August 26, 2026 —
OpenAI publishes the first performance results for its Jalapeño inference chip
Today’s North America AI brief is led by OpenAI’s first public performance results for Jalapeño, its custom inference chip and the opening move in a multigenerational compute platform. The rest of the issue tracks capital flowing toward licensed creative AI, lower-power robotics hardware, IBM’s new open reasoning models, the widening Windows on Arm ecosystem, and a growing emphasis on governance and verification as AI enters production workflows.
OpenAI publishes the first performance results for its Jalapeño inference chip
OpenAI released its first measured results for Jalapeño, the company’s custom inference accelerator. On SemiAnalysis’ InferenceX benchmark, OpenAI says the chip delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the commercial comparison systems across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 workloads.
The company rates Jalapeño at 700 watts and says sustained power stayed at or below 550 watts in the tested workloads. OpenAI plans to ramp the chip in the coming months as the first generation of a broader platform spanning accelerators, memory, networking, serving software, and models. Independent coverage from The Verge and TechCrunch confirmed the publication and benchmark claims.
Stability AI raises $76 million from entertainment and technology backers
Stability AI announced a $76 million Series B that brings funding raised under CEO Prem Akkaraju to $232 million, including equity rounds and convertible notes. Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group, AMD Ventures, and Pacific Alliance Ventures joined the round. The company said it will invest in creative-production products, applied research, and professional services.
NVIDIA doubles entry-level Jetson inference performance
NVIDIA introduced Jetson Orin Nano 2, an 8GB robotics computer rated at 78 trillion operations per second. The company says it provides twice the inference performance of Jetson Orin Nano Super in the same form factor and uses 40% less power at matched performance in 15-watt mode. The module and developer kit are expected in the first half of 2027.
IBM releases Granite 4.2 reasoning models for enterprise agents
IBM released Granite 4.2, a family of open dense reasoning models in 3B, 8B, and 30B parameter sizes under the Apache 2.0 license. The models support native tool calling, thinking and non-thinking modes, and long-context extension to 512K tokens. IBM says the 8B and 30B variants received agentic reinforcement learning in sandboxed coding, terminal, and web environments.
Microsoft says Windows on Arm now spans more than 7,000 verified apps
Microsoft published a new Windows on Arm ecosystem update, saying its compatibility catalog now lists more than 7,000 verified apps and games. The company highlighted growing native support across security, education, productivity, and creative workloads, alongside upcoming NVIDIA RTX Spark PCs from Surface, ASUS, Dell, HP, Lenovo, and MSI and the established Qualcomm Snapdragon X platform.
Bill Gates calls for national and international frameworks for the AI transition
Bill Gates published a wide-ranging essay arguing that AI will disrupt work, education, information integrity, and public institutions faster than earlier technology shifts. He called for national bodies that can coordinate policy across agencies, an international organization for cross-border risks, stronger safety nets for displaced workers, and broader public participation in deciding how AI’s benefits and costs are distributed.
GitHub maps an evaluation loop for production LLM systems
GitHub detailed how it tests LLM-assisted secret scanning against product outcomes, safety constraints, production-like data, error categories, and online experiments rather than relying on benchmark scores alone.
AWS brings verifiable observability views into agent conversations
AWS showed how OpenSearch MCP Apps return both structured text and interactive trace, log, metric, and topology views inside compatible agentic IDEs.
Microsoft replaces release waves with an always-on AI roadmap
Microsoft will move Dynamics 365, Power Platform, and Dataverse into its continuously updated AI at Work roadmap starting in September, with RSS, CSV export, and MCP access.
SK hynix positions hybrid bonding as a foundation for denser HBM
A new SK hynix technical note explains how copper-to-copper hybrid bonding can reduce interconnect distance and support thinner, denser memory stacks beyond conventional micro-bumps.
loveholidays reports broader software delivery with Codex
An OpenAI case study says loveholidays grew AI-assisted code changes from 7% to 79% in a year while deployments rose 73% without expanding engineering headcount.
Runable raises $21 million for an outcome-focused business agent
Runable announced a $21 million Series A to expand an agent that builds products, operates business workflows, and runs customer-acquisition campaigns from shared context.