Interview with Tian Yang of Yinpu: After the Crayfish Craze Fades, What's Next for Edge AI Hardware?
What will the next AI device in our homes actually look like?
The map of AI compute is being redrawn. NVIDIA bought Groq for $20 billion and unveiled a "training-inference unified" solution at its latest GTC; on the motherboard, everything from CPU and GPU to SSD and memory is being revalued — AI is no longer a GPU solo act.
Meanwhile, another thread appears frozen in place. Today's computers, Mac minis, and various home AI devices dubbed "lobster machines" — though all "edge hardware" — still have to phone home to the cloud when running AI. PC architecture hasn't seen real transformation in ages, while cloud innovation sprints ahead.
This gap is forcing a question: What will the next AI terminal in the home actually look like?
Yinpu aims to answer this. Its founder, Yang Tian, carries跨界 labels — Tsinghua University's Xinghuo Program, PhD in computational neuroscience, statistical physics, Huawei's 2012 Lab, a return to academia for a faculty position, then resigning to start a company.
Not long ago, Feng Li, founding partner of FreeS Fund, sat down with Tian for a deep conversation about Yinpu's "home AI hardware" — what exactly is it? How does it differ from today's computers, Mac minis, and various "lobster machines"? And why does a home need another such device at this moment?
Their main topics included:
-
Why "optimal control of compute and storage across the entire motherboard" became a viable entrepreneurial direction
-
As AI shifts from training to agents, is every component on the motherboard being revalued?
-
Chips entering the "mesoscopic scale" after several nanometers — what does this mean for edge hardware?
-
We experienced CPU overclocking in the 1990s; is edge AI hardware entering a "parts-bin" era?
-
Will the desktop become the next "compute router" — a default household utility like water, electricity, and gas?
-
After the lobster wave recedes, might hardware makers be the ones who most need to predict the next trend?
We've edited portions of the conversation below. For the full dialogue, search for "高能量" (High Energy) on Xiaoyuzhou App and Apple Podcast.
Reader Giveaway:
Have you bought home AI hardware (Mac mini, lobster machine, etc.)? What do you use it for most? Share your thoughts in the comments. By 5:00 PM, June 5, 2026, the 2 most thoughtful commenters will each receive a book recommended by Feng Shu.
01 When "Information" Is Also a Form of Energy
Feng Li: Your background is remarkably跨界 — Tsinghua Xinghuo, computational neuroscience PhD, statistical physics, Huawei 2012 Lab, faculty position, then entrepreneurship. What's the through-line connecting these twists?
Yang Tian: I just took that viral SBTI (spoof MBTI). My result was GOGO — "the Wayfarer" — values rationality but often follows curiosity. Pretty accurate for me.
Feng Li: What exactly does computational neuroscience study?
Yang Tian: The field has two camps. One is computer science backgrounds doing bio-inspired simulation. The other is physics backgrounds — that's me. We use physics methods to study how the brain encodes external information, the dynamics of neural activity, and what states those dynamics imply.
Feng Li: Today, whether VLA or VLM, everyone faces this problem — can we build different architectures on top of video or language models to solve physical-world interaction? This is the "world model" everyone's discussing. But underneath is a fundamental question: how do animals understand physical quantities and laws? Monkeys instinctively know a branch won't hold them. Yet "intuitive physics" sometimes fails — people intuitively think two iron balls won't land simultaneously. How do you see this?
Yang Tian: One reason I left computational neuroscience was that I don't consider the human brain perfect. Robotics' obsession with biomimicry has progressed slowly — we must admit humans and robots are different. People habitually borrow from themselves because everything we created used to be weaker than us. But AI may be something stronger than us, and subconsciously, that scares people.
So I believe robots can completely abandon biological imitation for more general, open approaches — looking not just at neuroscience but at self-organizing systems in physics and chemistry: molecular robots, self-assembling robots, swarm behaviors... These may be more worth learning from than simply copying the brain.
02 AI Moves to Applications, the Entire Motherboard Lights Up
Feng Li: What connected this path from academia to Huawei, back to academia, then to entrepreneurship?
Yang Tian: Landauer's principle tells us: erasing one bit of information consumes a definite thermodynamic quantity. This perspective made me realize computers themselves can be excellent physics research objects.
A motherboard has many controllable physical parameters — voltage, frequency, power consumption. These conversion relationships are what I want to control, enabling the entire computer to achieve higher energy efficiency.
Feng Li: Why choose to start a company a year ago?
Yang Tian: Several reasons. First, simple economics — AI can't stay in infrastructure mode forever. It will inevitably move to inference and applications. Once it does, CPU and memory makers become increasingly important.
Second, observations about hardware itself. All chips are pushing below 7nm. Physically, as chip scales shrink, they move from macroscopic to "mesoscopic" scale — where substantial non-equilibrium thermal fluctuations exist. In other words, a chip on a motherboard is increasingly affected by nearby heat and other physical quantities. If I can control the motherboard environment, I can create better conditions for it.
From the consumer trend perspective, consumers encounter products, not servers. The two sides are diverging — server-side has ever more new technology, while C-side hasn't seen genuine innovation in ages. So I believe the industry is at a node where demand-side transformation is needed but no clear direction has emerged. Any attempt now has a chance to get on board.
Feng Li: For example, NVIDIA bought Groq for $20 billion and unveiled a "training-inference unified" solution at the latest GTC. From the investment side, we've observed AI heterogeneous chip heat moving from cloud heterogeneous chips (like TPU), to inference chips between cloud and edge (like Xingyun and Yuanchuan Micro that we've invested in), to now edge chips just heating up.
These three chip types differ at the hardware level in their scheduling ratios of compute, communication, and storage resources. This difference becomes more pronounced as large models move to applications. But today this difference still manifests mainly on the server side; edge chips are only just beginning to heat up.
Yang Tian: Yes. When we first promoted new technology, we faced a classic problem — how do tech companies educate the market? We need to educate a market that's stagnated for nearly 20 years without technological transformation. Eventually you hit this question: how do you make people realize this thing is actually useful?
Under this trend, all tech companies must inevitably turn to doing technical or commercial demos on their own platforms first. Yinpu is no exception — building our own platform is extremely important.
03 From Modems to Wireless Routers: What Comes Next for Edge Hardware?
Feng Li: The definition of "AI-native hardware" today is actually quite chaotic. From Mac minis to various lobster machines, forms vary wildly. How do you view these concepts?
Yang Tian: Future consumer terminals bearing AI compute functions may not naturally separate from traditional computers — consumer psychology is "spend one price, get as many functions as possible." Form-wise, small devices go with you; large devices stay home — these are two categories.
The core of mobile is "interaction." What survives must have complete interaction plus compute power. A device without interaction, just providing compute — like a "power bank" form hanging off a phone — actually creates a terminal dependent on another terminal, and will soon be swallowed when that host terminal stuffs the compute chip inside itself. Desktop doesn't need strong interaction; it will become a "compute router" — a default household resource like water, electricity, and gas.
Feng Li: Like how modems evolved into wireless routers in the 1990s. What's driving the next evolution is "everyone needs more compute, storage, and data communication." By the same logic, it will emerge as an independent home device for scheduling. So where does your upcoming product sit in this picture?
Yang Tian: Most Yinpu products will be "high compute, low profile" — 4 desktop models and 1 pocket handheld this year. Yinpu's capability is making better products with higher performance from the same hardware. We want to build this era's AI-native computer.
Feng Li: Does desktop still need a monitor and keyboard?
Yang Tian: We'll always emphasize that it can have them, but increasingly de-emphasize that it needs them.
Feng Li: For consumers, what's the biggest experience difference?
Yang Tian: Stronger performance, greater stability. Specifically, this June Yinpu will release AI SSD — an SSD with special acceleration partitions and customized control chips, supporting both traditional KV cache offload and MOE model hot-cold expert prefetching. You'll soon see this product at major computer makers' launches and tech exhibitions, where we'll demonstrate how this AI SSD enables existing computer platforms to run larger-scale MOE models. For example, we'll show how a regular consumer GPU with 12GB VRAM and regular CPU with 16GB RAM, paired with this AI SSD, can run 35B or 80B MOE model-based coding agents at 35 tokens/s inference speed. Separately, Yinpu will release a near-memory computing coprocessor chip (FPGA version) by year-end, single-chip supporting high-speed inference for 397B to 480B scale MOE models (decode speed stabilizing at 300-500 tokens/s in common context scenarios), heavily emphasizing efficient coordination between compute units and DDR/Flash.
Feng Li: How do you make existing hardware combinations, plus Yinpu, achieve beyond-original capabilities? And why can you do this?
Yang Tian: Product planning splits into two categories. First is motherboard-level DVFS control (comprehensive regulation of performance, power, and thermal), pushing processing chips to极限 performance — generally used on Yinpu's own computers. Second is eliminating IO bottlenecks on the motherboard, serving AI inference and agent applications. The AI SSD and coprocessor chip I mentioned belong here; besides using on our own products, we'll supply to various computer manufacturers.
Feng Li: So it's adding a smarter scheduling layer among existing hardware resources? Letting each component be fully utilized, releasing existing resources without affecting cross-vendor hardware compatibility and resource scheduling conflicts?
Yang Tian: Yes. Taking AI agent applications: Yinpu's future AI-optimized products can be summarized this way — when we interact with agents, masses of data sit in SSD, heaps in VRAM. When VRAM fills, you must decide: discard entirely and recompute later, or unwilling to discard and stuff back into SSD. This always goes through something called the PCIe bus — loading data between VRAM, memory, and SSD. This is Yinpu's core concern. First, we'll collaborate with some storage companies to optimize SSD IO speeds, like the June AI SSD release. Meanwhile, the PCIe bus has many segments where short-term technical solutions are hard to push commercially. For these IO bottlenecks, we hope to directly replace with new products, like next year's "plug-in game card"-style AI inference accelerator. We're tracing the entire PCIe bus, seeing where AI data needs to flow, whether optimization is commercially viable under existing conditions. If yes, partner with manufacturers; if not, build our own chips to gradually eliminate that IO.
Feng Li: This resembles 1990s "overclocking." Personal computers were still expensive; ordinary users couldn't afford the best CPUs, so they found ways to push existing CPUs to maximum frequency. Overclocking lasted from early 1990s to early 2000s, fading only as purchasing power rose and chip prices fell.
Yang Tian: That's an excellent analogy. We live in an era where DDR and HBM cost like gold — insufficient capacity inevitably requires scheduling. For at least the next 1-2 years, such "compromise hardware solutions" will be mainstream. This year we'll openly acknowledge these are all existing parts, just that we have the capability to assemble something different. Only when Yinpu's own chips emerge will we emphasize essential differences.
Feng Li: You're doing a coprocessor chip next year. Coprocessors have two completely different historical paths. First, the GPS path — later merged into main chips. Second, the GPU path — NVIDIA twenty-plus years ago started as a gaming graphics coprocessor, but when parallel computing became crucial in the AI era, it moved to market center. Which path will you take?
Yang Tian: If people accept AI as a future foundational function, coprocessors will likely diverge too. One pursues generality — good compatibility but limited cost can only support smaller AI. The other abandons compatibility for extreme hardening, turning software directly into circuits, supporting larger-scale AI at the same cost. Given domestic industry chain realities, Yinpu takes the second path — abandoning generality, extreme hardening, dedicated ASIC.
Hardware Makers May Be the Ones Who Most Need to Predict Trends
Feng Li: Final topic — why was the "lobster" wave so much hotter in China than the US? Looking ahead, will lobsters keep spawning different versions to carry this forward, or slowly grow after the application layer stabilizes?
Yang Tian: China actually has no national-level AI application, yet we've been living in the AI era for 2-3 years. Subconsciously, people are waiting for a national-level app. When consumers see "lobsters," they think "maybe cloud providers will soon make this into a national-level app." Overseas doesn't need this — ChatGPT, Claude, Codex, already too many. So this heat expresses an expectation.
Feng Li: What lessons does this decline offer hardware makers?
Yang Tian: This exactly proves what I've always emphasized: hardware makers should predict trends, not chase them. Many lobster machines are traditional Mini PCs stuffed with Open LLM — "traditional hardware + pre-installed app" old thinking. Future should shift from "stuffing one app" to "stuffing a category of capability." Whoever can do better software-hardware integrated memory, making agents increasingly understand you, may be the biggest development direction for the next five years.
AI isn't yet smart enough for us to dare hand over phones containing massive personal information. We need devices where privacy matters less, more suitable for deep agent interaction. It can't be a phone, can't be a tablet — necessarily different in interaction from phones. We want it closer to pure agent interaction. An imperfect analogy: this machine can be understood as "AI-native Linux open-source handheld."
Feng Li: With this positioning, it's a professional, geek-oriented product.
Yang Tian: Yes. This isn't general consumer era. 99% of general consumers before downloading lobster didn't know what it could do, and still didn't after — core诉求 was just "gotta download a lobster." Hardware has hard costs; we need high-spending-power stable markets. Gamers are numerous, but as long as games run barely adequately, they won't upgrade. AI workers, graphics rendering professionals are different — they'll upgrade for any performance gain. Hardware makers should first secure professional users, then enter general consumer.
Feng Li: This actually responds to a more universal pattern — almost all consumer tech hardware started from professional users: cameras, drones, 3D printers... even mobile phones, earliest from maritime communications and other professional users. Only through dual下放 of technology and hardware capability did they slowly reach mass market. Final question: within 1 to 1.5 years, how do you预判 the market size for this category of home AI hardware terminals?
Yang Tian: Domestic B-side embraces AI extremely strongly; B-side sales may not lag overseas. Professional C-side in China is still relatively smaller, so almost all manufacturers' C-side volume is overseas-major — this is consumption habit difference. But domestic B-side has many seemingly traditional enterprises actually very strong in AI embrace, with substantial revenue themselves.
Feng Li: This aligns with our observations last year. Starting last April, DeepSeek all-in-one machines sold extremely well in domestic to-B, especially large-B and mid-B scenarios — good domestic industrial environment and digital infrastructure, plus sufficiently intense market competition where everyone fears others having what they don't, so to-B opportunity indeed emerged first.
Finally, wishing great success for Yinpu's June product launch and subsequent releases.
When Instruments and AI Rush Toward Each Other, How Should We Define Performance for the Next Era?