OpenAI Unveils First Custom AI Chip 'Jalapeño' to Challenge Nvidia on Inference

OpenAI Debuts In-House Inference Chip Jalapeño

OpenAI has released initial performance benchmarks for Jalapeño, a proprietary AI inference processor co-engineered with Broadcom. The custom silicon is designed specifically to execute trained models rather than handle training workloads, marking a major step toward reducing the company's reliance on external hardware vendors.

According to internal performance evaluations, Jalapeño outpaces Nvidia's top-tier accelerators in both execution speed and energy management:

• Performance and power: Operating at 700 watts, Jalapeño delivered response speeds up to 3.6 times faster than Nvidia systems rated at 1,200 watts, while achieving up to 1.9 times greater efficiency per watt.

• AI-assisted design: The hardware was developed in roughly nine months, utilizing OpenAI's upcoming Astra model and Codex to accelerate the design lifecycle.

• Internal deployment: The chip will not be sold to third parties. Richard Ho, OpenAI's VP of hardware, stated that internal compute demand fully consumes initial capacity. OpenAI will continue using Nvidia infrastructure for model training.

• Roadmap: Two subsequent silicon generations are currently under development. The first chips will roll out to data centers later this year, with broader production scaling slated through 2027.

Accelerated Understanding Aims to Model Physics Instead of Words

Caltech professor Anima Anandkumar and engineer Benedikt Jenik have established Accelerated Understanding, a venture focusing on physical AI. Rather than relying on traditional transformer architectures that predict the next token in a sequence, the startup's platform uses neural operators to track dynamic physical processes across three dimensions and time.

Key highlights of the initiative include:

• Extreme scale processing: Benchmark tests showed the architecture handling 5 trillion data points in a single computational run.

• Enterprise focus: Initial commercial applications concentrate on materials science for semiconductors, robotics, and severe weather forecasting.

• Strategic independence: Both founders previously declined executive roles and a 35% equity stake in Jeff Bezos-backed Prometheus to launch their independent venture.

Perplexity and Nvidia Introduce On-Device 'Portable Computer'

Perplexity, in partnership with Nvidia, announced Portable Computer, an on-device edition of its agentic AI software designed to prioritize local execution and user data privacy. The solution runs locally on Nvidia's $4,699 DGX Spark system without consuming cloud platform credits.

Users can select local base models, including Qwen 3.8 27B or Perplexity's own PPLX 27B, while retaining the capability to query more than 15 remote cloud models upon explicit approval. The launch coincides with reports of Nvidia in advanced talks to join a multibillion-dollar funding round valuing Perplexity above $30 billion.

Additional Industry Developments

• Unified Claude Memory: Anthropic introduced cross-environment persistent memory for Claude, allowing context to sync between standard chat interfaces and Cowork while capturing relevant topics in real time.

• Apple Mac Mini with M6: Apple unveiled an $899 Mac Mini powered by M6 silicon, positioning the machine for persistent, local agentic workflows with up to 4x faster execution speeds.

• Executive Departure at OpenAI: Chris Malone, head of data center operations at OpenAI, has departed the organization amid a broader series of leadership transitions.

• Specialized Gemini Editions: Google Cloud rolled out preview access to industry-specific Gemini Enterprise models tailored for legal and financial services sectors.

• Anthropic Market Projections: Ahead of a potential public offering, Anthropic is reportedly preparing to outline a total addressable market exceeding $30 trillion to prospective investors.