Why the Second Half of AI Won't Be Won by One Supermodel — Kevin Ding, Pyromind

# Why the Second Half of AI Won't Be Won by One Supermodel — Kevin Ding, Pyromind

Everyone keeps waiting for the model that kills all other models. Kevin Ding thinks that's the wrong bet.

Ding is the founder and CEO of Pyromind, an AI company focused on post-training and enterprise intelligence. In a recent episode of the *Crossing* podcast, he made the case that the second half of AI — the half we're just entering — won't be defined by a single dominant supermodel. It'll be defined by specialization, post-training, and agentic systems built close to real business problems.

That's a different thesis from the one that's been selling GPU clusters and foundation model hype for the last two years. And it has real implications for founders building right now.

The 'one model to rule them all' myth

The assumption has been simple: whoever builds the smartest, largest foundation model wins everything. OpenAI, Anthropic, Google DeepMind, and a handful of others compete at the frontier, and everyone else builds on top of whoever survives.

Ding pushes back on this directly. His argument: general-purpose models are hitting a ceiling on ROI for most enterprise use cases. A hospital system doesn't need GPT-5. It needs a model that understands its specific EHR format, its compliance requirements, its patient communication protocols — and can act on that knowledge autonomously.

That's a post-training problem, not a foundation model problem. And it's a problem Pyromind is building to solve.

This mirrors what's already happening at the infrastructure layer. The model war is real, but the application layer — where actual revenue gets generated — is increasingly about fine-tuning, retrieval, and agentic orchestration layered on top of whatever foundation model is cheapest or most capable for a given task.

Post-training is the moat

If the foundation model is a commodity, post-training is where defensibility lives.

Post-training includes everything that happens after a base model is trained: instruction tuning, RLHF, domain-specific fine-tuning, alignment work, and increasingly, the construction of agent frameworks that let models take actions rather than just generate text.

Ding's framing is that companies which own their post-training pipeline — which know how to shape model behavior for a specific industry context — will have a durable advantage that a model update from OpenAI can't erase overnight.

This is especially true in industrial and enterprise settings. A logistics company's AI needs to understand freight terminology, carrier constraints, and routing edge cases. A legal firm's AI needs to navigate jurisdiction-specific language and citation formats. These aren't problems you solve by waiting for the next GPT release.

For founders watching this space, the implication is clear: domain-specific post-training is a viable wedge. The market is large enough, the general models are generic enough, and the enterprise appetite for reliability over raw capability is strong enough that focused bets can win.

AI agents in the enterprise: where the money actually is

The consumer AI wave was loud. The enterprise AI wave will be larger.

Ding's work at Pyromind is centered on enterprise intelligence — AI systems that don't just answer questions but take actions, coordinate workflows, and operate with meaningful autonomy inside complex organizational contexts.

This is the agentic AI thesis playing out at scale. Not chatbots. Not copilots. Agents that own a process end to end: intake, analysis, decision, execution, reporting.

The challenge in enterprise isn't capability — current models are capable enough for most business tasks. The challenge is reliability, auditability, and integration. Enterprises won't deploy an agent that hallucinates 3% of the time into a customer-facing billing workflow. They need systems that fail gracefully, log their reasoning, and hand off to humans at the right threshold.

That's an engineering problem as much as an AI problem. And it's where companies like Pyromind, which are close to the enterprise deployment reality, have an edge over teams building in the abstract.

What this means for solo founders and small teams

Ding's company isn't a one-person operation — but his argument has direct implications for founders building lean.

If the action is in post-training, specialization, and agentic orchestration rather than foundation model development, then the capital required to compete drops dramatically. You don't need a $100M compute budget to fine-tune a model on a specific domain. You don't need a 50-person engineering team to build an agent workflow on top of an existing API.

The revenue per employee potential for companies in this layer is exceptional. A two-person team that owns a fine-tuned model and an agentic deployment for, say, insurance claims processing, is in a fundamentally different position than a two-person team trying to compete at the frontier.

This is exactly the structural shift that makes the one-person unicorn model credible. When the expensive part of building AI products (training from scratch, massive compute, huge research teams) is abstracted away, a small team with deep domain knowledge and strong agentic engineering skills can build real revenue at a fraction of the previous cost.

The funding and PMF question

The *Crossing* interview touched on Pyromind's funding status, revenue, and whether the company has found product-market fit. The specifics weren't fully public, but the framing of the conversation signals a company in early but intentional motion — building toward enterprise contracts rather than consumer scale.

For founders watching: enterprise AI PMF looks different from consumer PMF. It's slower, it's relationship-driven, and the feedback loops are longer. But when it lands, contract values are larger and churn is lower. A single enterprise deployment can generate more revenue than thousands of consumer subscriptions.

Ding's approach — focus on post-training, go deep on enterprise verticals, build agentic systems that solve real operational problems — is a blueprint worth studying. Not because it's the only way to build, but because it's grounded in where the actual business value in AI is accumulating right now.

The second half is fragmented by design

The first half of AI was about building the rocket. The second half is about figuring out where to fly it.

That means more specialization, more vertical focus, more companies owning specific slices of the post-training and deployment stack. The supermodel won't disappear — it'll become infrastructure, like cloud compute, something you build on top of rather than compete with.

For founders, that's good news. The opportunity isn't to out-OpenAI OpenAI. It's to build the AI-native company that owns a specific domain, a specific workflow, or a specific post-training pipeline — and make that the moat.

Kevin Ding is betting Pyromind on that thesis. Given where enterprise AI budgets are flowing, it's a reasonable place to plant a flag.

---

Is your company eligible? Submit to the leaderboard → onepersonunicorn.co/submit