The ceiling of Doubao, latest data on ASIC

Last week I wrote about Doubao and received some feedback. The most important point was regarding the prediction of MAU reaching 300 million by next year. The actual situation is that the MAU figure is not important; DAU is more important, specifically the DAU/MAU ratio is more important. Assuming that by the end of this year MAU reaches 100 million, the DAU might only be 15 million (just my guess), resulting in a ratio of 1/6. What if next year the DAU/MAU ratio increases to 1/3? From practical experience, it’s similar; for example, people are currently using 2-3 AI tools, and even if nothing changes next year, the market will naturally concentrate. This has happened many times in the internet space, from group buying, ride-hailing, to live streaming, from the hundred X battles to a mess, ultimately the winner takes all. MAU cannot reflect the intensity of competition; only DAU can indicate the final winner. The key question now is: is winning meaningful? This question may lead to a reverse selection; who would choose to win using the maximum resources?

If Doubao's DAU really reaches such a high number next year (as calculated above), it indicates that not only are users concentrated, but user demand is also concentrating towards the top. User usage scenarios, frequency, and the amount of tokens used per call will all increase, which means what? The average daily token consumption will not just be in the trillions per day, but possibly in the tens of trillions of tokens per day.

What does this mean? It roughly translates to each person calling tens of thousands of tokens per day, or 1 million words. I only realized what this means today (see the image below, the platform gave me the title of "Writing Superhero"). But note that this is not the number of tokens you need to read daily (most people can't finish reading that...). This is to meet your needs, the total number of tokens consumed during the entire reasoning process, including but not limited to: multi-step, long context, complex tasks, multi-modal factors, etc. Thinking about these broadly, it’s not impossible for each person to consume tens of thousands of tokens daily.

ASIC Supply Chain Update (from our guest AYZ):

1. The shipment of Tranium 2 chips in 2025 should be over 1.7 million units, which will bring over $3.4 billion in revenue to Marvell; Tranium 2 uses two active interposer dies, so the recent figure of 4 million units reported by the Asian supply chain is likely a misunderstanding, mistaking the number of interposer dies for the number of Tranium 2 units (i.e., the actual number should be divided by 2).

2. Starting in 2025, AWS has canceled all Inferentia chips, including Inferentia 2.5 and Inferentia 3. Therefore, AWS's third-generation ASIC chip will only be the Tranium 3 in collaboration with Alchip, which is currently expected to start mass production in Q3 2025. According to Alchip's reservations for 2025 CoWoS capacity at TSMC, if mass production goes smoothly, Tranium 3 will bring Alchip approximately $1 billion in revenue in 2025. The full lifecycle of Tranium 3 may bring Alchip approximately $4.5+ billion in revenue contribution.**

3. Due to the product transition from Google TPU V6e to V6p, the author expects Broadcom's AI ASIC revenue (which is Google TPU) to experience a nearly 20% quarter-over-quarter decline in CY4Q24/FY1Q25; however, this product transition period will be very short, lasting only two quarters in CY4Q24 and CY1Q25, and starting from CY2Q25, with the ramp-up of Google TPU V6p, it will return to a growth trajectory. Broadcom's AI ASIC revenue for the full year FY25 is expected to grow to approximately $13 billion.

4. From the supply chain, Google's next-generation TPU V7 will also have two versions. The V7p will still be fully managed by Broadcom, while the V7e will be designed by Google's in-house team for the ASIC die + MediaTek in Taiwan will be responsible for designing the I/O die. Currently, the V7e is planned to start mass production from CY3Q26, with an estimated total lifecycle of around 2 million units, which will bring MediaTek a revenue contribution of over $6 billion.**

Source: Information Equality, original title: "The Ceiling of Doubao, Latest Data on ASIC"

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this is at their own risk