Thursday, June 25, 2026
OpenAI unveils its first inference chip, built with Broadcom, joining the hyperscalers designing around Nvidia; Gemini folds computer use into its standard model; and an exploit breaks the SecureROM boot chain on Apple's A12 and A13.
OpenAI’s first silicon
OpenAI unveiled Jalapeño, its first custom chip, designed with Broadcom for AI inference. The company says early tests show better performance per watt than the parts it runs today; the figures are OpenAI’s own. The partnership dates to October 2025, and the chip is pitched as one piece of owning the stack from silicon to product. It puts OpenAI alongside Google’s TPUs, Amazon’s Trainium, and Microsoft’s MAIA: the largest buyers of AI compute now design around Nvidia rather than only buy from it, chasing both lower cost and an inference path they control.
Computer use goes native, open weights go geopolitical
Google added native computer use to Gemini 3.5 Flash, folding an agent that drives a screen into the standard model instead of a separate one, and says it added adversarial training against prompt injection plus enterprise gates that require confirmation for sensitive actions and halt suspicious tasks. Those safeguards are the company’s description, not an independent audit. Separately, Yann LeCun argued that open-source AI is the only path for most of the world, pitching a federated “Project Tapestry” in which countries contribute data without ceding it, and putting the economics starkly: a $200-a-month subscription against a $15,000 per-user cost he says proprietary models cannot sustain.
A break in the boot chain
A researcher published Usbliter8, an exploit of the SecureROM on Apple’s A12 and A13 chips. A mismatch in how the USB controller increments a DMA pointer yields a buffer underflow; on the A13 it overwrites critical memory to bypass Pointer Authentication and reach code execution at EL1, compromising the chain of trust from the device’s earliest boot code. SecureROM is fixed in silicon, so a flaw there cannot be patched on affected units. Apple was told before publication.
What to watch today
- A peer-reviewed critique in Nature argues Microsoft’s Majorana 1 evidence reflects quantum-dot artifacts and “basic Python errors,” not topological qubits; Microsoft disputes it, cites DARPA verification, and keeps its 2029 roadmap. Whether it answers the specific coding claims is the test.
- NVIDIA says its Rubin liquid-cooling design runs coolant at 45°C, cuts water use to near zero, and saves a 50-megawatt site more than $4 million a year; a vendor figure awaiting field data.
- The Guardian’s account of Indian factory workers filming themselves to supply egocentric training data, often uncompensated, is the human supply chain beneath the robotics push.