Community testing is live. A few are already in. More soon.

Cheap GPU hours are not cheap compute. A medical or agent workload dies on the run that never finishes. Hourly price is a listing. Completed-job cost is the bill. Buy runtime that starts, checkpoints, and settles. Not idle silicon.

Q3 or Q4?

Users don’t buy GPUs. They buy reliable runtime: job starts, job finishes, and usage is metered correctly.

Agent compute is not one big bill. It is hundreds of small executions that become visible only after they add up.

14,000 designs. 9 binders. 3 sub-nanomolar. The agent didn't win because it was smart. It won because the loop was closed: design → wet-lab validate → feed back. Autonomous execution only pays off when there's a checkpoint at the end.

A wallet is not an approval system. If an agent can spend, you need to define: what it can buy, how much it can spend, and when a human must approve.

Open models do not just create more builders. They create more agent attempts. More experiments. More tool calls. More failed runs. More inference demand. More need to verify what actually happened. The open-source wave expands the execution market, not just the model market.

Agents need receipts. Not screenshots. Not vague logs. A real one should show: intent → compute → verification → payment → result. PAN is building that execution record.

Training compute is easy to spot. Big clusters. Big budgets. Big announcements. Agent compute will be harder to see. Small jobs. Always running. Spread across tools, apps, and users. Paid for one action at a time. The biggest compute market may be the one nobody sees.

Training compute is easy to recognize. Big clusters. Big budgets. Big announcements. Agent compute will be harder to see. Small jobs. Always running. Spread across tools, apps, and users. Paid for one action at a time. That market may be much larger than it looks.

Official Announcement Thrilled to partner with @XAgent_official the zero-code platform empowering anyone to Speak to Build custom AI Agents. As the Demand Layer for AI Compute PAN will explore seamless compute access agent-to-agent payments and scalable infrastructure for the growing Agent Economy. Together were accelerating autonomous AI that actually works. Stay tuned for more.

We’re not trying to make smarter agents. We’re trying to make agent execution something you can actually verify and settle. Intent, compute, payment, result. PAN sits in that loop.

I don’t think the next AI compute wave will look like the last one. Training was big clusters and big headlines. Agents feel different. Smaller jobs. Running constantly. Happening in more places. Messier, but probably much bigger than people expect.

Once agents start paying for tools and compute, settlement stops being optional. Someone has to prove what was requested, what ran, and who paid. That’s the part we’re building around.

One question keeps showing up for us at PAN: if agents are going to work all day, where does that work actually run? Not the model. Not the demo. The execution behind it.

Every tool call and inference step can be a transaction. Agents will request, verify and pay for execution. PAN is building financial rails for the agent economy.

Training loves centralized clusters. Agents need continuous, low-latency compute. Idle GPUs will join the supply side. PAN is building the coordination layer for verifiable agent compute.

In preparation for what’s next, Pan new website is now live. We’ve updated the design, product pages, and onboarding flow. What do you think we should launch next?

Agent autonomy is rising. Without verification, one rogue action can cascade. Production agents need intent checks and execution proofs. Safety is an infrastructure layer.