have clear functional boundaries and short development cycles; they extensively use mature 12/16/28nm processes, keeping tape-out costs a fraction of advanced nodes; they don't require full compiler and driver ecosystems. A gaming chip can go from conception to mass production within a single SoC generation cycle, reaching the market one to two years faster than waiting for Qualcomm to update GPUs in the next Snapdragon.

This trend has bidirectional impacts on the chip industry. Demand for mature-node dedicated chips is rising, boosting utilization rates for 12/16/28nm fabrication lines. Simultaneously, Qualcomm and MediaTek must adapt: for OEMs' coprocessors to seamlessly integrate with SoC data pathways, more low-level interfaces must be opened, shifting cooperation from "selling a chip" to "providing a collaborative platform."

03 OpenAI Is Making a Phone Too

OpenAI plans to launch an AI-centric phone in 2028, partnering with Qualcomm and MediaTek for chip development. This choice is noteworthy: the world's largest AI company, when entering the phone market, didn't attempt to self-develop an SoC but instead chose two existing mobile chip platforms. This reaffirms that the SoC layer isn't the focus; securing the model layer, where Gemini already holds sway, is the true objective.

This aligns with the phone industry's ongoing transformation: SoCs are becoming infrastructure, with true differentiation emerging across three layers: AI model layer, coprocessor layer, and application layer.

Competition in the AI model layer hinges on whose on-device model runs longer within a 750mA power budget and whose orchestration logic users actually engage with; in the coprocessor layer, it's about who can excel in specific scenarios like gaming upscaling or imaging processing; in the application layer, it's about who can genuinely alter user habits through on-device AI.

Phone chip demand is being pushed toward efficiency by AI and toward integration by cost pressures. Both trends erode the exclusive value of flagship SoCs while opening new opportunities for mature-node processes and local players.

Future phone price differences will stem from manufacturers' innovation, AI implementation speed—the true breakthrough points that matter.