1. OpenAI Shows Its First Chip Results (AI)
On August 25, OpenAI published the first performance results for Jalapeño, its custom chip for running AI models. In OpenAI's tests, systems built around the chip completed 1.5 to 1.9 times more work per watt and delivered 1.7 to 3.6 times lower end-to-end latency than Nvidia GB200 and GB300 systems. 1
On highly interactive workloads, Jalapeño delivered 2.1 to 4.1 times higher performance. The tests covered three public models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. 1
OpenAI also used its models to help design and program the chip. It plans to deploy very small volumes at the end of 2026 and expand through 2027, while continuing to buy Nvidia hardware. 2
Why it matters
Inference is a recurring cost every time a user or agent calls a model. A faster, more efficient chip could lower OpenAI's costs and reduce its dependence on Nvidia for one of the most expensive parts of the business.
Reality check
The results come from OpenAI and compare a future chip with Nvidia systems available today. Jalapeño is built for inference, not training, and it will not reach meaningful volume until 2027. OpenAI still needs Nvidia.
2. Nvidia's $96 Billion Quarter (Markets)
Nvidia reported $96.2 billion in quarterly revenue on August 26, up 106% from a year earlier. Data-center revenue reached $89 billion and net income was $59.7 billion. 3
The company expects $108 billion in revenue this quarter, excluding data-center compute sales to China. Its gross margin remained at 75%. 3
Nvidia was also in talks to buy Hugging Face for about $13 billion. Hugging Face is a central platform for sharing models, datasets, and AI tools, but the companies had not announced a signed deal by August 31. 4
Why it matters
Nvidia already controls much of the hardware used to train and run AI. Buying Hugging Face would give it a larger role in how developers find and distribute models, just as major labs begin building chips of their own.
Reality check
The Hugging Face talks were not a completed acquisition. A deal could still change or fail, and ownership by Nvidia could push developers and rival chipmakers toward other platforms. The earnings show that Nvidia's current business remains far larger than the threat from custom chips.
3. OpenAI Cuts Off Cursor (AI / Markets)
On August 28, OpenAI notified SpaceX that it intends to stop providing models to Cursor. The proposed cutoff date is November 12, and OpenAI said Cursor will not receive its future models. 5
The contract allows OpenAI to cancel after a change of control. OpenAI said it could not be confident that SpaceX would use its technology within the agreed terms. 5
Cursor said discussions were continuing. Anthropic said it planned to add more capacity for Claude models in Cursor. 6
Why it matters
AI coding tools depend on upstream model suppliers. Cursor changed owners, and one of those suppliers became a strategic rival. The dispute shows that buying the application does not secure the models underneath it.
Reality check
The cutoff is proposed, not final, and the companies still have time to reach a deal. Cursor already supports other models, so losing OpenAI would narrow the product rather than shut it down. The dispute is also tied to the personal conflict between Sam Altman and Elon Musk.
4. XPeng Raises $900 Million for Humanoids (Robotics)
XPeng's robotics unit entered agreements to raise more than $900 million at a post-money valuation above $6.3 billion. XPeng described it as the largest private funding round in China's embodied-AI sector. 7
The money will fund hardware, software, model training, production facilities, and international expansion. XPeng expects its IRON humanoid to enter mass production by the end of 2026, with early use in its stores and industrial sites. 7
The company is targeting a production rate of 1,000 robots a month by the end of the year. Commercial deliveries in China and other markets are planned for 2027. 8
Why it matters
XPeng can reuse car factories, suppliers, chips, and autonomous-driving software for humanoid robots. That gives it a more direct path to production than a robotics startup building each layer from zero.
Reality check
The funding agreements still need to close. The production target comes from XPeng, and the first deployments will be inside its own business. Mass production does not prove outside demand or reliable performance.