The short answer
A complete compact private AI box runs $3,000 to $4,700 right now. A serious single-GPU workstation, $6,500 to $7,500. A dual-GPU machine, $8,000 to $14,000. Enterprise gear jumps to tens of thousands per accelerator and isn't what most companies need. But here's the part every hardware article skips: on a real deployment, the machine is often the cheapest line on the invoice. The integration, security, and training work around it is where most of the money goes, and where most of the value comes from.
What I actually paid for our machines
I'd rather show you my receipts than a pricing chart. Adapt runs its whole operation on private and cloud AI mixed, and the private side currently looks like this:
Our main GPU box is built around an RTX 5090. I paid about $7,000 for it a year ago. Pricing that same build today, I'd expect $10,000 to $12,000, because GPU prices went up, not down. That matches what the market shows: complete single-5090 workstations from name brands are reviewing at $6,500 to $7,300, but the card itself has been selling for well over its list price all year.
We also run a pair of NVIDIA DGX Sparks, about $10,000 for the two, and six Macs that we bought at ordinary retail. The Macs matter more than people expect. A Mac Studio with 96 or 128 gigabytes of unified memory sells for under $4,000 complete and runs surprisingly large models quietly on a desk. It's a legitimate private AI machine. Half our fleet is Macs for a reason.
That's the fleet that runs a company doing website builds, workflow automation, Google Ads management, and the sales pipeline I've described elsewhere. Nobody needs a data center to get real work out of AI.
The 2026 price bands, tier by tier
| Tier | What it is | Complete-system price |
|---|---|---|
| Compact box | DGX Spark / GB10 class, 128GB unified memory | $3,000 to $4,699 |
| Mac Studio | 96 to 128GB unified memory, quiet, desk-friendly | $3,700 to $4,000 |
| Single-GPU workstation | One RTX 5090, the workhorse tier | $6,500 to $7,500 |
| Dual-GPU workstation | Two 5090s, more users at once | $8,000 to $14,000 |
| Professional card | RTX PRO 6000, 96GB, the card alone | $12,000 to $13,350 |
| Enterprise | H100/H200 class, eight-GPU systems | Tens of thousands per card, $260,000 to $420,000 per system |
Two warnings about that table. First, watch the difference between a card price and a system price. A $13,000 RTX PRO 6000 is just the card. The $12,787 Dell workstation that contains one is a whole computer. Vendors mix these numbers constantly and it makes budgets wrong by half. Second, these prices move. NVIDIA raised the DGX Spark from $3,999 to $4,699 this year over memory shortages, and 5090 street prices have run 25 to 60 percent over list. Treat any number older than a quarter as history. Including these, eventually.
What about cheaper? People ask me if they can start with an older gaming card and a spare desktop. My honest answer: for a business, no. Small cards and old machines will run a demo, and then fall over the first time three employees hit it at once with real documents. I don't put anything below this table in front of a client.
And the enterprise row is there mostly so you know it exists. Adapt is deliberately small-time on compute: modest machines for real companies. If you genuinely need an eight-GPU H200 system, you also need in-house staff or a much bigger vendor, and the honest move is for me to tell you that instead of quoting it.
The hardware is not the expensive part
This surprises everyone. The majority of a private AI deployment's cost is off the hardware entirely: the harnesses and interfaces your team actually touches, the databases that hold your company's knowledge, the cybersecurity around all of it, and the networking that connects it to the systems you already run.
The public numbers back this up. Firms that publish integration pricing put a smaller retrieval-and-documents deployment at $10,000 to $35,000, production systems at $30,000 to $80,000, and department-scale platforms into six figures. In my experience, cleaning up a company's data can eat a third to half of the budget before the fun part even starts. Your $7,000 workstation can easily be the cheapest thing in the room.
This is exactly why our engagements start with consulting instead of a hardware quote. Most companies that come to us wanting AI aren't ready for it yet. Their data is scattered, their processes are undocumented, their security has holes. Selling that company a server first is taking their money to create risk. Fix the prerequisites, then the machine becomes useful the week it's plugged in.
The same logic explains what makes quotes go up once I see a company's actual systems. It's never the GPU. It's building the infrastructure around it, developing the processes, and training the people. You can hand a team the most state-of-the-art tools on earth, and if they don't know how to use them, they're meaningless. The flip side is just as true: modest AI plus a team that knows what to do with it produces a lot of value.
Ongoing costs: support, power, and the noise nobody warns you about
Published support pricing is all over the map because "support" means different things: $500 to $2,500 a month for basic hosting and maintenance, around $950 a month for small managed local deployments, up to $3,900 and beyond for enterprise packages. A useful rule from the consulting world: budget 15 to 25 percent of the build cost per year for maintenance.
What that money buys, in practice, is a person. Someone monitoring the system and its security, fixing things when they break, and helping you expand it. If you have an in-house AI person doing that all day, you don't need a retainer. Most companies under $50M don't have that person, and that's the actual product behind our managed service: we're that person, fractionally.
Now the part people really underestimate: energy and cooling. A giant compute rack with fifty GPUs draws power like a small factory and needs serious ventilation, but you don't have to go anywhere near that scale to feel it. Even a desktop with multiple GPUs is super loud and super hot. NVIDIA's own build guide recommends a 1,600-watt power supply for a two-5090 machine, integrators rate that configuration "very high" noise, and their rack versions require 200-to-240-volt circuits. A compact box or a single-GPU workstation can live in an office. A multi-GPU machine needs a plan for the room, the circuit, and your employees' ears.
Electricity itself is the small line. The room it forces you to think about is the big one.
The middle option: renting GPUs
Between cloud subscriptions and a box in your office sits a third option most owners haven't heard of: renting GPUs by the hour and loading your own models onto them. Current public rates run about $0.50 an hour for older cards up to $6 and change for the enterprise stuff, with a 5090 around a dollar an hour.
I'd describe it as semi-private. You control the model and the software, which is more privacy than a cloud subscription gives you. But your data still lives on somebody else's infrastructure, so it is not the on-premise security boundary either. It's useful for burst work and for testing whether a model earns a purchase, and even the rental providers themselves will tell you that regulated or truly sensitive workloads belong on your own hardware. Where that boundary sits for your data is the whole subject of the on-premise versus cloud comparison.
Can your budget be too big?
I get this question in reverse more often, but it's worth answering straight: right now, I don't think you can have an AI budget that's too big. The constraint is almost never money. It's readiness. A company with a huge budget and scattered data will burn the budget. A company with a modest budget and clean, consolidated data will beat them.
What I will say is that this industry is genuinely unsettled. I track it daily and test new models the day they ship, and I still don't keep a rigid tier sheet, because the right answer changes quarter to quarter. Hardware guides argue about Mac versus NVIDIA, about whether the DGX Spark is worth it, about minimum memory, and none of them agree, because the honest answer to all three is "it depends on your workload." Anyone who hands you a confident one-size price sheet for private AI in 2026 is selling you last quarter's answer.
How to read any private AI quote
Whether you buy from us or anyone else, make the vendor break the quote into separate lines: the complete hardware (card-only or full system, in writing), the setup and network integration, the data preparation and permissions work, security review, training for your team, and the monthly support with what it actually covers. Then ask one more question: what happens when the machine sits idle? A busy machine pays for itself. An idle one is furniture with a power bill. If a vendor can't tell you how they'll keep it busy, they're selling you the box, not the outcome.
Questions owners actually ask
What's the cheapest real private AI setup for a business? A DGX Spark class box or a high-memory Mac Studio, $3,000 to $4,700 complete. Both run capable open models for a small team. Below that tier I'd tell you to stay on cloud subscriptions instead. The guide to private AI for business covers who should and shouldn't buy at all.
How much does a private AI server cost per month after purchase? Electricity is modest, usually well under $100 for an office machine. The real recurring cost is a human keeping it monitored, patched, and secure: from a few hundred a month for a light retainer to a few thousand for full managed service. Skipping it doesn't save money. It converts your server into shelfware over about six months.
Do I need an H100? Almost certainly not. That's a $25,000-plus card built for data centers serving thousands of users. Companies in the $1M to $50M range do real work on machines from the first four rows of the table above.
Should I wait for prices to drop? They haven't. GPU prices rose through 2025 and 2026, and the machine I built for $7,000 would cost me $10,000 to $12,000 today. Buy for the workload sitting in front of you today.
Why do quotes for the same hardware vary so much? Because the hardware isn't what you're buying. The spread is in the integration, data preparation, security, and training bundled around it, which varies company to company for real reasons. Make every vendor itemize, and comparisons get easy.