On November 12, Tongxin Intelligent announced the completion of its seed and angel rounds totaling hundreds of millions of CNY. Investors include Tsinghua Alumni Seed Fund, Beijing Institute for General Artificial Intelligence (BIGAI), Songlin Technology, Pingzhi Information, Dark Matter Intelligence, and Yuzi Zhangquan.
Founded in April 2026, the company was incubated at BIGAI and initiated by Professor Zhu Songchun. Founder Wang Haiyang is a founding member of Loongson CPU, former lead architect for Hygon's fourth-generation chips, and co-founder of a GPU unicorn. Their goal is not merely to launch another AI chip, but to answer a fundamental architectural question: If AGI computational demands differ from today's Transformer inference, what should the chip look like? The following analysis explores this from five dimensions.
| Where Are the Inherent Shortcomings of GPUs?
Tongxin Intelligent's chip comprises three parts: Vector, Tensor, and AGI Core. The first two handle vector and tensor computations, compatible with mainstream models like CNNs and Transformers. The AGI Core is specifically designed for AGI algorithms, focusing on accelerating tasks such as logical reasoning, causal reasoning, value reasoning, and probabilistic computation.
GPUs excel at large-scale parallel matrix multiplication, offering high efficiency when data is regular and computation-intensive. However, for tasks like graph traversal, search, and planning, data access is often irregular with numerous branching decisions, causing many compute units to idle. Logical reasoning and causal deduction essentially involve searching and state comparison within graph structures, where GPUs do not hold an advantage.
In other words, the AGI Core aims not just to "compute faster," but to enable chips to better handle these irregular reasoning tasks. Its design focus may lie less in peak computing power and more in on-chip memory, data scheduling, and sparse computing capabilities. Crucially, it must determine whether it can natively treat "graphs" as first-class citizens rather than decomposing graph computations into matrix multiplications.
| "Hybrid" Is More Important Than "Processing-in-Memory"
Tongxin Intelligent also prioritizes hybrid processing-in-memory (PIM). Founder Wang Haiyang is considered one of the early proponents of this approach.
The key lies in the word "hybrid." Pure analog PIM sounds powerful but suffers from precision issues due to process variations, low yield control, and limited capacity. Pure digital PIM is stable but incurs significant area overhead, making it impractical for covering large model weights. "Hybrid" implies layering by task: high-bandwidth, small-capacity scenarios use PIM; large-capacity weights utilize near-memory or conventional DRAM. This tiered strategy optimizes overall energy efficiency.
This represents a pragmatic engineering stance: the memory wall cannot be broken by a single point technology; the storage hierarchy requires holistic reconstruction. For architects, this argument is more persuasive than claims like "our CIM improves energy efficiency by X times."
| The Cost of Using One Architecture for Cloud, Edge, and Device
The first-generation product plan utilizes Chiplet technology, employing a unified architecture and software stack to cover cloud, edge, and device ends through different chiplet combinations. The cloud segment targets NVIDIA's B300 series, while the device end targets the Thor series.
B300-level cloud performance implies large dies, high power consumption, HBM, and advanced packaging (such as CoWoS). Thor-level device performance implies low power consumption and functional safety. Covering both ends with a single architecture requires extreme scalability—Chiplets are virtually the only solution, but the costs are clear: UCIe die-to-die interconnect interfaces, advanced packaging capacity and cost, and yield/test complexity under multi-chiplet configurations. Domestic advanced packaging capacity is already tight, an external constraint that design alone cannot solve.
The greater challenge is actually the "unified software stack." While hardware can scale, achieving unified scheduling of heterogeneous units via compilers and runtimes is the true barrier, and where many new architecture companies have failed.
| The First Generation Product Is Essentially a "Seed"
The official statement that "the initial chip will retain the computing capability required for mainstream models, and the area proportion of the AGI Core will expand in subsequent iterations as adaptation and verification progress" translates to engineering terms as follows: The first-generation chip is more like "a seed of AGI Core embedded within a general-purpose AI inference chip." A complete AGI-native architecture will gradually emerge in the second and third generations. This does not contradict "targeting NVIDIA's B300 series for the cloud"—the latter primarily refers to cloud inference computing power and energy efficiency goals, not directly benchmarking against B300's AGI capabilities.
Regarding the timeline, the company completed architecture design and project initiation within three months of establishment and has now entered the chip design phase, planning to start FPGA prototype verification in early December 2026. Based on the company's estimate of "at least 20 months from initiation to mass production," tape-out is expected in 2027, with scale mass production approaching 2028.
However, a single advanced-process tape-out for a large chip costs tens of millions to hundreds of millions of CNY. While the current funding covers architecture and the first tape-out round, future phases will inevitably depend on new financing. Capital availability is a prerequisite variable for whether this path can succeed.
| Historical Lessons from Algorithm-Defined Chips
Past AI chip ventures have faced pitfalls: hardware over-optimized for specific models or operators quickly becomes obsolete as models iterate, leading to poor software stack coverage and rapid declines in chip utilization. A relatively successful example is TPU, which benefits from Google's own large-scale workloads driving development and a closed loop across architecture, compiler, and cluster deployment, enabling continuous iteration.
Tongxin's structure sits between these extremes: Professor Zhu Songchun's team provides the CUV framework—integrating cognitive architecture, capability systems, and value systems into a unified space to address how agents understand environments, invoke skills, and make trade-offs based on goals. Wang Haiyang and Shi Feng are responsible for engineering implementation and AGI Core architecture definition, respectively. The core team averages over 20 years of industry experience, with key members from Hygon, Baidu, ByteDance, and Alibaba, having fully experienced the entire process of over ten large-scale chips.
This configuration is structurally sound: theory provides direction, and mass production experience provides constraints. The critical next step is accurately translating CUV, a cognitive science framework, into implementable hardware specifications such as area, bandwidth, instruction sets, and memory access models. This translation process itself is what the project needs to continuously validate.
| Conclusion
The divergence in domestic AI chip routes is clear: one path focuses on making general-purpose GPUs closer to international mainstream standards, while the other develops Domain-Specific Architectures (DSAs) around specific workloads. Tongxin Intelligent has chosen a more aggressive variant of the latter—not building an accelerator for a specific model, but hardware for a class of reasoning methods.
The value of this path lies not in whether it can benchmark against B300, but in answering another question posed by the AGI Core: When computation shifts from "fitting" to "deduction," should hardware remain as it is today? This question deserves exploration. Whether it succeeds will be determined by the 2027 tape-out and the first successful execution of real-world workloads, which will be more convincing than any architecture diagram.
Disclaimer: This article is based on public information available as of November 12, 2026, intended for technical exchange and reference only, and does not constitute investment advice or endorsement of any project's prospects. Content regarding company plans, product positioning, and timelines is subject to official disclosures. The author's analysis and inferences are not guaranteed to be accurate; readers assume all risks for decisions made based on this text.