Go.AI's Go1 MAX packs eight Blackwell GPUs into one box to serve 8,000 on-prem AI users

Go.AI has launched the Go1 MAX, the most powerful appliance in its Go1 line, the Chicago company announced on November 2, 2026. One box serves up to 8,000 concurrent users, runs AI entirely on premises, and sells at a fixed price with unlimited usage and tokens. For banks, hospitals and other regulated organisations, the pitch is AI they can own, audit and keep inside their own walls.

The Go1 MAX is aimed squarely at the largest teams. Regulated organisations want AI they can own and audit, Go.AI co-founder and CEO David Moscatelli said in the announcement, and the largest teams need real capacity to run it. The new appliance is designed to give a bank or a hospital with thousands of users cutting-edge AI and on-box fine-tuning in one appliance, without data leaving their environment.

One box, 8,000 users#

Every Go1 MAX sets up in 15 minutes, requires no dedicated server room, and runs Go.OS, Go.AI's purpose-built operating system for managing models, indexing and data. Teams can train and serve on the same appliance, without GPU queuing or cloud quotas, and the system is air-gap ready with no API exposure.

The fixed-price, unlimited-usage model is the other half of the pitch. Instead of metering tokens or seats, Go.AI sells the box and lets the organisation run it as hard as it likes. The announcement does not disclose a price, so the value of that deal cannot be judged until Go.AI names one.

What is inside#

Each Go1 MAX ships with eight NVIDIA RTX PRO 6000 Blackwell GPUs, each with 96GB of GDDR7 ECC memory, for 768GB of GPU memory. Go.AI's product page says that is matched one-to-one by 768GB of DDR5-6400 ECC system memory, with more than 30 PFLOPS at FP4 precision and about 14 TB/s of GPU bandwidth. Storage is more than 30TB of NVMe across four 7.68TB U.2 drives, for model weights and logs, with dual 25GbE networking and optional 100GbE or 200GbE.

The box is a 4U rack unit with 3+1 redundant Titanium power, drawing about 5–6 kW on standard 208V circuits, according to the product page. Against the original Go1, launched in March with three GPUs serving 2,000 users, the MAX is a straightforward scale-up: eight GPUs, 8,000 users, and headroom for on-box LoRA and adapter fine-tuning. Clustered deployments of up to four units scale to 32,000 concurrent users with automatic failover and load balancing.

Built to pass an audit#

Go.AI says the Go1 MAX is audit-ready, SOC 2 Type II and ISO 27001 certified from day one, with redundant power and out-of-band management. Response speeds run as fast as 50ms, against 200–500ms for cloud AI on Go.AI's own product-page comparison. The company also points to one deployment: a major financial institution with thousands of employees put compliant AI into core workflows while keeping its data protected, and now saves hundreds of hours each month across administrative and back-office functions. That is Go.AI's account of a single unnamed customer, not an independently verified result.

“Regulated organizations want AI they can own and audit, and the largest teams need real capacity to run it,” Moscatelli said. “The Go1 MAX gives a bank or a hospital with thousands of users the ability to use cutting-edge AI and on-box fine-tuning in one appliance, without data leaving their environment.”

Where it sits in the Go1 line#

The announcement follows Go.AI's name change in September — Go Abacus Corporation became Go.AI — and its $85 million Series A, extending momentum that began with the launch of the original Go1 in March. The Go1 MAX is available to order now through Go.AI's Go1 MAX product page.

What remains unproven is everything the announcement does not say: the price, how the box performs outside Go.AI's own figures, and whether deployments beyond the one cited bank example deliver the same time savings. On the evidence published today, the Go1 MAX is a credible capacity play for organisations that cannot send data to the cloud. Whether it is a good deal depends on a number Go.AI has not yet published.