What does an AI rollout look like at a bank that cannot send data to the cloud?
For more than 200 institutions, the answer is a box in the server room. Go.AI sells an on-premises appliance that runs large language models inside the customer's own walls. No API call. No data egress. No per-token bill.
On September 22, the Chicago company said it raised $85 million in a Series A led by Updata Partners. Total funding now sits at $90 million. A round that size is rare. A Series A for hardware that customers own outright is rarer.
Regulated buyers pay for control, not benchmark scores. Go.AI's customers run 12.5 million queries a day on hardware they own.
The $85 million sits in the top 2% of US enterprise software Series A deals on record, according to Dealroom.
The risk sits in the refresh cycle: eight H100s age faster than a compliance checklist.
Round led by Updata Partners
Total funding reaches $90 million, including earlier seed capital. · BusinessWire, 2026
Regulated deployments
Banks, insurers, hospitals and manufacturers, per the company. · Company-reported, 2026
Annual recurring revenue
Growth measured against a small base; no audited figure released. · Company-reported, 2026
The outlier in a thin Series A market
Dealroom logs the round in the top 2% of US enterprise software Series A deals on record. Updata Partners, the lead, manages more than $3 billion in committed capital and writes checks between $20 million and $200 million.
The company behind the round is older than its name. It was founded in 2018 as Abacus Analytics LLP, became Go Abacus Corporation in 2022, and took the Go.AI name on September 1. The Go1 appliance shipped in March. Headcount is above 50.
That record matters to a buyer. Banks and insurers rarely hand production data to a startup with a two-year paper trail. A 2018 incorporation date and a 200-customer base belong in the sales deck, not in a footnote.
One box, eight H100s, and a 405B ceiling
| Parameter | Go1 appliance |
|---|---|
| GPU | 8× Nvidia H100 |
| Memory | 640GB HBM3 |
| Storage | 30TB NVMe |
| Latency | <80ms P95 time-to-first-token |
| Max model | 405B parameters, single node |
| Concurrent users | Up to 2,000 |
| Deployment | Under 15 minutes, on-premises |
| Pricing | Flat fee, unlimited queries |
Go.AI product specifications, company documentation, 2026
The Go1 is a single node. Eight H100 GPUs, 640GB of HBM3 memory, 30TB of NVMe storage. Go.AI claims sub-80ms P95 time-to-first-token and support for open-weight models up to 405 billion parameters.
Two choices define the design. The 405B ceiling fits the largest open-weight models in current circulation. The single-node layout removes the cluster fabric that usually moves data between racks. Liquid cooling handles the heat.
Go.AI says the box reaches production in under 15 minutes and serves 2,000 concurrent users. Those are vendor figures. No independent benchmark has tested them.
AI doesn't have to be something people fear or don't understand. Done right, it works alongside the people who use it.— David Moscatelli, co-founder and CEO, Go.AI
Compliance is the product
Compliance officers buy the audit trail. Model quality is table stakes. Go.AI says the platform is built for FINRA, HIPAA, NCUA and GDPR, with SOC 2 Type II, ISO 27001 and HIPAA certifications from day one.
The commercial model follows the compliance model. Customers pay a flat fee for unlimited queries, which removes the per-token invoice that finance teams cannot forecast. Data never leaves the building, so the vendor never becomes a data processor.
This is the same argument Oxide Computer made for owning your own cloud, backed by a $445 million Series D in October. As we wrote in October, the appeal is control and predictable cost. Oxide sells to platform teams. Go.AI sells to risk committees.
Go.AI built the AI stack that regulated institutions can actually deploy: software and hardware that live inside the customer's secure environment with no proprietary data going to a third party.— Carter Griffin, General Partner, Updata Partners
The hardware bill arrives later
Every growth number in the pitch comes from the company. Go.AI reports more than 8x year-over-year ARR growth and profitability. Hoodline noted in September that those figures have not been independently verified. Neither has the 12.5 million daily query count.
That is normal for a private company. It is also the first item a diligence team should price. ARR on a hardware-heavy product can include maintenance and support contracts, which behave differently from software subscriptions when a customer slows down.
The larger risk is timing. The Go1 is built on eight H100s. Nvidia has tended to ship a new data-center architecture every two years or so, and a customer who owns the box also owns its depreciation. A bank that installs in 2026 will face a refresh decision before two compliance review cycles close.
The 405B ceiling is the second constraint. Agentic workloads that chain long-context calls can exhaust single-node memory, and the appliance has no fabric to scale out. Go.AI's answer is that regulated tasks — AML checks, policy questions, document review — do not need frontier scale. That holds today. It may not hold after another two model generations.
Competition is not static. Hyperscalers sell private and sovereign regions, and systems integrators can now package open-weight models into a compliant rack. The durable assets are the audit software and the installed base. The GPUs are replaceable.
The $85 million buys time to widen that moat. Updata's Carter Griffin joins the board. Go.AI plans to hire engineers and push beyond regulated industries into what it calls compliance-minded organizations. The phrase admits the beachhead is finite.
The math the buyer runs
Across 200 customers, 12.5 million daily queries average roughly 62,500 per customer. That is a steady workload, not a spiky one, and steady utilization is where owned hardware starts to pencil out.
The procurement comparison has three lines. One is capex and support for the appliance. One is the cloud bill, which moves with tokens, model size and egress. The third is the compliance review, which has a cost even when it does not appear on an invoice.
Go.AI's flat fee wins when volume is predictable and the review burden is high. A bank running document review every business day gets a fixed bill and a known depreciation schedule. A bank with seasonal spikes pays for idle GPUs between peaks. The appliance model does not remove the trade-off. It moves the trade-off from the invoice to the balance sheet.
What breaks the appliance model
Probability: 60% — compliance review time, not model quality, decides where regulated inference runs.
🟢 Optimistic scenario (30%)
Consequence: the installed base doubles and the flat-fee model becomes the reference procurement pattern for regulated AI.
🟡 Base scenario (50%)
Consequence: a durable, profitable niche that never becomes a standard — and a valuation capped by hardware multiples.
🔴 Pessimistic scenario (20%)
Consequence: the appliance becomes a feature of a services business, and the next round prices as hardware.
The bet is narrow. A slice of enterprise AI will never leave the building, and someone will sell the box that keeps it there. Go.AI has a two-year head start and a profitable base. The test is whether it can refresh the hardware faster than a bank can approve a new vendor.