NVIDIA Bets $4B on Photonics as DeepSeek V4 Prepares Trillion-Parameter Launch

DeepSeek is preparing to release V4, a 1-trillion-parameter mixture-of-experts model timed strategically ahead of China's annual Two Sessions parliamentary meetings beginning March 4. Despite its massive total parameter count, the model activates only 32 billion parameters per token — a technique that keeps inference costs remarkably low, with projected pricing of $0.10–$0.30 per million input tokens, up to 50 times cheaper than GPT-5.

Architecturally, V4 builds on its predecessor with three key innovations: Manifold-Constrained Hyper-Connections for training stability at trillion-parameter scale, a new Engram Conditional Memory system for efficient retrieval over its 1-million-token context window, and an upgraded Lightning Indexer for sparse attention. The model also handles text, images, and video natively, and is optimized to run on Huawei Ascend chips — a clear response to U.S. export controls limiting access to NVIDIA hardware.

DeepSeek plans an open-source release, allowing developers worldwide to fine-tune and deploy V4 without licensing fees. Internal benchmarks suggest performance competitive with — or exceeding — leading Western models on long-context software engineering tasks, continuing DeepSeek's pattern of delivering frontier capability at dramatically reduced cost.

Around 500 protesters marched through London's King's Cross tech hub on February 28 in what organizers described as the largest anti-AI demonstration ever held globally. The march began outside OpenAI's UK office on Pentonville Road and wound past the headquarters of Google DeepMind and Meta, ending in a Bloomsbury church hall where attendees convened a People's Assembly to draft formal demands for the British government.

Organized jointly by Pause AI and Pull the Plug, the demonstration surfaced a wide range of grievances — from AI-generated online misinformation and synthetic abusive imagery to concerns about autonomous weapons and long-term existential risk. Participants are calling for a formal pause on frontier AI development, democratic oversight of the technology, and an immediate halt to new data center expansion in the UK.

The protest's scale reflects a growing disconnect between public sentiment and government policy. According to polling cited by organizers, 84% of British people believe the government prioritizes its relationships with large technology companies over public interests when regulating AI — a perception that is increasingly mobilizing civil society across the country.

A senior U.S. Defense Department official testified on March 3 that commercial AI contracts signed under the Biden administration contain sweeping operational restrictions that could paralyze military missions — including the ability to plan and execute combat operations. The remarks highlight a deepening tension between Silicon Valley's ethical guardrails and the Defense Department's operational requirements.

The contracts in question were signed with major AI vendors as the Pentagon rushed to integrate AI into defense workflows. Their terms reportedly prohibit uses of the technology that fall within normal military operational planning, creating situations where warfighters could be legally blocked from using AI tools they rely on. The Pentagon is now reviewing these agreements and pushing for revised terms that preserve ethical safeguards while restoring operational flexibility.

The disclosure comes as the Trump administration has broadly moved to loosen AI regulation across civilian agencies, and as OpenAI reportedly struck a new deal with the Pentagon — a potential model for how future defense AI contracts might be structured to avoid the restrictions currently flagged by military officials.