It's time again for YC's Demo Day - which means it's time to unpack the flood of AI and ML companies launching out of the startup accelerator.
This time around, 194 companies were related to AI or ML - up from 158 in the last batch and a whopping 80% of the current batch.
While the broader investor community is beginning to question AI’s ROI, the startup community appears to still be bullish on artificial intelligence.
Nowhere is safe from AI
AI startups are launching across a huge range of industries and use cases. Yet the top categories remained similar:
Developer Tools and AI Ops: Helping companies build, deploy, and monitor AI models. Examples include Outerport (efficient GPU usage), Patched (AI workflow automation), Zenbase (prompt engineering), Storia (codebase AI assistant), and many others.
SaaS and B2B Tools: AI-powered software for functions like sales, marketing, customer support, procurement, accounting, and more. Notable startups include Odo (winning government contracts), Tabular (AI accounting), and Callback (outsourcing repetitive tasks).
Healthcare & Biotech: Streamlining operations for healthcare providers and advancing drug discovery and biotechnology. Standouts include Ångström (AI wet lab simulations), Vera Health (clinical decision support), and AminoAnalytica (protein design models).
Nearly 40 categories were represented (depending on how you view it) from insurance and legal to manufacturing and mining to gaming and kids' entertainment. Clearly, no sector is safe from potential AI disruption.
From copilots to agents to employees
Another key shift is the move from generic AI assistants toward industry- and task-specific AI "agents" that aim to automate entire job functions. While some startups are still using "copilot" as a buzzword, clearly the name of the game is now "agents."
Or, in some cases, we're moving beyond agents to "AI employees": AI sales associates, financial analysts, designers, support reps, and wholesale buyers, to name a few.
The promise is that these AI agents can handle 80%+ of the repetitive, predictable parts of the role by fetching relevant information, drafting emails and documents, updating systems of record, and engaging in dialog. This would allow human workers to focus on higher-level strategy, edge cases, and relationship building.
It remains to be seen how well these AI agent products will actually work in practice, given the limitations and challenges of building with LLMs. Reliably embodying an "AI employee" is going to be extremely difficult, but seems to be the current pitch to investors.
The maturing AI stack
Building real-world AI systems is hard. From curating training data to prompt engineering to building evals and adding observability, modern LLM-driven apps require a lot of infrastructure and tooling.
As we've seen over the last few batches, the AI tech stack is continuing to mature. These days, we're still seeing several types of "AI Ops" companies:
Data curation platforms to prepare training data for specialized domains
AI observability tools to monitor and troubleshoot models in production
Testing infrastructure to evaluate model accuracy, bias, toxicity, and other key properties
Hallucination mitigation techniques to reduce model confabulation
Alignment-as-service to bake AI safety best practices into model development
It's still early days, and most AI systems don't even have "best practices" established yet. But
We're still in the early stages of the AI infrastructure ecosystem, and best practices remain to be established. However, an increasing number of startups are laser-focused on solving specific challenges AI developers face today.
AI Combinator
Paul Graham has long written about how the cost of starting a startup has decreased dramatically, enabling orders of magnitude more people to do it. In the past, those costs have been primarily in servers and infrastructure.
Now, training and serving ML models has dramatically decreased in cost - and YC appears poised to be going all-in on the AI boom. In the past three batches alone, YC has graduated nearly five hundred companies related to AI in some way. By backing such a large number of AI startups, YC is aiming to own a piece of as many winners as possible.
The recent decision to switch to four batches a year also enables this spray-and-pray approach. With AI advancing so rapidly, YC doesn't want to miss out on the next breakout company just because the timing didn't align with the traditional batch calendar.
But there's clearly tons of company overlap - even within a given batch. And with so many AI startups crowding into the same spaces, it remains to be seen which ones will break out and achieve the network effects and scale advantages to dominate their categories.
If you're a startup building yet another AI-enabled solution for sales or customer service, how do you think about differentiating yourself from your batchmates?
Looking ahead
Stepping back, the scale and diversity of the YC batch underscores that we are in the early innings of the "AI revolution." The core building blocks are in place (or are quickly being invented) to sustain this Cambrian explosion of AI applications.
If anything, the companies we're seeing aren't imaginative enough - with so many B2B SaaS startups and developer tools, I'm left wondering where all of the truly "AI-native" companies are.
Of course, many of these startups will fail. Some are too early, some will be outcompeted, some will simply not find product-market fit. But if YC's track record is anything to go by, a meaningful fraction will go on to create immense value and reshape industries over the next 5-10 years.
