News

OpenAI sidesteps Nvidia with unusually fast coding model on plate-sized chips

· Ars Technica

OpenAI has made a significant strategic move by launching its first production AI model to run on non-Nvidia hardware, marking a potential shift in the AI infrastructure landscape. The company released GPT-5.3-Codex-Spark, a specialized coding model that operates on Cerebras chips and delivers code generation at over 1,000 tokens per second—approximately 15 times faster than its predecessor. This development represents more than just a speed improvement; it signals OpenAI's effort to diversify beyond Nvidia's dominant GPU ecosystem. The new model is currently available to ChatGPT Pro subscribers ($200/month) through the Codex application, command-line interface, and VS Code extension, with API access being rolled out to select design partners. Codex-Spark builds upon OpenAI's full GPT-5.3-Codex model but has been specifically optimized for speed rather than comprehensive knowledge depth. The model features a 128,000-token context window and focuses exclusively on text-based coding tasks at launch. According to OpenAI's internal benchmarks, Spark outperforms the older GPT-5.1-Codex-mini on software engineering evaluations including SWE-Bench Pro and Terminal-Bench 2.0, while completing tasks significantly faster. The performance gains are substantial when compared to OpenAI's existing Nvidia-based infrastructure. Current models like GPT-4o deliver approximately 147 tokens per second, while o3-mini reaches about 167 tokens per second. However, Cerebras has demonstrated even higher capabilities, achieving 2,100 tokens per second with Llama 3.1 70B and 3,000 tokens per second with OpenAI's open-weight models, suggesting potential for further optimization. This release comes amid an intensifying competition in AI coding assistance. Previous testing revealed that Codex's speed had been a weakness, taking roughly twice as long as Anthropic's Claude Code to complete coding tasks. The new Spark model appears designed to address this competitive disadvantage directly. The broader implications extend beyond performance metrics. By partnering with Cerebras, OpenAI is exploring alternatives to the Nvidia-dominated infrastructure that currently powers most AI workloads. This diversification strategy could prove crucial as demand for AI compute continues to grow and companies seek to reduce dependency on single hardware vendors. While the model currently handles only text-based coding tasks, its specialized nature and impressive speed gains suggest a new approach to AI model deployment—creating purpose-built tools optimized for specific use cases rather than relying solely on general-purpose models. This trend toward specialization could reshape how AI companies develop and deploy their technologies, potentially leading to more efficient and cost-effective solutions for specific applications like code generation.

OpenAI has taken a notable step away from Nvidia's hardware dominance by launching GPT-5.3-Codex-Spark, its first production AI model designed to run on alternative chip architecture. The new coding-focused model operates on Cerebras processors and achieves remarkable performance speeds of over 1,000 tokens per second, representing a 15-fold improvement over previous iterations.

The strategic significance of this move extends beyond mere performance metrics. By diversifying its hardware partnerships, OpenAI is positioning itself to reduce reliance on Nvidia's GPU ecosystem, which currently dominates AI infrastructure. Sachin Katti, OpenAI's head of compute, emphasized the importance of this partnership, noting Cerebras as a valuable engineering collaborator in expanding platform capabilities.

Codex-Spark represents a focused approach to AI model development, prioritizing speed and efficiency for coding tasks over the broad knowledge base of larger models. The model features a 128,000-token context window and operates exclusively with text input at launch. This specialization allows for optimized performance in its target domain while maintaining practical utility for developers.

Access to the new model is currently limited to ChatGPT Pro subscribers, who pay $200 monthly for premium features. Users can interact with Codex-Spark through multiple interfaces including the dedicated Codex application, command-line tools, and Visual Studio Code extensions. OpenAI is also gradually expanding API access to selected design partners, suggesting a measured rollout strategy.

Performance comparisons reveal the substantial improvements achieved through this hardware transition. OpenAI's existing models running on Nvidia infrastructure deliver significantly lower token rates: GPT-4o processes approximately 147 tokens per second, o3-mini reaches 167 tokens per second, and GPT-4o mini operates at around 52 tokens per second. The 1,000 tokens per second achieved by Codex-Spark represents a dramatic leap in processing speed.

Interestingly, Cerebras hardware has demonstrated even higher capabilities in other contexts. The company has recorded 2,100 tokens per second with Meta's Llama 3.1 70B model and achieved 3,000 tokens per second with OpenAI's own open-weight gpt-oss-120B model. These figures suggest that the current Codex-Spark performance may represent conservative utilization of the available hardware capabilities.

The competitive landscape for AI coding assistance has intensified significantly, with speed becoming a crucial differentiator. Previous evaluations highlighted Codex's performance disadvantages, particularly in head-to-head comparisons with Anthropic's Claude Code, which completed coding tasks in roughly half the time. The new Spark model appears specifically designed to address these competitive weaknesses.

OpenAI claims superior performance on industry-standard benchmarks including SWE-Bench Pro and Terminal-Bench 2.0, which evaluate software engineering capabilities. However, the company has not yet provided independent validation of these performance claims, leaving room for third-party verification of the reported improvements.

The broader implications of this development extend to the AI industry's infrastructure evolution. As companies seek to optimize costs and performance while reducing vendor dependency, alternative hardware solutions like Cerebras chips may gain increased adoption. This trend could accelerate innovation in specialized AI processors and create more competitive dynamics in the hardware market.

The focus on specialized, task-specific models also reflects a maturing approach to AI development. Rather than relying exclusively on large, general-purpose models, companies are increasingly creating optimized solutions for specific use cases. This specialization strategy can deliver better performance and efficiency for targeted applications while potentially reducing computational costs.

Looking forward, the success of Codex-Spark could influence broader industry trends toward hardware diversification and model specialization. As AI workloads continue to grow and evolve, the ability to deploy models on diverse hardware platforms may become increasingly valuable for maintaining competitive advantage and operational flexibility.

Referenced links