Humans&, founded by researchers who left Anthropic, xAI, and Google, announced a $480 million seed round on January 20, 2026, valuing the three-month-old company at $4.48 billion. The startup has roughly 20 employees and nothing to sell yet.
The math on this is strange
At 9x the capital raised, Humans& joins a small club of AI startups commanding valuations that would have seemed absurd two years ago. Mira Murati's Thinking Machines Lab raised $2 billion at a $12 billion valuation last July. Ilya Sutskever's Safe Superintelligence hit $32 billion in April despite not releasing anything publicly. Humans& is smaller than both by comparison, but the pattern holds: investors are paying for pedigree, not products.
SV Angel led the round alongside co-founder Georges Harik. Nvidia wrote a check, as did Jeff Bezos. GV, Emerson Collective, and a roster of others piled in.
Who are these people?
The founding team reads like an AI lab alumni directory. Eric Zelikman left xAI in September after working on Grok 2's pretraining data and building the reinforcement learning infrastructure for Grok 3 and 4. Before that, he was a PhD candidate at Stanford, where he authored the STaR paper on teaching language models to reason using their own rationales.
Georges Harik was Google's seventh employee, give or take. He helped build AdWords and AdSense, worked on Gmail, and was involved in the Android acquisition. He's been angel investing since leaving Google in 2005, but this marks his return to company building.
Andi Peng worked at Anthropic on reinforcement learning and post-training for Claude models from 3.5 through 4.5. She's a PhD student at MIT on leave. Noah Goodman is a Stanford professor of psychology and computer science who also holds a research engineer position at Google DeepMind. Yuchen He rounds out the founding group, another xAI researcher who contributed to Grok.
The pitch
Humans& says it's building "human-centric AI," which sounds like marketing until you hear how the founders frame it. The idea is that current AI development optimizes for autonomous systems, models that can work for hours without human input. Humans& wants to go the other direction: AI that coordinates people rather than replacing them.
According to TechCrunch, the company's focus areas include long-horizon reinforcement learning, multi-agent systems, and something they call "user understanding." That last one is interesting. Current chatbots remember nothing between sessions. Humans& wants AI that asks clarifying questions, holds onto context, and learns what specific teams actually need.
Why did Peng leave Anthropic?
This is where it gets pointed. Anthropic has been pushing toward agentic AI, models that can operate autonomously for extended periods. Peng apparently wasn't sold. According to reporting from Tekedia, she described Anthropic's approach as highlighting how its models could churn for 8, 24, or 50 hours by themselves to complete tasks.
That's a philosophical split worth watching. OpenAI, Anthropic, and xAI are all racing toward autonomous agents. Humans& is betting that's the wrong direction, or at least not the only one that matters.
The valuation question
Is $4.48 billion reasonable for a company with no product? By traditional metrics, no. A SaaS company at this valuation would need $50 million in annual recurring revenue growing at 100% year over year. Humans& has zero.
But traditional metrics stopped applying to AI labs around 2023. The comparison set isn't enterprise software; it's other pre-product frontier labs. Thinking Machines hit $12 billion. Safe Superintelligence reached $32 billion. The pattern suggests investors are pricing in a small probability of building something that matters at massive scale.
The resource gap is real, though. Anthropic raised $7.3 billion in 2024 alone. OpenAI is reportedly seeking to raise at valuations approaching $500 billion. xAI has been building one of the largest GPU clusters in existence. Humans& has $480 million. That buys meaningful compute but not infinite runway.
What happens in 12 months?
The company says it'll launch a product early this year. Think of something like an AI-enhanced collaboration tool, a Slack-type product where the AI actually remembers your team's context and can help with research. That's the pitch, anyway.
Whether the technical approach works remains to be seen. Long-horizon reinforcement learning is genuinely hard. Multi-agent coordination is hard. Building AI that understands users over time rather than treating each interaction as isolated, that's hard too. The founders have relevant research backgrounds, but research papers and shipping products require different things.
The other question: can $480 million compete against labs spending billions? Zelikman has argued that most spending at major labs goes toward inference, multimodal systems, and product engineering rather than pure research. If Humans& is doing something genuinely different, he claims, they might not need the same scale.
Maybe. Or maybe that's what every well-funded startup tells itself.
