Abacus AI Just Dropped AI Supercomputer With Claude and Gemini

[music] >> Abacus AI just launched AI Supercomputer and the basic pitch is this. For around $10 a month, Abacus wants to give users an always-on cloud computer where AI agents can build, deploy, [music] host, and run real software. Not just generate code or create a prototype. Actually build the app, connect the database, expose it through HTTPS, keep it online, and let users access it through a real public URL. Now, the AI coding race is already crowded. What Abacus is trying to add is the missing infrastructure layer around those agents. According to them, Supercomputer gives users an Ubuntu Linux cloud environment with two vCPUs, 8 GB of RAM, persistent disk storage, built-in SQL databases, S3-compatible cloud storage, terminal access, SSH access, GitHub integration, AWS integration, custom subdomains, and one-click HTTPS deployment through abacus.ai.cloud. The gap is simple. AI can write the app, but someone still has to make it live. Abacus is trying to make that last step part of the agent's job, too. The platform gives users access to multiple AI coding agents in the same environment. In the launch material, Abacus talks about its own Abacus AI CLI, OpenAI Codex running GPT 5.5, Claude code with Sonnet 4.6 and Opus 4.7, and Google Antigravity running on Gemini 3.5 Flash. The exact model versions may change over time because these platforms update constantly, so it is worth checking the live Abacus page before using exact names in production. The broader point is that Supercomputer is designed as a place where different coding agents can work with a real file system, terminal, database, and deployment environment. One of the clearest examples is the open source AI hosting demo. In that demo, Abacus shows its CLI deploying Quen 2.5, a half billion parameter open source language model with a public chat interface. The user gives a natural language request and the agent handles the setup. It checks the supercomputer VM ingress settings to understand the available ports. It creates a Flask backend to serve the model. It configures Nginx with proxy headers so the app can be accessed through public HTTPS. It downloads the Quen model with progress tracking and automatic resume support in case the network interrupts. >> [music] >> It deals with disk space constraints. It verifies that the app works locally on localhost. Then it checks that the public abacusai.cloud URL is working. The final result is a working chat interface called Quen Chat with conversation history, >> [music] >> new chat creation, markdown rendering, and the model running on the user's own infrastructure. That is one of the strongest parts of the whole launch because it connects directly to the private AI trend. A lot of businesses want AI features, [music] yet they also want more control over where data goes. If you are building something around healthcare, legal work, internal company documents, customer records, financial data, or anything sensitive, sending every request through a third-party API can become a serious issue. With this kind of setup, the model runs inside your own cloud environment. You are not relying on OpenAI or Anthropic APIs for every response in that specific deployment. You are not dealing with per token API costs in the same way. You are not building the whole product [music] around external rate limits. And you have more control over the stack. Now, this does not automatically make every app compliant with HIPAA, GDPR, SOC 2, or anything [music] else. Real compliance still depends on on full product, the data flow, access controls, logging, policies, contracts, and how the organization handles user data. Still, the infrastructure direction makes sense for people who want private AI apps and more control than a normal hosted chatbot gives them. Abacus also publishes security claims around its broader platform. Their security material says customer data is encrypted at rest with AES 256 and encrypted in transit using TLS 1.2 or higher. They say customer prompts and responses are not used to train large language models and that customers retain ownership of inputs and outputs. They also claim GDPR and CCPA compliance, plus SOC 2 controls. Those are important claims for businesses, [music] especially if Abacus is trying to position this as more than a toy for hobby project. The second major demo is an always-on CRM. This one is less flashy than the AI model demo, yet it may be more useful for understanding what Supercomputer is supposed to be. In the demo, Google Antigravity builds a Django-based enterprise CRM from a single prompt and deploys it to mycrm.abacus.cloud. The app includes customer management, deal pipelines, task scheduling, analytics dashboards, [music] and role-based access control. It uses PostgreSQL as the database, Gunicorn as the app server, and Nginx as a reverse proxy. That stack matters because it is the kind of setup people actually use for real web apps. It is not just a pretty landing page. It has users, it has records, [music] it has database tables, it has state. It has uploaded files, it has sessions, it has logs. In the demo, users can add new customers, create deals worth tens of thousands of dollars, move deals across a visual Kanban board from new to qualified to proposal to closed one, schedule follow-up calls, log notes, and see analytics update as the data changes. The most important phrase here is always on. A normal local AI-generated app may work while your dev server is running. This CRM is presented as a live public server. Someone in one country can create a customer record, someone else can update a deal later, and someone else can check analytics from another location. The server keeps running around the clock. That is why Abacus keeps pushing the idea that supercomputer is not just a sandbox. >> [music] >> It is meant to be a hosting environment. The app remains deployed, the database persists, the storage persists. The public URL remains available. The third demo goes in a more visual direction, a 3D browser game. Abacus shows Cloud Code running Sonic 4.6 building a game called Stranger Runner. It is a Stranger [music] Things-inspired endless runner with three-lane Subway Surfers-style gameplay. Technically, it uses 3.js r160, WebGL, Unreal Bloom pass for neon glow, cinematic tone mapping, custom shadows, dynamic fog, and two shifting dimensions. One dimension is Hawkins at night with streetlights and abandoned cars. The other is the upside down with glowing tendrils and Demogorgon-style obstacles. The player dodges barriers, collects Eggo waffles for points, grabs shields for temporary invincibility, and the speed ramps from 17 to 50 [music] units per second as the difficulty increases. The AI also creates combo scoring, high score persistence through local storage, a CRT-style interface with scanline effects, and mobile responsive swipe controls. According to the demo material, the whole thing goes from prompt to live public game in 8 minutes and 52 [music] seconds. That is a very specific demo, and it is clearly designed to show speed. The point is not [music] that this game will beat a professional studio project. The point is that an AI agent can create rendering logic, movement, UI, scoring, mobile [music] controls, visual effects, and deployment in one short workflow. Then there is the macOS app [music] demo, which is a different category again. In that one, Codex CLI with GPT 5.5 builds a desktop app called Night Owl. It is described as a second brain workspace with a dark academia corkboard style. The app has digital sticky notes, textured paper, pinned corkboard visuals, subtle shadows, tape effects, and an animated AI owl companion. The AI does not only generate the app code. It also creates concept artwork to guide the design, scaffolds an electron and vite project, sets up TypeScript, resolves NPM dependency conflicts, adds semantic connections between notes, and includes rag-based contextual recall so the [music] user can chat with the owl about their own notes. The assistant can summarize themes, find patterns, >> [music] >> and surface forgotten ideas from the workspace. Then the agent runs a production build, takes headless Chrome screenshots to visually check the interface, and packages the project as a macOS release zip [music] file. Now, Supercomputer fits into the trend where AI systems are starting to take action because it gives the operator somewhere to work. >> [music] >> The model is one part of the system. The agent loop is another part. The tools, files, terminal, database, storage, [music] hosting, logs, and deployment pipeline are the rest of the system. A stronger model alone can write better code. A stronger agent inside a real cloud environment can potentially ship more complete products. That is the shift people should understand. At the same time, there are limits. Two vCPUs and 8 GB of RAM are useful for many projects, demos, internal tools, prototypes, smaller web apps, APIs, dashboards, [music] and lightweight AI services. They are not magic unlimited infrastructure. Heavy GPU workloads, large-scale production traffic, massive databases, enterprise-grade security reviews, complex devops setups, and serious customer-facing apps still require real engineering decisions. Also, most of the exact supercomputer claims right now come from Abacus's own pages, FAQs, [music] and launch demos. There are outside user reviews of Abacus as a broader platform with some users praising the value and all-in-one model access, while others complain about credits, bugs, reliability, or pricing clarity. So, it is smart to treat this as a promising launch and a strong direction, not as a fully proven replacement for every developer workflow. That said, the product is still interesting because it points at where AI tools are going. The next wave of AI coding may not be only about which model gets the highest benchmark score. It may be about which system can take an idea and move it further down the actual software pipeline. Let me know what you think about this one in the comments. Is this the kind of AI tool that actually helps builders, or is it still too early to trust agents with full software deployment? Like the video if you found it useful, subscribe for more AI updates, and thanks for watching. I'll catch you in the next one.