NVIDIA CEO Jensen Huang told the world at GTC 2025 that traditional software companies will evolve into agent platforms. He called it the shift from humans using software to AI agents using software on behalf of humans. That is not a research claim. It is a description of what is already happening in production systems.
I run two products on self-hosted infrastructure. ConnectEngine OS automates lead generation, content creation, and uptime monitoring. ProductShot generates AI product photography. Both products run agents that execute multi-step workflows without human intervention. The monthly infrastructure cost is roughly $45 USD. No venture capital. No cloud markup. Just a VPS, Docker containers, and agent orchestration built on n8n.
What Does Jensen Mean by Agent-First Software?
Huang's argument at GTC was structural, not speculative. SaaS requires a human to log in, click buttons, and execute tasks. Agents as a Service (AaaS) flips that model. The agent monitors your pipeline, identifies deals at risk, drafts outreach emails, updates the CRM, and flags the human only when judgment is required. The human becomes the exception handler, not the default operator.
This changes the economics of software. When an agent handles 10,000 tasks per month, you pay for compute and model inference, not 10,000 human hours. Pricing shifts from per-seat to per-task or per-outcome. Availability becomes 24/7 continuous instead of business hours. Scalability means spinning up more agents, not hiring more people.
NVIDIA positioned itself as the infrastructure layer underneath all of this. The company announced NIM (NVIDIA Inference Microservices), pre-packaged containers for running AI models in production. Enterprises can deploy customer-facing agents without managing model infrastructure. Huang also highlighted partnerships with Cadence, CrowdStrike, Dassault Systemes, Palantir, SAP, ServiceNow, Siemens, and Synopsys. These companies are building agents on NVIDIA infrastructure.
How ConnectEngine OS Implements the Agent Model
ConnectEngine OS runs three agent modules. LeadFlow discovers leads from Google Maps, Google Places, Etsy, and e-commerce stores. It scrapes business data, enriches contact information, and scores prospects based on website quality. ContentFlow generates platform-specific content from a single idea, rewrites it for blog, X, LinkedIn, Instagram, and Facebook, verifies claims against sources, and schedules publication. OpsFlow monitors uptime and sends Telegram alerts when sites go down.
These agents run continuously. LeadFlow fires on a schedule, finds new leads, and queues them for outreach. ContentFlow takes an idea, generates five platform-specific drafts, runs a verification pass, and publishes on schedule. OpsFlow pings monitored sites every few minutes and alerts immediately on failure. No human clicks a button to start these workflows. The agents execute autonomously.
The architecture is self-hosted n8n for workflow orchestration, Supabase for the database with Row Level Security enforced from day one, and Docker containers on a Hetzner VPS. Multi-tenancy is built into the schema. Each client's data is isolated via RLS policies. Credentials are stored in a vault with per-client API keys. Notifications route to Telegram or Slack depending on client preference.
Why Self-Hosting Beats Managed Platforms for Agent Infrastructure
The monthly cost for running both products is roughly $45 USD. That includes the VPS, offsite backups, and DNS. No per-execution fees. No per-seat licensing. No cloud markup on compute. The agent workflows run as often as needed without incrementing a bill.
Managed automation platforms charge per execution or per task. When an agent runs 10,000 operations per month, those fees compound. Self-hosting eliminates that variable cost. The infrastructure cost is fixed. The marginal cost of adding another workflow or another client is near zero.
Control is the other advantage. The n8n instance runs inside Docker on the VPS. Workflow JSON is versioned in git. Credentials are environment variables, never hardcoded. Backups run nightly to offsite storage. There is no vendor lock-in. The entire stack can be migrated to a different provider in hours if needed.
What the Shift to Agents Means for SaaS Founders
Huang's prediction is already visible in production systems. Salesforce ships Agentforce. ServiceNow ships AI agents. Vertical-specific platforms are selling agent capabilities as core features. The companies that adapt early will have a structural cost advantage over those that continue to sell per-seat software.
The technical barrier is lower than it appears. Open-source orchestration tools like n8n handle the workflow layer. Supabase provides a multi-tenant database with built-in auth and RLS. Claude and OpenAI provide the reasoning layer via API. The infrastructure cost is minimal if you self-host. The hard part is designing workflows that execute reliably without human intervention.
Start with one repeatable workflow. Identify a task your users perform manually every day. Build an agent that executes that task autonomously. Test it in production. Iterate until it runs without errors for a week straight. Then add the next workflow. Agent-first software is not a rewrite. It is an incremental shift from human-initiated actions to autonomous execution.
Questions readers often ask
Q: What is the difference between SaaS and Agents as a Service?
SaaS requires a human to log in and execute tasks. Agents as a Service means the agent performs the work autonomously and only escalates to a human when judgment is required. The pricing model shifts from per-seat to per-task or per-outcome.
Q: Can you build agent infrastructure without venture funding?
Yes. Self-hosted n8n on a VPS costs roughly $45 per month. Open-source tools handle orchestration, multi-tenancy, and auth. The marginal cost of adding workflows is near zero. Cloud platforms charge per execution, which compounds at scale.
Q: What is the biggest technical challenge in building agent workflows?
Reliability. Agents must execute without human intervention for days or weeks at a time. That requires error handling, retry logic, and monitoring at every step. The workflow must degrade gracefully when external APIs fail or rate limits are hit.
Q: How do you handle multi-tenancy in a self-hosted agent platform?
Row Level Security in Supabase isolates each client's data at the database level. Every query is scoped to the authenticated user's client ID. Credentials are stored in a per-client vault. Notifications route to the correct Telegram or Slack channel based on client configuration.
Q: What did Jensen Huang say about the future of enterprise software?
Huang said the era of humans using software is giving way to AI agents using software on behalf of humans. He positioned NVIDIA as the infrastructure layer for this shift and highlighted partnerships with SAP, ServiceNow, Palantir, and others building agents on NVIDIA infrastructure.
Top comments (0)