Should your business buy AI hardware? almost never

Most business owners who ask me whether they should buy their own AI machine do not need one. The question keeps coming up anyway, usually for two reasons: privacy, and the feeling that you want to be independent of OpenAI or Anthropic.

Both are understandable. Both are almost never a reason to put down thousands on hardware.

I sat down and worked through the numbers once, properly. Here is what came out.

There are three ways to run a model, not two

Most people think the choice is buy your own box or rent a GPU in the cloud. There is a third option that is almost always the best one, and it is the one most often skipped: rent the model itself.

Renting the model means you use an open model through a provider like OpenRouter or DeepInfra. You pay per use, no hardware. Renting a GPU means you buy raw compute by the hour and run the model yourself. Buying a box means hardware sitting in your own office.

Renting the model quietly solves every downside of owning hardware. No ceiling on capacity, nothing gathering dust when you are not using it, no electricity bill, and you can run several models at once. You give up exactly one thing: privacy. And that is the one real reason to buy instead.

The number that decides whether a model even runs

Before any of the hardware talk, one thing trips people up. What matters is not how much memory the machine has, it is how much of that memory the graphics chip can actually use. That is VRAM.

A gaming PC with 128GB of RAM and a normal graphics card cannot run a big model, because only the card's VRAM counts, and that is often just 16 or 32GB. This is why Apple machines and the new AMD mini PCs matter: their memory is unified, so the whole pool works as VRAM. Once the model fits, a second number, memory bandwidth, decides how fast it answers.

There is a real unlock here. A dense model, one big block of 70 billion parameters, crawls on any affordable box, around 2 to 5 words per second. But newer mixture-of-experts models only fire a few billion parameters per word, so the same cheap machine runs them at 40 to 86 words per second. That is the difference between unusable and genuinely good, and it is why the picture changed in 2026.

If you do buy, here are the three boxes

If you decide to own, three machines can hold a large model at a sane price. As of mid 2026:

• AMD Strix Halo, in a Framework Desktop or similar, 128GB, around $3,600 to $4,000. Best value, and best on the mixture-of-experts models you would actually run.
• Nvidia DGX Spark, $4,699 after a price hike. Held back by memory bandwidth, so it runs large models slowly. A well known developer who owns one says it "kind of just sits there" and is not worth using.
• Mac Studio, from $5,299 for the current high-memory model after Apple's price increases. The genuinely fast Mac for large models, the M5 Ultra, is expected in the autumn at $4,700 to well over $6,000.

The short version: if you must own, the AMD box is the value pick and the Mac is the performance pick. Skip the DGX Spark.

What renting actually costs

Now compare that to renting a raw GPU by the hour. For a few hours of use a day:

• An RTX 4090 runs about $31 to $37 a month.
• An A100 with 80GB about $74 to $125 a month.
• An H100 about $199 to $254 a month.

Run those same cards around the clock and the A100 lands near $600 to $1,000 a month and the H100 near $1,650 to $2,100. Owning starts to make sense only when you are genuinely using the hardware that hard.

The break-even is not close. A $10,000 machine spread over three years is about $278 a month in hardware alone, before power. Renting a capable GPU for a few hours a day is $31 to $125 a month. And owning is never free: one high end GPU running day and night burns roughly $5 a day in electricity, about $2,000 a year.

The discount that is not really there

One more trap, and it is the one that sells people on open models. They look dirt cheap. Often ten times cheaper per token than a top model like Opus.

But per token is not per job. Open models burn two to three times more tokens to finish the same task. Count it per completed job and that ten times cheaper shrinks to about two times. Still cheaper, but a very different story than the price per token suggests.

Runnable is not the same as good

The open models everyone is excited about, the genuinely good ones, are 200 gigabytes and up. They do not run on any machine that fits in your office. What you actually run at home is a stripped-down version that barely functions. Fine for a narrow task, weak for the rest. Even a ten thousand euro machine hits a wall there.

When buying is right

Buying is not always wrong. It is right in two clear cases.

Your data cannot leave the building. If customer data cannot go to a cloud for legal or contractual reasons, that alone is enough to consider owning hardware, whatever it costs.

Or you have one narrow task running around the clock. Think transcription or a simple classification that never stops. Then owned hardware pays for itself.

Outside those two: rent. Almost always.

The point

The question is rarely which machine to buy. The question is whether to buy anything at all. And the answer is usually no.

That is the kind of call I help business owners make. You run the numbers instead of chasing the hype, and most of the time the calm answer is also the cheapest.