Two storylines collided today: OpenAI became a chip company, and Nvidia walked into earnings with Wall Street holding its breath. Everything else on the board — Apple, Google, Intel, Xiaomi — pointed the same direction.
OpenAI's Jalapeño takes direct aim at Nvidia
OpenAI unveiled its first custom silicon, an inference chip codenamed Jalapeño, developed with Broadcom according to India TV News. The company published early performance results claiming the chip runs AI models faster while drawing less power than the hardware the industry currently relies on. OpenAI's own framing was about as subtle as the codename: we made a chip, and it is fast.
The important word is inference. Jalapeño is not built to train models — it's built to run already-trained ones, the step where a model actually answers your question. That's the unglamorous, enormous, every-single-day workload, and it's where the economics of AI increasingly live. A company that serves billions of queries has an obvious incentive to stop renting someone else's silicon to do it.
Nvidia's earnings day, and a busy week besides
Nvidia reports second-quarter results Wednesday afternoon. Seeking Alpha framed it as the headline event of a heavy week for traders, with markets already moving in the run-up and options pricing implying a swing of roughly 6% on the numbers. That is a lot of money riding on one afternoon — and it landed on the same day a major customer showed off a chip designed to need less of Nvidia's.
Nvidia wasn't standing still. It announced the Jetson Orin Nano 2, a small single-board computer pitched as a "brain" for robots, drones and camera-based AI systems that need to think locally rather than phone home to a data center, with roughly double the AI performance of its predecessor. Separately, TradingKey reported Monday that Nvidia has begun full mass production of the Groq 3 LPX, an AI chip built on Groq technology, with rack systems based on it expected to ship later this year.
Hot Chips 2026: everyone is specializing
The annual Hot Chips conference, where designers explain new silicon in unusual technical detail, produced the day's clearest theme. Intel detailed Crescent Island, a data center GPU built specifically for AI inference — and notably, its bet is on memory rather than raw compute muscle, which tells you where Intel thinks the bottleneck actually is.
Google presented its eighth-generation TPU and confirmed it is splitting its homemade AI silicon into specialized parts, building two chips where it used to build one. Wall Street is beginning to attach a very large number to what that in-house program could be worth.
The rest of the lineup was gloriously mismatched: Intel, AMD, IBM and Waymo sharing a stage, with talks spanning mainframes, robotaxi computers and budget laptop processors. It's a reminder that "the chip industry" is not one race.
Apple pulls AI onto the desk
Apple announced the M6 and M5 Ultra, alongside refreshed Mac mini and Mac Studio desktops built around them, already available for preorder per The Verge. News9Live reports the M6 is Apple's first chip on a 2nm process — a meaningful manufacturing step, since smaller transistors generally mean more performance per watt.
Both chips are aimed at running AI work on the machine in front of you rather than in a distant data center, with the M5 Ultra positioned to handle large models locally. Between Apple's desktops and Nvidia's Jetson board, the same idea showed up twice today from opposite ends of the market: inference wants to move closer to the user.
And a new entrant
Xiaomi unveiled the Xring O3, an in-house 3nm system-on-a-chip for AI and image processing on its phones, manufactured by TSMC — a direct move on Qualcomm and MediaTek's turf.
The through-line
OpenAI, Google, Apple and Xiaomi all shipped or detailed silicon they designed themselves, and Intel built a GPU for one specific job. The era of buying one general-purpose chip for everything is quietly ending. Nvidia's report tonight will be read as a verdict on how fast that's happening.