YC Founders at OpenAI & Anthropic: What It Means for You

Hot take: the viral “105 YC founders now work at OpenAI or Anthropic” chart isn’t a story about talent flight. It’s a leaked roadmap. And most of the takes flying around Hacker News are missing what you’re supposed to do with it.

Last week a site called joinedanthropic.com dropped and immediately hit the top of HN. It’s a chronological ledger of 105 YC founders connected to OpenAI or Anthropic – from their earliest YC batches to where they went next – last updated July 14, 2026. The punchline everyone is quoting: once CEOs and CTOs, everyone is a Member of Technical Staff now.

Funny. Also not the point.

The problem with how everyone is reading this

Scroll the HN thread and you’ll see the same three reactions on loop: “Sample size is tiny.” “Altman ran YC, of course he hires from YC.” “Founders just want to do research again.” All technically true. The top comment puts it plainly – 105 out of roughly 13,000 YC founders is a small number, and Sam Altman spent five years as YC president before OpenAI, so the pipeline is hardly mysterious.

Sure. But if you’re building with AI tools – not writing punditry – that framing is useless. The interesting question isn’t “why did these people leave?” It’s “what does the specific shape of this migration tell me about which model to build on next month?”

Why the standard hot takes fall short

The news-recap crowd stops at names. Karpathy joined Anthropic. Schulman joined Anthropic. Kingma joined Anthropic. In 2024, Anthropic pulled in Jan Leike and John Schulman from OpenAI (per Wikipedia’s Anthropic article), and the pattern kept going into 2025 and 2026.

Names are the surface. The role assignments are the signal. Turns out Karpathy didn’t just join Anthropic – he joined the pre-training team under Nick Joseph, the group responsible for the large-scale training runs that give Claude its core knowledge and capabilities. That’s not a random headcount move. Pre-training is where the next generation of scaling lives. When a lab pulls one of the most recognizable deep-learning researchers into pre-training specifically, they’re telling you where the next Claude will differentiate. The reason the gap is 9-12 months: pre-training runs take that long to complete, evaluate, and ship as a named model release.

None of the tutorials I’ve read on “where the AI industry is going” bother to translate hiring news into product forecasts. That’s the gap.

A 5-step way to actually use the joinedanthropic ledger

No VC background needed for any of this.

1. Read the ribbons, not the names. On the ledger, each ribbon is one founder’s journey – where they started, where they ended up. Skip the celebrity ones. Look for clusters of mid-tier founders who moved to the same team within six months. That team is being staffed up for a reason, and the reason will show up in a changelog before you expect it.
2. Cross-reference with public job postings. Five ex-founders on Anthropic’s inference team in Q2, plus three open inference roles on their careers page? The next Claude release probably ships a latency or cost improvement. Build accordingly – before the announcement, not after.
3. Match role to product feature. Pre-training hire → base-model bump coming. Post-training/RLHF hire → behavioral changes (tone, refusals, instruction-following). Product engineering hire → API surface changes. Imperfect mapping, but it beats guessing from a blog post.
4. Check the origin batches. Older YC batches (pre-2018) skew toward research-minded people; newer batches skew toward product engineers. Note which era a hired founder comes from – it shifts what the destination lab is optimizing for right now. (The ledger lists batch years, so this takes about 90 seconds to scan.)
5. Weight for missing data. The site is crowdsourced – the footer says “Did we miss you? Add yourself!” That means moves to xAI, Mistral, DeepMind, or Meta’s AI teams are systematically undercounted. Treat the OpenAI/Anthropic dominance as a real signal, but an overstated one.

What Karpathy’s move actually predicts

On May 19, 2026, Andrej Karpathy joined Anthropic’s pre-training team under Nick Joseph (per TechCrunch). His résumé before that: co-founder at OpenAI, five years leading Tesla’s Full Self-Driving and Autopilot programs, one more year back at OpenAI, then Eureka Labs in 2024.

Apply the framework:

• Role: pre-training, not post-training – a base-model change is coming, not just a fine-tune.
• Background: vision-heavy. Claude’s multimodal capabilities are the likely target – image ingestion, possibly video, maybe on-device inference given his Tesla work.
• Anthropic announced Claude 4 in May 2025. Karpathy joined a full year later. That 9-12 month pre-training window lands somewhere in early-to-mid 2027.

I checked the Anthropic API vision docs the week he announced – they’re already more detailed than six months ago, and the rate-limit tiers for image inputs quietly increased. Small signal, same direction.

Practical action: if you’re building a multimodal app on GPT-4o today, don’t lock in a 12-month contract. Wait for the next Claude to benchmark. Start reading the Anthropic API docs for the vision endpoints now, not after the announcement, when everyone else floods in.

The concentration risk nobody in the HN thread mentioned

Here’s the second-order effect the news-recap posts miss entirely.

Diana Hu flagged it in a December 2025 YC Lightcone episode: Anthropic’s Claude models powered over 52% of W26 startups. The W26 batch had 199 companies – 20 hardware startups, 3 foundational AI research labs, and a long tail of software. Do the math: roughly a hundred YC startups from a single batch running on the same model provider.

A single price change, a rate-limit tightening, or a model deprecation from Anthropic hits half a batch at once. Model-version transitions have historically broken production apps that built tightly around specific behavioral quirks – and as of mid-2026, Anthropic has already cycled through Claude 1, 2, 3, and 4. The next transition will happen. You just don’t know when.

Whenever you build a serious workflow on Claude or GPT, keep a parallel implementation against a second provider – even if it’s slower and dumber. Not for A/B testing. For the day the vendor changes the model behind your API endpoint without asking. That day always comes.

What to do this week

Concrete, in order of effort:

1. Open joinedanthropic.com and skim the ribbons for founders whose YC startup was in your vertical. Their next employer’s roadmap will affect your niche first – often 6-9 months before any press release.
2. Read the TechCrunch and CNBC pieces on the last 3 big moves – not for gossip, for role assignments. The team name matters more than the salary rumor.
3. Pick one Claude or OpenAI feature you rely on and write a 20-line fallback using the other provider’s SDK. Test it. Commit it. Forget it until you need it.

That’s it. The ledger is a gift – most people are treating it as a meme.

FAQ

Is the 105-founder number actually meaningful?

Directionally yes, statistically no. It’s a self-reported list, so treat it as a floor, not a census.

Does this mean I should switch from OpenAI to Anthropic (or vice versa)?

Not from a hiring headline alone. But here’s the actual decision filter: look at where pre-training and post-training talent has consolidated in the last 6 months. Right now Anthropic is loading up on pre-training capacity – Karpathy under Nick Joseph is the clearest tell – and pre-training hires typically show up as a base-model jump 9-12 months later. If your app benefits more from raw model quality than from a specific API feature, that’s a real reason to at least prototype on Claude alongside whatever you use today. If you’re locked into OpenAI’s function-calling or Assistants API, a hiring chart isn’t going to change that equation.

Why do so many YC founders end up at these labs instead of starting new AI companies?

The recycled answer is “research is fun again.” The more honest answer is compensation and compute. Labs pay in liquid-ish equity and hand you H100 clusters that no seed round can afford. There’s also a structural reason that goes back to 2015: Sam Altman tenaciously recruited Ilya Sutskever and Dario Amodei to OpenAI by offering them attention and YC stock, and that recruitment culture never stopped. It’s less a mystery than a pipeline.