As large language model parameters and data center power consumption continue to surge, the AI chip arena is undergoing a profound shift from “general-purpose computing” to “hardware-software co-design.” On one front, Google is reportedly developing a new server chip codenamed “Frozen v2,” which plans to permanently etch the architecture of its Gemini model into silicon to achieve exponential gains in energy efficiency. On another front, fundamental materials science has achieved a critical breakthrough: scientists have successfully synthesized a metallic crystal with theoretical thermal conductivity nearly three times that of copper—θ-phase tantalum nitride (θ-TaN)—offering a novel approach to solving heat dissipation challenges in future high-power chips.
Google’s “Frozen” project and the academic creation of θ-TaN single crystals may seem to belong to the distant realms of chip design and materials science, but they both point to the core contradiction of the post-Moore’s Law computing race: as traditional transistor scaling and general-purpose architecture optimization gradually hit physical limits, redefining everything from the chip’s underlying architecture to its physical heat-dissipation materials is becoming the key to breaking the computing power bottleneck.
Google’s Radical Path: Turning Chips into “Dedicated Machines” for Models
According to sources cited by The Information, Google is secretly developing a server chip informally named “Frozen v2.” The name derives from its core concept—permanently etching parts of the Gemini AI model’s computational logic, inference processes, and even decision-making mechanisms onto silicon.
This is fundamentally different from Google’s previously launched Tensor Processing Units (TPUs). Although TPUs have sought to reduce reliance on Nvidia’s general-purpose GPUs since their debut in 2016, they remain essentially general-purpose AI accelerators that require dynamic scheduling based on model structure during operation. The Frozen project, by contrast, adopts an approach closer to Application-Specific Integrated Circuits (ASICs), attempting to make the hardware “understand” the operational logic of a specific model, thereby drastically reducing runtime scheduling overhead.
Sources indicate that Google plans to deploy the chip as early as 2028. According to predictions from researchers involved, once released, the Frozen chip’s efficiency—measured by the number of tokens processed per unit of power—could be 6 to 10 times that of Google’s current latest AI chips. If achieved, this goal would significantly alleviate the internal resource strain caused by a severe shortage of AI computing capacity—a shortage that has reportedly forced Google Cloud to decline collaboration requests from some external customers.
However, this extreme hardware-software co-design carries significant risks. Chip design cycles typically span several years, while AI model iteration speeds far outpace the semiconductor industry’s rhythm. If the Gemini model architecture undergoes major changes in the future, the highly rigid Frozen chip could face a sharp performance decline or even failure. Sources revealed that Google is attempting to find a compromise between hardware rigidity and model updates, including allowing chip modifications to update parameters such as model weights.
In fact, “writing” a model into a chip is not unique to Google. Startup Taalas has already launched its HC1 chip, specifically optimized for the Llama 3.1 8B model. Built on TSMC’s N6 process, the chip spans 815 square millimeters and integrates 53 billion transistors. According to Taalas, the HC1 card can generate 17,000 output tokens per second, 73 times that of Nvidia’s H200 card, while consuming only one-tenth the power. Another company founded by Harvard dropouts, Etched, is dedicated to building inference chips specifically designed to run Transformer models; its valuation reportedly soared to approximately $20 billion in its latest funding round.
Materials Science Breakthrough: Creating a “Dream Metal” with 3x Copper’s Thermal Conductivity
While chip design pursues extreme efficiency, the physical bottleneck of heat dissipation also urgently needs to be overcome. Currently, copper is the workhorse material for chip heat spreaders, heat pipes, and packaging substrates, with a thermal conductivity of approximately 400 watts per meter-kelvin (W/m·K)—a record that has stood for over a century.
Earlier this year, a team led by Professor Yongjie Hu at the UCLA Samueli School of Engineering published a paper in Science, announcing the first successful synthesis of a single crystal of θ-phase tantalum nitride (θ-TaN), a material that previously existed only in theory. Experimental measurements showed that the material’s thermal conductivity at room temperature reaches approximately 1,105 W/m·K along the a-axis and about 928 W/m·K along the c-axis, nearly three times that of copper.
Since German chemist Brauer discovered the θ-phase in the 1950s, the scientific community predicted it possessed intrinsic thermal conductivity far exceeding conventional metals, but high-quality crystals remained elusive due to synthesis difficulties. Hu’s team, collaborating with Tohoku University in Japan, employed a flux-assisted metathesis reaction technique, using metallic sodium as both a reducing agent and flux to bypass traditional high-pressure synthesis routes, ultimately obtaining θ-TaN single crystals ranging in size from 10 to 100 micrometers.
θ-TaN’s exceptional thermal conductivity stems from its unique microscopic mechanisms. In conventional metals, heat conduction is primarily carried by free electrons, but electron-phonon (lattice vibration quantum) interactions consume transport capacity. θ-TaN is special because its phonon-phonon interactions are extremely weak, and electrons barely exchange energy with phonons—both scattering channels are simultaneously suppressed. Furthermore, tantalum is nearly monoisotopic (¹⁸¹Ta abundance as high as 99.988%), naturally weakening scattering caused by atomic mass differences, allowing heat to transmit through the lattice in a rare low-damping state.
It is important to note that while θ-TaN breaks the thermal conductivity record for metallic materials, non-metallic materials with even better thermal performance exist. Natural diamond can achieve thermal conductivity of 1,500 to 2,200 W/m·K, and ultra-pure synthetic diamond can reach 3,000 W/m·K. However, diamond suffers from drawbacks including slow growth rates, high manufacturing costs, electrical insulation, and extreme difficulty in machining. In contrast, θ-TaN’s metallic properties make it naturally compatible with existing chip packaging processes—tantalum nitride thin films are already commonly used as diffusion barrier layers in integrated circuit manufacturing. If deposited in thin-film form, it could seamlessly integrate into existing production lines and potentially even serve as an integrated interconnect layer combining both electrical and thermal conduction functions in high-power devices.
However, Hu’s team also noted that limited by production capacity and preparation difficulty, θ-TaN is unlikely to replace bulk heat sink materials like copper and aluminum in the short term. A more realistic application form is as a high-thermal-conductivity intermediate layer between the chip and the heat sink. Additionally, the θ-phase is metastable at room temperature and pressure, and its long-term thermal stability still requires further evaluation by industry.
The Computing Power Race Enters “Deep Waters”
From Google’s Frozen to θ-TaN, the AI chip industry is revealing a clear trend: the era of simply competing on peak computing power is passing. The holistic synergy of models, software, chips, and materials is becoming the core of the next round of competition.
Google is attempting to build a tightly closed loop from the Gemini model to TPUs and then to custom hardware through the Frozen chip, while the birth of θ-TaN reminds the industry that fundamental material innovations could physically redefine the ceiling of chip performance. As the research paper suggests, beyond θ-TaN, transition metal compounds such as tungsten nitride and molybdenum phosphide may also possess similar high thermal conductivity potential, opening the door for systematic screening of new heat-dissipation materials across the industry.
In the future, the competitive landscape for AI chips may no longer be a simple substitution between GPUs and ASICs, but rather a new division of labor: general-purpose GPUs will continue to support rapid model innovation, dedicated chips will enhance the deployment efficiency of mature models, and breakthroughs in new materials will provide the physical foundation for all of this. Models define chips, materials support computing power, and the boundary between software and hardware is further dissolving.