The Pentagon's AI Spending Spree, OpenAI Breaks Free From Microsoft, and a $1.1B Bet on AI That Learns Without Us

The Pentagon's Classified AI Shopping Spree Leaves Anthropic Behind

On May 1, the U.S. Department of Defense announced deals with seven companies — OpenAI, Google, Microsoft, Amazon, Nvidia, xAI, and the startup Reflection — granting them access to classified networks for AI work. The Pentagon simultaneously declared itself an "AI-first fighting force," signaling that autonomous AI systems will become central to military operations across every domain.

The notable absence: Anthropic, which previously held a $200 million contract for classified AI work. The DoD designated Anthropic a "supply-chain risk" in March after the company refused to allow Claude to be used in autonomous weapons systems. Multiple federal agencies began phasing out Claude models on a six-month transition timeline. This week's deal wave is the clearest signal yet that the Pentagon is building its AI infrastructure around companies willing to participate in military applications — and that principled refusal carries real commercial consequences.

For small businesses, this story matters in two ways. First, the companies that the Pentagon trusts enough for classified work will likely shape the enterprise AI tools you use in the next two years. The capabilities built for defense contracts tend to filter down into commercial products. Second, the Anthropic situation raises a question every business will face: what are your AI vendors willing to do with your data, and where are their red lines? Vendor alignment with your values isn't just ethics — it's risk management.

Source: The Verge (theverge.com, May 1, 2026); BBC (bbc.com, May 1, 2026); CNN (cnn.com, May 1, 2026)

OpenAI Ends Microsoft Exclusivity, Brings Its Models to AWS

On April 28, one day after OpenAI and Microsoft announced a revised partnership agreement, OpenAI said its models will be available through Amazon Web Services via Amazon Bedrock. The move ends years of exclusive cloud partnership with Microsoft, giving AWS customers access to OpenAI's models — including its Codex coding agent — through a managed service called Amazon Bedrock Managed Agents powered by OpenAI.

Under the new agreement, Microsoft retains a non-exclusive license to OpenAI's intellectual property through 2032, remains the primary cloud partner, and keeps its 27% equity stake. But OpenAI can now run its products on any cloud provider. AWS CEO Matt Garman called the deal something "customers have been asking us for for a really long time." OpenAI CEO Sam Altman, who was in court for the Elon Musk lawsuit the same day, sent a recorded message acknowledging the partnership.

For small businesses, this is unambiguously good news. More cloud providers hosting OpenAI's models means more pricing pressure, more deployment options, and less vendor lock-in. If your business runs on AWS but has been considering Azure just to access OpenAI, you no longer need to. The AI market is becoming genuinely multi-cloud, and that competition will benefit everyone except the incumbents who thrived on exclusivity.

Source: CNBC (cnbc.com, April 28, 2026); Microsoft Blog (blogs.microsoft.com, April 27, 2026)

DeepMind Legend Raises $1.1B to Build AI That Learns Without Human Data

David Silver, the researcher behind AlphaGo, AlphaZero, and AlphaStar at Google DeepMind, has raised $1.1 billion at a $5.1 billion valuation for his new startup, Ineffable Intelligence. Backing comes from Sequoia, Lightspeed, Nvidia, and Google. The company's goal: build a "superlearner" that discovers knowledge and skills through reinforcement learning — trial and error — without relying on human-generated training data.

This is the approach that produced AlphaZero, which taught itself to beat the world's best chess and Go programs from scratch, with no human games as input. Silver's thesis is that the current paradigm of training AI on vast human datasets has fundamental limits: it can only ever replicate what humans already know. A superlearner, by contrast, could discover strategies, solutions, and knowledge that no human has ever produced. The company's language is ambitious to the point of grandiose — comparing the potential breakthrough to Darwin's theory of evolution — but the funding and the founder's track record make it hard to dismiss.

For small businesses, the practical impact is medium-term. If reinforcement learning without human data works at scale, it could produce AI systems that solve problems no human could train them on — drug discovery, materials science, supply chain optimization. It also raises the question of how you validate AI outputs when there's no human baseline to compare against. The technology is promising; the governance questions are significant.

Source: TechCrunch (techcrunch.com, April 27, 2026); CNBC (cnbc.com, April 27, 2026); The Next Web (thenextweb.com, April 27, 2026)

NVIDIA Open-Sources Nemotron 3 Nano Omni: One Model for Vision, Audio, and Language

On April 28, NVIDIA released Nemotron 3 Nano Omni — an open-weight multimodal model that natively processes video, audio, images, and text in a single system. The model uses 30 billion total parameters with 3 billion active per forward pass through mixture-of-experts routing, making it efficient enough to run on a single GPU while topping six leaderboards for document intelligence, video understanding, and audio comprehension. NVIDIA claims it delivers up to 9x efficiency gains over systems that chain separate vision, speech, and language models together.

The key innovation is unification. Current AI agent systems typically route data through separate models — one for vision, one for speech, one for language — losing context and time with each handoff. Nemotron 3 Nano Omni handles all of these in one model, meaning an agent watching a video call can process the visuals, the spoken words, and any on-screen text simultaneously, with shared understanding across modalities.

For small businesses building AI-powered products or internal tools, this matters because it dramatically simplifies the architecture. Instead of stringing together three separate APIs and managing the context gaps between them, you can deploy one model that handles everything. It's open-weight, meaning you can run it on your own infrastructure without per-token pricing or vendor lock-in. If you've been waiting for multimodal AI to become practical and affordable, this is the model that makes it so.

Source: NVIDIA Blog (blogs.nvidia.com, April 28, 2026); Hugging Face (huggingface.co); The Next Web (thenextweb.com, April 28, 2026)

US and Allies Publish First Government Guidance on Securing AI Agents

On May 2, cybersecurity agencies from the United States, Australia, Canada, New Zealand, and the United Kingdom jointly published guidance on deploying AI agents securely. The document, co-authored by CISA, the NSA, and the Five Eyes intelligence partners, focuses on agentic AI — systems built on large language models that can plan, make decisions, and take actions autonomously by connecting to external tools, databases, and automated workflows.

The agencies' central message: agentic AI doesn't require an entirely new security discipline. Organizations should integrate these systems into existing cybersecurity frameworks, applying established principles like zero trust, defense-in-depth, and least-privilege access. The guidance identifies five categories of risk: excessive privilege (agents granted too much access), design and configuration flaws, data exposure through agent memory stores, supply chain vulnerabilities in agent tools, and insufficient monitoring of autonomous actions.

The warning is direct: AI agents capable of taking real-world actions on networks are already inside critical infrastructure, and most organizations are granting them far more access than they can safely monitor or control.

For small businesses, this is the first government document that treats AI agents as a concrete cybersecurity concern rather than a hypothetical one. If you're deploying AI agents that can send emails, modify databases, or execute code — and most AI-powered tools are heading in that direction — you need to treat them the way you'd treat a new employee with system access: with defined permissions, audit logs, and the principle of least privilege. The guidance is free, practical, and worth reading.

Source: CyberScoop (cyberscoop.com, May 2, 2026); CISA (cisa.gov)