Two years ago, everyone said the same thing: "AI startups? Just a thin UI slapped on top of GPT. Once the model gets better, the app disappears." The "GPT wrapper" mockery, basically.
Then in June 2026, SpaceX announced it would acquire AI coding startup Cursor for $60B — an all-stock deal set to close in Q3. So how did a "wrapper" end up here? a16z dug into the data to find out.
Here's what everyone predicted
The a16z authors (Kimberly Tan, Joe Schmidt, Marc Andrusko, and Olivia Moore) open by admitting the skepticism was reasonable. "Flashy demos are easy. Substantive products are hard," they write. And the skepticism had real grounding — model inference prices dropped from $30 to under $5 per million tokens in two years. OpenAI slashed o3 pricing 80% overnight in June 2025 (input $10→$2, output $40→$8).
If prices were collapsing that fast, what was the point of a "product" built on top? The closer the model gets to free, the less reason exists for an app that just calls the model — that was the consensus back then.
But the numbers went the other way
Cursor crossed $2B in ARR in February 2026 — double what it was three months earlier. Four months later, SpaceX moved to acquire it for $60B, just four days after SpaceX's own IPO. This happened in the exact category everyone said was "just calling the GPT API."
a16z reads this as proof the growth benchmark itself has shifted. Hitting $1M ARR at Series A used to be a strong result. Now it's below median. The reason: enterprises now carve out dedicated AI budgets and pull solutions toward them. Old-school SaaS meant salespeople chasing buyers. Now buyers are lining up.
But here's what's easy to miss: the easier demos get, the more the risk of production failures lands squarely on the company. Remember Air Canada's chatbot inventing a bereavement fare policy that didn't exist? A Canadian tribunal rejected the airline's argument that the chatbot was "a separate legal entity" from the company. The ruling was blunt: "everything on the website is the company's responsibility." It's the textbook case of a risk invisible in a demo that only shows up in production.
So what's the real moat, then?
a16z identified four things the surviving AI companies have in common.
| Moat type | Definition | Real example |
|---|---|---|
| System of record | Own the point where data is created, then build workflows on top | Eve, Salient, Toma — capturing voice calls and unstructured data at points software never reached before |
| Workflow lock-in | Embedded in the daily human-AI interaction loop | Decagon — AI handles support, humans monitor/tweak/analyze through the same screen |
| Deep legacy integration | Connected to messy legacy systems competitors don't want to replicate | Tennr (healthcare fax systems), HappyRobot (homegrown trucking TMS), Glean (enterprise tool sprawl) |
| Relationship entrenchment | Becoming a strategic partner shaping the customer's roadmap, not just a vendor | What a16z calls "relationship entrenchment" |
But that table shouldn't go unquestioned. Is workflow lock-in really safe? Some VCs argue AI agents are simply bypassing the defenses that UI and ecosystem integrations used to provide. Switching cost ultimately comes down to engineering time — and that's exactly what AI is collapsing.
Clouded Judgement takes it a step further: the era of defending data by where it's stored is over, but that doesn't mean defensibility vanished — it just moved up a layer, to who orchestrates the workflow. Lock-in didn't disappear. It relocated. The same shift shows up outside the US too — enterprise AI is moving past personal productivity tools and into core organizational infrastructure, with companies building AI they can trust inside their data, permission, and security boundaries.
Run your product through this moat check
- System of record — "If we disappeared, would that data exist anywhere else?"
Check whether you're capturing data at a point software never captured before — voice calls, unstructured documents. Eve, Salient, and Toma all built toward becoming the system of record this way. - Workflow lock-in — "Does the customer feel uneasy if they skip opening this screen?"
See whether your product sits inside the loop where AI does the work and humans monitor, tweak, and analyze it. Being the "checkpoint for AI output" makes you far harder to replace, the way Decagon is. - Legacy integration — "Are we plugged into something messy enough that competitors won't bother copying it?"
Look at whether you've built the unglamorous, time-consuming integrations — like Tennr's healthcare fax connections or HappyRobot's homegrown TMS work. The messier it is, the harder it is to clone. - Relationship entrenchment — "Does the customer invite us to roadmap meetings, or do they just call us a vendor?"
Ask yourself honestly: is this a relationship that only surfaces at contract renewal, or one where you're shaping next quarter's strategy together? The more it's the latter, the less you compete on price.
A caveat
Checking one box on this list doesn't mean you're safe. Even a16z is clear: "pure execution speed isn't enough." The companies that last are usually stacking multiple moats at once. Lean on just one, and the moment it's breached, everything shakes.
Want to go deeper?
From Demos to Deals The original a16z piece — full breakdown from the demo-to-product gap to all four moats a16z.com
SpaceX to acquire Cursor for $60B The full structure of the deal that closed four days after SpaceX's own IPO techcrunch.com
Cursor crosses $2B ARR How revenue doubled in three months, and the shift toward enterprise customers mlq.ai
Moffatt v. Air Canada case analysis The legal breakdown of the first ruling on chatbot liability mccarthy.ca
VCs Rethink Startup Moats As AI Compresses Time To Build The counterargument that workflow lock-in is eroding forbes.com
Workflows are King The case that defensibility moved from the data layer to the orchestration layer cloudedjudgement.substack.com
