Tech companies have a worrying grip on Artificial Intelligence
At a meeting with US senators in May, Sam Altman of OpenAI expressed a desire for people to use their ChatGPT AI system less. His reason: lack of GPUs. Altman’s statement illustrates the growing influence of large tech companies in the generative AI market, as they benefit from access to valuable and expansive infrastructures.
GPUs (graphics processing units) are specialized chips which were initially developed for video games, but are now essential components in the AI industry. These components are costly and scarce, with Nvidia Corp.’s market value reaching $1 trillion last month due to the surge in demand. For creating AI models, developers typically purchase cloud servers from firms like Microsoft Corp. and Amazon.com Inc., which are powered by GPUs.
The saying goes that during a gold rush, one should sell shovels. It is no surprise that today’s AI infrastructure providers are taking advantage of this. However, there is a distinct difference between now and the mid-19th century when Levi Strauss and Samuel Brennan were among the most successful participants of the California Gold Rush. For the next few years, most of the profits made from selling AI services will go to tech giants like Microsoft, Amazon and Nvidia who have already established their dominance in this space.
This is mainly because while cloud services and chips are becoming more expensive, it is becoming cheaper to access AI models. OpenAI slashed the cost of GPT-3 by a third in September 2022 and again 10 times in December 2022. In June they also reduced the fee for their embeddings model (which converts words into numbers for large language models) by 75%. Sam Altman has stated that the cost of intelligence is heading towards near-zero.
The cost of constructing AI models is increasing due to the difficulty in purchasing GPUs, similar to the challenge of buying toilet paper during COVID-19. Nvidia’s A100 and H100 chips are in high demand for machine learning applications, and their price has risen from less than $35,000 to over $40,000 in recent months. The shortage of these chips has caused many AI startups to wait behind larger customers such as Microsoft and Oracle for their orders. A Silicon Valley-based startup founder with connections to Nvidia reported that OpenAI was on the waiting list and wouldn’t receive H100 chips until Spring 2024. An OpenAI spokeswoman stated they do not disclose this information; however, Altman himself has spoken about his difficulty in obtaining chips.
Big Tech companies have an advantage over upstarts like OpenAI due to their access to GPUs and established customer bases. In 2022, Sam Altman gave up 49% of OpenAI in exchange for $1 billion from Microsoft. This seemed like a lot at first, but this partnership was necessary for AI companies to remain successful. Microsoft’s bet is paying off; CFO Amy Hood revealed that Azure OpenAI would bring in at least $10 billion in revenue. This product is more expensive than OpenAI’s but offers benefits such as better security and compliance for customers like CarMax and Nota.
AI model producers have difficulty keeping their products unique due to constant movement of personnel. They also face ongoing costs such as cloud credits for model training and inference, which AWS estimates can be up to 90% of operational expenses. Most of this money goes to cloud providers.
This has created a two-tier system for AI businesses; those at the top have access to funding and exclusive connections. For example, Y Combinator alumni have received computing credits worth hundreds of thousands of dollars from Amazon and Microsoft. Nat Friedman, a venture capital investor, has spent an estimated $80 million on GPUs to create his own cloud service, the Andromeda Cluster.
Second-tier AI companies face a long road to obtain the necessary connections and resources to train their AI systems, regardless of how clever their algorithms are. Smaller firms, however, have hope: Big Tech firms may eventually find their products and services becoming commoditized as the chip shortage lessens and GPUs become more accessible and affordable. Additionally, there will be increased competition between cloud providers as organizations such as Google develop TPUs and Nvidia creates its own cloud infrastructure to compete with Microsoft. Lastly, the development of techniques such as LoRAand PEFTwill enable AI models to require less data and computing power, reducing Big Tech’s hold on the market.