AI News

OpenAI's Jalapeno Chip Beats Nvidia, Apple's 2nm M6: General AI News for August 26, 2026

OpenAI's Jalapeno chip outperforms Nvidia, Apple's 2nm M6 appears, and a Transformer-free physics AI arrives—today’s general AI news.

Daniyal

26 Aug 2026 · 4 min read

OpenAI finally put real numbers on its first custom chip, and they are not normal. The inference accelerator actually beats Nvidia's next-generation silicon on performance per watt. That land grab is the lead story of a busy Tuesday that also brought Apple's first 2-nanometer chip, a physics AI that abandoned Transformers entirely, and the quiet end of Mechanical Turk. Here is the AI news from the last 24 hours that actually moved things.

Top story: OpenAI's Jalapeño chip outruns Nvidia Rubin on efficiency

SemiAnalysis published a deep dive on Jalapeño, the inference ASIC OpenAI taped out with Broadcom, after details surfaced at the Hot Chips conference on Tuesday. The B0 stepping delivers 13.4 petaflops of MXFP4 compute at 700 watts. Nvidia's upcoming Rubin pulls 900 to 1,150 watts for its class of work, so Jalapeño wins the efficiency comparison that matters for data-center economics.

The chip pairs HBM4 memory at 15.4TB/s and reportedly hits about 700 tokens per second per user on DeepSeek R1 and roughly 1,400 tokens per second per user on OpenAI's GPT-OSS line. OpenAI first introduced the part in June, claiming inference costs roughly 50% lower than comparable Nvidia GPUs. The Tuesday benchmark disclosure is the strongest sign yet that the custom-silicon bet is working.

Why it matters: this is the strategic endgame of the AI buildout. OpenAI, Google, Amazon, Meta and now Anthropic are all trying to stop paying Nvidia's toll on inference. A first-generation part beating the incumbent on efficiency compresses the whole roadmap.

Apple's first 2nm chip and the M5 Ultra

Apple unveiled the M6, its first chip built on a 2-nanometer process, alongside the M5 Ultra, its first quad-die design. The M6 packs a 12-core CPU, a 12-core GPU and a dual 16-core Neural Engine, with up to 32GB of unified memory at 170GB/s. It debuts in a new Mac mini.

The M5 Ultra is the headline for AI work. It combines up to a 36-core CPU and an 80-core GPU with up to 512GB of memory at 1.2TB/s, and Apple claims roughly 4.5 times the AI compute of the M3 generation. It ships in a refreshed Mac Studio. On-device inference keeps getting cheaper and faster for builders who run models locally.

A physics AI that skips Transformers entirely

Anima Anandkumar and Benedikt Jenik, two Caltech-based researchers, unveiled Accelerated Understanding Inc, an enterprise physics AI built on neural operators rather than the Transformer architecture behind ChatGPT. Reuters reported the duo walked away from a roughly $2 billion offer tied to Bezos-backed Project Prometheus to start it.

In tests the system processed about 5 trillion data points in a single prompt, roughly five million times the input flagship language models handle. The bet is that AI rooted in physical systems can solve problems natural-language models simply are not shaped for: materials, weather, chemistry, engineering simulation. It is a contrarian shift worth watching from the people who pioneered neural operators.

OpenAI bans a Russian influence campaign run through ChatGPT

OpenAI said on Tuesday it permanently banned a cluster of ChatGPT accounts originating in Russia that were promoting a fabricated "expert community" called the International Burke Institute. Investigators traced AI-generated posts to a site carrying copied and misattributed academic work. CNBC confirmed the takedown. The case shows how model-abuse investigations now start from AI-generated text fingerprints, not just network infrastructure.

OpenAI loses its data-center chief

CNBC, WSJ and Bloomberg confirmed that Chris Malone, OpenAI's head of data centers, has left. TechCrunch noted the timing stands out because Malone oversaw execution of the company's data-center strategy, the very buildout that the new Jalapeño chip is meant to anchor. It is the latest in a lengthening string of high-level departures at the AI lab.

AWS pulls the plug on Mechanical Turk

After 21 years, AWS is shutting down Mechanical Turk on September 30, the crowdsourcing marketplace Jeff Bezos once called "artificial artificial intelligence." The platform stopped taking new customers on July 30. It had been crowded out by AI-native data-labeling players like Scale AI, Mercor and Prolific, and a 2023 study found many Turk workers were quietly routing tasks through LLMs, undermining the human-judgment premise. The shutdown is a clean symbol of how much the labor model of AI changed in two decades.

Taiwan indicts nine over Nvidia server smuggling to China

The Wall Street Journal reported Taiwanese prosecutors indicted nine people, including one Nvidia Taiwan employee and two Super Micro Taiwan staff, for making 130 Nvidia B300 servers appear destined for a rented local facility. Prosecutors say 74 servers were rerouted to Chinese customers through direct shipments and trans-shipments via Indonesia, Japan and Hong Kong before customs stopped the rest. Seven defendants face up to five years in jail on breach-of-trust and forgery charges tied to US export controls.

The takeaway

The clearest signal of the day: the silicon is shifting under everyone's feet. OpenAI, Apple and a new generation of physics-first builders are all attacking the assumption that one vendor supplies the compute and one architecture does the thinking. If you build with AI, this is the week to watch cost curves, not just capabilities.

Follow all of our daily AI coverage via the AI news tag on this blog.

Related reading from earlier coverage, with the numbers behind each one: