Open-Weight AI Price War | AI Field Notes #93

AnalysisA developer who wants a small model that plans, calls tools, and writes code without renting a frontier API now has another free option. IBM released Granite 4.2 on August 25 in 3B, 8B, and 30B sizes under the Apache 2.0 license (free to use, modify, and sell on), with weights on Hugging Face and Ollama. Every model can show its reasoning before answering and switch that reasoning off to save time. The 30B version scores 57 on SWE-bench Verified, a test of fixing real software bugs, and its tool use was trained by letting the model practice on its own. Open weights keep closing the gap with the paid tier.

AnalysisA Chinese lab is trying to get paid every time an American cloud runs its model. Moonshot AI is in early talks with Microsoft, Amazon, and Google to host Kimi K3, its 2.8-trillion-parameter model, in exchange for up to 30 percent of what Azure, AWS, and Google Cloud earn from services tied to it, Reuters reported on August 26. It would be the first big revenue-sharing deal between a Chinese AI firm and a major US cloud, and the sticking points are the usual ones for this pairing: how to split the money, who sees the data, and how token usage gets audited. Distribution, not weights alone, is where the money now sits.

AnalysisThe slowest part of running a model on your phone is shuttling data between memory and processor, and Samsung wants to shorten the trip. At the Hot Chips conference on August 25, it detailed LPDDR5X-PIM, low-power memory with small logic units built into the chip so some of the math happens where the data already sits, a design called processing-in-memory. Samsung measured inference running 2.28 times faster and pushing 3.01 times the token throughput against standard LPDDR5X, with eight times the internal bandwidth. It targets devices rather than data centers, pitched as a cheaper, cooler alternative to the scarce high-bandwidth memory that AI servers fight over.

AnalysisOpenAI's first in-house chip beat the hardware it has spent years buying. In benchmarks published August 25, its Jalapeno inference chip, co-designed with Broadcom, ran 1.5 to 1.9 times more AI work at peak than Nvidia's Blackwell systems and used less power per unit of work, confirmed by independent tests from the analyst firm SemiAnalysis. Inference is the cost of running a model for users, separate from training it, and it is where the money leaks daily. OpenAI still says it will keep buying Nvidia, and Jalapeno ships in low volume late this year. Nvidia's margin, the fattest in the industry, just got a crack in it.

AnalysisThe cheapest way to run a near-frontier model in production got cheaper. On August 26, Z.ai (the Chinese lab formerly called Zhipu) released GLM-5.3-Flash, a 320-billion-parameter mixture-of-experts model that only wakes 18 billion parameters per request to cut cost, under the permissive MIT license with weights on Hugging Face. It reads text, images, and video, holds a one-million-token context, and lands within half a point of Claude Opus 4.8 on Z.ai's own coding benchmark while charging $0.15 per million input tokens against Opus-tier pricing many times that. The open-weight race is now a price war.

AnalysisVoice bots are creeping from the call center into the rest of the workday. Ringg, a Bengaluru startup, added $10 million to its Series A on August 25, led by Peak XV Partners, bringing the round to $15.5 million. Its agents handle lead qualification, onboarding, support, collections, and renewals across voice, WhatsApp, and browser sessions, handing off to a human when they stall, for clients including CRED, Flipkart, and the insurance marketplace Policybazaar. The pitch is a familiar one dressed in sharper capability: automate the repetitive conversations, keep people for the hard ones. Whether that second half holds is the open question.

AnalysisDriverless freight quietly went from pilot to payroll. Gatik, which runs autonomous box trucks on short regional routes between warehouses and stores, raised $200 million on August 25, led by the Qatar Investment Authority and Koch Disruptive Technologies. The company says it has completed 85,000 fully driverless deliveries with no human in the cab, at 99 percent on-time, for Walmart, Kroger, and PepsiCo, and plans to top 100 trucks by year-end. It sidesteps the hard robotaxi problem by sticking to fixed middle-mile lanes a system can learn cold. The demo phase for this one is over.

AnalysisThe man who spent last year predicting AI would replace doctors and teachers now wants some jobs walled off from it. In interviews on August 26, Bill Gates proposed 'Human Reserved' occupations, work society deliberately keeps for people, naming childcare and jury service, and said up to 40 percent of jobs could start out protected, though he called that his ceiling. He also floated a robot tax: hire a person and you pay payroll tax, buy a machine that does the same job and you write it off, so the tax code quietly favors replacement. He added that he would stake his reputation on AI being bad for the job market, and that he would love to be proven wrong.

AnalysisThe marketplace that helped train modern AI is being closed by it. Amazon will shut Mechanical Turk on September 30, ending a 21-year run in which hundreds of thousands of people did tiny tasks, tagging images, rating text, checking answers, for a few cents each. It stopped taking new customers on July 30 and is closing its sister annotation services, SageMaker Ground Truth and Augmented AI, on the same day. The premise cracked after a 2023 study found many workers were quietly using language models to do the very labeling meant to prove human judgment. The tool taught the machine, and the machine took the tool.

AnalysisThe systems tuned to keep teenagers scrolling just cost Meta up to $16.68 billion. In a court filing on August 26, the company agreed to settle claims by 29 states that it engineered Facebook and Instagram to hook minors and hid what it knew. Beyond the money, the terms reshape the product: teen accounts default to a two-hour daily cap across Meta's apps, a nighttime lockout runs from midnight to 6am, and switching off the limit needs a parent's sign-off, all policed by an independent auditor for ten years. The recommendation engine was the asset; regulators just made it a liability.

AnalysisThe city that wanted to be a showcase for self-driving cars will wait. Fully driverless commercial robotaxis are now unlikely to reach London this year, with a realistic launch pushed to late 2027, as Waymo, the Uber-backed British firm Wayve, and China's Baidu wait on rules that Transport for London (TfL) has yet to publish. On August 5, TfL licensed Wayve's fleet of Ford Mustang Mach-E cars, but required a licensed human safety driver in every one. The technology keeps improving while the paperwork sets the pace, a reminder that deployment is a regulatory problem as much as an engineering one.

AnalysisThe Chinese lab that spooked Silicon Valley with a cheap model last year is now raising money like an incumbent. DeepSeek is finishing a roughly $8 billion round at a pre-money valuation near $74 billion, or 500 billion yuan, with signing expected late in August, according to Bloomberg. Backers include China's national AI fund, JD.com, and the battery giant CATL, and the company has started early work toward a listing on Shanghai's STAR Market, the mainland's answer to Nasdaq. A team once praised for doing more with less is now sitting on serious capital, which tends to change what a lab optimizes for.