Dreamforce 2026, which closes in San Francisco today, produced a cluster of developments that together sketch the shape of the next enterprise AI architecture: domain models undercutting generalist APIs, CRM data becoming a live interface layer, and a sharp public divide on whether AI safety is a governance problem or an engineering one. Outside the conference, China's GPU build-out added another data point.
Koa: the domain model argument made concrete
Salesforce and Nvidia announced Koa on 15 September. Built by post-training Nvidia's open-weight Nemotron-120B with reinforcement learning (GRPO) on synthetic CRM data drawn from 27 years of Salesforce deployments, Koa is the first model Salesforce runs entirely within its own infrastructure as a customer-selectable option in Agentforce. No actual customer data was used in training; the dataset was generated synthetically to mimic enterprise patterns at scale.
The benchmark results are significant. On Tau2Bench, a multi-turn customer service evaluation across airline, retail, and telecoms, Koa scores 69.41 against GPT-4.1's 54.48. On BFCL, which tests agentic tool use across multi-step calling and stateful workflows, it scores 66.63% against GPT-4.1's 53.96%. On Salesforce's own CRM Bench, covering single-turn Salesforce and Agentforce workflows, Koa scores 0.86 against GPT-4.1's 0.81. The model surpasses GPT-4.1 decisively on the routine, well-defined tasks that make up most enterprise agent volume, while sitting below the strongest frontier models on general capability. The full technical paper is on arXiv.
The operator implication is clear: frontier APIs priced for general capability are overkill for repetitive, structured enterprise tasks. Open weights let a software vendor run a tailored model without pre-training costs, and Salesforce controls the weights, audits the inference, and prices it as a platform feature rather than a per-token API call. Customer pilots begin in October; general availability is planned for winter 2026 in the US. Expect other platform vendors — HR, ERP, legal — to follow this template within 12 to 24 months.
AIforce and Claudeforce: CRM data as a live interface
Alongside Koa, Salesforce unveiled AIforce, a live interface layer that makes Salesforce's data, workflows, business logic, permissions, security, and governance accessible from any AI platform, without forcing users back into the Salesforce UI. The premise is that the modal AI interaction point for employees is already an AI assistant, not a dedicated application screen, and the CRM should follow the user there.
The first three products are Claudeforce (Salesforce in Claude), Slackforce, and Agentforce Coworker. Claudeforce is now in beta for all customers: it packages Salesforce intelligence into a prebuilt MCP server inside Claude, with 37 prebuilt sales skills. Salesforce says service, marketing, commerce, and Tableau analytics skills follow in the near future. Claude Code gets a separate Salesforce Development plug-in with more than 40 skills and access to Salesforce's broader skills library on GitHub.
For operators: if your team's work is CRM-heavy, Claudeforce is worth piloting this week. More broadly, the "prebuilt MCP server as enterprise integration" pattern is now the preferred approach for major platform vendors, and it lowers the integration cost for any team already inside the Claude ecosystem. The harder question is governance: AIforce inherits Salesforce's existing permissions model, but operators will need to verify that their data-access policies translate correctly into the new interface layer before broad deployment.
Jensen Huang at Dreamforce: safety is an engineering problem, not a legal one
Speaking at Dreamforce on 16 September, Nvidia CEO Jensen Huang publicly rejected the case for AI regulation. His position: "safety is an engineering problem," not a legislative one, and existing market forces are sufficient to keep development responsible. He argued directly that "innovation, speed, and safe products" are not in tension.
His stance sits in explicit contrast to that of Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, who backed a more cautious pace and, as reported in yesterday's brief, have been in formal talks since July with Google DeepMind to create a self-regulatory AI standards body. The divide is structurally important: Nvidia supplies the compute that all three labs depend on, and Huang's public rejection of the safety-pause narrative signals that the infrastructure tier has no intention of slowing production or investment.
For operators: this disagreement directly determines how quickly your supply of affordable frontier capability grows, and how quickly safety standards become embedded in the model layer. The engineering-not-regulation camp argues the market and the tooling will handle it; the standards-body camp argues they will not. The next US AI executive order, expected before year-end, will likely be the first formal test of which frame the administration accepts.
Biren's third raise: China's GPU build-out accelerates
Bloomberg reported on 15 September that banks are sounding out investor interest in a roughly $1 billion share placement by Shanghai Biren Technology. This would be the company's third capital raise since it listed on the Hong Kong Stock Exchange in January 2026, following a $892 million secondary offering in July. The 90-day lock-up from that July placement expires in early October, and the new sounding comes within weeks of it doing so.
The cadence — three raises in under nine months — reflects both investor appetite and the urgency of scaling domestic GPU production in China. US chip controls have not choked off funding; if anything, they have concentrated capital into companies such as Biren that offer an alternative to Nvidia hardware. For operators running AI infrastructure with a global footprint, Biren's trajectory is a forward indicator of how quickly a second GPU ecosystem could become commercially significant, and what that means for pricing leverage and supply-chain resilience.
Taken together, today's developments point in the same direction: specialisation is displacing generalism at every layer. Koa beats generalist APIs on CRM tasks at lower cost. AIforce routes enterprise data into specialist agents rather than a monolithic UI. China is building specialised silicon that does not depend on Nvidia. Operators who start now mapping which of their AI use cases are routine enough to migrate to a domain model, and which genuinely require frontier capability, will find the transition cheaper than those who wait for the market to force it.