In recent years, the popularity of NVIDIA graphics cards has led to a group of AI cloud startups as "GPU secondary dealers", attracting heavy bets from venture capitalists. However, the profit margin of such companies is not high, and their business model may be difficult to maintain long-term high-speed growth.
NVIDIA graphics chips became the hottest technology product last year, attracting many people to hoard and hoard, with a group of startups selling second-hand graphics cards or providing serverless architecture graphics card services beginning to emerge.
These startups rent graphics cards from cloud service providers or other companies and institutions with their own servers, including universities and digital currency mining institutions, and then slightly increase prices by subletting graphics cards to AI developers. Such startups also provide additional software services to help train AI models or provide convenient access to open source models. So far, these companies believe they will definitely be favored by venture capital, and if we delve deeper into their business models, the preference of venture capital companies for second-hand graphics card subleasers may have been excessive.
These "GPU secondary dealers" have indeed seen rapid revenue growth, with one company named Together achieving an annual revenue of around $20 million just four months after its establishment. However, if investors learn a lesson from the financing trend in 2021, it is that this reckless revenue growth is often not the best strategy, and the second-hand graphics card subleasing industry should also apply this principle.
The basic principles of this business model are not very attractive. These startups usually find a large number of cloud service providers to rent graphics cards, such as LambdaLabs, CoreWeave, and Shadeform. Moreover, more than a dozen companies compete fiercely to rent graphics cards, forcing them to use the rented graphics cards together or sign long-term contracts. A founder of such a startup company stated that top cloud service providers such as Amazon Web Service and Microsoft Azure have more graphics cards, but the prices are too expensive, making it difficult for second-hand video card renters to afford them.
The founder also said that what's even worse is that second-hand video card subleasers need to keep a large number of servers and ensure that at least half of them are in a standby state, as users will increase or decrease usage without warning. Moreover, these users hope to pay less rent than cloud service providers such as CoreWeave, which means there is very little room for second-hand graphics card subleasers to increase their prices. Overall, the gross profit margin of second-hand graphics card subleasers is only 50% at most, compared to the gross profit margin of general software companies, which can reach over 70%.
Analysis suggests that even so, venture capital companies are still optimistic about such companies based on rapid revenue growth. However, once these companies go public, their meager profit margins will reprice the company, which may make them regretful.
Taking Together as an example, the company previously raised at least $50 million from Kleiner Perkins Caufield&Byers, a top Silicon Valley venture capital fund, with a valuation of over $300 million, marking the company's third fundraising in 10 months. At present, the company's valuation is approximately 15 times its annualized revenue. Moreover, as investors investing $50 million in a company with a valuation of $300 million, they need to increase the valuation of Target to at least $1 billion to achieve a return on investment of more than three times.
But this may not be easy. Given the low profit margin of such companies, at some point, a later stage investor may use lower revenue multiples or gross profit margin multiples indicators for valuation. For example, investors may draw inspiration from the valuation method of Twilio, whose market value is only twice the expected revenue for next year.
If estimated revenue is used for valuation, if Total wants to reach a valuation of $1 billion, its annual revenue will need to reach a staggering $500 million, and such a daunting task may pressure the company to take risks and accelerate growth. Moreover, this has not yet taken into account the situation where Nvidia will expand its production capacity to meet market demand, leading to a slowdown in the growth of second-hand graphics card subleasers.
Some second-hand graphics card subleasers have already foreseen the above situation and are starting to make astonishing moves: requiring investors to reduce the undervalue or directly choosing to accept the lowest valuation. The financing transaction between Together and Jin Kaipeng Huaying was like this, while Banana. dev, another second-hand graphics card sublease, only raised millions of dollars as a result.
It is certain that some venture capitalists are optimistic about the technological achievements of these second-hand graphics card subleasers. For example, the chief scientist of Together has invented a method to accelerate training and optimize large language models, and the background of the company's team members is also impressive. However, it is still these investors who believe that if these companies want to succeed in the long term, they need to abandon their graphics card subleasing business and focus on higher profit margin software services, which will undoubtedly significantly reduce the company's growth rate.