In AGNT Podcast Episode 15, Gemma Allen and Raphaëlle d’Ornano unpack what Dreamforce signaled for software: open foundation models tuned to enterprise workflows, the trusted infrastructure agents need to operate, and a contest over who captures the value of knowledge work. They also discuss why power delivery may constrain neocloud growth—and how consumer agents could change what it means for brands to be discovered.
Recorded in theCUBE’s New York studio with Gemma Allen, principal analyst at theCUBE Research.
Did Dreamforce change the software story?
The announcement d’Ornano singled out was Koa, Salesforce’s first CRM reasoning model. She sees it as part of a shift from writing every business process into software to post-training models on the workflows software companies already understand.
“Weights are the new code.” Raphaëlle d’Ornano · 01:29
Salesforce and NVIDIA announced Koa on September 15, describing it as post-trained from NVIDIA Nemotron 3 Super with synthetic scenarios based on Salesforce CRM workflows. The distinction matters: Nemotron is the open model underneath, while Salesforce says it controls Koa’s resulting weights and runs the model within its own trust boundary. D’Ornano’s larger point is that specialist models could give software companies a role in the new stack while still increasing demand for NVIDIA’s platform.
What does an AI harness need to do?
An agent needs more than a capable model. In d’Ornano’s view, the enterprise AI harness includes grounded context, security and observability: the controls that let agents work inside a company without losing track of its data, rules or processes. Dreamforce’s emphasis on the harness reflects a broader reworking of the enterprise stack, with software providers competing to supply the trusted environment around agents.
Security remains an open question. D’Ornano says traditional cybersecurity vendors are adapting their products as companies try to deploy agents, while the full threat surface and the right protections are still emerging. The episode treats security as part of the operating infrastructure for agentic software, rather than an afterthought.
Is the AI prize bigger than software?
The conversation widens from Salesforce’s product market to the value of knowledge work. D’Ornano argues that software companies, cloud platforms and AI labs are all competing to help complete work—often through teams of agents—with humans involved to varying degrees.
“We’re contesting how work gets done in the agentic enterprise.” Raphaëlle d’Ornano · 16:06
She says Anthropic’s roughly $30 trillion framing is defensible as a way to describe the scale of economic activity AI could touch. Allen adds an important caveat: a market can be addressable without being attainable. Anthropic’s own September 2026 economic scenario page describes more than $30 trillion in value created by the tasks that make up the U.S. economy over the prior year; that is not the same as near-term revenue available to any one AI company.
Why might power become more important than chips?
Discussing neoclouds, d’Ornano argues that the key question is not only how many GPUs a provider can offer, but how quickly it can bring powered capacity online. Grid access, turbines, electricity, and the people needed to build and operate data centers can all slow deployment.
“Not all megawatts are equal.” Raphaëlle d’Ornano · 23:55
Nscale is the episode’s main example. D’Ornano describes strong demand and potential, but also points to losses and an element of its Anthropic-related backlog that she says remained unfinanced. Those remarks need a current check: Nscale filed for a proposed U.S. IPO on September 18 and announced $3.36 billion in pre-IPO convertible financing on September 25. The episode’s wider point is that a backlog only becomes useful capacity when power, equipment, financing, and delivery all line up.
Will consumer agents become a new storefront?
D’Ornano describes using Instinct to plan a trip to Japan with her son. She says she gave the agent the goal, rather than a list of hotel brands or bookings, and it selected and booked hotels around that intent. Her example points to a different kind of competition: brands may need to be visible to the assistant that interprets a customer’s goal, not just to the customer browsing a search page.
She calls proximity to user intent a potential moat, alongside context and workflow. That idea applies in both consumer and enterprise settings: whoever owns the interface where people state a goal may shape which products and services get considered. Allen remains cautious about how quickly these agents will mature, while d’Ornano says the commercial model is still to be worked out.
What to watch
Specialist models: Whether workflow-tuned reasoning models deliver useful agent actions at lower cost, and how software firms package them.
The enterprise AI harness: How vendors combine grounding, security and observability as agents gain access to company systems.
Neoclouds: Whether providers can turn booked demand into powered, operational capacity; update Nscale context with its latest filings and financing.
Consumer agents: Whether assistants such as Muse and Instinct earn user trust, develop a business model and change how brands reach buyers.
Timestamps
00:53 Dreamforce and the Salesforce keynote
01:29 Koa, open foundation models and “weights are the new code”
06:12 Why software providers and NVIDIA can both benefit
08:31 Model release pace and enterprise security
12:21 Knowledge work and the scale of the AI opportunity
16:06 The agentic enterprise and the work software must do
19:00 Specialist graphs, agents and the future software stack
22:16 Neoclouds, compute access and grid power
23:55 Power delivery and Nscale’s operating questions
27:50 Meta’s Muse and the Instinct assistant
31:32 A Japan trip, consumer agents and proximity to intent
35:10 What d’Ornano is doing next