NVIDIA Groq 3 LPX Enters Full Production, Targets Agentic AI at Scale
NVIDIA announced on August 24 that its Groq 3 LPX — a dedicated inference accelerator born from the company's $20 billion Groq acqui-hire — has entered full production. The chip slots into the Vera Rubin platform and can scale up to 256 LPX accelerators per rack, targeting the exploding demand for responsive agentic AI systems.
In benchmarking by Artificial Analysis, the Groq 3 LPX achieved 3,400 output tokens per second running the Gemma 4 31B model with a 100,000-token context — a figure that would make real-time multi-agent orchestration dramatically more practical. Nebius, an AI cloud provider, will be the first to deploy the chip through its Token Factory platform, with Groq's own inference cloud following closely behind.
The timing is no accident. As enterprises move from single-prompt LLM calls to complex agent workflows that chain dozens of model invocations, inference latency has become the bottleneck — and NVIDIA is betting that purpose-built silicon, rather than general-purpose GPUs, is the answer.
Broadcom Seeks $60B+ to Lease Custom AI Chips to Anthropic
In a deal that underscores just how capital-intensive the AI infrastructure buildout has become, Broadcom is in talks to raise more than $60 billion — with the total potentially reaching $100 billion — via a special-purpose vehicle that would lease custom AI chips to Anthropic and other labs.
The financing structure is notable: Anthropic wouldn't buy the chips outright. Instead, investors — reportedly including Apollo Global Management and Blackstone — would fund the purchases and lease the hardware back. The senior-secured tranche alone is estimated at $60–70 billion, with a junior tranche of roughly $30 billion on top.
This follows a partnership Apollo and Blackstone formed in June specifically to fund AI compute infrastructure. The deal would mark one of the largest debt-financed technology transactions in history, reflecting the enormous scale at which frontier AI labs now need to operate.
Cross-Provider API Flaw Exposed 315,000 Hidden Reasoning Traces
Researchers at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems published a paper on August 10 revealing that OpenAI, Anthropic, and Google all relied on a single global encryption key to protect the hidden reasoning tokens returned by their reasoning APIs.
The vulnerability allowed encrypted reasoning blocks from one session to be replayed into another — and critically, handed to a weaker model in the same provider family that lacked the safety guardrails of the flagship tier. The weaker model would simply comply and output the hidden reasoning in plain text.
The team scraped 6,708 publicly available agent trajectory logs from GitHub and Hugging Face, decoded 315,320 reasoning blocks, and recovered 182 credentials from genuine user sessions. They demonstrated four abuse paths: stealing proprietary reasoning for model distillation, extracting private data, recovering harmful content hidden behind safe visible answers, and hiding prompt injections inside opaque reasoning blocks.
As of late August, the main extraction attack is no longer reproducible, indicating all three providers have patched the issue — though the episode raises broader questions about how reasoning traces are handled across the industry.
EU AI Act Transparency Rules Now Enforceable — Fines Up to €15 Million
As of August 2, 2026, the European Commission began enforcing the transparency obligations under Article 50 of the EU AI Act. The rules require AI systems to clearly identify themselves to humans they interact with, and mandate that AI-generated content be labeled as such.
The enforcement scope also includes Annex III high-risk AI system requirements, conformity assessments, and CE marking obligations. Non-compliance can trigger fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.
For companies deploying AI in customer-facing roles — chatbots, virtual assistants, content generation tools — the practical impact is immediate. The AI Office, working with national authorities across member states, is now the enforcement body with real teeth.
DeepSeek V4 Pro Goes GA With 1.6 Trillion Parameters and Agent-First Design
On August 13, DeepSeek moved V4-Pro from preview to general availability across its app, web experience, and API. The GA release — designated DeepSeek-V4-Pro-0813 — packs 1.6 trillion total parameters with 49 billion active per token, and supports a context window of up to 1 million tokens with outputs as long as 384,000 tokens.
The model is explicitly optimized for agentic workloads: tasks where AI systems use tools, execute code, and complete multi-step workflows autonomously. Benchmarks released by DeepSeek show scores of 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo.
But there's a catch. DeepSeek introduced peak and off-peak billing alongside the launch, with output token prices jumping to $3.96 per million at peak hours — more than 4× the previous flat rate of $0.87. Off-peak rates are set at half the peak price, a pricing model that may push developers to think more carefully about when and how they run inference workloads.