All briefs Older briefs
Artificial Intelligence Daily Brief · June 25, 2026 · preview

AI Compute Wars Intensify: From Nanostacks to Specialized Silicon, Hardware Advances Drive Agentic Capabilities

2 min read 5 sources Every claim cited

The race for advanced AI capabilities is accelerating across hardware and software layers, with major players unveiling specialized chips like OpenAI's Jalapeño and IBM's sub-nanometer nanostack architecture. Concurrently, model intelligence is advancing through agentic features—such as Google integrating 'Computer Use' into Gemini 3.5 Flash—while new benchmarks (RIFT-Bench) emerge to test these complex systems.

Model Releases

  • Google has integrated 'Computer Use' directly into Gemini 3.5 Flash, allowing the model to see, understand, and interact with computers, browsers, and mobile devices autonomously [25]. This capability allows developers to build agents for tasks like office automation or software testing across various environments using existing tools such as Search and function calls [25]. On the OSWorld benchmark, Gemini 3.5 Flash achieved a score of 78.4, surpassing models including GPT-5.4 mini (72.1) and Gemini 3 Flash (65.1) [25]. The feature is accessible via the Gemini API and the Gemini Enterprise Agent Platform [25]. [25]
  • In a hands-on benchmark comparing code generation, Snowflake CEO Sridhar Ramaswamy noted that GLM-5.2 is competitive with Opus 4.7 while costing significantly less [44]. The test covered 103 tasks requiring models to write working code for both DuckDB and Snowflake, where the two models achieved similar success rates of 66% (GLM-5.2) versus 67% (Opus 4.7) [44]. While Opus 4.7 was deemed the better model overall, GLM-5.2's pricing—$1.40 per million input tokens and $4.40 per million output tokens—presents a major cost advantage over Claude Opus 4.7’s $5.00 input and $25.00 output rates [44]. [44]

Hardware & Compute

  • OpenAI and Broadcom unveiled Jalapeño, an Application-Specific Integrated Circuit (ASIC) designed specifically for large language model (LLM) inference in data centers [46, 50, 40]. The chip was developed from initial design to manufacturing tape-out in nine months, leveraging OpenAI’s deep insights into LLM fundamentals and its roadmap of models and products [46, 50, 40]. Early testing indicates that Jalapeño will deliver performance per watt substantially better than current state-of-the-art, positioning it as the first step in a multi-generation compute platform expected to be deployed at gigawatt scale by the end of 2026 [46, 50]. [46][50][40]
11 more stories in today's full brief

Every claim cited to its primary source.

Log inCreate free account

Sources

  1. 25The Decoder · 2026-06-25 — Google bakes computer control directly into Gemini 3.5 Flash, letting the model see and operate your screen
  2. 40Ars Technica · 2026-06-24 — OpenAI and Broadcom announce chip designed for LLM inference at scale
  3. 44The Decoder · 2026-06-24 — Snowflake CEO finds GLM-5.2 competitive with Opus 4.7 at a fraction of the cost
  4. 46The Verge · 2026-06-24 — OpenAI reveals its first AI processor: Jalapeño
  5. 50OpenAI · 2026-06-24 — OpenAI and Broadcom unveil LLM-optimized inference chip