NVIDIA, AMD, and Intel Rarely Join Forces: Nine Companies Form SGLang Open-Source Inference Alliance
Beating AI News Flash: At the inaugural SGLang Summit, NVIDIA, AMD, and Intel—normally fierce competitors—unusually aligned, forming the nine founding members of the SGLang Alliance alongside Google, Meta, Thinking Machines Lab, Cloudflare, Modal, and RadixArk. The alliance will help shape the future roadmap of the open-source AI inference engine, with over 2,000 in-person attendees at the summit.
SGLang is foundational software for large model runtime, responsible for processing user requests, GPU scheduling, memory management, and computation result reuse. It directly determines how fast and cost-effectively a chip can run AI models. Previously, SGLang already supported multiple hardware including NVIDIA, AMD, and Intel. The alliance’s establishment means chip manufacturers will engage more systematically in joint development, contributing engineering and testing resources, and driving performance optimization for their respective hardware. Alliance members hold two roadmap meetings annually, but final technical decisions and code merge rights remain with the SGLang maintenance team.
SGLang officials stated that 332 institutions have adopted the project so far. Meta also disclosed at the summit that migrating to SGLang cut GPU resource usage in its large-scale production environment by more than half compared to its previous inference solution. The summit also announced the development direction of SGLang v1.0, which focuses on improving stability and compatibility for production environments; the official version is still in development. Additionally, SGLang has launched public preview support for AWS Trainium chips, and is collaborating with Intel to optimize its CPUs, GPUs, and accelerators. Reinforcement learning framework Miles v0.2 has also been released.
10 minutes ago
U.S. Democrats are drafting a data center power reform plan, which may require operators to pay an additional $35 billion over the next 10 years.
After the Democratic Party won control of both chambers of Congress, it is preparing a new federal licensing rule reform that could accelerate federal environmental reviews for various energy projects, making it easier for data center operators to access power. Democratic Senator Sheldon Whitehouse stated that once the bill is passed, data centers’ days of free-riding on the power system will come to an end. Going forward, when connecting to the power grid, data centers must bear the additional costs incurred themselves and cannot pass these extra expenses onto utility companies, avoiding shifting the burden to other paying customers. A report from consulting firm Grid Strategies shows that federal power grid rule reforms alone could force data centers to pay an additional $35 billion over the next 10 years. Additionally, if states fully utilize other provisions in the bill, a total of $13.6 billion in extra fees could be collected.
10 minutes ago
Aguila Trades triggered stop-loss on BTC long positions worth $99.1 million, which had earlier been on the verge of liquidation.
According to TradingBeats' monitoring, over the past 10 minutes, trader Aguila Trades liquidated roughly $99.18 million in BTC long positions, logging a total loss of around $478,300 for a return of -19.2%. Earlier, the address held about $145 million in BTC long positions in the evening, just 1.8% away from its liquidation price. Following this position reduction, the address still holds approximately $29.49 million in BTC long positions, with an unrealized loss of roughly $228,400, and its liquidation price has fallen to $57,865.
10 minutes ago
Acala Network is set to shut down, with users required to complete asset migration by the end of 2026.
According to official announcements, the Acala community has voted to shut down its network, with block production expected to cease after February 10, 2027. The project team is reminding users to complete asset transfers via the Astar Portal cross-chain bridge by December 31, 2026: users holding aSEED, LDOT, or ACA on the Astar network must move their assets to Acala; users holding ASTR on the Acala network must transfer their assets to Astar. The Astar Portal is expected to remove the Acala cross-chain channel in early January 2027, after which a governance proposal will be passed to shut down the cross-chain connection between Astar and Acala. Any ASTR remaining on the Acala network at that time will be non-withdrawable.
10 minutes ago
AMD completes approximately $2.9 billion acquisition of Taalas: Embedding AI models directly into chips, enabling 17,000 tokens generated per second.
Beating AI News Flash: AMD announces the completion of its acquisition of Canadian AI chip company Taalas. The two sides reached an agreement in August this year, with a total transaction value of up to approximately $2.9 billion, of which up to $1.4 billion will be paid contingent on meeting agreed conditions.
Taalas has taken a completely different approach from traditional GPUs: when manufacturing its chips, it directly hardcodes AI model parameters into the circuits. Conventional GPUs need to repeatedly read model parameters from video memory, while Taalas integrates storage and computing, reducing data movement and eliminating the need for expensive high-bandwidth memory (HBM) and related complex packaging.
In February this year, Taalas launched its first chip, the HC1, designed specifically to run Meta’s Llama 3.1 8B model. According to official tests, with 1000 input and output tokens each, the single-user generation speed reached approximately 17,000 tokens per second, nearly 9 times that of Cerebras at the time. However, this design comes with a tradeoff in flexibility. Conventional GPUs can load different models at any time; while the HC1 supports LoRA fine-tuning, switching base models still requires modifying manufacturing masks and re-producing chips, a process Taalas says takes about two months.
The HC1 also uses aggressive 3-bit and 6-bit quantization, which while significantly compressing model parameters, also results in some loss of model quality. The current 17,000 tokens per second figure only applies to Llama 3.1 8B; the actual performance on larger models remains to be verified.
AMD plans to integrate this technology into its future AI accelerator products, pairing it with its existing Instinct GPUs, so that the dedicated chip handles high-frequency inference tasks for fixed models.
10 minutes ago