Today's News
SemiAnalysis
SemiAnalysis analysis of Unitree's financial disclosures reveals the Chinese roboticist has slashed G1 pricing from $50K+ to $27.3K in 12–18 months while maintaining industry-leading margins, with sub-$20K pricing already appearing in select deals. This cost structure mirrors the BYD and DJI playbooks and appears structural rather than promotional. The company has also overcome early thermal limitations in its QDD actuator design—prior generations suffered 5-minute work cycles followed by 30-minute cooldowns—through iterations on motor efficiency, copper fill density, and active cooling, moving toward deployment-grade reliability.
Critically, ~250 Unitree robots are already deployed in labor settings with labor cost analysis showing hourly deployment costs approaching breakeven against human labor across realistic utilization and mean-time-to-failure scenarios. This marks a crossing of the threshold from research and hobbyist markets into real-world commercial viability at scale.
SemiAnalysis
SemiAnalysis InferenceX tracking reveals AMD's SGLang engineering team, led by HaiShaw, achieved a more than 100x performance improvement on DeepSeek V4 between Day 0 (April 25) and Day 26 (May 27) by replacing PyTorch fallbacks with native AITER, Triton, and TileLang kernels—moving from unusable 1–2 tokens/sec to production-viable performance on MI355X. This demonstrates the iterative software engineering required to unlock hardware potential post-launch and highlights the maturity gap between CUDA and ROCm ecosystems at model release.
Infrastructure economics favor dense GPU packaging: Nvidia's GB300 NVL72 (72-GPU rack with full NVLink connectivity) achieves $0.156 per million output tokens with multi-token prediction, dominating cost-per-token across all SKUs tested. The 72-GPU NVLink domain enables wide expert parallelism that keeps DeepSeek V4's MoE all-to-all operations on fast NVLink rather than slower scale-out fabric, whereas B200/B300 in 8-GPU islands hit throughput walls earlier. This validates the infrastructure economics of dense GPU packaging for large MoE models.
Techmeme
OpenAI has confidentially filed for an IPO with no set timing, citing preference for remaining private to pursue "certain strategic goals," though the filing gives the company the option to go public sooner if needed. This signals a major shift in OpenAI's long-term capital and governance strategy amid intense competition with Anthropic and others.
Separately, Google has ordered 3M+ TPUs from Intel for 2028 delivery, with Nvidia testing Intel's 18A process technology as both companies diversify from TSMC. Intel stock jumped 11% on the news, signaling a major shift in AI infrastructure sourcing and suggesting potential capacity constraints at existing foundries. This reflects the accelerating capital intensity of AI infrastructure buildout and competitive dynamics in custom silicon manufacturing.
Cognition's FrontierCode benchmark reveals coding is far less solved than prior evals suggested; Opus 4.8 scores only 13% on hardest tier
AINews
Cognition has launched FrontierCode, a new coding benchmark showing that state-of-the-art models remain far from solving real-world code tasks. Opus 4.8 scores only 13% on the hardest tier despite 50%+ performance on traditional SWE-Bench, with tasks built by open-source maintainers over 40+ hours each and graded on mergeable code quality rather than just test-passing. This directly challenges the narrative that agentic coding has been "solved" and establishes a new rubric standard for measuring code maintainability, regression safety, and real-world usability.
