This week's AI news cycle wasn't about a flashier chatbot — it was about the plumbing underneath the entire industry. Three stories that broke within the last two days show how the center of gravity in AI is shifting from model quality alone toward chips, compute distribution, and the human labor that built the systems in the first place.

OpenAI takes a swing at Nvidia's inference crown

At the Hot Chips conference, OpenAI unveiled benchmark results for Jalapeño, its first in-house inference chip, co-developed with Broadcom. According to independent testing firm SemiAnalysis, the chip reportedly outperforms both Nvidia's Blackwell and Rubin in throughput per watt and token latency, delivering 1.5x to 1.9x more AI work per watt at peak throughput with 1.7x to 3.6x lower end-to-end latency across models including GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.

The caveats matter as much as the headline. SemiAnalysis noted the comparison was "somewhat incomplete and unfair" because Jalapeño uses newer HBM4 memory, meaning Nvidia's Rubin platform, which also uses HBM4, is the fairer benchmark. Analysts also stressed this is an inference-only chip — it could reduce OpenAI's reliance on Nvidia for inference over time, but Nvidia GPUs remain important for training and other compute-intensive frontier workloads thanks to their programmability and mature software ecosystem. Deployment is modest at first: OpenAI plans a very small-scale deployment by the end of 2026, with a larger rollout following in 2027.

Amazon closes the book on "artificial artificial intelligence"

In a quieter but symbolically loaded move, Amazon announced it will shut down its Mechanical Turk crowdsourced labor platform on September 30, 2026, following an internal assessment of its programs and services. Launched in 2005, the service let companies farm out so-called Human Intelligence Tasks — data labeling and audio or video transcription to survey completion — to workers paid a few cents apiece, a setup Jeff Bezos famously described as "artificial artificial intelligence."

The closure isn't isolated: SageMaker Ground Truth and Amazon Augmented AI will close the same date, meaning Amazon is exiting the human data-collection infrastructure market entirely. The irony is hard to miss — a marketplace built to supply human judgment that computers couldn't replicate is being retired partly because AI models have advanced and a new crop of data labeling startups, including Scale AI, Mercor, and Prolific, have entered the market. Muddying the picture further, a 2023 study found that up to 46% of Mechanical Turk workers used AI models to complete tasks — meaning the "human" data pipeline had already been partially automated from the inside.

Moonshot AI wants a seat at America's cloud table

Meanwhile, China's Moonshot AI is reportedly negotiating with Microsoft, Amazon, and Google to host its Kimi K3 model on their clouds. Per Reuters, Moonshot is seeking up to a 30% share of revenue generated from K3-related services on Azure, AWS, and Google Cloud. If finalized, the deal could mark the first big revenue-sharing pact between a Chinese AI firm and a major US cloud company — notable given that China's leading AI models, often far cheaper than Western offerings, are gaining traction in the US despite national security concerns in Washington. Kimi K3's scale is also striking: it's a 2.8-trillion-parameter open-weight model, making it the largest publicly available model of its kind, which is precisely why cloud distribution matters — few organizations can run it themselves.

Taken together, these three stories describe an industry less obsessed with leaderboard bragging rights and more focused on who controls the chips, the compute, and the labor pipelines that make frontier AI possible at all.

Sources

AmazonArtificial IntelligenceDaily AI UpdateMoonshot AINVIDIAOpenAI