The Signal — May 23, 2026
The intelligence community is spending billions to close its AI gap, NVIDIA is rethinking how language models generate text, and California is bracing its workforce for what comes next.
White House Approves $9B for Spy Agencies to Catch Up on AI; Anthropic Finalizing Classified NSA Contract
The White House has approved a $9 billion request to acquire advanced AI chips for the CIA, NSA, and other intelligence agencies struggling with a critical shortage of computing power. The funding addresses a widening gap between the classified systems used by U.S. intelligence and the frontier AI models now standard in the private sector.
The core issue: intelligence agencies cannot run the latest AI models on their existing hardware. Classified networks are air-gapped from the commercial cloud, and the chips powering those systems are generations behind what companies like Google and OpenAI deploy. The $9 billion is intended to bring classified computing infrastructure closer to parity with the private sector, enabling agencies to run large language models and other AI tools on sensitive data without exposing it to outside networks.
Separately, Anthropic is finalizing a classified contract with the NSA to provide access to its Claude models for intelligence analysis. The deal would make Anthropic one of the first major AI labs to operate directly inside the classified environment, joining a small group of companies trusted with that level of access. The convergence of massive hardware investment and direct AI lab partnerships suggests the intelligence community now treats AI as operational infrastructure, not a side experiment.
Sources: NYT · WION News · NW Arkansas Online
NVIDIA Released Nemotron-Labs-Diffusion: A Tri-Mode Language Model
NVIDIA released Nemotron-Labs-Diffusion, a family of open-weight language models spanning 3B to 14B parameters that unify three distinct text generation methods into a single architecture. The models can switch between autoregressive decoding, diffusion-based generation, and self-speculative decoding depending on the task, a design NVIDIA calls "tri-mode."
The throughput gains stand out. The 8B parameter model generates 5.9 times more tokens per forward pass than Qwen3-8B while maintaining better accuracy on standard benchmarks. On real hardware, that translates to roughly 4x higher throughput, a real improvement for anyone running inference at scale. The models are open-weight, meaning researchers and developers can download, modify, and deploy them without licensing restrictions.
The architectural implications matter more than the speed gains. Most production language models today use purely autoregressive decoding, generating one token at a time. Diffusion-based text generation has been a research curiosity with clear theoretical advantages but limited practical payoff. Nemotron-Labs-Diffusion suggests the two approaches are not mutually exclusive and that hybrid architectures can outperform either method alone. If the results hold up under independent testing, this could shift how the next generation of models is built.
Sources: HuggingFace Blog · NVIDIA Research · MarkTechPost
Newsom Signed Executive Order to Prepare California Workers for AI Disruption
Governor Gavin Newsom signed Executive Order N-6-26, directing state agencies to study the workforce impact of AI and explore policy responses including expanded severance requirements, employment insurance reform, universal basic capital, and job retraining programs targeted at white-collar roles. The order comes days after Meta announced roughly 8,000 layoffs, many in roles increasingly automated by AI tools.
What makes the order unusual is its specificity. Rather than vague language about "preparing for the future," it directs agencies to evaluate concrete mechanisms: mandatory severance for AI-driven layoffs, expanding California's employment insurance system, and a concept called "universal basic capital" that would give displaced workers equity stakes or investment accounts rather than direct cash transfers. The focus on white-collar retraining reflects a shift from earlier AI workforce discussions that centered on manufacturing and logistics.
California is not the first state to address AI workforce disruption, but it is the largest economy to put this level of policy specificity into an executive order. Whether the studies lead to legislation remains an open question, but the order establishes a framework that other states will likely reference or adapt.
Sources: NYT · CA.gov · Mashable
On the Editor's Desk
Killed this week: an AI model benchmarks and pricing roundup that read more like SEO content than news, and a Cerebras Kimi K2.6 item that had gone stale. Meta's 8,000 layoffs broke earlier in the week and appear here as context for the Newsom story rather than as a standalone item. A Virgin Atlantic partnership with OpenAI Codex was held as corporate PR without broader implications. These may resurface if developments warrant a second look.
