← /editions
#138

Agents Are Eating CPUs

// Modal adds Nvidia Vera CPUs to host agent workloads, Meta puts its Wonder models behind the Muse agent starting with subscribers, and GM brings an action-taking OnStar AI to the 2027 Cadillac Lyriq.

$ today_in_10_seconds

Agents are reshaping the hardware stack. Modal is adding Nvidia's Vera CPUs to host agent workloads, a signal that the agent loop, not just inference, is now a compute market of its own. Meanwhile, Meta and GM are pushing agents out of the chat window and into subscriptions and steering wheels.

$ lead_story

Modal brings Nvidia Vera CPUs into its agent-hosting fleet

AI infrastructure startup Modal says it will start using Nvidia's Vera CPUs to host agents, alongside the Intel and AMD chips it already runs. Until now it relied entirely on other vendors' processors. The detail that matters is that this is a CPU deal for agents. Inference happens on GPUs, but everything around the model runs on CPUs: sandboxed code execution, tool calls, browser sessions, file I/O, and orchestration. Long-running, multi-step agents multiply that overhead. For builders, the cost of an agent is increasingly measured in sandbox density and CPU-seconds, not just tokens. Profile where your agent's wall-clock time actually goes before you optimize the prompt.

// why_it_matters: The agent bottleneck is moving off the GPU and into everything around it.

read more →

$ also_worth_attention

  • Meta will power its Muse agent with Wonder models, subscribers first

    Meta says its Wonder models, which it claims compete with the industry's best, will power the Muse agent, starting with paying subscribers. Specialized models will follow for everyone using Muse. The tiering is the real story: the most capable agent behavior goes behind a paywall, and the free tier gets narrower, cheaper models.

    link →
  • GM debuts OnStar AI in the 2027 Cadillac Lyriq

    GM's new in-car assistant can execute vehicle functions and run diagnostics through voice commands. GM is pitching it as a personal assistant and errand-runner. This is tool use with physical actuators, so permission scoping and failure handling will matter far more than conversational polish.

    link →

$ bigger_picture

The trend behind the headlines

Today's three stories cover three layers of the same shift. Agents now have dedicated hardware underneath them, proprietary models inside them, and real-world tool surfaces at the edge. Each company is trying to own one layer outright: Nvidia the runtime silicon, Meta the model-plus-agent bundle, GM the physical endpoint. For anyone building agents, the durable advantage sits where those layers meet: the execution environment, the tool integrations, and the guardrails that decide what an agent is allowed to touch.

$ tool_of_the_day

Modal Sandboxes

Given today's lead, it's worth looking at what Modal actually offers agent builders. Sandboxes are isolated containers you spin up programmatically to run LLM-generated or untrusted code, which makes them a practical execution layer for coding and data agents. It's a good way to see firsthand how much of your agent's work is CPU-bound tool execution rather than inference.

explore Modal Sandboxes →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime