Character.AI users spent over two hours per session on average through 2024-25 — engagement that beat most consumer social apps and still produced a $150M Series A and a Google distribution deal rather than an IPO trajectory. Replika ($11.2M Series A) hit the wall faster: Italy's Garante banned the app for minors in 2023, the EU AI Act's 2025 consumer provisions reopened the same questions, and the product is now running at a fraction of its 2022 traffic. The structural tension across 50 published chatbot startups with $321B in cumulative funding is that engagement and revenue keep refusing to line up.
Voice is the new wedge
ElevenLabs ($922M Series D, UK) and Deepgram ($109M Series C) are the venture-backed voice pure-plays. Cartesia ($65M Series A) competes on sub-100ms latency for real-time voice. The bet across all three is that voice quality and turn-taking — not text quality — is where the next round of consumer chat differentiation gets fought, especially as OpenAI's Realtime API, Google's Gemini Live, and Anthropic's voice surfaces commoditize the basic chat layer. Sierra ($1.27B Series C) and Decagon ($296M Series B) take the same wedge into enterprise: voice-first customer support agents with integration depth that horizontal chat can't replicate.
Companion economics keep breaking
Inflection AI ($650M raised) was effectively absorbed into Microsoft mid-2024 — the team and Pi went, the corporate shell stayed. Character.AI struck a Google licensing-and-distribution deal in late 2024 that monetized the model and team while the consumer app continued. Replika and Janitor AI sit at the regulated edge: Italy's Garante action, EU AI Act consumer provisions, and ongoing debate about adolescent users have made pure-companion business models structurally difficult to fund. Moonshot AI ($300M Series B, Kimi) is the published Chinese long-context leader and operates in a regulatory environment western labs can't enter.
What survives
The horizontal "ChatGPT alternative" tier without a wedge keeps thinning. Survivors have one of three things: voice differentiation (ElevenLabs, Cartesia), enterprise integration depth (Sierra, Decagon, Kore.ai), or geographic distribution moat (Moonshot). Frontier labs — OpenAI, Anthropic, xAI ($56B Series E) — own the default consumer surface through OS, browser, and phone placement. Model commoditization is the only force that meaningfully widens the survivable middle, and 2026 is when we find out whether DeepSeek-class price compression actually loosens frontier-lab pricing power.