The AI Stack One Team Runs, By the Numbers

One integrated team powers the complete Hax AI stack that delivers fast image generation on substantial GPU resources, maintains thousands of active agent memories, tracks hundreds of live crypto assets, serves LLM responses in milliseconds, and publishes growing blog content, all measured transparently for clarity.

In short: This snapshot from bench_harness and ops telemetry shows how lean operations can still deliver robust capabilities across multiple AI services, with the numbers revealing the concrete scale of what runs behind the scenes.

A transparent report like this bridges the gap between abstract ideas of "running AI" and the actual coordinated systems required to make them available reliably to users.

One team runs all of this?#

Yes, Hax is operated by one team (moche-ai) that simultaneously keeps several distinct technical services alive and useful. Rather than separate groups specializing in isolation, this setup requires cross-functional skills to maintain the GPU pool for image generation, the memory backend for agents, the data ingestion for crypto quotes, the inference layer for the LLM gateway, and the web serving plus writing for the blog. This kind of unified operation is what turns a collection of tools into a cohesive platform. The team must monitor hardware health for the ComfyUI pool, manage storage growth for nearly ten thousand agent records with most staying active, curate reliable feeds for broad coin coverage, tune models and serving for low latency, and steadily add to the published pieces while keeping page loads snappy. It demonstrates practical DevOps for AI where everything runs under one roof.

What does each number mean?#

To make the scale concrete, here is the exact measured snapshot captured through bench_harness combined with ongoing ops telemetry:

Each row grounds the description of the stack in verifiable observations rather than estimates. The image generation hardware row, captured on 2026-07-04, indicates that the ComfyUI pool uses four GPU cards with 95.6GB VRAM per card. For a beginner, 95.6GB VRAM per card is a far bigger "workbench" than a high-end gaming PC — room to place a whole large model at once. This capacity supports loading and running sophisticated generation models without constant swapping or downscaling. The companion metric from 2026-06-30 compares two approaches: z-image-turbo finishes a generation in just 6 seconds whereas qwen-image requires 73 seconds for the same task, highlighting practical tradeoffs between speed and model choice.

Turning to agent memory measured on 2026-07-12, the system stores 9898 items in total while 9603 of them remain active. The high ratio of active to stored records suggests that the memory actively supports ongoing agent sessions rather than sitting as a passive archive. The live crypto quotes service on the same date tracks 718 distinct coins drawing information from 3 data sources, providing broad market visibility while limiting the operational surface to a manageable number of upstream providers.

The LLM gateway entry from 2026-07-04 shows a first-response latency of 120.8 milliseconds, the time until the initial reply begins streaming back, which feels nearly instantaneous and supports fluid interactions. Finally, the blog metrics from 2026-07-12 record 255 total published pieces alongside an HTTP P95 latency of 182 milliseconds, confirming that readers experience quick page loads even at the 95th percentile of requests.

Why does this scale matter?#

For anyone new to AI infrastructure, these numbers move the conversation from vague notions of "it uses AI" to a tangible picture of the resources and coordination involved. A beginner might imagine AI stacks as simple chatbot wrappers or small scripts on a laptop, yet the Hax measurements reveal multiple layers working in parallel on dedicated hardware and data systems. The GPU allocation, with its substantial per-card VRAM, underscores the compute intensity behind image generation. Sustaining a large set of agent memory entries with most remaining active demonstrates stateful, long-running AI agents rather than stateless one-off queries. Broad crypto tracking across 718 coins from a compact set of sources shows how data pipelines must be robust yet efficient. The measured gateway and blog latencies prove that user-facing performance stays a priority even as the backend grows more complex. Understanding this scale sets realistic expectations and shows what becomes possible when one team owns the full picture: quicker debugging across components, a consistent user experience, and the ability to evolve all parts together.

Note: measured dates differ per service (noted in table); bench_harness re-measures periodically.

Reference links

• ComfyUI
• vLLM
• Ollama