In recent days, Broadcom has seen a surge in its shares, driven largely by the growing importance of Application-Specific Integrated Circuits (ASICs). Conversely, the stock price of Nvidia has been on a downward trend since it peaked on November 21. This shift in the market dynamics poses intriguing questions: Is the resistance to Nvidia from downstream tech giants reaching a breaking point? What implications does this hold for China's AI industry?

With a staggering market capitalization of $1.03 trillion, Broadcom is now the second-largest semiconductor company, trailing only Nvidia, which boasts a market cap of $3.29 trillionThe intense focus on Broadcom and its ASIC offerings has even outshone Nvidia's upcoming production of the Blackwell series

So, what factors are contributing to this shift, and what can we learn from this situation?

Last Thursday, Broadcom shared its fourth-quarter and annual earnings report, highlighting robust growth across all important revenue metricsA standout point is that $12.2 billion of its projected semiconductor revenue for 2024 is derived from AI, reflecting a remarkable 220% year-over-year increaseThis dramatic growth is attributed mainly to their market-leading XPU AI chips and a comprehensive range of Ethernet networking products.

Furthermore, Broadcom revealed that it is currently collaborating with three very large clients on AI chip development

• Why Did Bitcoin Suddenly Plummet?
• Record Losses for the Federal Reserve!
• Huawei Chip Breakthrough Threatens ASML
• Domestic Production of Solid-State Lithium Batteries
• Bitcoin Could Reach $120,000

The company's market for AI chips is projected to reach between $15 billion and $20 billion next year.

Market analysts have begun to recognize that demand for custom AI chips and network devices from large cloud service providers is accelerating, posing a competitive threat to Nvidia, which has traditionally dominated the AI accelerator chip space.

This understanding directly contributed to a remarkable 38% surge in Broadcom's shares over two consecutive trading days, establishing it as the second publicly traded chip company to cross a trillion-dollar market capThe last firm to experience such a substantial rise was Nvidia itself, following its surprise performance report in May of the previous year.

The parallel narratives of both companies underscore a shared theme: the explosive demand for AI chips

However, Nvidia has emerged as the “dragon” in Broadcom's quest, once the uncontested champion of the AI chip realm.

Nvidia is widely recognized for its premium pricing and difficulty in procurementFor large cloud service providers, the cost of transitioning away from Nvidia’s offerings is not simply a matter of developing proprietary chips; it entails the optimization of total cost of ownership (TCO) associated with computing infrastructure and cloud service provisioningThis is a vital factor influencing competitive dynamics among cloud service providers.

As a result, while tech manufacturers competed fiercely for Nvidia chips, many began to develop their chips

Google pursued a partnership with Broadcom for its Tensor Processing Units (TPU), while Amazon entered a five-year agreement with MarvellMicrosoft has made significant strides over the past year in building its own CPU-GPU-DPU architecture.

Collectively, these tech giants agreed on a unified goal: the demand for cost-optimized custom silicon solutionsBroadcom's offerings are poised uniquely to address these emergent requirements.

ASIC, or Application-Specific Integrated Circuit, serves a highly customized function compared to general-purpose GPUs, designed for specific tasks and operationsThis includes optimizing performance with faster processing speeds and reduced energy consumption.

Typically, when there is a market demand from large downstream companies, they can engage chip design firms for assistance, which extends their development timelines

However, once they create their ASICs, they no longer have to pay Nvidia high license fees, effectively enabling them to tailor chips precisely for their algorithmsFor instance, Broadcom's collaboration with Google on its TPU series perfectly illustrates this.

In the collaboration model surrounding XPU, chip design firms are playing increasingly critical roles, managing both front-end design and back-end processes including wafer fabrication and packaging, thereby helping major firms strike a balance between performance and cost.

Currently, Broadcom dominates the ASIC market with a 55%-60% share, with Marvell trailing at a 13%-15% share.

China also has several companies developing ASIC chips, including Cambricon, YITU Technology, Beijing Junzheng, and Guoke Microelectronics.

Cambricon, which has recently surpassed a market cap of $280 billion, follows the ASIC model with its AI chips and launched its first Neural Processing Unit (NPU) back in 2016. The company is currently developing the Siyuan 590, directly competing with Huawei's Ascend 910B and Nvidia's H100.

Although the company's current performance may not fully justify its valuation, the substantial increase in inventory and advance payments in the third quarter suggests strong future production investment

As demand rises for self-controlled advanced process chips, the certainty of future revenue growth increases.

In fact, analysts have projected that the growth rate for customized AI ASICs will likely outpace that of GPU computing.

In addition to Nvidia, Broadcom currently holds a remarkably advantageous position within the chip ecosystem.

This is primarily because providing AI infrastructure is more than just developing chips.

Nvidia's dominance in the AI training market rests on its proprietary technology, primarily the combination of its chips with CUDA and NVLink, a specialized networking protocol that ensures high-speed low-latency operation.

While custom silicon reduces cloud service providers' reliance on CUDA, Broadcom excels not only in ASICs but also offers networking protocols and related chip intellectual property, thus enabling customers to establish XPU + ultra-high-speed interconnect server clusters.

On the other hand, Nvidia's recent stock decline reflects market concerns regarding the challenges it might face in maintaining future GPU demand.

First, speculation surrounds whether the Blackwell chip will continue to dominate the market

Slated for a slight shipment this fourth quarter, it is anticipated to become Nvidia's primary product in the coming yearHowever, as technology giants increasingly adopt customized ASICs, the demand for inferencing calculations is set to multiply, putting Nvidia’s GPUs at risk of not sustaining their reign in the inference market.

Secondly, Nvidia grapples with the high expectations set by its unprecedented performance growth over the past two years.

Analysts surveyed by FactSet generally project that Nvidia's revenue will increase by 55% in 2025, reaching $191.45 billion, while for 2024, a staggering doubling is expected to reach $123.37 billion.

The competitive landscape for AI chips is witnessing a significant recalibration, hinting at profound shifts in the market dynamics that could emerge as soon as next year.

Broadcom's ascendance in ASIC solutions can be likened to TSMC's historic challenge to traditional Integrated Device Manufacturers (IDMs), marking a transformative moment in the semiconductor field.

The next year is anticipated to emerge as the year of inference.

Since the launch of ChatGPT over two years ago, the battlefield for AI has shifted away from training to a more expansive and competitive arena: inference.

In recent months, the AI application sector has witnessed a multitude of impressive developments.

OpenAI has been conducting a series of impactful product updates, and simultaneously, Google has been unveiling exciting advancements, including Gemini 2.0, Veo 2, and quantum chips

High-ranking Google executives view Gemini 2.0 as an AI model tailored for the age of intelligent agents.

Turning back the clock three months, OpenAI launched the o1 series, which enhanced the logical reasoning capabilities of Large Language Models (LLMs) through reinforcement learningAs additional reinforcement learning and thinking time are invested, the performance of o1 consistently improves.

OpenAI has embraced Scaling Law, which has attracted significant capital investments, as they aim to represent the growth signaled by this principle.

Yet, the data fueling this race is akin to fossil fuels for model training and is bound to deplete

This transition from pre-training to inference is increasingly pivotalRecently, Ilya Sutskever, a former executive at OpenAI, publicly addressed this likelihood, suggesting that the next generation of AI models will operate as genuine AI agents, equipped with reasoning capabilities.

Additionally, Microsoft CEO Satya Nadella highlighted that a new Scaling Law will emerge, defined primarily by testing or inference time.

Thus, o1 is not merely a simple upgrade; it signifies the industry’s exploration of a brand new set of regulations: the longer the model “thinks,” the more accurate the answers it provides.

From this perspective, the inference scenario is set to manifest a tremendous demand for computing power, with the inference costs anticipated to decline rapidly alongside the momentum generated by various AI-native applications.

A supplementary report from Barclays posited that AI inference demand may constitute over 70% of total computing needs for general artificial intelligence, potentially reaching 4.5 times the training computing requirements.

In October, AI startup unicorn Anthropic launched Claude 3.5 Sonnet, which outperformed OpenAI’s o1 in performance tests, also introducing automated interactions that enable AI to directly operate computers and execute complex commands.

In the mobile domain, Chinese startup Zhipu introduced AutoGLM, which bypasses mobile operating systems and simulates user operations in the UI, effectively allowing the AI to “take over the phone.” Media reports indicate that OpenAI is set to unveil the Agent “Operator” next January.

What about Google? It has launched Agents across various platforms, including mobile devices, AR glasses, and developer and research-focused environments, all based on Gemini 2.0.

Reflecting on this evolution, two years ago, ChatGPT merely functioned as a chatbot

Now, as a crucial component in complex multimodal interactive reasoning situations, AI agents must learn and reason from user-end data to execute precise actions—a focus that is pivotal for the refinement and expansion of inference models in the days ahead.

Key to the data flywheel's movement is the development of appropriate software and hardware to satisfy user needs.

During the latter half of this year, the commercialization races for “AI + applications” have garnered attention again, with companies like Applovin, Shopify, and Palantir in the US demonstrating strong stock performance buoyed by AI-driven business spikes.

China is also focusing intensively on the development of AI applications

ByteDance's large models, albeit new to the scene, have quickly scaled to reach close to 9 million daily active users (DAU), experiencing a growth rate exceeding 15%, and claiming the second position globallyBeyond launching a suite of applications, they are also investing in AI hardware, actively seeking consumer-facing scenariosOther internet giants, such as Baidu and Xiaomi, have begun to engage in developing AR glasses.

As for the future, experts predict that Nvidia's GPU currently boasts around 80% market share in the inference domain, but this percentage is expected to decline to approximately 50% by 2028 due to the rise of customized ASIC chips from major tech companies.

Nevertheless, the ascendancy of ASICs does not signify the obsolescence of GPUs