Econify.AI
Roundtable recap, Nov 13: NVIDIA's leadership change, AI in front of customers, and which model to pick
BY Econify
•Fri, Nov 13, 2026

Three questions came up this week. What does Jensen Huang's exit from NVIDIA mean for open models? Who is putting AI in front of customers? And how does anyone decide which model to use for which task?
What NVIDIA's leadership change means for open models
Our host opened with the week's biggest story, Jensen Huang stepping down. NVIDIA has been one of the strongest backers of open source models, so the question was what happens to them now. He also wondered whether frontier models are nearing diminishing returns, and whether pressure on Chinese open models could leave the field to three or four closed labs.
One attendee thinks Apple is betting on local models. If compute and power stay expensive, running smaller models on-device is how the work keeps going. They also wondered whether non-coding work like writing will move to local models while developers stay on frontier ones. A report of a $300 a month Grok plan using about $14,000 a month in tokens came up as a sign current pricing can't last.
Another attendee finds open models usable for some tasks but still reaches for frontier models for practical dev work, since they get it right the first time. Their worry is hardware. GPUs are already expensive, and new leadership could pull back from consumer parts. A third said teams that built internal infrastructure on open models are starting to ask whether they'll fall behind.
Our host expects models at the Haiku 5.5 level to become practical to run locally. He also sees OpenAI leaning toward search and ads while Anthropic goes after the office suite, which suggests the model alone isn't a business plan. The group agreed the labs are burning cash, and that developers alone are too small a market to cover it.
Is anyone putting AI in front of customers?
Mostly no. Most of the group uses AI on the back end. The customer-facing features attendees have seen are simple, like listening to an article, and one agent went through a vendor partly for legal reasons. News organizations are cautious for the same reason. Google skipping AI overviews on election night sent direct traffic to news sites way up.
The patterns that work sit behind the scenes. One team lets editors configure a video platform's CMS by asking for the change in plain language. A fintech attendee described an LBO calculator written as deterministic Python and exposed to the AI as an MCP server. The AI reads an analyst's memo, maps it onto the model's inputs and asks follow-up questions, but never does the math. Another attendee said these tools work best when they have a schema, a spec and rules, as when teams build Datadog and Splunk dashboards by describing the query.
Which model for which task?
Nobody had a rule. One organization can't cap token use per person or restrict groups to certain models, and one user is running up a very large bill. Others move between Haiku and Opus depending on deadlines, and one admitted to some guilt about using a big model on a small task.
The ask was guidance in the tools themselves (Raycast shows speed and intelligence, Copilot in VS Code shows cost) and evals that teams can plug in to answer whether a new model is better, cheaper and faster. Our host said these are eval problems and that Econify has dealt with this one. One attendee suspects providers keep model guidance vague on purpose. Another said teams inside large companies work in silos, and that the useful thing is hearing from someone further down the road. One suggestion: keep a weekly journal of lessons and have an LLM pull out the common threads.
Keeping logs, and keeping work and personal separate
One attendee recommended sending Claude Code, Codex and T3 telemetry to an OpenTelemetry-compatible server you own. Later you can ask an agent to find patterns, like which models you actually use for what. Saved session history is also useful for reflection and even training. Our host's habit is to ask the model, after a problem is solved, what he could have done to guide it better.
That led to mixing work and personal use. Some attendees keep separate work and personal accounts, plus a throwaway tool for questions they don't want in their history. Admin dashboards for Codex and ChatGPT show message counts but not content. Tools exist for scoping MCPs and memory to specific repos, and Anthropic has been public that routing multiple accounts through a proxy is against its terms.