🚀 The Startup Khan Newsletter

September 23, 2026 — Snorkel AI Hits $3.5B as Better Models Make Better Data More Valuable

🧠 Top Story: Snorkel AI Raises $350M at a $3.5B Valuation

Here’s one of AI’s more interesting paradoxes:

The smarter models become, the harder their next training data becomes to create.

Snorkel AI has raised $350 million at a $3.5 billion valuation, nearly tripling the $1.3 billion valuation attached to its previous financing in May 2025.

The round was led by Insight Partners and S32, with existing investors including Addition, Greylock and Wells Fargo participating.

But the valuation may not be the most interesting number.

CEO Alex Ratner says Snorkel’s annualized revenue run rate has climbed from roughly $20 million a year ago to more than $350 million today.

That’s company-reported rather than independently audited—but it points toward extraordinary demand for an increasingly important AI infrastructure layer.

What changed?

Snorkel started by helping companies automate data labeling.

Now it’s moving deeper into:

Finished training datasets + Reinforcement-learning environments

Its system combines human domain experts with thousands of specialized models and agents.

Experts create:

Problems → Scenarios → Rubrics → Grading criteria

AI increasingly handles:

Generation → Testing → Quality assurance → Scale

The company works with specialists across areas including coding, medicine and law.

Why founders should care

The AI data market is evolving.

Internet data

Human-labeled data

Synthetic data

Expert-designed environments + Synthetic generation + Human verification

Early models could learn enormous amounts from the open internet.

Frontier systems increasingly need harder material:

Expert reasoning.

Rare edge cases.

Difficult coding problems.

Medical knowledge.

Legal judgment.

Realistic simulations.

That means human expertise doesn’t necessarily become less valuable as AI improves.

The valuable part simply moves upstream.

👉 Founder takeaway: As intelligence becomes abundant, high-quality judgment becomes scarce.

And scarcity creates markets.

---

🔥 Funding Watch: Firecrawl Raises $75M to Give AI Agents Reliable Access to Reality

An AI agent can be brilliant and still fail because it has the wrong information.

Firecrawl wants to fix that.

The company has raised a $75 million Series B led by Smash Capital, with participation from Altos Ventures, Nexus Venture Partners, Y Combinator, Freestyle and Offline Ventures.

Firecrawl started by making websites easier for AI systems to crawl and convert into structured information.

Now it’s launching Alexandria, a broader knowledge-access layer combining:

Live web → Official data providers → Site-specific connectors → Firecrawl indexes

behind a common retrieval system for AI agents.

Firecrawl says agents using Alexandria produced 21% better answer quality than agents using built-in web tools across an internal 845-task evaluation.

That result comes from Firecrawl’s own testing, so treat it accordingly.

The bigger opportunity

A language model knows a lot.

An agent doing real work needs something different.

It needs to know:

What happened five minutes ago?

What does this customer’s account say?

What does the authoritative database say?

What information exists on this obscure website?

And perhaps most importantly:

Can I trust it?

That creates an increasingly important agent stack:

MODEL

Reason.

RETRIEVAL

Find.

KNOWLEDGE

Verify.

ACTION

Execute.

👉 Founder takeaway: AI agents don’t merely need intelligence.

They need reliable access to reality.

---

🛡️ Security Watch: AI Is Starting to Attack Itself

As AI gets better at writing software, it’s also getting better at finding ways to break it.

Palo Alto Networks has unveiled Unit 42 Continuous Frontier AI Defense, a system designed to continuously test web applications, APIs and cloud infrastructure for vulnerabilities using frontier and open-weight models.

The concept points toward a fundamental change in cybersecurity.

Traditional security:

Scan → Find vulnerability → Patch → Repeat

Agentic security:

Attack continuously

Reason about exploit paths

Identify weaknesses

Recommend remediation

Attack again

The result is effectively an AI red team that never clocks out.

Founder signal

This creates another emerging startup category:

Agents attacking systems before malicious agents can.

And it fits a broader pattern we’ve tracked all month.

Every increase in AI capability creates additional demand for:

Evaluation → Security → Permissions → Monitoring → Verification

AI doesn’t only create new products.

It creates new problems that require products.

---

🧰 Tool of the Day: Firecrawl

Today’s funding story also happens to highlight a useful founder tool.

Firecrawl converts websites into clean, structured information that AI applications and agents can consume.

That becomes useful when you’re building:

• Research agents
• Sales intelligence
• Competitive monitoring
• AI search
• Market research
• RAG applications
• Autonomous workflows

Instead of manually building and maintaining extraction logic for every website, developers can give their AI systems a more standardized way to access web information.

👉 Best use: When your agent is smart enough to reason—but keeps failing because its information pipeline isn’t.

---

📊 Trend Check: AI Is Creating Scarcity Through Abundance

Here’s a useful framework for understanding where startup value may move next.

Ask:

If models become abundant:

Reliable data becomes scarce.

→ Training-data infrastructure.

If generated content becomes abundant:

Verification becomes scarce.

→ Evaluation systems.

If AI agents become abundant:

Trust becomes scarce.

→ Governance and permissions.

If inference becomes abundant:

Electricity becomes scarce.

→ Energy infrastructure.

If synthetic information becomes abundant:

Authoritative knowledge becomes scarce.

→ Retrieval and provenance.

If machine intelligence becomes abundant:

Expert human judgment becomes scarce.

→ Expert-data networks.

This is one of the most useful startup frameworks in the current AI cycle.

👉 Don’t only chase what’s growing. Look for what that growth makes scarce.

---

🌟 Founder Spotlight: Build One Layer Away From the Model

Snorkel and Firecrawl point toward a useful founder strategy.

You don’t necessarily need to compete with frontier model companies.

Instead, ask what those models increasingly require.

Better training environments?

Build them.

Reliable real-time knowledge?

Build it.

Evaluation?

Build it.

Security?

Build it.

Identity and permissions?

Build them.

Energy?

Build it.

The biggest opportunity might be one layer away from the obvious technology.

That’s where constraints often become businesses.

---

🔴 Founder Opportunity Radar

🚨 September 25 — TechCrunch Disrupt Pricing Deadline

Current ticket savings for TechCrunch Disrupt 2026 expire September 25 at 11:59 p.m. PT.

The conference runs October 13–15 and expects more than 10,000 founders, investors and operators.

📣 September 30 — Disrupt Exhibitor Deadline

The startup exhibitor program has reopened, with companies able to book an exhibition table until September 30 at 11:59 p.m. PT.

For B2B founders actively fundraising or selling, exhibiting may offer more targeted value than simply attending.

🇨🇦 October 12 — Montréal Startup en résidence

Canadian founders have until October 12 at 11:59 p.m. ET to apply for Startup en résidence.

Selected companies receive:

C$15,000 grant

100 hours of personalized support

Eight months of programming

Workspace at Ax.c

Potential VC investment of up to C$250,000

An information session runs tomorrow, September 24 at noon ET.

👉 Founder reminder: Don’t evaluate programs by perks alone. Ask whether they materially improve your access to customers, capital or distribution.

---

⚡ Quick Win: Run a Scarcity Audit

Take the fastest-growing technology in your market.

Then ask:

If this becomes 10x cheaper, better and more abundant, what becomes harder?

More agents might create demand for better permissions.

More generated code might create demand for better verification.

More AI infrastructure might create demand for more electricity.

More synthetic content might create demand for trusted information.

Write down five answers.

Then ask:

Which constraint becomes painful enough that customers will pay to remove it?

That’s where the startup idea starts.

---

🧩 Final Thought

The AI market is revealing a powerful paradox.

Models are becoming:

More capable.

More plentiful.

Cheaper to access.

Yet their abundance makes adjacent resources more valuable:

Expert data.

Reliable knowledge.

Evaluation.

Security.

Energy.

Trust.

Snorkel AI and Firecrawl are two versions of the same strategy.

They aren’t trying to build the smartest model.

They’re building scarce infrastructure around increasingly abundant intelligence.

That’s the founder opportunity worth watching.

When the core technology becomes abundant, look for what its abundance makes scarce.

— The Startup Khan

Your 5-minute startup edge — funding news, tools, and founder playbooks.