Has YC hit peak AI? (F24)
The latest batch is up to 86% AI startups.
Yesterday was the latest YC Demo Day and the first-ever Fall batch of the esteemed accelerator. In case you missed it, YC doubled the annual batches (from two to four) but halved the batch size in response. And we've seen that here, with 93 companies in the F24 cohort (compared with nearly 250 in previous batches).
But even with the smaller batches, YC isn't letting up on its AI focus. In fact, it's actually managed to increase the overall percentage of AI companies. 80 out of the 93 companies are related to AI or ML, or 86% - in contrast, previous batches were 79% for S24 and 65% for W24. I don't know whether we can call this "peak AI," but there isn't much room for AI to keep growing - only 1 out of every 7 YC companies isn't AI-related at this point.
With fewer companies presenting at Demo Day, drawing major trendlines from the data was a bit harder. But a few interesting patterns still emerged.
From text to speech
In the generative AI wave, most startups have focused on building with large language models like ChatGPT and Claude. Because so much of white-collar work is done via text, businesses have raced to extract value from text-based use cases: writing emails, summarizing content, generating code. Despite plenty of viral moments, image generation hasn't been nearly as widespread in its enterprise adoption.
However, voice models are emerging as a second modality with major B2B potential. This batch had a number of startups focused on voice and speech generation, each targeting different pieces of the stack:
Direct Customer Interaction: Companies like telli and Helpcare AI are automating outbound calls and patient scheduling, betting that voice AI has reached the quality threshold for customer-facing roles
Training and Simulation: Symphony is creating voice AI simulations for employee training, representing a shift from text-based scenarios to more immersive learning experiences
Specialized Applications: Lightscreen AI is tackling interview automation, while ISSEN focuses on language learning through real-time voice companions
The interesting thing here is the productization of text-to-voice, not just the generation. One year ago we saw plenty of models only focused on creating speech (i.e. ElevenLabs), with API credits being the sole business model. Now, startups are trying to turn those models and APIs into full-fledged products.
It took a lot of work to get there - quite a few advances were needed, like real-time audio processing, the ability to handle interruptions, and various audio-based guardrails - but seeing these companies suggests we're reaching a tipping point where voice AI is becoming practical for business use cases.
It's also notable that we're seeing startups focus on voice-specific infrastructure. Just as LLMs require an ecosystem of deployment, monitoring, and optimization tools, voice AI comes with unique technical challenges. Companies like fixa ("Sentry for AI voice agents") and Vocera are building specialized monitoring tools, while Protegee is developing payment infrastructure specifically for voice interactions. What remains to be seen is whether AI infrastructure for speech will carve out its own niche, or whether it will eventually merge with existing AI Ops platforms.
The differentiation dilemma
A recent TechCrunch analysis pointed out an uncomfortable truth about YC companies: uniqueness isn't actually a prerequisite for success.
A deep dive into the data from all of the nearly 5,000 companies YC has backed to date reveals a surprising truth: YC startups don’t have to be unique. Far from it.
YC commonly accepts startups that are building similar or nearly identical products to previous YC grads. Some of them are direct competitors; others differ slightly by targeting a new geography (Asia or Latin America), or are a subset of a larger market (point-of-sale software for bars versus coffee shops).
More than a dozen startups building AI code editors went through the YC program between 2022 and 2024 — some in the same batch with the same YC partner.
PearAI, which was part of this batch, was a particularly pointed example of this. The AI coding editor went viral for all the wrong reasons after being called out as a clone/fork of a different YC-backed AI coding editor. To make matters worse, one of the founders admitted to cloning the original repo and papering over the open-source license with a ChatGPT-written replacement.
Yet, it still points to a critical question: where are the moats?
YC produces many similar companies, especially when you look at cohorts over time. After looking at over a year's worth of YC AI startups, I've seen countless repetitions of AI code editors, AI sales reps, AI customer support, AI drug discovery, AI email clients, AI for lawyers, AI for doctors, AI for accountants, and more. Surely, these are not all destined to become billion-dollar companies - but how are YC founders thinking about breaking through the noise and differentiating themselves from their batchmates, let alone the broader market?
We've seen some clear differentiation strategies in other batches: targeting specific niches (e.g., AI for dentists), workflows, geographies/countries, or even training your own models. But it's unclear whether that's ultimately enough to build a moat.
Waiting for GPT
Like I said at the top - this wasn't a huge batch, so it's hard to glean too many insights. But I was still left feeling like the AI startups have started stagnating. I don't have any data to support this, but besides the improvements in voice infrastructure, most of the companies wouldn't have been out of place in any other recent batch. It all feels a bit... same-y.
It may be because LLMs, in general, have stagnated a bit. For 18 months, the state of the art in LLMs didn't change all that much - after GPT-4, newer models added features like bigger context windows, prompt caching, and structured outputs, but not as much on the raw intelligence front.
Of course, that changed with o1-preview, a new model that brought advances in reasoning capabilities in September. That might be opening the door for more diverse startups in the coming months as they incorporate o1 into new ideas. Everyone, including its maker, is trying to figure out what o1 is best at - so far, the answer seems to be complex problems that require considering multiple solutions at the same time. But startups may be more adept at finding newly unlocked use cases.
Similarly, Google's recent Genie 2 demo points toward another frontier: world modeling. Genie 2 shows how AI can move beyond processing individual frames or sequences to understanding and recreating the underlying rules and dynamics of physical systems by generating complete interactive environments from text or image prompts. I can already imagine plenty of companies trying to democratize similar systems outside of Google's kingdom or building infrastructure to apply models like Genie 2 to robotics or multimodal LLMs.
We're now two years into our post-ChatGPT world, and we may be closing one chapter of AI startups and opening another. But for founders and investors alike, the challenge will be recognizing which of these new directions represents genuine paradigm shifts versus incremental improvements. As YC's new seasonal format suggests, the pace of innovation isn't slowing - if anything, it's accelerating. The next few batches should prove fascinating to watch.