On the All-In Podcast at GTC 2026, Jensen Huang was asked what makes a company defensible when AI levels the playing field. His answer was two words: “deep specialization.”

Not a bigger model. Not more data. Not a proprietary algorithm. Know your vertical better than everybody else.

The question I asked myself immediately: what does that mean for a company with 50 employees, not 43,000?

The horizontal platform and the vertical moat

Jensen described a dynamic that’s already happening. Platform software companies — the big SaaS players selling horizontal licenses — are going to become value-added resellers of AI tokens. They’ll take models from Anthropic, OpenAI, or open source and repackage them for their customers.

It’s the same thing the big systems integrators did during the ERP wave. SAP sold the platform. Accenture and Deloitte sold the customization. The margin was in knowing the client’s industry, not in the base software.

Jensen put it this way: “Know your vertical. Know it as deep and as better than everybody else.” Then he added the key part: wait for these tools because they’re catching up to you and now you can imbue them with your knowledge.

For a mid-market B2B company, this is good news in disguise. You can’t outspend Microsoft on compute. You’re not going to train a better model than Claude or GPT. But Microsoft also doesn’t know how your supply chain works, or what exceptions your operations team handles every Tuesday at 7am. That knowledge is the moat AI can’t replicate — it can only amplify.

The flywheel nobody is spinning

The second thing Jensen said was equally direct: the sooner you connect your agent with customers, that flywheel is going to cause your agent to get better.

That describes a self-reinforcing cycle:

1. Your domain knowledge trains the agent
2. The agent interacts with real customers
3. Those interactions generate data and feedback
4. The data deepens your domain knowledge
5. Back to step 1, but with a more capable agent

The problem is that most B2B companies are stuck at step zero. They’re experimenting with AI internally — meeting summaries, email drafts, document queries. All fine. But no agent is touching a workflow that involves a customer.

Say you run an industrial distribution company. Your sales team quotes between 80 and 120 orders per week. Each quote requires checking availability, calculating volume discounts, reviewing customer history, and applying rules that exist only in the heads of two or three people.

An agent connected to that process — not a generic one, but one trained on your rules, your exceptions, your catalog — could pre-fill ~70% of each quote. Speed is just a byproduct. The real value is that every completed quote feeds the agent more context about how your team makes decisions. In six months, the agent knows patterns that even your best rep hasn’t articulated.

That’s the flywheel. And every month that passes without spinning it is a month of advantage your competitor can accumulate first.

$250K in tokens per engineer

One of the most concrete moments in the podcast was when Jensen talked about token budgets. He said that if a $500,000-a-year engineer consumes only $5,000 in tokens, he’d be “deeply alarmed.” His expectation: at least $250,000 per engineer per year in AI tools.

That number sounds like an NVIDIA data point — a company with 43,000 employees and $350 billion in revenue. But the logic behind it applies at any scale, especially when AI costs dropped ~80% in 18 months.

If you have 10 business analysts earning $60,000 a year and you’re giving them $0 in AI tools — or $200 for a basic ChatGPT subscription — are you actually equipping your team? You don’t need to reach $250K per person. But if the AI fluency gap is already open, investing in tools is what closes it.

And here’s where vertical specialization connects to the budget: spending on tokens without vertical context produces generic answers. Spending on tokens with your catalog, your customer history, and your business rules loaded as context — that produces results that compound.

Spending $10K on specialized tokens instead of generic ones doesn’t just double your output — it turns a basic utility into a real competitive advantage.

What Jensen didn’t say

Jensen speaks from the perspective of a CEO selling the infrastructure. His advice is real but incomplete for mid-market companies. There are three questions he didn’t answer:

What’s your actual specialization? Many companies confuse their product with their expertise. “We sell industrial valves” isn’t a specialization. “We know exactly which valve fits every plant configuration in the petrochemical industry, including the regulatory exceptions by country” — that is. Specialization lives in the decisions your team makes without thinking, the exceptions they handle from memory, the tacit knowledge that never made it into a manual.

Is your data structured so an agent can use it? The knowledge in your best employees’ heads is useless if an agent can’t access it. That means turning tacit business rules into structured data — not an 18-month initiative, but a practical exercise you can start this week by documenting the 20 most frequent decisions your team makes manually.

How do you know if the flywheel is spinning? Forget “how many agents did we deploy” or “how many tokens did we consume.” Track this instead: does the agent make better decisions today than it did three months ago on the same type of case? If the answer is no, the problem isn’t the AI — it’s the judgment behind what you’re asking it to do.

Your vertical is your best asset

Jensen described a future where every industry needs specialized models — sub-agents trained on vertical knowledge that only companies in that sector can build. Platforms will compete to be the base layer. The value will sit in the specialization on top of that base.

For a mid-market B2B company, here’s what remains: your vertical knowledge is the only asset that appreciates when AI gets better. Every new model that ships makes your expertise more useful, not less. But only if you structure it so an agent can actually use it.

AI will rewrite your industry regardless. When it does, your knowledge will either be wired into a system that multiplies it or trapped in the heads of three people who can walk out the door tomorrow.

---

If you want to identify what your real vertical specialization is — the tacit knowledge an agent could multiply, not your product catalog — that’s exactly what we do in a focused AI maturity assessment. We map where your advantage lives, what data you need to structure, and how to connect the first agent to a real customer workflow.

Identify my vertical advantage