The US debt of $39 trillion has raised long-term concerns, with analysts saying the risk of an unsustainable fiscal path is rising.
Related
US debt tops $40T stoking debate on what it means for Bitcoin
Analysts say the $40 trillion debt milestone could bolster Bitcoin’s long-term case, even as Treasury yields, dollar strength and liquidity remain key near-term drivers.
Meituan open-sourced its trillion-parameter large-scale model LongCat-2.0, and simultaneously released the inference code for domestically developed Chinese card processors.
According to Beating's monitoring, Meituan has officially open-sourced its trillion-parameter large-scale model, LongCat-2.0, with a total of 1.6T parameters and an average activation of approximately 48B, designed specifically for real-world agentic coding tasks. Architecturally, it innovatively introduces LongCat sparse attention and N-gram embedding. The former reduces fragmented memory access through flow-aware indexing and hierarchical indexing, accelerating training and inference with millions of contexts; the latter, while achieving nearly 97% sparsity in MoE, invests 135B parameters into the embedding layer, balancing parameter gains and structural stability. Post-training employs multi-teacher online distillation, categorizing experts into Agent, Inference, and Interaction types, seamlessly integrating them on a domestic computing power cluster through the MOPD architecture. As the industry's first trillion-parameter model to complete inference on a 50,000-card domestic computing power cluster, LongCat-2.0 validates the mature capability of domestic chips to handle complex large-scale model tasks. To address the multiple limitations of domestically produced Chinese chips in terms of memory, bandwidth, and interconnects, Meituan has made breakthroughs in three areas: model, chip adaptation, and deployment. At the model level, ScMoE leverages the core control capabilities of domestically produced chips to achieve physical core-level parallelism for Dense and MoE branches, combined with KV-cache partitioning to alleviate the pressure on ultra-long context memory. At the chip adaptation level, Super Kernel reduces operator startup overhead, and Weight Prefetch hides I/O latency, maximizing hardware utilization under constrained conditions. At the deployment level, PD separation is adopted to balance TTFT and TPOT, along with asynchronous Expert-Parallel load balancing to solve load unevenness under high EP (efficiency level). This open-source release simultaneously provides multiple precision versions, including BF16, FP8, and INT8, and fully opens up inference results optimized for domestic computing power, aiming to enable existing domestically produced cards and even older cards to smoothly deploy trillion-model inference services. --------------------------------- Click the original link below to join the Beating · Lark AI news channel and monitor global AI hot topics and news 24/7.
Kingsoft Cloud accelerates GPU computing power construction, securing a 10 billion yuan budget from Xiaomi and a multi-billion yuan long-term contract from Alibaba.
According to a report by Jiemian News, as reported by Mars Finance, Kingsoft Cloud will accelerate the construction of its GPU computing power clusters in the second half of the year to meet the explosive growth in computing power demand from leading clients. Xiaomi's GPU computing power requirement for Kingsoft Cloud has been upgraded from a 10,000-GPU cluster, with the related budget increasing significantly from the initial nearly 4 billion yuan to over 10 billion yuan. In addition, Alibaba's big data model team has signed a 5-year computing power leasing contract with Kingsoft Cloud, involving more than 3,000 eight-GPU servers. Based on the contract price, the annualized revenue after full delivery will exceed 4 billion yuan. To meet the surging customer demand, Kingsoft Cloud has increased its 2026 capital expenditure plan to 15 billion yuan, with a full-year revenue target of 12.5 billion to 13.5 billion yuan. It is understood that due to tight upstream supply, Kingsoft Cloud is currently only accepting long-term contract customers for 3 to 5 years for its GPU computing power, and some orders are facing delivery delays. Due to concerns about the risk of asset impairment from stockpiling high-priced computing cards, Kingsoft Cloud is currently suspending its aggressive expansion of hardware, anticipating that computing hardware prices may reach a turning point in the third quarter of this year.
Michael Saylor: If Bitcoin's long-term price increase exceeds 3.3%, BTC capital gains could fund STRC dividends indefinitely.
According to ChainCatcher, Bitcoin Treasuries.NET posted on the X platform that Strategy's Michael Saylor stated that if Bitcoin's long-term price increase exceeds 3.3%, BTC capital gains can fund STRC dividends indefinitely. Even if BTC's annual price increase is 0%, Strategy will still have dividend funds for 31 years.
Vanguard Group publicly advertised for a head of digital assets, having explicitly stated that crypto assets were inconsistent with its long-term investment philosophy.
According to Mars Finance, Vanguard Group is hiring a Head of Digital Assets for its Personal Wealth business. The job requirements include over 10 years of relevant experience, a deep understanding of digital assets (tokenization, stablecoins, custody, settlement, etc.), and innovation and risk management capabilities in a regulatory environment. This position will be responsible for developing Vanguard's strategy, roadmap, and implementation in the digital asset space, including assessing digital asset capabilities, product development, operating models, and cross-functional collaboration with product, technology, operations, risk, legal, and compliance departments. It will also require representing Vanguard in external communications with industry players, regulators, and clients. It is understood that Vanguard began allowing brokerage clients to trade crypto ETFs and mutual funds last December, but the company has explicitly stated that it has no plans to launch its own crypto investment products, believing that digital assets are still inconsistent with its long-term investment philosophy.
MIRA, an open-source 5B multiplayer world model, uses DINOv3 representation to mitigate long-term drift and can simulate 2v2 battles in Rocket League in real time.
According to Beating, AI research firm General Intuition, in collaboration with French AI lab Kyutai and Epic Games, has launched MIRA, a multiplayer interactive world model. As a generative game simulator supporting real-time multiplayer interaction, MIRA can simulate 2v2 battles in Rocket League in real time, based solely on historical footage and player button presses, without requiring a physics engine, rendering engine, or explicit 3D representation for inference. Unlike the "decoupling of logic computation and image rendering" approach adopted by companies like Odyssey, MIRA takes a generative simulation approach based on video latent space. MIRA boasts 5 billion parameters, and its core design builds the latent prediction space on a frozen general-purpose visual encoder, DINOv3-L. Leveraging pre-trained visual features, the generated latent states can more stably fall within the effective representation space, significantly mitigating image drift and divergence during long-term prediction. For multi-screen alignment, MIRA stitches the latent images from four player perspectives into a unified grid, enabling spatial attention mechanisms to operate naturally across viewpoints and improving the spatial consistency of vehicles, the soccer ball, and key events across multiple perspectives. The Action Dropout introduced during training also helps the system complete the game behavior of vehicles not controlled by commands when parts of the motion flow are missing. Currently, MIRA can run in real time at 20 frames per second on a single NVIDIA B200 graphics card. The team has open-sourced the training and inference code and released the Rocket Science dataset, which contains 1,000 hours of matches, approximately 4,000 hours of video, motion flow, and physics data from four perspectives; the complete training of the model used approximately 10,000 hours of clean match data.