The Spring of AI Cannot Save the Winter of Hardware

[baoshi.rao]

In an era where AI is reshaping everything, stories of salvation and revival are repeatedly unfolding in the consumer electronics sector.

Over the past few years, mobile phone and PC manufacturers have been struggling through the winter of their respective markets. As the AI wave surges once again, these players seem to have found a lifeline, rushing to enter the fray:

Looking back over the past six months, mobile phone manufacturers like Huawei, Xiaomi, and vivo have successively unveiled their large model strategies; PC players like Lenovo and HP are heavily betting on AI PCs; and upstream tech giants such as Intel, Qualcomm, and NVIDIA have also joined the hardware transformation, enhancing the AI computing power of CPUs and SoCs—a coordinated effort aimed at stimulating consumer demand for device upgrades is already underway.

However, whether this oasis is real or just another mirage remains unknown for now.

From memory, the first time I heard the term 'AI PC' dates back to the era when Acorn International was still frequently active on TV screens.

It was a netbook priced at 1,699 yuan, running on Windows CE with a resolution of only 640*480. The host emphatically promoted its selling points—listening to music, watching movies, virus-free operation, and long battery life—repeatedly during the nearly ten-minute advertisement. Yet, throughout the entire ad, he never explained the meaning of 'AI' in the product's name.

Times have changed, and the classic Windows CE system has recently bid farewell to the stage of history. What were once tedious moments during TV commercial breaks have now been reborn as viral meme videos on platforms. Meanwhile, the term 'AI PC,' which was merely a codename back then, has now become a buzzword in the tech world over a decade later.

Even from a more pragmatic perspective, AI in consumer electronics is not a new story. Long before the explosion of large model hype, AI had already been widely integrated into consumer electronics by various manufacturers in different forms. For example, Face ID on the iPhone X, which 'killed' fingerprint unlocking, relies on machine learning capabilities powered by the neural engine embedded in the A11 Bionic chip.

Another example is computational photography, which smartphone manufacturers have been keen to highlight in recent years. At its core, computational photography uses AI algorithms to achieve functions like adjusting shooting parameters, enhancing image quality, and removing artifacts. Before the frenzy around large models, computational photography was one of the closest integrations of AI with consumer electronics.

However, in the early days of computational photography, hardware players and app developers were competing in the same arena—smartphone manufacturers introduced algorithms into their camera systems, while app developers launched a series of photography apps based on these algorithms. Without differentiated competition, hardware-focused players, whose main business is selling products, naturally couldn't gain an advantage against the rapid iteration of software applications.

Against this backdrop, targeted hardware upgrades became the solution for smartphone manufacturers to break free. In 2017, at IFA 2017 in Berlin, Huawei unveiled the Kirin 970, which featured a built-in NPU (Neural Processing Unit). With the NPU's support, the Kirin 970 significantly improved its image processing capabilities.

Since Qualcomm's upgrades to ISP (Image Signal Processing) units and the recent wave of dedicated imaging chips from smartphone manufacturers, the competition in computational photography has evolved from pure algorithm optimization to hardware-level advancements.

With AI becoming increasingly integrated with hardware, smartphone makers have successfully avoided endless battles with app developers. Under computational photography's influence, mobile imaging has become more professional, ultimately sounding the death knell for once-popular DSLR cameras. App developers, meanwhile, have gradually retreated from professional photography, instead differentiating themselves through lighter, more personalized filters and templates.

The current trend of on-device large language models among smartphone manufacturers appears to follow a similar logic to computational photography's evolution.

In industry terms, large models represent consumer electronics manufacturers' hope to emerge from the current market downturn. In the smartphone sector alone, since August this year, Huawei, Xiaomi, Honor, OPPO, and vivo have all unveiled their large model strategies. Even Apple, typically slower to adopt industry trends, is actively recruiting talent to develop its own large model.

This means that with tech giants making significant inroads, large models have become an area smartphone manufacturers cannot afford to ignore.

Although the 'thousand-model war' began over half a year ago, practical applications of large models have primarily focused on the B2B sector, remaining somewhat disconnected from consumer markets. Smartphones, PCs, and other hardware devices serve as ideal entry points for consumer-facing large models, naturally attracting numerous AI model developers.

Since ChatGPT entered the mobile market, domestic large model products like Wenxin Yiyan, iFlytek Spark, and Zhipu Qingyan have quickly followed, successively appearing in mobile app stores. With tech giants rushing in, mobile manufacturers must act or risk being left with only scraps from application developers.

However, the battle of large models differs significantly from previous mobile imaging competitions. As the hottest sector in current tech discourse, this field is dominated by industry giants, creating unprecedented competitive pressure for hardware manufacturers—most of whom lack the capability to compete with tech giants in terms of model parameters, computing power, or data resources.

In this context, on-device large models have become the key for consumer electronics players to leverage their 'home field advantage.'

Examining mobile manufacturers' strategies and Apple's job posting about 'implementing compressed and accelerated large language model functionality in Apple device inference engines,' on-device models combined with cloud computing appear to be an unavoidable discussion point.

From a technical perspective, AI applications like ChatGPT and Midjourney deploy models in the cloud, with smartphones and PCs serving merely as display mediums. This approach inevitably faces limitations such as long response times, inability to function offline, and privacy security concerns.

In contrast, the cloud-device hybrid approach incorporates localized, lightweight large models into terminal devices—handling simple requests through on-device models for real-time offline inference and quick responses, while delegating complex problems to the cloud.

In response, Qualcomm and MediaTek have proactively upgraded their processors and established close collaborations with smartphone manufacturers. Taking Qualcomm's recently released Snapdragon 8 Gen3 as an example, thanks to upgrades in NPU, APU and other components, its AI performance has improved by 98%, sufficient to support the operation of Meta Llama 2 and Stable Diffusion on mobile devices.

This approach resembles the hardware-focused strategy adopted by smartphone manufacturers during the early days of computational photography. Clearly, facing the ambitions of tech giants in the AIGC era, consumer electronics manufacturers still firmly believe in the competitive philosophy that "the best defense is a good offense."

Clayton Christensen, Harvard Business School professor and author of "The Innovator's Dilemma," once said: "Innovation always starts as something unremarkable, a joke even. But when it suddenly addresses consumers' unmet needs, it grows explosively and becomes dominant."

The logic of "quantitative change leading to qualitative change" behind this statement has always been a creed in the tech world. Often, however, this is merely self-indulgence by players - clinging to so-called technological romanticism that ultimately fails to resonate with users, leaving only themselves moved.

For the consumer electronics industry, large AI models seem to be the "lifesaver" that could stimulate a replacement wave and reshape the product cycle. At this stage, however, no one can answer: Where is the killer application? And how much are users willing to pay for current large model applications?

The fluctuations of products like Wenxin Yiyan and iFlytek Spark in the consumer market highlight that the perception and actual demand for technological evolution among consumers may not be as strong as industry players imagined. If users remain unwilling or unable to perceive innovation, so-called 'AI hardware' might end up no different from past gimmicks like 'nano-level home appliances.'

On the other hand, whether it's AI-powered smartphones or Lenovo's AI PC, the underlying logic is to leverage AI to stimulate replacement demand in the existing market.

This is a pitfall players have already stepped into with foldable screens. At that time, the smartphone market was highly homogenized, and with the iterative upgrade of flexible screen technology, eager manufacturers quickly launched foldable devices, packaging them with new demands. In the early stages of foldable screens, aside from widely criticized issues like creases and hinges, there were also numerous software and application compatibility problems. Such immaturity led the consumer market to view them as 'white elephants.' Yet, with Pandora's box opened, players had no choice but to push forward with foldable screens alongside the supply chain, gradually resolving issues like hinges, compatibility, and weight.

Even today, after multiple iterations, foldable screens have seen significant growth in shipments. However, they still haven't reversed the decline of the smartphone market. In the PC sector, even Nvidia's continuously upgraded DLSS technology can only prolong the life of PCs rather than save them, failing to halt the market's downward trend.

Perhaps smartphone and PC markets are unlikely to witness another 'iPhone moment'—consumer electronics manufacturers aim to breed a faster horse, but what the market truly wants might be a car.

In other words, if the development of large models in the consumer market remains tightly bound to hardware, the replacement demand in the existing market will struggle to sustain the popularity and prosperity of AI. The next entry-level hardware that can fully capitalize on demographic dividends might be the real silver bullet.

Fortunately, the close correlation and cross-application of consumer electronics technologies have infused hardware manufacturers committed to large model development with deeper and more diverse value. Although tangible results may not yet be visible in the shallows of time, this initiative has undoubtedly planted forward-looking seeds, laying a solid foundation for the future.