Summary in One Sentence
Recently, chip giant Broadcom released its best financial report ever: total revenue in the third quarter approached $30 billion, a year-on-year increase of 86%. Revenue from custom AI chips more than doubled, resulting in substantial profits. This isn't due to some sudden miracle for Broadcom; rather, it reflects a global trend among leading AI companies—from Google, Meta, OpenAI to Amazon, as well as domestic players like Huawei, Alibaba, and Baidu—all turning to custom or self-developed AI chips. They no longer want to rely entirely on NVIDIA's off-the-shelf chips. The underlying motivation is clear: to save money, gain more control, and ensure their own survival. The monopoly held by NVIDIA in the AI chip market is being challenged by these collective efforts.
---
Detailed Analysis
1. Why Is Broadcom Suddenly Making So Much Money?
Many people previously viewed Broadcom as a company that mainly produced network switching chips for data centers. Its impressive financial results this quarter are due to its ability to meet the critical needs of all major AI companies. Most large model companies lack the capability to develop their own chips from scratch. Hiring hundreds of chip engineers, coordinating with foundries, and handling all aspects of design is not only costly but also time-consuming. By the time they complete the process, competitors' models might have already gone through several iterations. Broadcom, on the other hand, offers a “turnkey” solution: they simply need to know the performance requirements and optimization goals for their models, and Broadcom takes care of the rest, from chip design to production. For example, OpenAI’s “Little Pepper” inference chip took only 9 months from design to successful production, demonstrating exceptional efficiency. Now, orders for custom chips from top companies like Google, Meta, and OpenAI are largely in Broadcom’s hands. If Broadcom doesn’t make money from these orders, who will?
2. Custom Chip Development by Giants Is No Longer Just Talk: Most Are Already in Action, and Some Are Making More Profit Than Selling Models
The development of custom chips by AI companies is no longer just theoretical; many have already implemented them commercially:
- Google was among the first to move this direction, developing its own TPU chips in 2016. Its eighth-generation chips are about to be mass-produced, and it has even created dedicated chips for training and inference purposes. The $200 billion in computing power contracts signed with Anthropic also involve Google’s own TPU chips, resulting in no revenue for NVIDIA.
- Meta has already incorporated its first custom chip into its Facebook and Instagram content recommendation systems. The new chips being developed will be optimized for generative AI models.
- Amazon’s self-developed Trainium AI training chip has reached its third generation, and even before its release, demand has outpaced production. The future revenue from these contracts alone exceeds $225 billion, putting its custom chip business in the top three globally, generating more profit than many specialized chip companies.
- Domestic companies like Huawei’s Ascend, Alibaba’s Zenwu, and Baidu’s Kunlun chips are also being widely used. Only Tencent has not yet made its moves in this area, but other companies are actively planning to do so.
3. Custom Chips Can Save Big Money
The cost savings from developing custom chips are significant. For example, NVIDIA’s chips can be used out of the box, but manufacturing them in-house saves a lot of money. Imagine running a chain of 100,000 supermarkets: using custom shelves that fit the products, scan faster, and consume less power can result in annual savings of hundreds of millions. OpenAI’s “Little Pepper” chip, optimized for large model inference, reduces inference costs by 50% compared to NVIDIA chips. With billions of user queries processed daily, the annual savings in computing power costs can amount to hundreds of millions. The investment in chip development is quickly recouped.
4. Taking Control of Hardware
Previously, AI companies were at the mercy of NVIDIA’s chip supply. Now, they can dictate the hardware that supports their models. NVIDIA controlled over 70% of the global AI chip market and the associated software ecosystem, determining product schedules and features. This meant that companies had to wait for NVIDIA to release new chips to upgrade their models. With self-developed chips, companies can add the necessary hardware features directly during the design process, ensuring that their hardware matches their needs. For example, Google’s TPUs enable faster model execution than NVIDIA’s chips of the same performance level.
5. Ensuring Dependability and Reducing Risks
Having custom chips provides a safety net. If NVIDIA were to raise prices significantly or stop supplying chips, companies would be at a disadvantage. With self-developed chips, they have a backup option, allowing them to negotiate prices more effectively and reducing the risk of supply disruptions. This is especially crucial for domestic companies, as there are export restrictions on high-end NVIDIA chips. Developing custom chips is not just about profit but about survival.
In summary, the shift towards self-developed chips represents a major shift in the AI industry, aimed at saving money, gaining more control, and ensuring business continuity.