TLDR
- OpenAI has been seeking chip alternatives to Nvidia since last year for roughly 10% of its inference computing needs due to speed concerns
- The $100 billion Nvidia investment in OpenAI has been delayed months past its expected completion date as negotiations continue
- OpenAI secured chip supply deals with AMD, Broadcom, and Cerebras Systems to reduce dependence on a single supplier
- Nvidia spent $20 billion to license Groq’s technology and hire its chip design team to strengthen its inference product lineup
- Both Sam Altman and Jensen Huang publicly dismissed reports of friction between the two companies
OpenAI has been exploring alternatives to some Nvidia chips for more than a year. The effort centers on finding hardware better optimized for inference tasks, when AI models produce responses to user queries.
The company needs improved speed for certain applications. Code generation and AI-to-AI communication require faster processing than current solutions provide.
OpenAI intends to obtain alternative chips for about 10% of its future inference operations. Several sources confirmed dissatisfaction with Nvidia’s existing hardware speeds for specific workloads.
$100 Billion Deal Timeline Extended
Nvidia announced plans in September to invest up to $100 billion in OpenAI. The deal was expected to close in weeks but has remained incomplete for months.
OpenAI’s shifting product plans have changed its computing needs. These modifications have prolonged negotiations with Nvidia.
The ChatGPT maker signed chip agreements with AMD, Broadcom, and Cerebras Systems during this timeframe. These suppliers offer products designed to rival Nvidia’s hardware.
Issues became apparent in OpenAI’s Codex product for code generation. Staff attributed some performance problems to Nvidia’s GPU-based infrastructure.
Sam Altman told reporters on January 30 that coding users value speed highly. He said the Cerebras partnership would help meet this demand.
Memory Architecture Requirements
OpenAI has focused on companies building chips with extensive SRAM memory. This memory is embedded directly in the chip’s silicon structure.
The configuration delivers speed advantages for chatbots serving millions of users. Inference requires more memory than training since chips retrieve data more often than performing calculations.
Nvidia and AMD GPUs depend on external memory. This design adds latency and slows chatbot interaction speeds.
Competing services like Anthropic’s Claude and Google’s Gemini use different hardware. They rely more on Google’s tensor processing units built for inference calculations.
OpenAI held discussions with Cerebras and Groq about faster inference chips. Nvidia’s $20 billion licensing deal with Groq ended OpenAI’s negotiations with that startup.
Licensing and Talent Acquisition
Nvidia hired Groq’s chip designers as part of the licensing agreement. Groq had discussions with OpenAI and attracted investor interest at a $14 billion valuation.
Jensen Huang called tension reports “nonsense” on Saturday. He stated Nvidia plans a large investment in OpenAI.
Nvidia said in a statement that customers choose its chips for inference because of performance and cost benefits. The company called Groq’s technology complementary to its development plans.
Sam Altman posted on social media following the reports. He said Nvidia makes “the best AI chips in the world” and that OpenAI expects to remain a major customer.
An OpenAI spokesperson said Nvidia powers most of the company’s inference systems. The statement highlighted Nvidia’s performance-per-dollar advantage for inference work.
OpenAI infrastructure executive Sachin Katti posted Monday that Nvidia remains “the core of our training and inference.” Both companies stressed their ongoing partnership despite the reports about chip alternatives and investment delays.
Get 3 Free Stock Ebooks
Discover top-performing stocks in AI, Crypto, and Technology with expert analysis.
- Top 10 AI Stocks - Leading AI companies
- Top 10 Crypto Stocks - Blockchain leaders
- Top 10 Tech Stocks - Tech giants
