Hey there 👋 I read three announcements this morning that had nothing to do with each other on paper. A chip keynote, a Korean telecom report, and an ERP company's developer conference. Different rooms, different continents, same week. By the second one I could see they were telling one story. For about two years the whole fight was whether agents could do the work. That question is boring now, because the answer is mostly yes. So the industry quietly moved on to the next one. If an agent is going to do the work, somebody has to pay for it, supervise it, and be able to shut it off. This week three different companies each shipped a piece of that answer, and all of it lands on whoever runs the fleet. - Ron
The Big ThingNVIDIA wants about $2,000 a year per GPU to run your agentsJensen Huang floated a number this week that should go straight into your 2027 planning doc. NVIDIA is weighing an annual software license of roughly $2,000 per GPU for the software that runs AI agents on its chips. Do the math: every million GPUs running it is $2B a year in recurring revenue for NVIDIA. That figure is arithmetic, not a company forecast. That is a new line item sitting on top of hardware you already paid a fortune for. Read the framing carefully before you panic or dismiss it. This is a signal, not a SKU. Huang said it out loud around the GTC Berlin stretch, but NVIDIA has not named which products the fee would cover, what you would get for it, or when the meter starts. There is no invoice. There is a direction of travel, and the direction is clear: the company that sells you the shovel now wants a cut of the digging. Here is why it belongs at the top of the issue even without a price tag you can wire money against. For two years the smart play was to own your agent runtime. Buy the GPUs, self-host the engine, keep the models close, and your marginal cost per agent trends toward electricity. A per-GPU software fee breaks that math. If it lands, every agent you run carries a recurring tax that grows with the size of your fleet, whether or not the thing is busy. The teams this hits hardest are the ones who did the "responsible" thing and built their own infrastructure instead of renting it by the token. If I had to bet, this is the clearest reason yet to keep an escape hatch in your architecture. The CUDA lock-in cracks we covered last week, Google open-sourcing its tooling, the push to run models on anything but NVIDIA, suddenly have a dollar figure attached. DeepSeek open-sourcing its inference engine on October 25 reads differently when the alternative is a two-grand-per-chip annual bill. You do not need to rip anything out today. You do need one deployment path that does not route through NVIDIA's software stack, so that when the meter turns on you have a choice instead of a renewal notice. Why it matters: a per-GPU agent software fee turns your own hardware into a subscription, so budget for it now and keep one inference path that does not depend on it. Verdict: WATCH, and plan around it. Sources: TechSpot, TweakTown, FourWeekMBA.
Tour de Headlines📋 SK Telecom published its first AI Governance Report this week, and it reads like an HR manual for software. The carrier treats agents as a kind of staff. Each one gets an employee ID. The report lays out tracking for each agent's role, its supervisor, its status, its token use, and its work history. An agent with more autonomy has to report to a human manager, a team lead or an executive. An agent judged less useful at review can be suspended or retired, the same way you would manage out a bad hire. The employee-ID idea goes back to SKT's AX Innovation 2.0 push in June, but this is the first time they have written down the full operating model. Steal it even if you run five agents. You already need to know who owns each one and how you turn it off. Verdict: BORROW. Seoul Economic Daily 🚪 SAP put real mechanics behind its agent push at TechEd on Tuesday. Joule Work now speaks MCP and Agent2Agent, so agents can pull context and run multistep workflows across tools. The piece worth your attention is the governed gateway. It is a single controlled entry point where an outside assistant or agent can invoke SAP's Joule Agents, and where your org gets to enforce policy and watch what happens. SAP's own line is "open protocols need governed doors," which is the most honest thing anyone said about MCP all month. There is also a new hook to trigger n8n workflows from Joule Studio, so an agent decides when a job runs and the deterministic workflow runs the fixed steps. No GA dates or pricing yet, so pilot the gateway, do not promise it to anyone. Verdict: PILOT. SiliconANGLE 🎤 Block showed off Dessa, a voice model that does something the usual stack cannot. It is full-duplex, so it talks and listens at the same time, and it was trained only on human speech with zero text tokens. The emotion is baked into the model itself, instead of a tag you bolt onto a text-to-speech layer. The headline for builders is where it runs: locally, on a consumer machine like a MacBook M4 Max, no cloud round-trip. That changes the cost and privacy math for any voice agent you were planning to pipe through a hosted API. The catch is a real one. There is no public access yet, no stated license, and no API. Block has asked safety researchers to poke at whether a voice agent this natural could fool the person on the other end. Watch it, you cannot deploy it. Verdict: WATCH. TechCrunch
Sponsor Your team's calls are full of signal. RapportScore reads it. RapportScore scores how your team communicates on real recorded calls and coaches them on it, so you can see where rapport holds and where it slips. Built for B2B teams who live on the phone. See your team’s score → |
Tool of the Day🌳 agent-treeAn MIT-licensed plugin that drops a live tree of your Claude Code subagents into your terminal. agent-tree is the dashboard you have been missing if you run Claude Code subagents. It shows every subagent, its status, the step it is on, which model it is using, how many tokens it has burned, and how long it has run. There is a timeline view with the critical path, a trace of each agent's tool calls in order, a board grouped by status, and a history of your last 150 runs. Install is one line. Be honest with yourself about what it is: a brand-new project with almost no stars yet, so treat it as a sharp utility, not load-bearing infrastructure. One note, the optional AI report mode makes Haiku or Sonnet calls on your own account, so leave it off if you want zero surprise spend. Verdict: TRY. Get agent-tree on GitHub →
Worth a Click- Claude Code 2.1.311 landed October 26. The one that matters: sessions that blow past 200K tokens now compact and keep going instead of dying on you. It also lets subagents use MCP servers that finish connecting late, and fixes SubagentStop hooks so they fire once per subagent instead of on a loop.
- Bubble MCP is in beta, and it lets you point your own Claude Code or Codex agent at a Bubble app to take most build actions. It cannot read logs, merge branches, or deploy, so you stay in the driver's seat while it does the grunt work. The feature shipped October 6 and got a fresh demo in Bubble's October 26 founder AMA.
Capability was the argument we spent two years having. This week agents got a meter, a manager, and a bouncer, all in the span of seven days. NVIDIA wants to charge per GPU to run them. SK Telecom files them under HR with a badge and a termination path. SAP built the door they have to knock on. Three layers of the stack, one move: an agent is now a worker you pay for and supervise. The bill, in every version, goes to whoever runs the fleet. Go put one non-NVIDIA inference path in your architecture doc. Future-you, reading the invoice, will be grateful. - Ron |
