OpenAI takes a swing at Nvidia
The day's biggest story is a challenger, not an incumbent. OpenAI unveiled Jalapeño, a chip it designed itself and built for one job: inference. That's the stage where a trained model actually generates an answer for you, as opposed to the enormous training runs used to create it in the first place — a distinction Livemint spelled out for readers encountering it fresh. OpenAI also published Jalapeño's first performance results, and the claim is blunt: it beat Nvidia's best on both speed and power consumption.
Treat that as a vendor's own numbers until independent testing lands. But the direction of travel matters more than the benchmark. The company that did more than anyone to make Nvidia indispensable is now shipping silicon designed to run its models without it.
Intel bets on memory, not muscle
OpenAI isn't alone in reading inference as the opening. At the Hot Chips 2026 conference, Intel detailed Crescent Island, a data center GPU aimed squarely at running trained models rather than training them. The pitch is a genuine departure from the usual spec-sheet arms race: Crescent Island leans on memory rather than raw compute horsepower. That's a sensible read of what inference actually demands, and a tacit admission that Intel isn't going to out-muscle Nvidia head-on.
Nvidia answers with volume
Nvidia's response to all of this is, essentially, scale — and today it came from four directions at once.
The company said Monday it has begun full mass production of the Groq 3 LPX, an AI chip built on Groq technology, with rack systems based on it due to ship later this year, per TradingKey. At Hot Chips, Nvidia opened up the technical details of Vera, the custom 88-core CPU anchoring its next-generation Vera Rubin systems, which Tom's Hardware reports leans on a new technique called spatial multithreading.
Demand for those systems is already showing up in writing. Bloomberg reports that Indian data center firm AM Intelligence has placed a binding order for 9,000 Vera Rubin systems — one of the largest publicly reported orders yet for the hardware, and a reminder that the AI buildout is no longer a US-and-China story.
Cisco, meanwhile, is deepening its Nvidia partnership to chase the same boom, according to an Axios exclusive also carried by Yahoo Finance. The expansion centers on Cisco's Secure AI effort and reflects where the market has moved: customers increasingly buy AI capacity by the rack, not by the chip.
And then there's orbit
The most improbable Nvidia headline of the day: SpaceX plans to launch satellites powered by Nvidia processors, in what Tom's Hardware describes as one of the first serious attempts to run heavy-duty AI workloads off the planet. The two companies have built a space-optimized version of the Vera Rubin NVL72 system, with a launch targeted for the fourth quarter of 2027, Elon Musk said in a post resharing Nvidia's announcement. The intended workload is Grok. File it under ambitious rather than imminent — Q4 2027 is a long way out — but the engineering work is apparently already done.
Apple moves the model to your desk
Apple went the opposite direction: not orbit, not the data center, but the machine in front of you. It announced two new processors, the M6 and the M5 Ultra, alongside refreshed Mac Mini and Mac Studio desktops built around them. The M6 is Apple's first chip on 2nm process technology, according to The Verge — a meaningful manufacturing milestone. News9Live frames both chips as aimed at running AI work locally rather than in a distant data center, with the M5 Ultra positioned to handle serious models on a desktop.
Hot Chips' odd couples
One last note from the conference floor. Hot Chips, where designers explain new silicon in unusual technical detail, drew a strikingly varied cast this year — Intel, AMD, IBM and Waymo all took the stage, with talks spanning mainframes, robotaxis and budget laptops. When IBM's mainframe engineers and Waymo's robotaxi team share a program, it's a decent sign the definition of "interesting silicon" has stretched well past the GPU.