The Signal — August 26, 2026

Two of today's stories come from OpenAI, and they sit at opposite ends of the same company: the silicon it now builds for itself, and a Russian operation that used its chatbot to invent an academic institution. The third is Amazon setting a closing date for the platform that supplied this field with its first human labor.

OpenAI's first custom chip posted its numbers, and they beat Nvidia's

On August 24 we covered SemiAnalysis open-sourcing AgentX and its InferenceX suite, and noted that the firm's comparative chip results sat behind a paywall, so we would not report figures we could not check. Some of those figures are now public. At the Hot Chips conference, OpenAI presented the first benchmarks for Jalapeño, the inference accelerator it built with Broadcom, measured on that same suite.

Across three tested models, GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, OpenAI reports 1.5x to 1.9x more work per watt at peak throughput and 1.7x to 3.6x lower end-to-end latency than the best commercially available systems, rising to 2.1x to 4.1x on interactive workloads. On GPT-OSS the chip reportedly reached about 1,400 tokens per second per user. It handles inference only, and it is a general LLM accelerator, not one tuned to OpenAI's own models.

OpenAI supplied those numbers and SemiAnalysis verified some of the runs on-site, which is a different thing from an independent test. Blackwell is also the flattering comparison. SemiAnalysis says the fair one is Vera Rubin, since both parts use HBM4, and on total cost of ownership per token those two land roughly even. Rubin is shipping to customers while Jalapeño reportedly has not moved past engineering samples, and Nvidia and AMD have posted results on larger models nobody has run on Jalapeño yet. Cutting the other way, the chip hit these figures without multi-token prediction or speculative decoding, optimizations some of the comparison systems did use.

The timeline is the part worth keeping. Design work started in mid-2024, the chip taped out in November 2025, and OpenAI says nine of those sixteen months covered first design to finished blueprint, with its own models helping on design and optimization. SemiAnalysis wrote that "the CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon." That is a firm that got invited into the lab, so weigh it accordingly, but the claim underneath is checkable and it has nothing to do with watts. Nvidia's durable advantage was supposed to be the cost of moving software anywhere else. A team doing this for the first time appears to have paid that cost in under two years.

Sources: The Decoder · SemiAnalysis · TechCrunch · ServeTheHome

Amazon set a closing date for Mechanical Turk

AWS will close Mechanical Turk on September 30, ending a twenty-one year run. The notice explains nothing: "We regularly evaluate our programs, tools and services and make adjustments based on those assessments. Following an assessment, we've made the decision to close AWS Mechanical Turk." Launched in 2005, MTurk matched workers with Human Intelligence Tasks paying a few cents each, labeling data, transcribing audio, filling out surveys and whatever else was too small to automate. Jeff Bezos called it artificial artificial intelligence. The name came from the eighteenth-century chess automaton that was operated by a chess master hidden inside the cabinet.

It would be tidy to say AI killed it, and Amazon has not said that. What is on the record: the platform served more than 500,000 workers at its peak, Amazon stopped accepting new customers last month, and Krista Pawloski of the data-worker advocacy group Turkopticon told CNBC that Amazon put fewer resources into MTurk as Scale AI, Mercor and Prolific moved in on the same work. There is also a 2023 study from Swiss researchers finding that up to 46% of MTurk workers were using AI models to do the tasks, which means the platform built to supply human judgment had been quietly handing model output back to the people buying it.

The workers get five weeks. Pawloski started turking in 2008 while on maternity leave, then went full time in 2012 after losing her job so she could care for her son, and she is describing people who assembled the corpora these models learned from at a few cents a task. Some insurance and travel firms still route data work through the platform and are now looking for somewhere else to put it. "There's some people that still pretty much still do it full time," she told CNBC. "They're concerned now."

Sources: CNBC · Amazon Mechanical Turk · AWS documentation · Techmeme

A fake think tank, built with ChatGPT and other people's papers

OpenAI banned a cluster of ChatGPT accounts it says very likely originated in Russia, used to promote the International Burke Institute, a self-described expert community listing a street address in Israel. The site was registered in February 2025 and published a sovereignty index that ranked Russia favorably against Western countries, scoring the Russian military half a point above the American one.

The model was doing the least interesting job in the operation. Of 36 expert-attributed IBI articles published between September 2025 and May 2026, OpenAI found 34 were copied from elsewhere, several with invented credits: a Cambridge University Press article reassigned to a University of Nottingham professor, a Migration Policy Institute piece credited to an Australian food chemistry academic. The site also listed Francis Fukuyama and Noam Chomsky. ChatGPT handled the outer layer, reached over VPNs because OpenAI does not serve Russia, prompted in Russian, asked for English posts with the linguistic tells stripped out. Those went to X, LinkedIn, Facebook, Substack and Telegram, including a German-language channel called Lahme Ente that attacked Ukraine, the EU and the German government.

Manufacturing the appearance of institutional authority, a masthead of famous names and a body of published work with an index attached, used to be the expensive part of an operation like this. It is not anymore. By OpenAI's own accounting this one still did not work especially well: individual posts drew very few views and the official IBI accounts had low follower counts, though the linked Telegram channels each reached ten to twenty thousand. OpenAI rates it category three of six on the Brookings Breakout Scale, meaning cross-platform spread with early signs of reaching real people, and says it could not establish the relationship between every named individual, the institute and the operators. The defense that caught this is the boring one, checking whether the papers were written by the people whose names are on them.

Sources: The Decoder · Anadolu Agency · Security Online

On the Editor's Desk

Liquid AI released Pipette, an open benchmarking platform for running models on phones and laptops, with Artificial Analysis validating the methodology. It is a real artifact with real data behind it, over a thousand configurations of model, quantization, runtime, device and context length. We held it anyway. On August 20 we skipped an Artificial Analysis benchmark for reading more like a buyer's guide than news, and this is the same shape from the same corner. It becomes a story the first time somebody disputes what it measures.

IBM shipped Granite 4.2, a family of Apache-licensed reasoning models with an agentic training stage, reporting 57.00 on SWE-Bench Verified for the 30B. Every one of those numbers is IBM's, run by IBM, and nobody outside has reproduced them yet. We also left the Taiwanese indictments over AI server exports where they were yesterday, since an indictment is still an allegation and the export-control story underneath it ran here on August 13.