Chai AI

Verdict

CHAI (chai-research.com) is run by Palo Alto-based Chai Research Corp., founded by former quant trader William Beauchamp. It is one of the largest mobile AI roleplay apps in the world: $120M annualized revenue run-rate as of October 2026 ($5.1M/month in net iOS and Google Play revenue, plus 40% from ads), 1M+ daily active users, and only 15–20 employees.

Most companion apps wrap one third-party LLM. CHAI runs its own 1.4 Exaflop GPU cluster (~5,000 NVIDIA A100/L40S and AMD Instinct MI300X/MI325X GPUs, backed by AMD Ventures and CoreWeave). It serves 1.2 trillion tokens a day across 51,000+ blended LLMs, including post-trained 2x 235B A22B MoE, DeepSeek-V4, and Kimi-K2. The catch: CHAI is mobile-only (iOS and Android). There is no browser chat on chai-research.com, and it is text-only, with no selfies or video.

---

Oct 2026 Scale: $120M ARR, $55M+ Funding and Compute

chai-research.com and /company-facts/ (reviewed October 2026) share more financial and infrastructure detail than most companion companies:

• $120M run-rate: CHAI passed $100M ARR in July 2026 and hit $120M on October 1, 2026. Reuters covered it on October 14, 2026.
• 60% subscriptions, 40% ads: Sensor Tower logged $5.1M in worldwide net app-store revenue from September 10 to October 9, 2026 across Apple and Google Play. Ads shown to free users make up about 40% of revenue.
• 15–20 people in Palo Alto: The same small team handles model post-training, custom CUDA/ROCm kernels, and the mobile apps (Company Overview table):

---

Inside the 1.4 Exaflop Cluster and 51,000+ Blended LLMs

1. Hardware: NVIDIA A100 / L40S and AMD MI300X / MI325X

The /cluster page and homepage explain why CHAI left third-party inference providers:

• 5,000+ GPUs, 1.2T tokens a day: About 5,000 NVIDIA A100, NVIDIA L40S, AMD Instinct MI300X, and AMD Instinct MI325X accelerators, run with custom quantization, Flash Attention, and KV-cache compression instead of stock vLLM.

2. Model Blending: 2x 235B A22B, DeepSeek-V4, Kimi-K2 and GSPO

CHAI doesn't send every turn to one model. It uses Model Blending ("Blending Is All You Need", arXiv:2401.02994) and reinforcement learning on real user signals ("Rewarding Chatbots for Real-World Engagement with Millions of Users", arXiv:2303.06135):

• 2x 235B A22B MoE: Adding 235B A22B MoE models to the production blend raised user screen time and revenue by +25%.
• June 2026 post-training: In CHAI AI: Post Training to Achieve SOTA MoE, DeepSeek-V4 and Kimi-K2 post-trained with GSPO beat GRPO in live A/B tests across 1M+ daily users and lifted Day-30 retention.
• CHAI Memo 2026: The ROADMAP link opens an 8-page investor and engineering memo (chai_roadmap_2026.pdf) on Revenue Growth, Scaling Compute, and Cohort Retention Smiles:

---

Mobile App, Safety Filters and the 2026 Policy Shift

CHAI is on the iOS App Store (apps.apple.com/us/app/chai-chat-with-ai-bots/id1544750895) and Google Play (play.google.com/store/apps/details?id=com.Beauchamp.Messenger.external).

• Fast mobile chat and bot creation: Swipe through a feed of community characters, start a text roleplay, reroll replies, or build a bot in under 60 seconds.
• NSFW and safety in Oct 2026: CHAI made its name with a loose filter on spicy user bots. Section 08 Safety of /company-facts/ now says:
  • CHAI is rated 18+ on Apple and Google Play, and some regions require a paid subscription.
  • Every message is scanned in real time for self-harm and minor-sexualization content.
  • CHAI sells mainly through Apple and Google (where it earns $5.1M+/mo), so its store policy says safety systems refuse restricted material, and users can toggle filters to hide mature content. For zero-filter web roleplay outside app store rules, Janitor AI or SpicyChat AI are less exposed to store crackdowns.

---

Pricing: Free (Ads) vs CHAI Premium and CHAI Ultra

---

What It Does Well

• Real in-house research: CHAI post-trains its own models (DeepSeek-V4, Kimi-K2, 235B A22B MoE) with GSPO and RLHF on retention data from 1M+ daily users.
• Fast replies: 5,000+ NVIDIA and AMD GPUs serving 1.2T tokens a day mean near-instant streaming on iOS and Android.
• Huge bot library: Millions of community characters across romance, fantasy, anime, and dark roleplay.

Where It Falls Short

• No web chat: The CHAI WEB link on chai-research.com does nothing. You must install the iOS or Android app.
• Text only: No image generator, selfies, or video.
• Ads and a pricey top tier: Free users get frequent ad breaks (40% of CHAI's revenue), and the Ultra models cost $29.99/mo.

---

Who Should Skip It

• You want to chat in a browser without an app: Try SpicyChat AI, which also runs DeepSeek V4 and Kimi K2 on the web, or Janitor AI.
• You want selfies, voice calls, and NSFW video: Try EroLove, Candy AI, or FantasyGF.
• You want a long-term companion with visual memory maps on web and mobile: Try Nomi AI.

---

FAQ

Can you use CHAI AI on a web browser in 2026?

No. chai-research.com shows a disabled CHAI WEB button in the footer. All discovery and chat happen in the CHAI iOS and Android apps.

What AI models does CHAI use?

CHAI blends 51,000+ in-house fine-tuned LLMs on a 1.4 Exaflop cluster (~5,000 NVIDIA and AMD GPUs), including post-trained 2x 235B A22B Mixture-of-Experts, DeepSeek-V4, and Kimi-K2 models tuned with GSPO and RLHF.

How much does CHAI Premium and CHAI Ultra cost?

The Free tier has ads. CHAI Premium is $13.99/month ($134.99/year) for ad-free unlimited messages. CHAI Ultra is $29.99/month ($269.99/year) for priority access to the largest MoE models and longer context memory.