Frontier Capability Developments Deep-Dive

Thursday, March 05, 2026

Top Line

• Anthropic hits $20B revenue run rate — doubling in months through military applications and enterprise adoption, now outpacing OpenAI's commercial trajectory despite public safety disputes Bloomberg

• Google ships Canvas workspace to all US users — integrated AI-powered document creation and planning tool in Search represents direct move against Microsoft Copilot and Notion AI in productivity layer TechCrunch

• US military operationally dependent on AI for Iran targeting — Central Command confirms AI systems managing "enormous amounts of data" for strike coordination, marking first public acknowledgment of AI in live combat operations Bloomberg

• Broadcom projects $100B+ AI chip revenue for 2027 — CEO Hock Tan signals major inroads into Nvidia territory through custom silicon for hyperscalers, reshaping competitive dynamics in AI infrastructure Bloomberg

• NotebookLM launches cinematic video generation — Google upgrades research summarisation beyond slideshows to fully animated videos using Gemini 3 stack, extending capability frontier in multimodal synthesis The Verge

Key Developments

Military AI Adoption Accelerates Amid Live Combat Deployment

The US military's operational reliance on AI for the Iran conflict represents a watershed moment in capability deployment. Central Command publicly confirmed AI systems are managing "enormous amounts of data" for targeting and operational decisions, while Bloomberg reports this marks the first large-scale use of AI in active combat coordination. Meanwhile, Wired documents companies like Smack Technologies explicitly training models "to plan battlefield operations," moving beyond intelligence analysis into direct operational planning.

This deployment is occurring despite significant friction between AI labs and the Pentagon. Anthropic initially terminated its military contract over safety disagreements, with CEO Dario Amodei calling OpenAI's messaging around replacing them "straight up lies" according to TechCrunch. However, Bloomberg and the Financial Times now report Amodei has reopened discussions with the Pentagon, suggesting commercial pressure and the reality of Claude's continued military use through third parties may be forcing a pragmatic recalibration.

OpenAI's Sam Altman told employees Tuesday that the company "does not control how the Pentagon uses their artificial intelligence products," per The Guardian, while simultaneously rushing to add surveillance safeguards according to the Financial Times. The gap between stated policies and operational realities is widening rapidly.

Why it matters: The capability frontier has shifted from research benchmarks to live deployment in high-stakes military operations, forcing labs to confront the control problem in practice rather than theory.

What to watch: Whether Anthropic's reopened Pentagon talks result in technical controls that actually restrict military applications or merely public relations frameworks that legitimise existing deployments.

Anthropic's Revenue Explosion Reshapes Commercial Landscape

Anthropic is on track to generate nearly $20 billion in annual revenue, more than doubling in just months according to Bloomberg. This represents a faster commercial scaling trajectory than OpenAI achieved in comparable timeframes, despite—or perhaps because of—Anthropic's continued military applications. TechCrunch reports Claude is being used for "many targeting decisions" in Iran operations even as some defence tech clients flee over reputational concerns.

The competitive dynamics are shifting dramatically. While Nvidia CEO Jensen Huang announced his company is "pulling back from OpenAI and Anthropic" for future investments according to TechCrunch, he simultaneously ruled out the $100 billion maximum investment in OpenAI that had previously been pledged per Bloomberg. This suggests Nvidia is hedging its bets across multiple model providers rather than backing a single winner.

The Trump administration's designation of Anthropic as a potential national security risk, which tech groups are now urging be reversed according to Bloomberg, adds regulatory uncertainty that could constrain Anthropic's growth despite its commercial momentum.

Why it matters: Anthropic is demonstrating that aggressive enterprise and government adoption can outpace consumer-focused deployment strategies, reshaping what successful AI monetisation looks like.

What to watch: Whether the national security designation sticks and how it affects Anthropic's ability to serve both commercial and government customers simultaneously.

Google Extends Gemini Reach Through Canvas Integration

Google rolled out Canvas in AI Mode to all US users in Search, creating a dedicated workspace for planning, project development, and document drafting according to TechCrunch and The Verge. The feature pulls current information from Search into a side panel, representing Google's attempt to defend its core search business while moving users into longer-session AI-assisted workflows that compete directly with Microsoft Copilot and Notion AI.

Separately, NotebookLM received a major capability upgrade with "cinematic" video generation using "Gemini 3, Nano Banana Pro" models according to The Verge. This moves beyond the previous slideshow-style video overviews to fully animated content, extending Google's lead in research synthesis and multimodal generation for knowledge work.

Why it matters: Google is leveraging its search distribution moat to embed AI capabilities directly into existing user workflows rather than requiring adoption of standalone tools.

What to watch: Whether Canvas drives meaningful increases in search session duration and whether enterprises adopt it for collaborative work, potentially threatening Notion and Coda's positions.

Custom Silicon Competition Intensifies in AI Infrastructure

Broadcom CEO Hock Tan projected AI chip sales exceeding $100 billion in 2027, marking "major inroads into territory dominated by Nvidia" per Bloomberg. This reflects hyperscalers' push toward custom ASICs that bypass Nvidia's ecosystem, with Meta explicitly planning "to develop custom chips to train its AI models" according to Bloomberg.

Meanwhile, Nvidia is responding by reallocating TSMC capacity away from China-bound H200s to its latest Vera Rubin products as export controls constrain Chinese sales, according to the Financial Times. The competitive pressure is real enough that the Financial Times questions whether "Nvidia's margins can last" given its "symbiotic relationship with TSMC" represents both an advantage and "key vulnerability."

Why it matters: The shift from merchant silicon to custom ASICs could fragment the AI training ecosystem, potentially slowing capability diffusion but also reducing Nvidia's architectural influence over model development.

What to watch: Whether Broadcom's projected growth materialises through actual deployments at scale or represents speculative pipeline rather than committed production.

Signals & Trends

Military AI deployment is outpacing governance frameworks by orders of magnitude. The gap between Anthropic's public safety disputes and the reality of Claude's use in targeting decisions, combined with US Central Command's open acknowledgment of AI-managed operations, suggests regulatory and ethical frameworks are functionally irrelevant to actual deployment timelines. Companies that positioned safety as a competitive moat are discovering it provides little leverage against government customers operating under wartime imperatives.

The productivity layer is now the primary AI battleground. Google's Canvas rollout, Meta's custom chip plans, and the broader shift toward embedded AI in existing workflows rather than standalone chatbots indicates the industry has moved past the "build a better ChatGPT" phase. The winners will be determined by distribution moats (Google Search, Microsoft Office) and vertical integration (custom silicon for specific workloads) rather than raw model capability. This explains why Anthropic's enterprise revenue is scaling faster than OpenAI's consumer subscriptions despite OpenAI's brand recognition advantage.

Infrastructure costs are creating a new capability bottleneck. Seven tech giants signing Trump's "rate payer protection pledge" per The Verge and the Financial Times highlighting Trump's admission that "AI companies need PR help over data centre backlash" suggests power availability—not algorithmic innovation—may be the binding constraint on capability deployment through 2027. The fact that data centres have become military targets in Iran conflicts per Bloomberg adds geopolitical risk to infrastructure buildout that wasn't present in earlier technology cycles.