The surge of ChatGPT has ignited an AI race among tech giants and startups alike. In this AI gold rush, NVIDIA, provider of top-tier AI chips, has emerged as the winner, with Wall Street dubbing NVIDIA "the sole arms dealer in the AI war."
Over the past month, NVIDIA has been on a frenzied spree in the AI venture capital scene. It appears to be capitalizing on its lead in GPU technology to reinforce the strongest AI industry alliance it has built.
Two AI Unicorns Announce NVIDIA Investments in a Single Day, Inflection AI's Valuation Rises to Third Highest Globally In June of this year, NVIDIA took part in the financing rounds of three high-profile AI unicorns, all of which announced new rounds of funding. On June 9, Canadian AI company Cohere, which creates ChatGPT-like chatbots, announced the completion of its $270 million Series C funding round, in which NVIDIA, Oracle, and Salesforce participated. This round raised Cohere's valuation to approximately $2.2 billion.
The two startups that announced large-scale funding almost simultaneously last Thursday are: Inflection AI, which launched the AI chatbot Pi, and AI video creation startup Runway.
Inflection AI, co-founded and led by DeepMind co-founder Mustafa Suleyman, raised $1.3 billion in new funding, placing it fourth in terms of the scale of AI financing rounds, according to Crunchbase data. The lead investors in Inflection AI's latest round were Microsoft, LinkedIn co-founder Reid Hoffman, Bill Gates, and Google chairman Eric Schmidt, with NVIDIA being the only new face among the investors.
After completing the funding round, Inflection AI's valuation rose to approximately $4 billion, making it the third-largest generative AI unicorn globally, trailing only OpenAI and Anthropic.
Runway raised $141 million in new funding, with new investors including Google, NVIDIA, and Salesforce. This round brought Runway's valuation to around $1.5 billion, tripling in less than half a year.
Inflection AI Introduces 22,000 NVIDIA H100s to Create a Superior Supercomputer, Outpacing Meta's Supercomputing Cluster in Chip Quantity Inflection AI recently introduced its first proprietary language model, Inflection-1, which it says is trained on a very large dataset using thousands of NVIDIA H100s and is the best model in its computing category. In various benchmark tests commonly used to compare large language models (LLMs), Inflection-1 performs better than GPT-3.5, LLaMA, Chinchilla, and PaLM-540B.
Last Thursday, Inflection AI also announced that, in partnership with NVIDIA, it would expand its supercomputer to include 22,000 NVIDIA H100 chips to support the training and deployment of a new generation of AI mega-models. The number of integrated AI chips surpasses the supercomputing cluster of 16,000 A100s announced by Meta this May.
In addition to NVIDIA, the other partner in the super GPU cluster project is cloud service provider CoreWeave. It claims to offer "80% cheaper computing power" than traditional cloud providers. NVIDIA previously invested $100 million in CoreWeave. In June, some media outlets reported that Microsoft agreed to invest several billion dollars in CoreWeave over the next few years for cloud computing infrastructure construction.
In the latest authoritative AI performance benchmark test MLPerf, the cluster built by NVIDIA and CoreWeave, boasting 3,584 H100s, trained the large language model GPT-3 in less than 11 minutes.
NVIDIA Leads Rivals by Two Years, But Its Dominant Position Is Not Impenetrable; CUDA Won't Always Be Its Moat Jon Peddie Research's GPU market data report shows that NVIDIA shipped 30.34 million PC GPUs last year, nearly 4.5 times that of AMD. As of the fourth quarter of last year, NVIDIA held 84% of the independent GPU market, significantly surpassing its industry competitors.
NVIDIA's revenue for the first quarter of this year, released last month, greatly exceeded expectations. The revenue of its AI chip business hit a historic high, maintaining a year-on-year growth rate of over 10%. The revenue guidance for the second quarter was a blowout, not only did it not decline for the fourth consecutive quarter as the market expected, but it also surged 33% year on year, demonstrating the red-hot demand for AI chips.
Following the release of the financial report, Jay Goldberg, founder of chip consulting company D2D Advisory, stated, "Currently, the AI chip market still appears to be one where NVIDIA eats all."
Last month, when AMD released its most advanced AI chip, the MI300X, analysts did not jump on the hype train, but instead pointed out pragmatically that AMD still has a long way to go to challenge NVIDIA's leading position in the AI chip industry, and this chip alone won't cut it.
Karl Freund, founder and chief analyst of Cambrian-AI Research LLC, believes that apart from NVIDIA having the largest software and researcher ecosystem in the AI industry, and the MI 300X not having a significant memory or cost advantage, a key challenge AMD faces is the lack of a Transformer Engine like that of the H100 (a library for accelerating Transformer models on NVIDIA GPUs), which can double the performance of large language models (LLMs).
This means that if it takes a year to train a new model with thousands of NVIDIA GPUs, it may take two or three more years or triple the number of GPUs to train it with AMD's hardware.
Following the release of NVIDIA's financial report, some commentary noted that Nathan Benaich, a partner at AI startup investment firm Air Street Capital, estimated that NVIDIA was two years ahead of its competitors. He described NVIDIA's success as follows:
NVIDIA saw the future ahead of everyone else. They shifted focus to making GPUs programmable, spotted the opportunity, made a big bet, and consistently outpaced competitors.
At the same time, industry insiders believe that Wall Street's enthusiasm for NVIDIA is somewhat overly optimistic. Benaich pointed out that NVIDIA's hardware and software are far from impervious to attack.
As previously mentioned by Business Times, NVIDIA's parallel computing platform and programming model CUDA, which is based on its GPU production, has helped it build a robust ecosystem and increase competitive barriers.
The company behind the popular AI painting model Stable Diffusion, Stability AI, recently agreed with Benaich's assessment of NVIDIA's competitive situation. They said, "Google, Intel, and other companies' next-generation chips are catching up, and even CUDA is no longer a moat as software becomes standardized."
