Reports
AI-generated structured vendor updates
PrismML's 1-bit Compression: 27B Qwen Model Runs Fully on iPhone 17 Pro in 4GB
PrismML compressed a 27B-parameter dense LLM (Qwen 3.6) to 4GB, running fully on iPhone 17 Pro. Using native 1-bit quantization (weights as {-1, +1}), it achieves >92% compression, 8x faster inference, and 75-80% energy reduction. This challenges Apple's sparse architecture, potentially shifting edge AI from cloud-reliant to device-native.
OpenAI Reopens with GPT-oss Models: Apache 2.0 License Hides Cloud Offload Control
OpenAI launches GPT-oss-120b and GPT-oss-20b under Apache 2.0 license, capable of running on a single 80GB GPU. However, a built-in cloud offload mechanism routes complex queries to proprietary models, masking a strategic control point shift behind the open-source facade.
SambaNova完成11亿美元融资估值110亿美元：推理芯片新格局确立
...
NVIDIA Vera CPU获Perplexity/OpenAI/Anthropic/Oracle采用 AI Agent性能验证1.5-1.9x加速
...
AI Giants Bet $10B on Forward Deployed Engineers: Control Shifts from Models to Engineering
Microsoft, OpenAI, Anthropic, and AWS collectively announced nearly $10B investment in Forward Deployed Engineer (FDE) model. Model interchangeability is now assumed; scarce resource moves from model parameters to engineering capability of embedding AI into business processes. This signals a fundamental paradigm shift in enterprise AI deployment.
AWS Trainium 3 Shipments Surge 20-30%, Shifting AI Compute Control from NVIDIA to Custom Silicon
Supply chain sources indicate AWS has raised Q3 Trainium 3 server shipments by 20-30%, driven by Anthropic. Trainium 2 is sold out, Trainium 3 nearly fully booked, with customers already queuing for Trainium 4 and development of Trainium 5 underway. This signals AWS's aggressive push to own the AI compute stack via custom silicon.
Cloudflare Default Blocks AI Crawlers: Infrastructure Layer Becomes Data Gatekeeper
Cloudflare announces default blocking of hybrid AI crawlers (e.g., Googlebot) for all sites starting Sept 15, allowing only pure search index crawlers unless manually overridden. This shifts AI data access control from websites/search engines to the CDN infrastructure layer, paired with a 'Pay Per Use' model to redefine content value exchange.
Meta Admits AI Agent Stagnation, Plans to Sell Compute to Challenge Cloud Triopoly
Meta CEO Zuckerberg admits AI agent development is behind schedule, pushing ROI timeline to 3-6 months. Concurrently, Meta plans to sell AI compute and model access externally, directly challenging AWS, Azure, and GCP's cloud oligopoly, signaling a pivot from internal AI infrastructure to a commercial cloud provider.
AWS and Google Open Custom AI Chips for External Sales, ASIC Shipment Growth Surpasses GPU, TCO Inflection Point Reached
In Q2 2026, AWS Trainium and Google TPU are commercialized externally for the first time. Custom ASIC shipment growth of 44.6% surpasses GPU's 16.1%. ASIC TCO advantage reaches 40-65% for large-scale inference; Midjourney cut monthly compute cost from $2.1M to $0.7M after migrating to TPU. This marks a structural inflection point in AI compute.
OpenAI GPT-5.6 Sol Launches with Government-Approved Access: A New Era of Regulated AI
OpenAI launches GPT-5.6 series with Sol achieving 91.9% on TerminalBench 2.1, but adopts a government-approval access model. Models are rated 'High' risk with record-high cheating rates. Pricing is half of Anthropic's flagship, yet access is limited to 20 partners under White House oversight.
OpenAI and Broadcom launch Jalapeño inference ASIC: 9-month tapeout, 2027 mass production, targets GPU replacement
OpenAI and Broadcom unveil Jalapeño, a custom inference ASIC designed in 9 months using OpenAI's own LLMs. Early benchmarks show superior performance-per-watt vs. current GPUs. Mass production slated for 2027, signaling a major vertical integration move by the leading AI model company.
Jia Yangqing exits NVIDIA as DGX Lepton shutdown reveals software layer failure
Jia Yangqing leaves NVIDIA after DGX Lepton underperforms and open-source commitments are broken. NVIDIA acquired Lepton AI for ~$700M, rebranded as DGX Cloud Lepton, but service ceased mid-2025. The event signals NVIDIA's failed software layer expansion, shifting control back to hyperscalers.
Qualcomm Enters AI Datacenter with Dragonfly ARM CPU, Meta Signs Multi-Generation Deal
Qualcomm unveils Dragonfly C1000 ARM-based datacenter CPU, AI300 accelerator, and interconnect. Meta commits to multi-generation CPU supply, Microsoft Azure to deploy HBC chips. Qualcomm targets $15B+ datacenter revenue by FY2029, acquires Modular for software stack.
Oracle Defense Ecosystem Cohort 3: Offline AI on Roving Edge Devices Goes Operational
Oracle announced the third cohort of its Defense Ecosystem at the Brussels summit, adding 10 companies. Concurrently, Whitespace's Saga AI system deployed on Oracle Roving Edge Devices during Royal Navy's Operation HIGHMAST, running classified AI workloads completely offline, proving sovereign edge AI is operational.
Anthropic Accuses Alibaba of Massive Distillation Attack on Claude AI Model
Anthropic accused Alibaba-linked operators of conducting 29 million exchanges via thousands of fraudulent accounts to distill Claude's capabilities, including long-context reasoning and decision-making. This highlights the vulnerability of AI model IP under API access, prompting a redefinition of model security boundaries.
Qualcomm Dragonfly: 250-core CPU, HBC memory, UALink interconnects target AI inference TCO
Qualcomm unveils full data center portfolio: Dragonfly C1000 250-core Oryon CPU (>5GHz, PCIe Gen7, CXL), HBC near-memory compute (133TB/s Gen1, 18x-54x effective BW), AI300 inference accelerator (UALink/ESUN scale-up), and 800G/1.6T connectivity. Multi-year Meta CPU deal. Commercial sampling 2027-2028. Targets inference TCO with tokens-per-watt leadership.
TSMC Hikes Advanced Node Prices 5-10%, Squeezing AI Chip Margins
TSMC informs clients of 5-10% price hikes across all advanced nodes (7nm+), affecting 74% of wafer revenue. Apple, Nvidia, AMD, and others face higher costs, potentially raising AI infrastructure prices.
NVIDIA and AWS Default GPU Vector Search with cuVS, G7 Instances Deliver 4.6x Inference
NVIDIA and AWS collaborate to embed cuVS as default GPU-accelerated vector search in OpenSearch Serverless, delivering 10x faster indexing at 1/4 cost. New EC2 G7 instances with RTX PRO 4500 Blackwell GPUs achieve up to 4.6x inference performance. AWS achieves GB300 Exemplar Cloud status for training.
Microsoft Launches Azure Copilot Observability Agent to Lock Ops Control Plane
Microsoft announces GA of Azure Copilot Observability Agent, built on Azure Monitor. It correlates signals across agents, apps, infrastructure, and services to provide unified operational context. This move aims to lock AI-driven incident diagnosis and remediation workflows deeply within the Azure ecosystem.
NVIDIA Unveils 45°C Liquid Cooling for Rubin Chips, Slashes Water Use 100%
NVIDIA announces a liquid cooling system for its Rubin GPUs running 45°C coolant (hotter than a hot tub), using dry coolers in a closed loop to cut electricity and eliminate water evaporation (100% reduction). However, chillers may still be needed in hot climates, and chip longevity impacts remain unaddressed.