Researchers at UC Berkeley unveiled the Sky-T1-32B AI model, trained for under $450 using DeepSeek’s R1 architecture, achieving 89% on MMLU benchmarks, signaling a seismic cost shift in AI development.
A sub-$450 AI training breakthrough by UC Berkeley and DeepSeek challenges industry cost norms, leveraging efficiency gains to rival premium hardware solutions.
Berkeley Team Unveils Budget AI Benchmark
In a press release published Monday, researchers at the University of California, Berkeley revealed their Sky-T1-32B model trained using DeepSeek’s R1 architecture for just $450. The system scored 89% on the Multitask Language Understanding (MMLU) benchmark, approaching the performance of models requiring 500x greater investment.
TechCrunch analysis notes this follows DeepSeek’s January demonstration of a $5 million training run that matched capabilities of systems previously costing over $50 million. ‘We’re seeing compounding efficiency gains,’ said a Berkeley spokesperson, citing optimized parameter allocation strategies.
Cloud Providers Rethink Hardware Roadmaps
Industry analysts report major cloud providers are accelerating evaluations of cost-efficient AI training solutions. AWS and Google Cloud representatives confirmed ongoing tests with DeepSeek’s framework during earnings calls last week.
Nvidia’s data center GPU sales grew just 12% year-over-year in Q2 2025 compared to 34% growth in 2024, according to SEC filings. TechCrunch attributes this slowdown to emerging alternatives like R1 that enable training on consumer-grade hardware clusters.
Open Source Community Mobilizes
The Linux Foundation announced plans to integrate R1 architecture specifications into its AI Ethics Toolkit by Q4 2025. Meanwhile, GitHub repositories related to budget AI training models saw contributor activity spike 217% month-over-month.
DeepSeek confirmed via blog post that R1 documentation will remain open-access through 2026, though commercial implementations require licensing. Microsoft’s AI chief called the move ‘a pragmatic balance between innovation and sustainability’ in a LinkedIn statement Wednesday.
