Why Vertical AI Founders are Ditching Model Labs
with Zain Jaffer, Founder & CEO of Blazel and Vungle (acq. for $780M)
Zain Jaffer is the founder and CEO of Blazel, an AI-powered marketing platform that helps companies build authentic presence on LinkedIn and generate leads. Before Blazel, Zain built Vungle into the fastest-growing company in mobile ads, scaling it from zero to >$400M in ARR before Blackstone acquired it for $780M — after which it went public at a $4B valuation. After a few years on the VC side, he jumped back into the founder seat with a $10M pre-seed for Blazel. Tune in to hear why he’s betting that Vertical AI companies can (and should) claw back control from the model labs by building or adapting their own capabilities. He makes the case that the economics of renting intelligence may not hold up as long as founders assume.
Today’s Episode
Every Vertical AI founder faces a version of the same question: how much of your intelligence stack should you own? The default answer — build on Anthropic or OpenAI, move fast, ship product — has been the right one for most early-stage companies. The models are excellent, API credits are abundant, early gross margins can be forgiven, and nobody needs to reinvent inference infrastructure at Day Zero.
But a growing number of founders are hitting a wall. Costs escalate unpredictably. Latency from multi-agent architectures compounds. The frontier labs continue their saber-rattling about moving into the application layer as every API call you make teaches them more about your workflow. Meanwhile, open source seem to catch up faster every cycle. Zain lays out why he decided to build Blazel’s own model capabilities, pivot away from an AI-Native Services model, and what his decisions reveal about where value will actually accrue in Vertical AI.
The $10M Trap
Zain’s journey toward model independence started with a pivot. Blazel began as an AI native agency — one of the first, before Sequoia, a16z, and YC started championing the category. The agency grew fast. Replacing existing marketing budgets was a frictionless sale: buyers already have a purchasing workflow, there’s no new vendor approval process to navigate, and you don’t even need a UI. Zain’s take on agency-style AINS: “It’s very easy to get to $10M ARR. I’m not kidding.”
The problem is what happens at $10M. You’ve hired a team of humans. Any AI you introduce threatens their jobs, and they become a barrier to the very automation that should be your competitive advantage. More agencies flood in because the barrier to entry is so low. The allure of hiring more humans before the LLM cost structures warrant them — to maintain growth and raise more capital — is hard to deny. And if you go to your VCs and propose cutting revenue to pivot toward a platform model, you face a high likelihood of resistance, even disillusionment.
We’ve discussed this dynamic in the SaaS-to-AI transition: incumbents are generally terrified to cannibalize their own revenue, and it gets harder with scale. Zain saw it forming and made the call to pivot before the agency got too big. “It’s very hard to pivot when you’re 10M ARR.” Zain also emphasized the difficulty of running a people and technology business simultaneously — at Vungle, he ultimately felt their decision to split focus between brand advertising and app installs was a major misstep. Competitor AppLovin kept singular focus on one ICP and one problem. Vungle exited at $4B. AppLovin is worth $150–250B.
The agency, in Zain’s telling, wasn’t the business — it was the product roadmap. Humans doing the work revealed exactly which workflows to automate and where domain-specific AI could outperform general-purpose models.
A quick word from our partner and sponsor on the Verticals show, Parafin. They power embedded lending across Amazon, DoorDash, Walmart, and others — with >$35B in financing extended to date. Purpose-built for vertical SaaS & AI platforms serving SMBs.
Explore a custom program for your platform today →
When the API Bill Hits
Blazel’s content pipeline initially ran on Anthropic’s Sonnet. Multiple agents would generate content, other agents would edit out AI slop, then the output fed back as labeling data to train evals. It worked. Until the bills came.
“How did I spend $70K last month on Claude?” Zain recalls. “I wasn’t monitoring my tokens because it’s free credits.” The cloud credits that seed-stage companies get from AWS, GCP, or Azure mask the real unit economics. At API pricing — roughly 20x what you’d pay through consumer products — the math gets ugly fast for any company processing high volumes of domain-specific content.
Cost was one forcing function. Latency was another. Every agent call is a round-trip to the model provider. Chain four or five agents together — generate, edit, check for slop, evaluate, reformat — and the latency compounds. For a product that needs to feel responsive to time-strapped CMOs, that kills the experience.
But the existential forcing function is competitive. Frontier labs are building applications — or at least threatening to. Just by existing on their rails, you’re helping them accelerate your obsolescence.
The model providers themselves aren’t content sitting in the model layer. They want to go into the application layer… You are basically training the model provider. — Zain
Fine-Tuning Your Way Out
So Blazel did something that sounds, on its face, absurd: a pre-seed startup started building its own model capabilities.
The approach wasn’t training de novo models from scratch. The team fine-tuned open-source models using reinforcement learning and pre-training techniques, then built a harness of prompts and tools around the result. The labeling data came directly from the agency’s human editors: months of expert corrections that encoded what “good” content actually looks like in specific domains. This is why the agency phase mattered — it generated the proprietary training data that no frontier lab could easily replicate.
We got to the point where our own AI model was producing better content than what our team could produce. And we hired the best content people you can imagine. — Zain
A domain-specific model, trained on domain-specific evals by domain experts, beat a frontier model at a specific task — not because it’s a better model overall, but because it’s been shaped by data the frontier labs don’t have. Your content editors’ judgment, your customers’ voice patterns, your vertical’s quality standards. These artifacts can’t be scraped from the open internet.
We made this argument in Dude, Where’s My Moat? — workflow and data are the immutable primitives of software defensibility. That principle applies equally to the model layer. The companies with the deepest workflow integrations, that win them access to the most relevant proprietary data, will build the most defensible model capabilities. They do it not by outspending the labs or going toe-to-toe with them on compute, but by outlearning them on narrow, high-value tasks.
To work toward model independence, evals are key. “Labeling is a fundamental layer when it comes to AI and that’s where humans are absolutely important. For the LLM to be a good judge, it needs to have some intuition, and intuition comes from hiring humans.” They’re the mechanism through which domain expertise gets translated into both moats, whether at the application or AI infrastructure layers.
A Trillion Parameters Under Your Desk
While Zain’s bet seems quite contrarian at the moment, it’s increasingly plausible as we consider where compute is heading. Nvidia’s DGX Station, announced at GTC 2026, puts 20 petaFLOPS of AI performance and 748GB of coherent memory in a deskside form factor — enough to run a trillion-parameter model locally.
If running large models locally becomes cheap enough, the economic argument for API intelligence dependency weakens substantially. Companies can fine-tune and serve open-source models on their own infrastructure, keeping proprietary data in-house and eliminating per-token costs at the margin. The labs’ moat narrows to whatever capabilities can’t be replicated with open-source architectures, custom training data, and increasingly affordable compute.
Zain frames it as an existential risk for the frontier labs: “A model itself could actually become a commodity. I think that’s a threat to OpenAI and to Anthropic, which is why they’re trying to go” app-side. If the model layer commoditizes, value accrues to whomever owns the workflow and the data — and the remainders will be either have to make their bones on supply ownership, become Dispatchers to the cloud provider oligopoly, or get competed to death. The irony is that the frontier labs, continuously making noise about their path into the application layer, seem to agree. Even SpaceX — a business that has basically no B2B apps — claimed in their S-1 that the enterprise application layer represents >90% of their TAM.
The Takeaway for Vertical Founders
None of this means you should rush to fine-tune an open-source model at pre-seed or seed stage. The API-first playbook is still the fastest way to product-market fit. Use the credits. Ship the product. Learn what your customers actually need.
But be intentional about what you’re learning. Every human edit, every customer correction, every eval your team runs — that’s labeling data. It’s the raw material for the domain-specific model you may need to build sooner than you planned. Zain’s team didn’t set out to build their own model. They set out to build a marketing agency, and the data they generated while doing it became the foundation for something far more defensible.
Founders who want to best position themselves should consider treating API dependency as a strategic phase, not a permanent architecture — at least to maintain optionality and set the right product culture. Build on the frontier labs today, but capture the data and develop the evals that would let you peel off high-volume, domain-specific workloads when / if the economics demand it. The compute to run your own models is getting cheaper by the quarter. It’s too soon to imagine every startup with it’s own LLM. But the question is not whether Vertical AI companies will own more of their model layer — it’s when, and in which form factor.
Zain has lived the cycle. He built a company to a $780M exit by disrupting the mobile ad incumbents. He watched from the VC seat as SaaS companies failed to adapt to platform shifts. Now he’s back, betting that a startup with domain-specific data, proprietary evals, and fine-tuned models can build something the trillion-dollar labs can’t match in his vertical. The labs have scale. Vertical founders have specificity, earned insight, and (in the best cases) data frontier model providers can’t scrape or buy. In a world where compute is commoditizing fast, the model landscape of the future — at least in Vertical AI — might not be as homogenous as recent history has trained us to believe.
See you next week.
Key Moments from this Episode
00:00 — Intro
01:30 — Building an AI marketing platform from scratch
08:05 — The future of AI-native marketing teams
17:14 — Why data flywheels are the new moat
20:31 — Choosing the right ICP in the AI era
25:36 — The hidden trap of AI-native agencies
33:25 — Human labeling, evals, and building better AI
37:42 — Should every startup build its own model?
42:10 — Are OpenAI and Anthropic worth a trillion dollars?
49:23 — What an $800M founder had to unlearn about startups