The Second Half of AI Chips: The Siege Against NVIDIA
-
NVIDIA is currently thriving, with its growth rate over the past few years surpassing any previous period, driven by the cryptocurrency boom and the era of large AI models. This has propelled the chipmaker's market value to surpass the $1 trillion mark.
However, unlike the speculative virtual economy of cryptocurrencies, the 'real demand' brought by large AI models is the core driver behind NVIDIA's trillion-dollar valuation. Reports indicate that the lead time for NVIDIA's H100, from order to delivery, has stretched to several months, with spot market premiums once nearing 100%.
But NVIDIA's golden days may not last long. As large AI models are confirmed as a 'promising path,' companies are accelerating purchases of NVIDIA's GPUs to build their own training servers. Yet, watching funds flow out like a flood, they are also making their own plans.
Recently, OpenAI announced it would begin developing its own AI chips to reduce reliance on NVIDIA. Coincidentally, Microsoft, which is building large AI servers, also unveiled its own AI chip plans. Interestingly, although OpenAI is nominally part of Microsoft's camp (Microsoft previously completed its acquisition of OpenAI), the two do not seem to be sharing chip development plans.
Beyond OpenAI and Microsoft, many other manufacturers are also gearing up.
01. Besieged on All Sides
Supporting a large data center is no small expense, with initial hardware investments alone reaching hundreds of millions. Microsoft's recently announced European data center plan involves an initial investment of $500 million, excluding subsequent maintenance costs. Of this $500 million, aside from infrastructure, a significant portion is allocated to purchasing NVIDIA's professional computing cards.
Recent analyses suggest that NVIDIA's chip costs may be over 10 times lower than their selling prices. For example, the H100, highly sought after by large enterprises, costs around $2,000–$2,500 to produce but sells for over $25,000.
Whether to cut costs or carve out a share in this emerging market, developing in-house AI chips has become urgent. From current information, semiconductor giants like Intel and AMD have already announced new AI chip development plans. Intel is taking a unique approach by leveraging CPUs to create a different type of AI chip, even releasing a first-generation product. Meanwhile, AMD aims to challenge NVIDIA's dominance in the GPU space.
It’s no surprise that traditional semiconductor giants are vying for a piece of the pie. What’s more concerning for NVIDIA is OpenAI and Microsoft’s announcements to develop their own AI chips. As two of its most critical customers, their departure would significantly impact NVIDIA’s ecosystem and revenue.
OpenAI’s chip plans were only recently revealed. For an AI company, I have doubts about OpenAI’s chip development capabilities. Moreover, based on recent job postings, OpenAI is building a team from scratch, meaning initial results may take at least a year—and likely won’t rival NVIDIA’s flagship chips.
In contrast, Microsoft’s chip plans are more noteworthy. Microsoft has long invested heavily in chip development and has released several products in recent years. A recently leaked chip codenamed 'Athena' is reportedly in trial production after being in development since 2019.
Sources say OpenAI has secretly tested the Athena chip. Designed specifically for training and running large models, its performance is said to be impressive, at least on par with mainstream chips from Amazon and Google.
Of course, Athena's performance cannot match Nvidia's flagship chips, but it allows Microsoft to gain more autonomy and slightly curb Nvidia's pricing power in chip supply. Moreover, Athena is only Microsoft's first professional AI chip, and the over $2 billion R&D investment clearly won't yield just one outcome.
As OpenAI's major backer, Microsoft will likely require OpenAI to provide testing and deployment environments for the Athena chip, following the examples of Amazon and Google. Long before Microsoft, Amazon and Google invested heavily in AI companies. Amazon, while providing $4 billion in funding to Anthropic, also mandated the use of its two self-developed AI chips.
When leading AI companies start adopting alternative or self-developed chips, it will inevitably significantly impact the hardware choices across the AI industry—precisely what Nvidia doesn't want to see. How will Nvidia respond?
02 Nvidia's Countermeasures
The allure of large AI models has captivated numerous tech companies, with some even considering it the dawn of the next industrial revolution. Regardless of how many technologies have been labeled as such, large AI models are undoubtedly one of the most impactful technological advancements for the general public in recent years.
Their close relevance to everyday life means these technologies have vast application markets and can be rapidly commercialized for profit. Few technologies have progressed from inception to commercialization as swiftly as large AI models. From ChatGPT's release to the proliferation of various AI models, the entire process took less than a year.
From productivity to entertainment, consumption, transportation, and education, large AI models have already spawned numerous applications. As a result, capable companies are accelerating the construction of their own data and computing centers to deploy and train larger AI models, securing a competitive edge.
As the AI market enters a phase of fierce competition, companies are seeking more efficient training methods and more powerful models. Beyond optimizing algorithms, more powerful specialized computing cards are essential. Thus, Nvidia's strategy is straightforward: stabilize its R&D team and launch AI chips that far surpass competitors' offerings.
Hardware performance is Nvidia's greatest advantage. Whether it's Amazon or Microsoft, as long as they seek an optimal balance between performance and energy efficiency, Nvidia remains their primary choice. The reasons driving companies to develop their own chips are twofold: Nvidia's chips are too expensive, and limited supply delays expansion plans.
Currently, Nvidia's production capacity is gradually increasing, and as procurement volumes decline, supply and demand should soon reach equilibrium. The main issue then becomes pricing. Given the nearly 10-fold gap between Nvidia's costs and selling prices, there's ample room for price reductions.
Personally, I believe that if Nvidia is willing to lower prices, purchasing its professional computing cards to build high-performance data centers would still be a cost-effective choice for many companies. As for self-developed chips? In reality, data centers require different types of chips based on scale and purpose. Self-developed chips are suitable for centers with lower performance demands.
In short, training and development centers can use Nvidia's professional computing cards to enhance efficiency, while data centers serving general users can employ self-developed or other chips to reduce construction and maintenance costs. As AI model applications expand, companies will need to build more data centers globally to meet user demands locally.
Therefore, the competitive advantages NVIDIA has accumulated will not easily diminish in the future. However, as other companies enter the market, NVIDIA's influence may gradually weaken, potentially requiring it to relinquish some profits in areas like product pricing to maintain its market share.
Unlike past scenarios where 'when gods fight, mortals suffer,' this time the collective challenge by AI firms against the industry leader might actually enable small and medium-sized AI companies to access more cost-effective data center deployment solutions.