Vol.213 Industry Watch 40 | Thinking Through Why CPUs Are Heating Up and the New AI Intelligent Terminal
Summary
- As AI moves beyond training into inference, applications and Agents, performance bottlenecks are spreading from a single GPU to the CPU, memory, SSD, VRAM and PCIe data movement, putting the entire motherboard back in the value-allocation equation. 田阳 believes that as chips move to 7nm and smaller process nodes, they may become more sensitive to thermal disturbances and surrounding conditions, making power delivery, cooling and I/O scheduling more than engineering details. 银普’s goal is: “Given a fixed amount of energy and a fixed amount of heat generated, compute as much of the information that should have been computed as possible.”
- Personal AI terminals may split into 2 categories: handheld devices will compete for the interaction layer, while home devices will become “compute routers.” If the former is merely an external compute box for a phone, the phone will eventually absorb its capabilities; the latter can quietly connect robots and different kinds of AIoT over the local network, shifting competition toward stability, high compute, small size, low power consumption and a reasonable price.
- 银普 will first sell system integration this year—using the same components to deliver stronger performance—then move the differentiation into control chips and coprocessors next year. The roadmap includes a first desktop device, followed by 4 more desktops and 1 handheld; this June, it will launch an AI SSD at Compute X that lets the same computer run AI models far larger than before. Next year’s products aim to accelerate specific models through replaceable AI cards, much like “plugging in a game card,” while making clear that 银普 “will not make processor chips.”
- If a Lobster device or Mac mini-style hardware merely ships with 1 popular app preinstalled, its commercial life may be shorter than the hardware development cycle; the capabilities that can truly compound are Memory, retrieval, the OS and storage. 田阳 sees the core opportunity over the next 5 years in software-hardware Memory: local data accumulates over time, allowing an Agent to “understand you better and better.” This would elevate AI from an app into a system-level capability akin to graphics rendering; whether vendors call it AIOS or AIPC is beside the point.
- Over the next 10 years, the edge-cloud relationship is more likely to be organized by task value than by 1 side eliminating the other. Private, frequent, important but uncomplicated daily tasks are suited to local execution, while complex programming and task orchestration will continue to draw on cloud capabilities; 银普 hopes to take hardware priced roughly like an iPhone, which typically runs only 8B or 16B models, to support 235B, 397B and larger models. 田阳 also acknowledges that compromise-style system integration will remain mainstream over the next 1-2 years.
- 银普’s coprocessor strategy sacrifices general-purpose compatibility, using extreme software hardening and ASICs to achieve lower costs and support larger models. The products may be plug-in modules through interfaces such as M.2 and U.2, or may enter embedded devices; they need not be integrated into the main processor to become a dedicated capability layer on the motherboard. 田阳 believes the traditional PC supply chain cannot beat Lenovo and Asus head-on, but OS, chip and hardware-software coordination could create a new brand that does not replace the old PC yet owns a new category.
- China’s “Lobster craze” looks more like collective anticipation in the absence of a mass-market AI application, while hardware companies should still begin with professional C-side and B-side users. 李丰’s personal observation is that roughly “99% of General C” users did not know what the Lobster could do either before or after installing it; 田阳 says professional users will upgrade when they get a generational performance improvement, and the heat has already faded as “Lobster festivals” have become less frequent. Over the next 1-1.5 years, China’s B-side market may be no weaker than overseas, while professional C-side adoption is more likely to come from overseas: “Hardware builders are not well suited to chase a fad; they need to anticipate one.”
Deep dive
1. Interdisciplinarity Is Not Résumé Decoration; It Is the Starting Point for Treating the Computer as a Physical System
田阳 attributes his shift from computational neuroscience and statistical physics to Huawei’s 2012 Lab and entrepreneurship to “doing it for fun.” His SBTI result was “GoGo” and his MBTI type is INTP; what looks like risky randomness is fundamentally interest-driven, following his instincts.
Computational neuroscience has 2 paths: researchers with computer science backgrounds tend toward simulation and biologically inspired computing; 田阳’s physics-background path uses mathematical tools to explain how the brain encodes external information and what dynamical states neural activity exhibits, without rushing toward applications.
He noted that the founder of Anthropic also pursued a PhD in statistical physics and complex systems, under William Ballek. The common imprint of this training is an attempt to use concrete mathematics to describe complex systems that were previously difficult to quantify.
2. Robots That Only Mimic Humans May Miss the More Powerful Forms of Intelligence
Looking at VLA, VLT A, VLM and world models, 李丰 asks: animals do not study physics, yet they can judge whether a branch will bear weight, whether a rock can serve as a foothold and whether they can clear a puddle. Could this “intuitive physics” become the entry point for robots to understand the real world?
田阳’s counterpoint is that “all biological organisms are extremely flawed.” Humans do have certain senses and subconscious capabilities that are stronger than today’s AI, but that does not mean robots should return to biomimicry; if humans may create something stronger than themselves, “robots should be different from humans.”
Natural intelligence worth borrowing is not limited to the brain: at the chemical scale there are molecular robots and self-assembly; at the physical scale there is collective self-organization. Human intelligence is only 1 part of natural intelligence. The real target is the rules left by long evolution, not a replica of human form and mechanisms.
3. After 2015, Neuroscience Shifted From Labeling Brain Regions to Distributed Systems
田阳 sees 2015 as the inflection point: traditional fMRI research tended to assign specific functions to specific brain regions, but a growing body of work shows that cognitive processing often engages the whole system—“pull 1 thread and the whole body moves”—making distributed coding important.
李丰 compares this mobilization of multiple regions to jointly process inputs and outputs with MoE in large models. 田阳 did not explicitly endorse the analogy here, instead pivoting to his entrepreneurial experience.
The brain-computer interface chain has also become clearer: first identify signal circuits associated with a limb or cognitive phenomenon, then extract neural signals, use mathematics to isolate the effective components and finally map neural activity to human intent. 田阳 cautiously expects applications for movement, speech and communication among people with disabilities to “possibly leave the lab within the next 10 years.”
4. Information Thermodynamics Turns “Compute More Per Joule” Into an Entrepreneurial Problem
田阳 encountered information thermodynamics in statistical physics: information is not merely a computer science concept but can also be treated as a fundamental physical quantity. The core relationship revealed by Landauer’s principle is that erasing information requires a definite thermodynamic cost.
This led him to view the motherboard as a physical system carrying out information-processing tasks: voltage, frequency, power consumption and heat are controllable variables. The entrepreneurial problem therefore becomes how to schedule them so the computer achieves optimal energy efficiency under real-world constraints.
银普’s entire technology stack ultimately aims to let a computer “using a given amount of energy and generating a given amount of heat” complete as much of the information it should compute as possible. 李丰 translates this as releasing the compute resources of each motherboard while reducing wasted energy.
5. Agents Put the Entire Motherboard Back Into the Value-Allocation Equation
When 田阳 started the company in 2024, he was betting that “AI cannot stay in the infrastructure era forever, spending money without producing output.” The move from training to inference, then from inference to applications and ultimately Agents, was a path he believed would happen but could not yet prove.
Training relies mainly on GPUs, with the CPU playing more of a supporting role. Inference raises the CPU’s importance, while Agents require frequent data movement among the SSD, memory, CPU and GPU. As a result, every compute- and storage-related component on the motherboard becomes a performance variable again.
李丰 cites Nvidia’s acquisition of Groq, unified training-inference architectures, TPU, Groq’s LPU, RISC-V and various edge chips as examples: chip momentum is moving from heterogeneous cloud computing along the inference chain toward the endpoint.
6. Below 7nm, Thermal Conditions Become a Performance Variable Rather Than an Engineering Detail
The second entrepreneurial bet comes from physical scale: as chips move to 7nm and smaller nodes, they are gradually entering the mesoscopic regime. Energy fluctuations from non-equilibrium thermal disturbances may become comparable to the scale of the devices themselves, increasing the impact of surrounding heat on chip operation.
田阳’s conclusion is that controlling the chip’s external environment—power delivery, cooling and the condition of the motherboard as a whole—can create more favorable conditions for the processor and bring it closer to peak performance.
Demand is showing a corresponding gap: servers continue to deploy new technology, while the laptop that consumers actually use has seen no architectural innovation for “nearly 20 years.” The old market has stagnated as demand changes, meaning any effective attempt “may have a chance to get on board.”
7. An In-House Platform Is the Commercial Answer to Educating a Stagnant Market
New technology entering a market that has been stagnant for nearly 20 years must first answer not whether it works, but “how to make people realize that this thing is actually useful.” 田阳 believes technology companies generally need to complete the technical and commercial demo on their own platform first.
He originally expected the Agent-terminal opportunity to emerge in 2-3 years, but it arrived early in 2025. The acceleration pushed 银普 away from the classic semiconductor-company route and toward using its own platform and complete systems to prove its capabilities.
The shift does not mean abandoning chips: 银普 still has chip R&D and product plans, but it will first establish demand through platforms that can be sold directly, then gradually move mature capabilities into control chips and coprocessors.
8. AI Terminals Will Split Into “High-Interaction Handhelds” and “Quiet Desktops”
田阳 believes consumers naturally want to “pay once and get as many functions as possible,” so AI computing terminals may not remain separate from traditional computers. The 2 are more likely to be integrated into the same device, with products taking 1 large and 1 small form.
A Pocket device must have its own complete interaction capabilities. If it is merely an external module for a phone, like a power bank, it becomes “a terminal dependent on another terminal”; once a phone maker puts the same compute back into the phone, the add-on loses its reason to exist.
A fixed endpoint does not need to compete for the interaction layer and can use the local network to connect distributed AIoT devices. It handles powerful functions while tolerating inconvenience in keyboards and screens; the mobile endpoint leans more toward entertainment and portable interaction, leaving the 2 with a long-term division of labor similar to the phone and PC.
9. The Home Compute Box Will Ultimately Look More Like a Router Than Another PC
A desktop device can become a hub that centrally supports household robots and AI functions—a “router for compute.” Like water, electricity, coal or a wireless router, it would simply be there by default. Users would care less about its appearance than about stability, strong compute, compact size, low power consumption and price.
李丰 compares this with the history of home internet access: during the dial-up era, a “modem” was separated from the telephone line; as fiber, Wi-Fi and connected devices proliferated, it evolved into an independent router. As household demand for compute, storage and communications continues to expand, compute may undergo the same device separation.
10. 银普 Will Sell System Integration This Year and Move the Capabilities Into Chips Next Year
银普 will first release 1 desktop device, with plans for 4 more desktops and 1 handheld that can fit in a pocket within the year. The desktops can connect to a monitor and keyboard, but the product design will gradually reduce the need for them.
The immediate product experience this year is stronger performance and greater stability. In June, 银普 will launch an AI SSD at Compute X alongside several international giants, allowing the same computer to run AI far larger than before. “The same parts assembled into something different” is the current public selling point (“同样的零件拼出来不一样的东西”).
Once its in-house chips can run independently, next year’s positioning will shift toward computer control chips and coprocessors: users will swap AI cards like game cartridges to accelerate different specialized models. 田阳 says 银普 “absolutely will not make processor chips,” allowing it to work with all major processor vendors.
11. The 2 Technical Lines Aim Separately to Exhaust Chip Performance and Clear I/O Bottlenecks
The first product category controls motherboard power delivery, cooling and related information to keep the processing chip in a more favorable operating environment, releasing the limits of general-purpose compute, floating-point operations, gaming and graphics rendering.
The second targets the I/O bottleneck in Agents: data resides on the SSD, memory and VRAM; when VRAM is insufficient, the system must decide whether to discard the data and recompute it or download it over the PCIe bus for storage on the SSD. That transfer path is precisely where 银普 is most focused.
Where commercial coordination is possible, 银普 will work with storage vendors to optimize SSD ingress and egress speeds; for the PCIe bottleneck, which is harder to move in the short term, it plans to replace the path directly with its own accelerator card or chip. “Wherever the AI data needs to flow, the product follows it along the bus” (“AI的数据需要流动到哪”，产品就沿总线追到哪).
12. “Overclocking-Like” Optimization Works Because DDR and HBM Are Still Priced Like Gold
李丰 compares 银普’s resource scheduling with early PC overclocking: when consumers could not afford the most powerful CPU, they found ways to run an existing CPU at a higher frequency; today, the goal is to extract stronger AI performance from an affordable combination of CPU, memory and SSD.
田阳 agrees with the analogy but gives its failure condition: if DDR eventually becomes as cheap as rice and an entire computer can be filled with HBM without Flash, today’s scheduling work will look thin. For now, however, DDR and HBM remain “as expensive as gold,” making capacity constraints and scheduling unavoidable.
李丰 maps this path onto the capital markets: storage-related companies have gained “more than 5x” over the past 6 months, and he expects some consumer electronics products to become more expensive starting in March as storage prices rise. His explanation is that as AI moves from training to inference and applications, demand for storage and high-speed communications is amplified at every layer.
13. Lobster Devices Merely Preinstall an App; the Next Stage Is to Preinstall Capabilities
田阳 reduces today’s Lobster device to a mini PC that has existed for years: from Intel NUC to Asus’s NUC and then domestic manufacturers such as GMK, the hardware is not new. The only change is that Open Cloud or a variant comes preinstalled.
The commercial essence of this model is downloading a preconfigured program onto traditional hardware. The real transition is from “putting an app inside” to preinstalling Memory, retrieval and system scheduling capabilities that can serve different Agents.
Hardware may continue to resemble a Mac mini for the next 5 or even 10 years because physical products iterate slowly. The competition is not about creating an unfamiliar enclosure, but about using the existing supply chain to build an Agent computer better suited to running, storing and continuously understanding its user.
14. Memory May Be the Main Battleground for Edge Agents Over the Next 5 Years
田阳’s clear observation for investors is to watch companies entering Memory from either software or hardware. Phone data is not all pushed to the cloud because personalization matters; an Agent seeking personalization likewise needs to store continuously accumulating data locally.
The most valuable software-hardware Memory would allow an Agent to “understand you better and better” over time. Retrieval, OS-kernel changes, new operating systems, motherboard design and storage innovation will grow around it; 田阳 calls this “possibly the biggest development direction of the next 5 years.”
Once Memory and Agent execution become infrastructure, they will no longer be pure software apps but system-default capabilities, much like graphics rendering. Whether vendors call them AIOS or AIPC is irrelevant; what matters is that any Agent running on the software-hardware stack becomes better.
15. The Edge-Cloud Debate Will First Become Task Segmentation, Not a Fight to the Death
田阳 first acknowledges the cost constraint: every infrastructure technology is expensive when first invented, and AI is still at a stage where only some people can afford to use it. If compute eventually becomes cheap enough, the edge-cloud debate may disappear naturally, but for now users will inevitably segment tasks by importance.
More private, frequent, important but uncomplicated daily tasks are suited to the edge because they are used every day without requiring continuous payments. Programming and complex task orchestration will continue to purchase cloud capacity. He expects this division to “possibly remain in place for the next 10 years.”
The local AI advertised on many current PCs can run only 8B or 16B models. 银普 hopes to support 235B, 397B and larger models at hardware priced roughly like an iPhone. 田阳 emphasizes that this first depends on better integration; it does not mean a new processor has already appeared.
16. 银普 Is Trading Generality for ASIC Economics and Larger Model Support
AI coprocessors may move toward 2 extremes. One prioritizes generality, can accelerate different AI workloads and is suitable for integration with powerful processor vendors, but at a limited cost can typically support only smaller models and may ultimately be absorbed into the main processor.
The other hardens the software to an extreme, gives up compatibility and supports larger models at the same cost. It is better suited to a plug-in capability layer on the motherboard, replaceable through M.2, U.2 and similar interfaces like a game card, and can also enter embedded devices where cost matters.
Given the realities of China’s industrial chain, 银普 has chosen the second path: dedicated ASICs, extreme hardening and lower costs. The trade-off is that it is difficult to become part of a general-purpose processor; the payoff is the ability to push performance further in clearly defined models and scenarios.
17. Legacy PC Brands Are Hard to Beat Head-On, but New Categories Still Offer a Reset
田阳 describes the market structure this way: Lenovo has long maintained the world’s highest share; Asus represents a younger, faster-expanding force; Acer, MSI and others are also changing. At the same time, new OEMs and ODMs continue to emerge and build their own brands, while Lenovo’s share has already declined to some degree.
He does not believe a new company can replace Lenovo in traditional laptops; its supply-chain management has a “crushing advantage.” If a newcomer continues with the brand-plus-assembly-factory model, it cannot compete with Lenovo and Asus on the same field.
The opportunity lies in capabilities legacy vendors lack: OS development, hardware-software scheduling and chips. 田阳 uses Apple as an example: a new brand does not need to become the world’s first computer company. If its new product category is important enough, it can establish a position outside the existing giants.
18. The Case for a Handheld AI Terminal Is That Users Still Do Not Trust Agents With Their Phones
The fundamental reason a handheld device would remain separate from the phone is not compute but trust. Today’s AI may be intelligent, but not intelligent enough for users to hand over full control of a phone containing extensive private data or a laptop holding work materials.
A new device cannot become a second phone, nor can it remain a glass slab for users to tap. 银普 wants to optimize it around Agent interaction; even if non-Agent tasks are “extremely inconvenient,” sufficiently natural Agent use would give the device independent value.
The technology stack comes from nearly 10 years of accumulated experience in the open-source Linux handheld industry, including Clockwork Pi. 田阳 describes the planned product as an “AI-native Linux open-source handheld,” leading 李丰 to conclude that it will initially be professional, niche and geared toward enthusiasts.
19. Hardware Startups Should Target Professional Users Before Chasing General C
李丰’s personal observation after attending a Lobster festival is that “99% of General C” users did not know what the Lobster could do before installation and still did not know afterward. 田阳 adds that the user’s core demand throughout the process was simply: “I want to download a little Lobster.”
Software can potentially work economically even with a 1-yuan purchase, but hardware carries hard costs from the moment it starts up. It therefore needs users with purchasing power and a stable market, not a huge General C population with random purchasing behavior.
田阳 frames the comparison as: “Are there more gamers in the world, or more people who can write AI code?” Gamers are more numerous, but gaming makes the smallest contribution in Nvidia’s earnings. Professional AI and graphics users will upgrade when performance improves by a generation, so the logical sequence is professional C-side or B-side first, and General C last.
20. China’s “Lobster Craze” Looks More Like a Projection of the Missing Mass-Market AI Application
田阳 believes China has been in the AI era for 2-3 years but still lacks a mass-market AI application comparable to ChatGPT or Cloud. Consumers and cloud vendors alike see Open Cloud as a potential answer, while the latter also hopes to break through the limitation of the cloud serving only as a home for chatbots.
Overseas markets already have a large number of tools, including Code X, Open Code and Cloud Code. Open Cloud may create a shock, but it is unlikely to generate a sustained nationwide craze in the Chinese style. 田阳 views the declining number of domestic Lobster festivals as a direct signal that the heat is fading.
Hardware needs at least 1 month to “make a small machine,” while an app’s fad may last only 1 month. So “hardware builders are not well suited to chase a fad; they need to anticipate one” (“做硬件的人不太适合去赶风潮，而要去预测一个风潮”). What the Lobster leaves behind is a demand signal, not necessarily the final product form.
Chinese users are more accustomed to TikTok-style super-entrances and low-agency interaction; some people born after 2005 do not even know how to unzip a file. That creates a market for RMB39.9 “licensed Python,” RMB100 Lobster installation, RMB200 uninstall and RMB300 data-clearing services. Truly mass-market AI may have to do everything for the user.
21. China’s B-Side Will Monetize First; Professional C-Side Adoption Is More Likely Overseas
田阳 still views the Lobster as a potential To C inflection point, but not as evidence that similar products will become mainstream. It is better understood as a public expression of demand for a “TikTok-level AI application.” Open Cloud is too open and may not fit the low-barrier, low-agency format preferred by China’s mass market.
For hardware sales over the next 1-1.5 years, he observes that domestic B-side companies are highly capable of adopting AI and may sell no less than overseas markets. Many seemingly traditional companies with already substantial revenue are also willing to purchase.
Professional C-side demand remains small in China, so C-side volume for vendors other than 银普 is more likely to come from overseas. 李丰 cites DeepSeek all-in-one machines, which sold “extremely well” to large and midsize domestic B-side customers from April through December last year, as evidence that domestic revenue opportunities do exist—they simply arrive first through enterprises rather than mass consumers.