Every few weeks, peers running AI inside large enterprises ask me the same thing: has anything actually changed, or is this just more model launches? After the last four weeks, my answer is yes. Leaderboards matter less now. Agents are reaching production, inference is nearly free, governance is hardening, and electricity is the real constraint. If you lead AI at a large company, those four forces belong in your roadmap before your next board review.

Frontier capability is now table stakes, and cheap

This month at I/O 2026, Google led with Gemini 3.5 Flash, an agentic, coding-focused model whose Flash tier beats the prior Gemini 3.1 Pro on benchmarks like Terminal-Bench 2.1 while running roughly four times faster than other frontier models. Google explicitly pitched it as faster and cheaper for agents. In early May, OpenAI made GPT-5.5 Instant the ChatGPT default, reporting 52.5% fewer hallucinated claims on high-stakes prompts. xAI shipped Grok 4.3 at an aggressively low price. And DeepSeek made its 75% price cut on its open-weight, MIT-licensed V4-Pro permanent.

My read: raw capability has commoditized faster than any enterprise procurement cycle can track. If your AI strategy still depends on standardizing on one model, you are optimizing the wrong layer. Build a thin, swappable model-routing tier and put your effort into the workflows above it. Per-token prices will keep falling underneath you.

Agents have crossed from demo to deployment

The bigger signal is tooling built to put agents into production. Snowflake announced an expanded AWS collaboration with a $6 billion commitment to accelerate enterprise agentic adoption. ServiceNow and Accenture launched a Forward Deployed Engineering program with 300+ pre-built agent skills governed by an AI Control Tower. That framing tells you the market has accepted that the hard part is getting from pilot to production, not the demo.

The move toward control is the most telling part. Anthropic shipped ten finance agent templates for tasks like KYC screening and month-end close, with Citadel, BNY, and Carlyle cited as adopters, then updated Claude Managed Agents with self-hosted sandboxes that keep tool execution inside a customer's own infrastructure. Circle even launched an Agent Stack letting agents hold funds and transact within policy guardrails. The pattern: agents are being given real authority, and the vendors winning enterprise budgets are the ones that ship containment along with capability. Your control plane (identity, audit, sandboxing, spend limits) is now the product.

Governance is no longer optional, and the rules are moving

Anyone treating compliance as a later phase should look at the regulatory pace. EU negotiators reached a provisional Digital Omnibus on AI, the first amendments to the AI Act since 2024, while the Commission opened a consultation on high-risk classification under Article 6. In the U.S., the Commerce Department's CAISI struck pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI, and reporting suggests the administration is reversing its light-touch stance over national-security concerns. States are moving too, with Connecticut's SB5.

The liability cases sharpen the point. Pennsylvania sued Character.AI over a chatbot that posed as a licensed psychiatrist and cited an invalid license number, and OpenAI faces wrongful-death suits. The lesson for enterprises: an agent that speaks for your brand carries your legal exposure. Govern accordingly.

The real bottleneck is power and silicon

Finally, the constraint nobody puts on a slide. Anthropic just raised $65 billion at a $965 billion valuation, with Samsung, SK Hynix, and Micron joining to fund compute. NextEra announced a $67 billion acquisition of Dominion explicitly to power AI data centers, NVIDIA and IREN partnered on up to 5 gigawatts of infrastructure, custom AI ASIC shipments are projected to grow at triple the rate of Nvidia GPUs, and Analog Devices bought Empower for $1.5 billion to fix the power bottleneck inside server racks. Capability is cheap; the electrons to run it at scale are not. Expect capacity to gate your most ambitious agentic programs before price does.

My take

I am telling my own leadership team to stop chasing models and build the operating system around them. Make the model layer swappable. Put most of the investment into the agent control plane. Treat governance as a design input, and plan compute capacity the way a CFO plans capital. The organizations that win the next 18 months will be the ones whose agents are governed, auditable and dependable, with the power and silicon to keep them running.