Báo cáo AI Trends — 29/08/2026

1. New LLMs & Model Breakthroughs

• GLM-5.3-Flash Released under MIT License: Z.ai launched GLM-5.3-Flash, a 320B-parameter native Mixture-of-Experts (MoE) model (activating 18B parameters/token). It features a 1-million-token context window and a hybrid sparse-linear attention mechanism for fast multimodal processing.
• Alibaba Unveils Qwen3.8-Flash-Next: Alibaba rolled out Qwen3.8-Flash-Next, featuring Qwen Sparse Attention (QSA) and gated residual layers. Designed specifically to slash long-context inference costs, it supports native 262k tokens (extendable to 1M).
• DeepSeek-V4 Speculative Decoding Deployment: Widespread adoption of DSpark speculative decoding on DeepSeek-V4-Pro builds across enterprise inference clusters, alongside surging GitHub traction for its open-source DeepSeek Harness (`dsh`) agent runtime.

Sources: [1] [2] [3]

2. AI Companies & Mega Deals

• Nvidia Eyes $12.9B Hugging Face Acquisition: Reports confirmed Nvidia has reached an agreement to acquire Hugging Face for $12.9 billion (~86x revenue multiple). The deal represents Nvidias largest acquisition since Mellanox ($7B).
• Defensive Moat Against Custom Silicon: Analysts view the acquisition as a strategic play to dominate the open-source AI distribution pipeline, hedging against major clients (OpenAI, Meta, Google) spinning up in-house ASICs.
• Community Scrutiny Over Neutrality: Open-source AI developers and consortiums are voicing concerns over platform neutrality and ecosystem lock-in should the primary model registry fall under GPU vendor ownership.

Sources: [4] [5] [6]

3. AI Hardware & Silicon Infrastructure

• OpenAI Details Jalapeño Custom ASIC: At Hot Chips, OpenAI disclosed deep architectural specs for its custom inference chip Jalapeño (co-developed with Broadcom).
• Efficiency & Power: 700W peak rating running at 550W sustained, delivering 1.5x to 1.9x higher throughput per watt than GB200/GB300 systems.
• Memory Bandwidth: Equipped with 216 GB HBM4 memory pushing 15.4 TB/s bandwidth.
• Scaling: Supports 128 accelerators per local rack and scales up to 2,048 ASICs in 16-rack pod clusters for internal ChatGPT and API serving.
• Inference-Centric Datacenter Pivot: Hyperscaler capital allocation is aggressively shifting toward custom inference ASICs as multi-turn agentic workflows overtake standard training runs in total compute share.

Sources: [7] [8] [9]

4. AI Policy, Governance & Regulation

• EU AI Act Article 50 Enforcement Live: The European Commission and the EU AI Office have initiated active compliance audits for transparency rules:
• AI Interaction Disclosure: Chatbots and virtual agents must explicitly inform human users they are interacting with AI.
• Machine-Readable Provenance: Providers must embed cryptographic watermarking / C2PA metadata in all synthetic audio, video, and image outputs.
• Deepfake Labeling: Mandatory disclosures for synthetic media concerning matters of public interest.
• Grace Period Countdown: Existing generative AI models deployed prior to August 2026 have until December 2, 2026, to integrate compliant watermarking pipelines. Frontier labs (Anthropic, OpenAI) have already rolled out signed provenance stamps across all active endpoints.

Sources: [10] [11] [12]

Consolidated Sources List