How much does Apple need to invest annually in generative AI to catch up with the competition? A reasonable estimate starts at several billion US dollars / Apple需要每年投資多少錢在生成式AI來追趕競爭對手？合理推估為數十億美元起跳
2 min readOct 23, 2023
- My latest survey indicates that Apple plans to purchase 2,000–3,000 and 18,000–20,000 units of AI servers in 2023 and 2024, respectively. It represents about 1.3% and 5% of worldwide AI server shipments in 2023 and 2024, respectively.
- I think it’s reasonable to assume that the AI servers Apple purchases are primarily this year’s most popular specification: Nvidia’s HGX H100 8-GPU, designed for generative AI training and inference. It’s also expected that some shipments in 4Q24 will likely upgrade to the B100 solution.
- Each HGX H100 8-GPU server is priced at around $250,000. Therefore, it’s estimated that Apple will spend at least about $620 million in 2023 and $4.75 billion in 2024 on AI server purchases.
- Apple’s AI server purchases are significantly lower in 2023 than in 2024 due to a shortage of Nvidia AI chips starting from 2Q23, compounded by Apple placing its orders later than other major customers.
- However, even when the supply of Nvidia AI chips improves in 2024, Apple’s AI server purchases still lag behind its competitors. Taking Meta as an example, its AI server purchase in 2024 will be about 40,000 units. Not to mention, Meta’s AI server count before 2024 already far exceeds that of Apple.
- Assuming that Apple doesn’t provide services such as cloud hosting, it’s not appropriate to compare Microsoft (with AI server purchases of about 80,000–100,000 units in 2024) with Apple.
- I believe that when Apple trains its large language models, it prefers to use AI servers it has purchased and installed itself rather than virtual hosts from other cloud service providers for security, privacy, and design flexibility.
- Apple’s AI infrastructure computing power is less than its major competitors, which means that it must have superior software development capabilities to catch up.
- The above does not consider other costs (such as labor costs, infrastructure operating costs, etc.). Thus, it’s reasonable to estimate that Apple would have to invest at least several billion US dollars annually to have a chance of catching up with its competitors. If Apple really intends to spend only a billion dollars a year on generative AI development, it won’t matter much if my survey is wrong, but I am genuinely concerned about the future of Apple’s generative AI business/service.
- I know some will argue that Apple can develop its own AI server chips to save on AI server procurement costs. While I agree this is a noteworthy trend, the visibility of the development is not clear at present.
- 我最新的調查指出，Apple預計在2023年與2024年分別採購2,000–3,000與1.8–2萬台AI伺服器，分別約佔全球2023與2024年AI伺服器出貨量的1.3%與5%。
- 我認為可以合理假設Apple採購的AI伺服器規格主要為今年最常見、用於生成式AI訓練與推理用的Nvidia HGX H100 8-GPU，也可預期可能4Q24出貨部分會升級成B100方案。
- 每台HGX H100 8-GPU約25萬美元，故估算Apple至少需在2023與2024分別支出約6.2億與47.5億美元在AI伺服器採購上。
- Apple在2023年採購AI伺服器數量顯著低於2024年的原因是Nvidia AI晶片自2Q23開始短缺，Apple又較其他主要客戶晚下單。
- 但即便是2024年Nvidia AI晶片供應較為緩解，Apple採購的AI伺服器數量仍低於其他競爭對手。以Meta為例，僅2024年採購的AI伺服器數量約為4萬台。更不要說Meta在2024年前擁有的AI伺服器數量就已經遠高於Apple。
- 假設Apple不提供類似雲端虛擬主機的服務 (Cloud hosting service)，故不適合拿Microsoft (2024年AI伺服器採購量約8–10萬台) 與Apple比較。
- 我相信Apple在訓練其AI大型語言模型時，基於安全性、隱私與設計彈性考量，應該會較偏好用自行採購與建置AI伺服器，而非用其他雲端服務商提供的虛擬主機服務。
- Apple的AI基礎建設算力低於主要競爭對手，若要趕上競爭對手則必須具備更優異的軟體開發能力。
- 以上均不計算其他費用 (人力成本、基礎建設營運費用等)。故合理推估Apple需要每年至少投資數十億美元才有機會追趕競爭對手。如果Apple真的打算只願意每年花10億美元用於開發生成式AI，我的研究錯了不要緊，但我真的很為Apple的生成式AI事業/服務前景擔心。
- 我知道一定會有人說Apple可以自行開發AI伺服器晶片以節省AI伺服器採購費用。我也同意這是一個值得注意的趨勢，但是目前此事仍沒能見度。
