Owning AI hardware now makes commercial sense for ordinary UK businesses — but not always. A single inference workstation costs £4,855 ex VAT and runs today’s strongest 27B to 30B models. Whether that beats an API comes down to three things: utilisation, number of users, and control over data. This guide shows the real machine, the real prices and the real sums.

For a UK SME, owning AI hardware now makes commercial sense in some very ordinary situations.

If 10 or more people use AI every day, or your team works with client information that cannot reasonably be sent to a US hosted API, a single local inference machine deserves serious consideration.

Not because local AI is always cheaper.

It is not.

The case comes down to three things, utilisation, number of users, and control over data.

Here is the machine we built and the arithmetic behind it.

Hardware: £4,855 ex VAT.

That buys:

• One RTX 5090 with 32 GB of VRAM
• 64 GB of system RAM
• A 16 core Ryzen processor
• 4 TB of NVMe storage

Those prices came from SCAN UK in February 2026.

It comfortably runs the strongest current models in the 27B to 30B range. Qwen3.8 27B, released on 13 August 2026 under Apache 2.0, needs roughly 14 to 16 GB at 4 bit. NVIDIA Nemotron 3.5 Lightning 30B A3B also runs on one card. Both permit commercial use.

Power costs work out at roughly £400 a year, assuming a UK small business electricity rate of 26.8p/kWh.

Over three years:

Spread across 10 people, that is roughly £17 per user per month.

A comparable rented GPU is more than £400 a month. We have paid those bills ourselves. At that rate, buying the workstation catches up with renting in roughly 12 months.

The comparison with APIs is less dramatic.

A heavy Sonnet class user works out at around £262 a year using the assumptions later in this article. Ten users therefore cost about £2,620 a year.

The owned machine costs £2,018.

That is a saving. It is not a landslide.

Move to 20 users, high token volumes or Opus level pricing, and the calculation changes quickly.

So this is the point to remember:

Control is different. You get that from day one.

Your model runs on your infrastructure. Your documents stay on your network. Your costs stop moving every time usage increases.

That is why self hosting has moved from enthusiast territory into a credible SME infrastructure choice.

There are really three ways for a business to buy AI inference.

1. Pay an API

You send requests to OpenAI, Anthropic or Google.

The attraction is obvious.

No capital expenditure. No server administration. Access to the strongest reasoning models available.

You also pay for every token, your information leaves your own estate, and costs increase as adoption grows.

Published rates in August 2026 were:

For a small team with modest usage, this is often the cheapest answer.

Simple.

2. Rent the GPU

The middle option is dedicated cloud compute from providers such as RunPod, OVHcloud or Vast.ai.

You avoid buying the hardware and retain much more control over deployment.

It is also easy to stop.

That makes renting ideal for testing.

The problem starts once the workload becomes permanent.

OVHcloud lists an L4 with 24 GB at roughly $1.00 per hour on demand. Committed European pricing for a GPU in the 32 GB range comes in at more than £400 per month.

That is £4,800 every year.

Over three years, you have spent £14,400 and still do not own the machine.

3. Own the hardware

You pay first.

After that, every extra request costs almost nothing beyond electricity and administration.

You also control the entire data path.

But ownership only works financially when the GPU has work to do.

A lightly used £5,000 machine is poor economics.

A heavily used one is extremely cheap compute.

For many SMEs, the sensible architecture is therefore neither pure cloud nor pure local.

It is hybrid. We set that trade-off out in more detail in our comparison of self-hosted AI and cloud AI.

Use local inference for regular work, internal knowledge and sensitive information. Send the minority of requests that genuinely require frontier reasoning to an external model.

A business insisting that there is one correct answer for every workload is usually trying to sell you that answer.

The hardware existed before.

What changed was the software.

Two years ago, running AI locally usually meant accepting a noticeable drop in capability.

That gap has shrunk.

For a large share of normal business work, model quality is no longer the main reason not to self host.

Three changes got us here.

Dense models became much better

Qwen3.8 27B arrived on 13 August 2026.

It has:

• 28 billion dense parameters
• Hybrid attention
• 262k native context
• Context extensible to 1 million tokens
• A vision encoder
• Apache 2.0 licensing
• A SWE bench Pro score of 61.7

Its predecessor, Qwen3.7 27B, is already being used in production by us and many others.

Qwen3.8 is not a completely different generation.

It is simply better.

And that matters.

Sparse models changed the throughput equation

NVIDIA released Nemotron 3.5 Lightning 30B A3B on 11 August 2026 under the OpenMDW 1.1 licence.

It combines Mamba 2, attention and mixture of experts.

There are 30 billion parameters in total.

Only 3 billion activate for each token.

That is the important part.

Its published scores include:

It runs on a single H100 or DGX Spark, and NVIDIA reports output speeds up to four times those of comparable models.

Its Mamba layers also stop memory consumption growing in the same way as a conventional KV cache as context expands.

That makes very long context, including million token use cases, much more practical on one card.

Commercial licensing got easier

This used to require much more scrutiny.

Today, several major open model families come with commercially usable licences:

There are no revenue thresholds or per seat reporting requirements attached to those licences.

You can build commercial systems around them.

The rough arithmetic is straightforward.

At 4 bit, model weights require around 0.5 bytes per parameter.

At FP16, think closer to 2 bytes per parameter.

Then leave roughly another 10 to 20 per cent for framework overhead and KV cache. Long context windows and concurrent users push that requirement higher.

Some useful reference points:

That makes the practical dividing line quite clear.

For an SME buying one workstation, 27B to 30B is the useful range.

Those models fit comfortably on a consumer GPU with 32 GB at 4 bit, while leaving enough memory for a meaningful context window.

Once you move beyond roughly 200B parameters, you are no longer planning an office workstation.

You are planning data centre infrastructure.

This is not a suggested parts list.

It is the actual configuration we priced at SCAN UK in February 2026.

GPU

This is where almost all the value sits.

LLM inference is heavily limited by memory bandwidth.

The RTX 5090 provides 1.79 TB/s of GDDR7 bandwidth.

On the same 30B class model, published vLLM benchmarks show aggregate performance of roughly:

That is close to double.

The 32 GB capacity also matters. A 27B model at 4 bit fits without forcing you into an impractically small context window.

System memory

There is little reason to jump automatically to 128 GB.

The model weights sit in VRAM.

System memory supports loading, document processing and the retrieval system around the model.

For this configuration, 64 GB is enough.

And DDR5 is not cheap.

Storage

Four terabytes sounds generous until you start storing multiple models, several quantisations and a vector database.

It fills.

Power supply

A 1200 W supply is not excessive here.

The 5090 alone carries a 575 W board power rating and can create transient loads well beyond that.

Poor power delivery is one of the easiest ways to turn a reliable inference server into a machine that crashes for apparently random reasons.

Cooling

This machine does not behave like a gaming PC.

It may sit under heavy load for hours.

So cool it accordingly.

A full tower, oversized air cooling and industrial fans help prevent thermal throttling deep into a batch job.

They also make the machine tolerable in an office.

One warning on GPU pricing

Waiting does not necessarily save money.

The current pressure is coming from GDDR7 and HBM memory supply.

Prices are moving up.

NVIDIA has increased RTX 5090 pricing as memory costs have risen.

The RTX PRO 6000 Blackwell is an even clearer example.

It launched below $8,000.

NVIDIA’s own marketplace now lists it at around $13,250.

That is roughly a 55 per cent increase.

Current expectations do not point to supply normalising before 2027 to 2028.

So if the plan is to wait specifically because GPUs should become cheaper soon, the evidence does not currently support that assumption.

What if one GPU is not enough?

There are three obvious routes.

People often quote one inference speed.

That hides what matters.

What one user feels

For a dense 27B model at 4 bit on a 5090, the weights occupy about 15 GB.

With 1.79 TB/s of memory bandwidth, the theoretical ceiling lands near 110 tokens per second.

Real performance is lower.

A sensible expectation is around 70 to 90 tokens per second.

That is roughly three times normal reading speed.

For interactive use, it feels immediate.

What the whole machine produces

Now look at batching.

Published vLLM results show:

Sparse mixture of experts models batch much better than dense ones.

For a dense 27B model, plan for aggregate throughput in the high hundreds rather than several thousand tokens per second.

That aggregate figure determines how many people the machine can actually support.

How many users?

VMware published a capacity study in April 2026 using gpt oss 120b.

Its results included roughly:

The workload mixed short, medium and long requests rather than assuming that every user behaved identically.

Scale that down to a well configured 5090 running a 27B to 30B model and 10 to 20 simultaneously active users is a reasonable planning number for chat and retrieval work.

The total number of seats can be much higher.

People are not all generating tokens at the same instant.

Two findings from the VMware work are particularly useful.

First, VRAM capacity was not the main constraint.

The limiting factors were attention compute and memory bandwidth.

KV cache utilisation reached only 10.5 per cent, largely because prefix caching removed duplicated context.

More VRAM is therefore not automatically the best investment.

Second, configuration changed the result substantially.

Capping --max-num-batched-tokens around 1,024 to 2,048 helped stop large prefill requests from damaging inter token latency.

A badly tuned server can waste most of the performance you paid for.

Three years of ownership

The power calculation assumes roughly 1,385 kWh each year.

That comes from approximately three hours a day under sustained load at around 650 W, plus continuous idle consumption of 110 W.

Electricity is priced at 26.8p/kWh, the current UK small business average used in this model.

Administration is not included.

Add your own number.

A reasonable planning allowance is half a day each month from somebody competent, or the cost of a support arrangement.

Cost per person

The hardware cost does not grow with the seat count until you actually exhaust its capacity.

That is where ownership starts to become financially interesting.

Renting versus buying

Take £400 a month as the cost of a 32 GB class rented instance.

That gives:

Against continuous rental, purchase cost is recovered in roughly 12 months.

Renting still has a place.

Use it when:

• You are proving that the use case exists
• Your demand moves sharply up and down
• Capital expenditure is difficult to approve

Once the workload becomes stable, continuing to rent starts looking expensive.

The API comparison

This is where simplistic self hosting arguments often fall apart.

Take one heavy user.

Assume:

• 80 interactions per working day
• 5,000 input tokens for each interaction, including context and retrieval
• 700 output tokens each time

That works out at approximately 104 million input tokens and 14.6 million output tokens per year.

At published August 2026 pricing and an exchange rate of $1.3534 to £1 on 15 August 2026, annual costs are approximately:

Now compare that with owning the server.

Ten heavy Sonnet users:

The server wins.

Barely.

For a smaller team or lighter use, the API can easily be cheaper.

That is not a failure of self hosting.

It is simply arithmetic.

Scale changes the result

The hardware cost still does not move.

Then consider a workload that really uses the GPU.

Suppose the machine spends three hours each day producing 800 tokens per second in aggregate.

Annual output is roughly 2.2 billion tokens.

At £2,018 a year, effective compute cost becomes approximately £0.90 per million tokens.

Compare that with roughly £7.39 per million output tokens for Sonnet class API pricing.

Around eight times cheaper.

But the condition matters.

You only get that economics if the machine stays busy.

That is the argument in one sentence.

Money is not the only column

Three other differences matter.

You are not dependent on one vendor continuing a model, maintaining a price or accepting your use case.

We sell this work.

There are still several situations where we would tell a company not to self host.

For some UK SMEs, cost is secondary.

The real question is what they can confidently say to a customer about their information.

UK GDPR and the Data (Use and Access) Act 2025

Self hosting does not make data protection obligations disappear.

You still need:

• A lawful basis for processing
• A DPIA where processing is high risk
• Privacy notices that disclose your use of AI

What changes is the evidence.

If the processing happens locally, the explanation of where the information went becomes much simpler.

It stayed inside your environment.

NCSC AI cyber security guidance

The NCSC guidance issued in April 2026 provides a practical security baseline that many buyers now expect.

That includes:

• Current Cyber Essentials certification
• MFA for AI tool accounts
• Protection against prompt injection
• Out of band verification for financial actions

These controls matter whether you host locally or use an external provider.

Data processing agreements

Use a third party API and procurement teams are likely to ask for a DPA.

They may also want a clear no training clause.

Keep inference inside your own infrastructure and you remove that processor relationship altogether.

The EU AI Act

UK companies can still fall within the Act where they have EU customers, employees or products.

The Article 4 AI literacy requirement has applied since February 2025.

Documented staff training is not optional.

High risk applications, including employment, credit and education, bring additional requirements around conformity assessment and data governance.

Sector rules do not disappear

Other requirements may sit on top.

Examples include:

• FCA Consumer Duty requirements around explainability
• ICO guidance covering healthcare and special category data
• Professional confidentiality rules from the SRA
• ICAEW requirements
• Bar Standards Board requirements

For businesses handling controlled technical information, MoD supply chain material, patient records or privileged client information, one statement is particularly powerful:

It answers a difficult procurement question quickly.

Buying the machine is the easy part.

Configuration decides whether it works well.

Serving

For experimentation or one person, use Ollama.

Installation takes about ten minutes and its model library is excellent.

Once several people need the machine, move to vLLM.

Continuous batching and paged attention change the number of users identical hardware can serve.

A server that comfortably handles two people through a simple setup can support twenty when the serving layer is designed properly.

Do not keep an entire production team on Ollama simply because that was the easiest place to start.

Quantisation

For most business deployments, 4 bit should be the default.

AWQ, GPTQ and NVFP4 can reduce memory use to roughly one quarter of full precision with a quality difference that many ordinary business tasks cannot detect.

If you have enough VRAM to run FP8, test it.

Do not assume that higher precision is automatically worth the extra memory.

Evaluate the outputs first.

Retrieval

The general model is not where most of your business value sits.

Your internal knowledge is.

That includes:

• Company documents
• Price structures
• Procedures
• Product information
• Technical conventions
• Customer history

Keep that corpus current and put a retrieval system in front of the model.

That is what turns a model anyone can download into something specific to your company. It is exactly what Ascentis Cortex was built to do.

Logging and controls

Record every prompt and completion.

Tie activity to a user.

Retain those records under your normal retention rules.

You will need good records when producing a DPIA.

You will also want them the first time somebody challenges an AI generated answer.

Routing

Design external routing from the beginning.

Even if 95 per cent of requests stay local, make it possible to send the remaining five per cent to a frontier API.

That gives you a much more robust system.

The local model does not need to win every benchmark.

It needs to handle the work it is good at, cheaply and reliably.

Days 1 to 5. Measure what people are doing now. Count actual AI requests. Do not count enthusiasm. If nobody can tell you how much the company currently spends on APIs each month, buying hardware is premature.

Days 6 to 10. Rent compute first. Run Qwen3.8 27B and Nemotron 3.5 Lightning against your actual documents and workflows. Use a rented card for a fortnight. Define what counts as “good enough” before reviewing the results. That avoids moving the target after you see the model.

Days 11 to 15. Make the decision using three questions:

1. How much utilisation do you expect?
2. How many people will use it?
3. How sensitive is the information?

If two of those three strongly favour local deployment, buying starts to make sense. If they do not, stay with the API. Review the decision again when headcount or usage changes.

Days 16 to 25. Build. Install vLLM. Configure batching. Connect the retrieval layer. Add monitoring. Expect software tuning to make more difference than assembling the components.

Days 26 to 30. Pilot one measurable workflow with one real group of users. Compare it against the baseline captured in week one. Do not roll it out company wide simply because the server is running. Start with work you can measure.

How much does a self hosted AI server cost in the UK?

A capable single GPU inference workstation costs around £4,855 excluding VAT, or £5,826 including VAT, using the RTX 5090 machine priced at SCAN UK in February 2026.

Electricity adds roughly £400 per year at 26.8p/kWh.

That produces a three year cash cost of around £6,055.

Is owning AI infrastructure cheaper than OpenAI or Anthropic?

Sometimes.

Not automatically.

Below roughly 10 regular heavy users, a frontier API is often cheaper and will certainly be simpler.

At 20 or more seats, large token volumes or top tier model pricing, ownership becomes much more compelling.

A busy machine can reach an effective cost near £0.90 per million tokens, compared with approximately £7.39 per million output tokens under Sonnet class pricing.

Which model should a UK SME run locally in 2026?

The most useful range is roughly 27B to 30B.

Two strong options are Qwen3.8 27B, Apache 2.0, released August 2026, particularly for general business work and coding; and NVIDIA Nemotron 3.5 Lightning 30B A3B, OpenMDW 1.1, particularly where throughput and long context are important.

Both can run on a single 32 GB card at 4 bit.

How many users can one GPU support?

A properly tuned RTX 5090 running a model in the 27B to 30B range at 4 bit can support roughly 10 to 20 simultaneously active users for chat and retrieval work.

It can support considerably more registered users because they will not all generate at once.

Configuration matters enormously.

Batched token limits and latency tuning can determine whether exactly the same hardware feels excellent or unusable.

Is renting a cloud GPU sensible?

Yes, for the right stage.

Rent while evaluating the idea.

Rent when demand is irregular.

Rent when capital expenditure is difficult.

For an established workload, £400 or more every month becomes hard to defend when an owned workstation reaches cost parity in around 12 months.

Does self hosting make a company UK GDPR compliant?

No.

You still need the correct lawful basis, a DPIA where the processing presents high risk, and privacy notices that explain the use of AI.

Local deployment does make the processing chain easier to demonstrate and removes the external AI provider from the data processor relationship.

Should I wait for GPU prices to fall?

There is currently little evidence that waiting will produce lower prices.

GDDR7 and HBM shortages are pushing costs upward, and current expectations suggest supply may not return to normal until 2027 to 2028.

The RTX PRO 6000 Blackwell provides a useful example, moving from a launch price below $8,000 to around $13,250, an increase of roughly 55 per cent.

---

Self hosting is not the cheap answer by default.

For a lightly used system, it can be the expensive answer.

For a busy one, the economics become very strong.

The decision becomes particularly compelling when three conditions start appearing together:

• 10 or more regular users
• Significant and predictable AI volume
• Data that you would rather not send to an external API

At 20 users, the financial argument becomes much easier.

For controlled technical information, customer confidential material or regulated data, the control argument can outweigh cost entirely.

And there is no need to make this an ideological choice between open models and frontier APIs.

Use both.

Run the repeatable, high volume and sensitive work locally.

Send the genuinely difficult reasoning elsewhere.

That is the architecture we would choose for most established SMEs today.

Hardware pricing came from a SCAN UK basket dated 5 February 2026.

GPU performance figures use CloudRift vLLM benchmarks published during 2025 and 2026.

Concurrency figures come from the VMware Cloud Foundation capacity study published 30 April 2026.

Model specifications use Qwen3.8 27B, 13 August 2026; NVIDIA Nemotron 3.5 Lightning 30B A3B, 11 August 2026; and the Onyx self hosted LLM leaderboard, 20 July 2026.

API pricing reflects published rates on 14 August 2026.

UK business electricity pricing uses AquaSwitch data from August 2026.

Currency conversion uses $1.3534/£ on 15 August 2026.

The regulatory position is current as of August 2026 and is not a substitute for legal advice.

Single stream dense model throughput figures are derived from memory bandwidth arithmetic. Validate them against your own workload before making a capacity decision.

All cost calculations are illustrative.

Your result will depend on how much you actually use the machine.