By The Core Concept: Every enterprise will operate a hybrid AI model — the debate is not whether to self-host, but when and how intentionally to make that transition as token costs and data exposure risks compound.
The Guest: Rob Hirschfeld, CEO at RackN
The Bottom Line:
• Hybrid AI is not a strategic option — it’s the destination every enterprise is heading toward, and the organizations that plan for it now will have a significant cost and control advantage over those that react to the token bill
Speaking with TFiR, Rob Hirschfeld of RackN, defined the current state of the self-hosted vs. SaaS AI debate — and argued that framing the decision as a binary choice is itself the mistake most enterprises are making.
WHAT IS THE HYBRID AI MODEL?
Hirschfeld’s core argument is that the question “should I run my own AI model or use an API?” has a single correct answer: both. Every enterprise, regardless of where they are in the AI adoption curve, will ultimately operate a hybrid infrastructure — some workloads running on SaaS frontier models, others on self-hosted open models.
“Everybody will ultimately be hybrid. Being hybrid frees you up to decide how to start that journey—but it also lets you prepare for that work.”
This framing matters because it dissolves the pressure to make an all-or-nothing infrastructure bet early in the AI journey. Organizations can start aggressively with SaaS APIs, absorb the token costs as a learning investment, and begin planning their self-hosted inference capacity in parallel.
THE TOKEN SPEND PHASE IS INTENTIONAL — NOT A MISTAKE
One of Hirschfeld’s more counterintuitive points: early, expensive token spend on frontier models is not waste — it’s necessary organizational learning. Teams need to develop fluency with AI tooling, build agents, and internalize new ways of working. Trying to optimize token costs too early slows down that capability development.
“Teams need to learn how to work and think in this model, and it will cost money and consume tokens—but doing that is absolutely essential.”
The shift comes as proficiency grows. Once teams are using AI at scale, the distribution of workloads becomes clear: a small fraction genuinely requires frontier model reasoning, while the majority — background automation, log analysis, routine correspondence, batch processing — can be handled by open models running on self-hosted infrastructure at a fraction of the cost.
BROADER CONTEXT: WHY SELF-HOSTING BECOMES INEVITABLE
In this interview, Hirschfeld expanded on the data sovereignty dimension of this decision — arguing that the information flowing through AI systems (including code and business processes) is inherently sensitive, and that enterprises that delay thinking about self-hosting are quietly accumulating risk alongside their token spend.
He described the operational reality at RackN’s largest AI inference customer — a company running over 40,000 servers for AI workloads — where servers must be racked and onboarded within hours of delivery to stay competitive. That urgency, he argued, is the direction the entire industry is heading.
“The appetite for improving your AI throughput is bottomless at this point. That level of urgency should be shaping your AI decisions now.”
Watch the full TFiR interview with Rob Hirschfeld here.
