ChatGPT and Claude are the new operating systems, but the moat is not in the interface
HubSpot's co-founder says models are becoming interchangeable infrastructure; no one will strangle entrepreneurs' supply chains. The real moat is accumulated workflows and data, not the chat interface users enter.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Startup ideas' quality can only be validated, not analyzed beforehand
Dharmesh says the common founder mistake is thinking detailed market analysis (he jokes, ‘eighteen columns in an Excel spreadsheet’) can predict whether an idea will work. But it's almost impossible to know an idea's quality or success probability beforehand. The only way is to build it fast and test it in the market. Even after success, the validated product often looks nothing like the original vision, retaining only ‘genetic’ similarities from the concept. Conclusion: whoever iterates fastest and learns cheapest wins long-term.
— Dharmesh ShahBreaking the ‘do one thing’ rule because customers fear fragmentation
HubSpot launched with content management, SEO, landing pages, forms—an entire marketing toolkit—violating the startup rule to master one thing. Dharmesh explains it wasn't greed: customers' real pain wasn't a missing feature but being overwhelmed and terrified by incompatible tools. The only solution was an ‘all-in-one’ package—a principle HubSpot kept for 20 years: first understand the customer's core problem, then define the product boundary.
— Dharmesh ShahOpen platforms eventually close; Twitter is the cautionary tale
Fifteen years ago, early Twitter was radically open—APIs spawned hundreds of third-party clients and companies—but Twitter gradually tightened and turned hostile to developers, shutting down products relying on its ecosystem. Dharmesh says it's not unique to Twitter but a platform disease: eventually platforms deliberately or accidentally crush the developers they cultivated. HubSpot's response: intentionally practice radical transparency with partners, saying upfront what it won't build.
— Dharmesh ShahHow to spot a dangerous startup idea: add the word ‘yet’ and see if it still holds
Simple test: if your startup thesis is ‘ChatGPT can't do this yet’ or ‘Claude Cowork can't do that yet’—and the claim still holds after the ‘yet’—it's a red flag because frontier labs will probably patch it soon. Second common mistake: comparing products, not companies. ChatGPT can't do it now, but that doesn't mean OpenAI can't ever. Safe ground is where frontier model companies face structural constraints (regulation) or incentive constraints (not worth their effort).
— Dharmesh ShahIn 2023 he feared model companies would strangle startups; now he doesn't
In 2023 Dharmesh's rule was ‘never compete with Sam Altman’—he felt frontier model companies held an unteachable ‘magic formula’ and controlled supply-chain chokepoints. But he says that was wrong. Now models are becoming infrastructure: price and replaceability trend toward commodity. Tools like Open Router make model switching easy. Startups won't be strangled by any one company. Building AI-native now is actually less scary than in 2023.
— Dharmesh ShahChatGPT and Claude are the new browsers; the interface isn't the endpoint
Take Uber Eats: the app itself isn't complex; the real value is restaurant partnerships and delivery logistics. Same with most B2B software—the interface is just how users do work, not the value itself. ChatGPT and Claude Cowork play the browser role: Google didn't monopolize the application layer; countless software companies thrived ‘running in the browser.’ Same now: companies like HubSpot stay as plugins and connectors plugged into this new ‘operating system’ instead of being replaced.
— Dharmesh ShahUse agentic coding to replay your codebase's architectural history
A newly possible operation: HubSpot has 20 years of GitHub history and complete evaluation sets. Spend, say, $200k in tokens to have AI rewind and replay an architectural decision—assume you'd picked a different data structure or framework back then, rewrite that part, validate with existing evals whether the new version truly optimizes quality, latency, cost. Incumbent companies can now cheaply answer ‘what if we'd chosen differently’—turning legacy weight into testable assets.
— Dharmesh ShahThe solo founder wave is usePot's real market soil
U.S. Census Bureau data: new company registrations hit historic highs—2025 saw approximately 5.6 million, 2026 through August surpassed 4 million, with no slowdown. Critically: the bureau predicts whether new companies will hire in two years. Those staying one-person for two years now exceed 70% and keep rising. This is the real market HubSpot Next's solo CRM usePot (launched 30 days ago) targets.
— Dharmesh ShahIn their own words · checked verbatim
Like Twitter had this in the early days, right? The original Twitter was this kind of very open ecosystem. We're a platform. And literally a thousand flowers bloomed back then.
Dharmesh Shah20:41
if you can just append the word yet to it, and it still holds water, that's a problem. It's like, oh, ChatGPT can't do that yet. Does that ring true?
Dharmesh Shah23:54
I have very few rules for business. One is never own a restaurant. Number two is don't ever be a landlord and charge rent. And number three is never compete with Sam Altman.
Dharmesh Shah28:14
the actual thing that's valuable is the work that's happening. The interface is a means to an end.
Dharmesh Shah32:17
if you just look at the kind of chat GPT and cloud co-working, whatever, they're effectively the new OS.
Dharmesh Shah34:17
anything that can be solved with a structured query deterministically, solve with a structured query deterministically. Don't use inference to solve something that a query can solve.
Dharmesh Shah35:19
in 2025, we had the most number of new company formation that the United States has ever seen.
Dharmesh Shah50:38
Figures
| U.S. new company registrations, 2025 | approximately 5.6 million | 50:38 |
| U.S. new company registrations, 2026 (through August) | over 4 million | 50:38 |
| New companies projected to remain solo beyond two years | over 70% | 51:40 |
| Days since usePot launch | 30 | 43:28 |
Glossary
- Yellow Brick Road
- The visible highway frontier AI labs are rapidly advancing.
- second brain
- Organized knowledge base of personal or enterprise unstructured information available for AI access.
- context graph
- Graph-structured storage of unstructured data for AI reasoning and agent execution.
- evals
- Automated test suite measuring whether model or code changes meet output quality standards.
How to listen
SaaS founders and product/engineering leads tracking how to choose startup directions in the AI era, or debating whether to compete with ChatGPT on interface.
Opening small talk and HubSpot's early founding history (0:00-6:00) have low information density and can be skipped.