AI Signals Worth Watching: March 16, 2026
Three things happened this week that, taken together, say something important about where AI is actually heading.
Apple Chose Google. Not OpenAI.
iOS 26.4 — targeting release this month — makes Google’s Gemini the default intelligence layer across 2 billion Apple devices. ChatGPT isn’t gone. It’s still available for opt-in queries. But in AI, defaults are everything.
This is structurally similar to the search default deal Apple struck with Google two decades ago — the one that now generates roughly $20 billion annually for Google. The current deal is estimated at $1–5 billion per year. More importantly, it means Google’s Gemini now runs on the vast majority of smartphones worldwide, across both Android and iOS.
For OpenAI, losing the Apple integration is a real strategic setback. Distribution is now the competitive front that matters, and OpenAI doesn’t control it. Their response — a stealth hardware device in collaboration with Jony Ive — signals they understand the problem. Whether that device becomes a real answer is a different question.
The deeper point is this: the AI race is no longer primarily a model race. Gemini 3.1 Pro leads 13 of 16 major benchmarks. But benchmark supremacy didn’t win Apple’s deal — deployment scale, privacy architecture, and ecosystem fit did. That’s a meaningful shift in what “winning” looks like.
Enterprise Agent Adoption Just Crossed a Threshold
Gartner now projects 40% of enterprise applications will include AI agents by the end of 2026. At the start of 2025, that number was under 5%.
What’s changed isn’t capability — it’s readiness. Companies now have the governance models, orchestration infrastructure, and security posture to put agents in production rather than pilots. The data confirms it: 57% of organizations are already running multi-step agent workflows. 16% have deployed cross-functional agents operating across teams.
The pattern that keeps surfacing: enterprises that invested in AI governance aren’t just more compliant — they’re shipping 12x more projects to production. Governance is revealing itself as a forcing function for speed, not a brake on it.
On March 13, Galileo launched Agent Control, an open-source governance layer for standardizing AI agent behavior across organizations. The market for governing agents — not just building them — is forming quickly.
New pricing structures are following: Salesforce introduced AELAs (Agentic Enterprise License Agreements) to replace unpredictable consumption models. Other SaaS vendors are building similar structures. When enterprise finance teams start budgeting line items for agentic AI, you know the technology has crossed from experimental to operational.
NVIDIA Built the Model Family That Multi-Agent Systems Have Been Waiting For
At GTC 2026 on March 11, NVIDIA released the Nemotron 3 family — Nano, Super, and Ultra — built specifically for agentic AI workloads at scale.
The architecture is worth paying attention to. Nemotron 3 Super has 120 billion total parameters, but only 12 billion are active at any time via a hybrid mixture-of-experts design. That’s what makes it fast enough to run complex multi-agent pipelines without the cost profile of dense models. Nemotron 3 Nano delivers 4x higher throughput than its predecessor — the number that matters for multi-agent systems running thousands of parallel calls.
It’s deploying on AWS Bedrock now, with Google Cloud, Azure, and CoreWeave following. NVIDIA is also releasing training datasets and RL environments alongside the models — making the package not just a model but a starting point for building specialized agents.
The trend line is clear: the model tier is evolving toward efficiency over raw scale. MoE architectures, sparse activation, and per-task specialization are where the serious infrastructure investment is going. The foundation model benchmarks will continue mattering for headline coverage. The real competition is happening at the agentic infra layer.
Taken together, these three signals point in the same direction: AI has moved from a research competition to an infrastructure competition. The layer that matters most now is not the model — it’s what coordinates the model, routes context to it, and acts on the output in a way that’s trustworthy and personal. That’s the space that’s just beginning to be built.
Eliran Keren — Founder of Deeplica, building the coordination layer for humans who’d rather direct than operate.
