Agents Go Local. Silicon Gets Personal. Physical AI Moves Out.

What happened
Perplexity launched Portable Computer, a local-first version of its Computer agent that packages models, orchestration, tools, connectors, and a sandbox onto NVIDIA powered hardware. Local work stays on-device and carries no per-token charge; cloud escalation requires user permission.

Why it matters
This attacks two agent-deployment headaches at once: sensitive-data movement and the recurring cost of long, multi-step workflows. The more consequential idea is the hybrid architecture—do routine work locally, then selectively call frontier models when the extra capability is worth the privacy and cost tradeoff.

What’s next
Perplexity is starting with NVIDIA/Linux systems and expanding hardware support, making the key question how much real knowledge work smaller local models can absorb before users need a cloud handoff.

What happened
OpenAI introduced an Admin plugin for ChatGPT Work and Codex that lets workspace administrators inspect usage, manage members and permissions, change spending limits, and execute supported administrative actions conversationally. It can also automate recurring workflows such as routing usage approvals through Slack or Microsoft Teams.

Why it matters
Enterprise agents are crossing an important line from answering questions to changing operational systems. OpenAI is pairing those write actions with existing roles, permissions, approval requirements, structured results, and review for broader impact changes exactly the governance layer enterprises need before agents become control-plane interfaces.

What’s next
The bigger test is whether conversational administration becomes a standard interface for IT operations: not just generating reports, but continuously monitoring policies, resolving routine requests, and escalating exceptions to humans. OpenAI says its own IT team already uses related agent workflows for support-ticket triage and execution.

What happened
Google Cloud launched Gemini Enterprise packages for Legal and Financial Services. The finance product includes a managed research agent, more than 50 specialized skills and 13 connectors; the legal package adds domain skills, MCP integrations, partner agents, data isolation, and centralized governance for workflows such as contract review, diligence, regulatory monitoring, and legal research.

Why it matters
The enterprise agent race is moving beyond “best general chatbot.” In regulated industries, competitive advantage increasingly comes from connecting proprietary systems, preserving permissions, grounding outputs in authoritative data, maintaining provenance, and packaging complete workflows—not simply serving a stronger model.

What’s next
Google explicitly calls Legal and Financial Services the first in a series of packaged industry solutions, pointing toward a market where major AI platforms compete vertical by vertical with prebuilt agents, connectors, controls, and partner ecosystems.

What happened
OpenAI published the first measured results for Jalapeño, its first custom inference chip. On the public InferenceX benchmark across GPT‑OSS 120B, DeepSeek R1, and Kimi K2.5, OpenAI reports 1.5–1.9× more AI work per watt at peak throughput and 1.7–3.6× lower end-to-end latency than its comparison systems; these are OpenAI reported benchmark results, not independent production measurements.

Why it matters
Inference economics increasingly determine how fast and cheaply AI products and especially agents can operate. Agent workflows execute many sequential steps, so latency compounds; owning silicon optimized around models, memory, networking, and serving software gives OpenAI another lever over responsiveness and cost-to-serve.

What’s next
OpenAI plans to begin deploying Jalapeño in its own compute infrastructure by the end of 2026, while Gen 2 is already “deep in development” and Gen 3 is taking shape. The company also says it will continue deploying NVIDIA and other third-party accelerators, suggesting custom silicon is an additional layer rather than an immediate replacement strategy.

What happened
Apple unveiled the M6, its first 2-nanometer chip and the M5 Ultra, its first quad-die Apple silicon design. The M5 Ultra supports up to 512GB of unified memory and 1.2TB/s of memory bandwidth, which Apple says can run LLMs with hundreds of billions of parameters entirely on-device; M6 is positioned for smaller local and agentic workloads.

Why it matters
Large-memory desktops are becoming credible private AI infrastructure. That creates a middle tier between laptops and cloud GPU clusters where enterprises and developers can run, fine-tune, and orchestrate substantial models without automatically sending sensitive workloads off-device.

What’s next
Apple’s Core AI, Core ML, Metal, and Xcode tooling can run and fine-tune models locally, so the adoption question moves from raw hardware capability to whether developers build enough serious local-agent and enterprise workflows around that memory advantage.

What happened
Accelerated Understanding, founded by Anima Anandkumar and Benedikt Jenik, unveiled an AI system built around neural operators rather than Transformer language-model architecture. The company says one test processed 5 trillion pieces of data in a single prompt, while its initial commercial targets include chip design, robotics, weather prediction, and energy-related geological analysis.

Why it matters
The important signal is architectural, not the headline-size data claim: frontier AI may increasingly split into specialized model families optimized for language, video, physical dynamics, biology, and other data regimes instead of converging on one universal Transformer stack. Accelerated Understanding’s performance claims still need broader independent validation.

What’s next
The company is prioritizing enterprise deals rather than a consumer product, making real deployments—and evidence that one physics model can generalize across substantially different scientific and industrial problems—the next meaningful proof point.

What happened
NVIDIA announced Jetson Orin Nano 2, an edge-AI computer with 78 TOPS of AI compute, 8GB of memory, and an eight-core Arm CPU. NVIDIA says it delivers 2× the inference performance of Jetson Orin Nano Super and, at equivalent performance in 15-watt mode, consumes 40% less power.

Why it matters
Smaller language and vision-language models are making sophisticated reasoning practical inside robots, drones, and machine-vision systems rather than requiring constant cloud calls. Lower edge latency, power consumption, and connectivity dependence directly expand where embodied AI can operate reliably.

What’s next
Cognex, Doosan Bobcat, Matic, and Wing are among companies adopting or evaluating the platform, but hardware availability is still ahead: NVIDIA expects the module and developer kit in the first half of 2027.

What happened
Autonomous-trucking company Gatik raised $200 million in a Series D led by Qatar Investment Authority and Koch Disruptive Technologies. Gatik operates dozens of fully driverless commercial trucks, while the company reports it has more than $600 million in contracted revenue and plans to use the financing to expand its regional freight network.

Why it matters
Gatik’s traction highlights where physical AI is commercializing fastest: constrained, repeatable workflows with obvious labor and logistics economics. Its focus on middle-mile routes between distribution centers and stores narrows the autonomy problem compared with trying to solve every passenger-driving scenario at once.

What’s next
Gatik says it intends to move from dozens of driverless trucks operating across North America to thousands in the years ahead. That jump from proven deployments to fleet-scale operation will be the real test of autonomous freight’s economics and reliability.

AI is moving closer to where the work happens. Agents are running locally, admins are turning conversation into action, industries are getting purpose built agent stacks, and more intelligence is moving onto custom silicon and edge hardware. The next phase of AI may be less about sending everything to the cloud and more about putting the right intelligence, with the right permissions, in the right place.

Move one AI workflow closer to the work.

Pick a task you normally send to a cloud model…research, document analysis, coding, data cleanup, or another repeatable workflow and test a more local or specialized version.

Try one of these:

Then compare:

💡 Insight
The best architecture may not be cloud or local. It may be local by default, specialized where useful, and cloud only when the extra intelligence is worth the handoff.