Summary of Key Points
Recently, ASIC chips have become the "hot topic" in the AI community: cloud giants such as Amazon, Google, and Microsoft are no longer treating their own ASICs as "internal playthings" but are beginning to sell them externally. OpenAI collaborated with Broadcom to develop its own inference chip in just 9 months, while Meta encountered significant challenges in its efforts to do the same. All these developments point to one conclusion: ASICs are moving from the periphery of AI computing power to the center, and in the future, they may split the market with GPUs. The underlying motivation is that everyone wants to control the AI infrastructure and utilize the cost-effective advantages of ASICs to reduce operational expenses.
1. Cloud Providers Are No Longer Keeping Their Chips to Themselves: Selling ASICs to Improve Profitability
In the past, cloud providers used custom-designed ASIC chips for their own needs, such as Amazon's Trainium and Google's TPU, primarily to meet their internal AI computing requirements. However, things have changed in 2026: they are now selling these chips externally because AI has shifted from training to inference tasks, making ASICs much cheaper than general-purpose GPUs.
For example, after adopting Google's seventh-generation TPU, the monthly computing cost for the AI image generation platform Midjourney was reduced by two-thirds, from $2.1 million to $700,000. Google's Gemini service is also 60% more cost-effective per million tokens compared to GPT-5.5. This cost advantage has opened up significant business opportunities for cloud providers:
- Amazon's Trainium chips are nearly sold out, with orders totaling $22.5 billion, and discussions are ongoing to sell them to data centers of other companies.
- Google and Blackstone have formed a $5 billion joint venture to sell TPU solutions, providing computing nodes for Anthropic, including a guarantee of $3.2 billion for a project in New York.
- Microsoft's Maia 200 chip is also being sought by Anthropics as its first external customer.
By selling their ASIC chips, cloud providers can not only earn more money but also strengthen their position in the AI infrastructure landscape. After all, those who control affordable and efficient computing power will attract more AI companies.
2. OpenAI Develops Its Own Chip: Jalapeño in Just 9 Months to Overcome Computing Power Challenges
As one of the largest GPU buyers in the world, OpenAI has long been struggling with insufficient computing resources. Therefore, it collaborated with Broadcom to develop its first inference chip, Jalapeño, in just 9 months.
Their approach was strategic: OpenAI designed the chip architecture (since they best understand the requirements of their models), while Broadcom handled production and hardware manufacturing, with TSMC being the manufacturer. This approach allowed them to avoid the time and effort associated with building a new chip development team from scratch, resulting in a chip that performs better than existing solutions in terms of performance and power efficiency.
OpenAI's goal is clear: to create a complete technology stack from models to chips, reducing its dependence on a single GPU supplier. This chip is expected to be deployed on a large scale, with a total power consumption of up to 10 GW—equivalent to the electricity demand of a medium-sized city, highlighting their immense computing needs.
3. Meta Hits a Roadblock: Spending $2 Billion on Chip Acquisition Goes Awry Due to Internal Issues
Not every company can successfully develop its own chips, and Meta is a prime example of this. In 2025, Meta spent $2 billion to acquire the chip company Rivos in an attempt to accelerate its AI chip development efforts. However, the project failed within half a year due to internal conflicts:
- The Rivos team and Meta's existing employees clashed over salary and technical approaches (should they use Meta's existing technology or Rivos' new innovations?).
- Multiple projects were delayed and eventually canceled.
This highlights a critical issue: software can be modified at any time, but once chips are produced, mistakes can result in significant financial losses and time delays. Microsoft's Maia chip also experienced delays, demonstrating that developing chips in-house requires long-term investment and team coordination, even with substantial funding.
In contrast to OpenAI's collaborative approach, Meta's all-in-house strategy was too aggressive. It may be more efficient to leave specialized tasks to professionals.
4. Will ASICs Compete with GPUs for Dominance? The Semiconductor Industry Is About to Change
The rise of ASICs is reshaping the semiconductor industry:
- Rapid Market Growth: Broadcom's AI chip revenue in the second quarter of 2026 increased by 143% year-over-year. World Semiconductor Market Research predicts that ASIC revenue will grow from $13 billion in 2024 to $150 billion by 2030, at an annual growth rate of nearly 50%.
- Outpacing GPU Growth: ASIC shipments increased by 44.6% in 2026, compared to just 16.1% for GPUs—this is the first time ASICs have outpaced GPU growth.
- Diversified Chip Strategies Becoming the Norm: AI companies are no longer relying on a single chip; for example, Anthropic uses chips from AWS, Google, and Microsoft, while OpenAI utilizes multiple chips, including its own.
In the future, GPUs will still dominate model training, but in large-scale commercial applications such as chatbots and image generation, ASICs will become increasingly popular. Goldman Sachs predicts that by 2027, the demand for ASICs and GPUs will be equal. The semiconductor industry's division of labor may completely change: model companies will define chip architectures, cloud providers will sell chips, and traditional semiconductor companies will focus on manufacturing, with everyone sharing the benefits.
In Conclusion
The surge in ASIC popularity is not accidental but a natural outcome of AI moving from the laboratory to commercial use. Those who can make computing power more affordable and efficient will gain an advantage in the next generation of AI competitions. And this chip battle has just begun.