What $20,000 buys you in 2026: a full local AI setup for a 30-person business

A CFO at a Denver contracting firm asked us a blunt question last month: "If I spend $20,000 on AI hardware instead of paying for subscriptions every month, what do I actually get, and when does it pay for itself?" That deserves a straight answer with real, sourced numbers, not a vendor slide deck. Here it is, researched and priced as of July 2026.

The short version: you get a server that runs a genuinely capable model privately, on your own network, with no per-token metering, no vendor guardrail silently refusing legitimate work, and no mandatory data-retention policy to negotiate around. You own the hardware and the model weights. It runs your 30-person team's everyday AI load 24 hours a day. What you do not get, no matter how the budget is sliced, is Kimi K2.7 or DeepSeek V4 Pro running natively on that box. We will explain exactly why, and what to build instead.

The honest answer on Kimi K2.7 and DeepSeek V4 Pro

Both models get asked about constantly, so let's settle it with numbers. Kimi K2.7 is a 1-trillion-parameter Mixture-of-Experts model from Moonshot AI, 32 billion of those parameters active per token. Even at the most aggressive usable quantization, roughly 2-bit, it needs about 325 to 350 GB of combined RAM and VRAM just to load, and a quality-preserving Q4_K_M quant pushes that to somewhere near 640 GB once you account for the full context window. DeepSeek V4 Pro is bigger still: 1.6 trillion total parameters, and a Q4_K_M quant needs roughly 517 GB of VRAM, climbing past a terabyte at higher precision. Neither fits on a single 8×80 GB node, let alone a workstation. Running either one properly means 8 H100 or H200-class GPUs, a rack, and a six-figure hardware bill before you've bought a single server chassis.

This is not a failure of the $20,000 budget, it's a correct read of what these models are for. Frontier-scale open-weight models exist to do the work nothing smaller can: end-to-end agentic software projects, research synthesis across huge context windows, reasoning chains that need to hold a million tokens at once. A 30-person SMB needs that rarely. What it needs constantly is fast, private AI for drafting, summarization, classification, document review, chatbot responses, and code assist, and for that tier of work a $20,000 box running a well-chosen mid-weight model beats a cloud API on cost, latency, and data control simultaneously.

So the architecture we build is hybrid by design: on-prem inference for the routine 90%, a cloud model as burst capacity for the rare 10% that genuinely needs frontier scale. You never pay per-token for the bulk of your AI load. You get frontier capability on the occasions that actually call for it. And you choose, deliberately, what data ever leaves the building.

The $20,000 line-item budget

All prices are in USD, researched in July 2026. GPU and memory prices in particular have moved sharply this year, an AI-driven DDR5 shortage has pushed server memory prices up an estimated 50 to 110% since early 2025, and Nvidia's professional-tier cards have climbed with it. Line items marked [VERIFY] should be confirmed against current vendor pricing before you order. We present two GPU options; everything else is shared.

Component	Item	Option A — Dual RTX 5090 (USD)	Option B — RTX PRO 6000 (USD)
Inference GPU(s)	NVIDIA RTX 5090 ×2 (32 GB each, 64 GB total) or RTX PRO 6000 Blackwell ×1 (96 GB GDDR7 ECC)	$7,400 – $8,600	$11,150 – $14,500 [VERIFY]
CPU	AMD Threadripper PRO 9955WX (16C/32T, Zen 5, 8-channel DDR5)	$1,700 – $2,300 [VERIFY]	$1,700 – $2,300 [VERIFY]
Motherboard	WRX90/TRX50-class workstation board, dual or single x16 slot	$650 – $850	$500 – $700
RAM	256 GB DDR5 ECC RDIMM (8×32 GB)	$2,200 – $3,300 [VERIFY, rising]	$2,200 – $3,300 [VERIFY, rising]
Storage	2× 4 TB PCIe 5.0 NVMe SSD (model weights + execution data)	$550 – $800	$550 – $800
PSU	1600 W Titanium (Option A) / 1200 W Titanium (Option B)	$400 – $460	$300 – $400
Case + cooling	Full-tower case with high-airflow front intake + 360 mm AIO	$370 – $470	$370 – $470
n8n server	8-core mini workstation, 32 GB RAM, 1 TB NVMe	$700 – $950	$700 – $950
UPS	Line-interactive UPS, 2200 VA	$850 – $1,050	$850 – $1,050
Networking	10 GbE managed switch + Cat6A patch + cabling	$400 – $650	$400 – $650
Install & commissioning	OS hardening, Docker Compose, Ollama, n8n, Postgres, Qdrant setup	$500 – $750	$500 – $750
Total		$15,720 – $20,180	$19,220 – $25,870

Option A, dual RTX 5090, is the reliable way to land under $20,000 in mid-2026. It also lets you run two models simultaneously, one per card, which is useful for routing different task types without queue contention. Option B, a single RTX PRO 6000 Blackwell, gives you 96 GB on one card and a simpler build, but Nvidia has raised the card's price roughly 55% over the past year to a $13,250 MSRP, and street prices from Newegg, PNY, and B&H Photo currently run $11,150 to $14,500. A build that comfortably fit inside $20,000 a year ago now stretches to $19,220–$25,870. If your workload genuinely needs the extra VRAM headroom for long-context tasks or a larger single model like Mistral Large 123B, Option B is defensible. If you just want the most capability per dollar under a hard $20,000 ceiling, Option A wins.

What this stack can actually run in parallel

The realistic model mix

With 64 GB of combined GPU memory (Option A) or 96 GB on one card (Option B), here is what loads cleanly in Ollama or vLLM and performs well for business tasks:

Model	VRAM (Q4)	Tokens/sec (est.)	Best for
Qwen3-72B Q4_K_M	~40–45 GB	20–35 tok/s	General drafting, long documents, multilingual work
Llama 3.1 70B Q4_K_M	~40 GB	22–38 tok/s	RAG, instruction-following, English-first support
Mistral Large 123B Q4_K_M	~72 GB	14–22 tok/s	Complex reasoning, contracts (needs the 96 GB card)
DeepSeek-Coder-V2-Lite Q4	~10 GB	60–90 tok/s	Code review, debugging, SQL generation
nomic-embed-text (local)	<2 GB	—	Embeddings for RAG and semantic search

For a 30-person SMB, the practical setup is Qwen3-72B as the primary workhorse on one card, handling the bulk of drafting, summarization, and support tasks, alongside DeepSeek-Coder-V2-Lite on the second card for developers needing fast code assist. That split uses roughly 50–55 GB of the 64 GB Option A pool and leaves headroom for KV cache on longer documents. A marketing manager drafting outreach and a developer reviewing a pull request hit two separate models with no contention between them, that is the "run in parallel" promise fulfilled on a workstation, not a data center.

When you actually need frontier scale

When does a workflow need Kimi K2.7 or DeepSeek V4 Pro? When the task genuinely needs frontier-scale reasoning: a multi-file refactor across a large codebase, a fully autonomous research agent, or a synthesis job that needs the strongest reasoning available regardless of cost. Route that task to a cloud model API as a burst endpoint inside the same n8n workflow, the change is a single URL swap in the credential node. The rest of the pipeline, and your data, does not move.

The 2026 self-host rig, part by part

CPU, power, and the PCIe bottleneck

The Threadripper PRO 9955WX is the current generation of AMD's workstation line, Zen 5 cores with 8-channel DDR5 and full PCIe 5.0 across enough lanes to run two GPUs at full x16 bandwidth simultaneously. That matters for tensor-parallel inference across two cards, a PCIe-starved bus quietly costs you a meaningful chunk of the GPUs' theoretical throughput. At full load a dual RTX 5090 system draws close to 1,150 W from the GPUs alone (575 W each), plus roughly 350 W from the CPU and 150 W for the rest of the board, pushing total draw toward 1,650 W. A sensible move is capping each GPU around 450 W via nvidia-smi -pl, which costs a few percent of throughput and keeps a 1,600 W Titanium PSU running with real headroom instead of at its ceiling.

Memory, storage, and why RAM is the volatile line

256 GB of DDR5 ECC RAM is not excessive for this build. Ollama and vLLM spill to system RAM when VRAM fills, and having fast RAM in reserve keeps a model running instead of crashing when context gets long. It also gives Postgres, Redis, and Qdrant room to run comfortably without fighting for memory headroom. ECC matters for a machine running continuously, it catches single-bit memory errors that would otherwise silently corrupt model output or database state. The catch, and it's worth saying plainly, is that this is the single most volatile line in the whole budget right now: 32 GB DDR5 modules have jumped from roughly $149 to $239 in recent months as data centers absorb the majority of global memory supply, and analysts expect prices to keep climbing through the second half of 2026. Price this line the week you order, not the week you plan.

The 2× 4 TB NVMe drives hold model weights, a 70B model at Q4 runs about 40 GB on disk, and you will accumulate a small library of models over time, plus execution data. Mirror them if you can; you can always re-download model weights, but Postgres data and execution logs are not recoverable without a backup plan. A high-airflow tower case matters more than it sounds: two GPUs running near their power limit for hours need real intake and exhaust, not a compact case that traps heat. And get the UPS. A 2200 VA line-interactive unit buys 15 to 20 minutes of runtime, enough to shut down cleanly instead of finding out the hard way what a mid-write power cut does to a Postgres database.

n8n as the orchestration layer

The inference server runs the models. n8n is the nervous system that connects those models to your actual business: an inbound phone call, a new CRM lead, an uploaded invoice, a website chat message. It receives the trigger, routes it to the right model with the right prompt, and dispatches the output to the right destination, your CRM, your calendar, your accounting software, your Slack channel.

n8n runs on the dedicated mini workstation in the budget above, not on the inference server. The two have different bottlenecks, n8n is CPU-and-I/O bound handling webhooks and API calls, the inference box is GPU-bound, and they should not compete for resources. The n8n box runs Docker Compose with four containers: n8n itself, Postgres for the workflow database and execution log, Redis in queue mode for parallel execution, and Qdrant as the vector store powering RAG workflows. That stack needs roughly 8 GB of RAM and 4 CPU cores at busy SMB scale, comfortably inside the 32 GB, 8-core spec in the budget. The exact same Docker Compose stack we document in our self-hosting n8n playbook deploys identically on-prem, swap the Caddy reverse proxy for an internal Nginx instance and the Let's Encrypt certificate for your internal CA, and the configuration difference is three environment variables.

What 30 people can actually do with it

The hardware is a means to an end. Here is what this stack maps onto for a typical SMB, using NexFlow's six standard service lines:

• AI Receptionist. An n8n workflow answers inbound calls through a local Whisper transcription node, routes the transcript through Qwen3-72B for intent classification and response drafting, and books jobs into Google Calendar or ServiceTitan. The model runs on the local server; call transcripts never leave the network. For a trades or field-service business taking 30–50 calls a day, this alone reclaims two to four hours of admin time daily.
• CRM Automation. Every new lead, from a web form, ad platform, or inbound email, is enriched by the local model (intent scoring, category classification) and upserted into HubSpot, GoHighLevel, or Salesforce via n8n. Draft follow-ups are generated locally; a human approves before anything sends.
• Invoice Processing. Supplier invoices land in an inbox or network share. An n8n OCR node extracts the fields, the local LLM validates, categorizes, and flags anomalies, and Xero or QuickBooks Online gets a coded draft ready for one-click approval. For 200–500 invoices a month, that's roughly a week of staff time recovered a year.
• AI Chatbot (RAG). Your knowledge base, products, pricing, policies, FAQs, is chunked and embedded locally with nomic-embed-text, stored in Qdrant on the same server. The chat workflow retrieves matching chunks, grounds Qwen3-72B in that context, and answers with a citation. Full architecture in our chatbot playbook. What customers ask, and what your pricing is, stays on your server.
• Smart Documents. Contracts, SOWs, and supplier agreements route through an n8n workflow that extracts key clauses, flags non-standard terms, and writes a plain-English summary to your document system or a Slack thread. Not a substitute for legal review on complex deals, but an excellent first-pass filter.
• Workflow Automation and reporting. The unglamorous, highest-value category: cross-system sync (Shopify → Xero → Slack), a morning KPI digest assembled from Stripe, HubSpot, and ad platforms and posted before 8 a.m., automated purchase orders triggered by inventory thresholds. n8n runs the logic; the local model adds language and anomaly detection at the nodes that need it.

None of this requires the server running flat-out around the clock. A 30-person team generates AI demand in bursts during the business day and near-idle overnight, which is exactly why the electricity math below is so forgiving.

Running costs: the full picture

Electricity by region

The $20,000 spend is the headline. The honest conversation is what it costs per month over the life of the asset, and that number depends heavily on where you are. The average US commercial electricity rate sits around 13.5¢/kWh as of mid-2026, but that hides real spread: Texas runs closer to 8.35¢/kWh, while California sits around 26–28¢/kWh, roughly double the national average.

Line item	Low (TX)	Mid (national avg)	High (CA)
Electricity — inference server + n8n box	$30	$50	$100
Business fiber internet (1 Gbps)	$80	$110	$150
Hardware depreciation (3-year, ~$18k build)	$500	$500	$500
Maintenance & support reserve	$100	$150	$200
Total monthly	$710	$810	$950

The payback math

Now compare that to the cloud alternative. Thirty employees each using roughly four hours of active AI time per workday, at a conservative $2–3 per hour per concurrent session for equivalent GPU cloud compute, works out to $5,000–$9,000 a month before you even factor in per-token API charges for a chatbot or document pipeline running continuously. At that rate, the $15,700–$20,200 hardware investment (Option A) pays back in two to four months. If your team is not doing four AI-heavy hours a day yet, and most aren't on day one, the payback stretches, but the cost structure changes permanently: your AI infrastructure cost stops scaling with usage.

Plan for a hardware refresh around the three-year mark; GPU architectures move fast enough that a 2026 build will still run well in 2029 but will be a generation or two behind. The $500/month depreciation line above is a reserve fund for that refresh, not a sunk cost.

Data sovereignty and compliance

We covered the data-retention problem with cloud frontier models in depth in a recent Field Note: mandatory 30-day retention of every prompt and output, no opt-out, which the ARC Prize Foundation cited as its reason for declining to even benchmark the model. A local AI stack inverts that entirely. When the model runs on your hardware, prompts and outputs never leave your network, full stop. There is no third-party retention policy to audit, no system card to parse for hidden safeguards, and no vendor who can change the rules with a server-side update.

For US businesses, that removes the CCPA question for the AI inference layer entirely, no personal information flows to a third-party processor for that workload. State privacy laws in Virginia, Colorado, and Texas reach the same conclusion. For clinics, insurers, and other businesses under HIPAA, on-prem inference is a meaningful step toward satisfying the technical safeguard requirements under 45 CFR § 164.312, since patient data processed by the model never transmits to a covered third party. You still need a Business Associate Agreement for downstream tools handling PHI, but the AI layer itself is clean. For UK and EU operations, the analysis under GDPR Chapter V is even simpler: no international transfer occurs when nothing leaves the building. Our AI governance checklist maps each workflow step to the right compliance treatment if you want the full framework. (Australian businesses: the Privacy Act's APP 8 cross-border disclosure rules are satisfied the same way, though this remains the fourth priority market behind the US, UK, and Europe.)

When this setup is the wrong answer

Not every SMB should spend $20,000 on local AI infrastructure. Three cases where we say don't:

• Under 10 employees. The economics only work with enough usage to justify dedicated hardware. Start with n8n Cloud against a privacy-respecting API, measure real usage for 12–18 months, then revisit.
• No one in-house who can own the infrastructure. The box needs an owner: patches, disk health, the occasional container restart, roughly one to two hours a month after a solid initial setup. If your team is entirely non-technical with no managed IT arrangement, a well-configured cloud setup is lower risk than a server nobody notices went offline.
• Your workload is mostly frontier-scale tasks. If nearly everything your team does genuinely needs Kimi K2.7 or DeepSeek V4 Pro-class reasoning, a local box running Qwen3-72B does not replace that. The hybrid architecture still applies, but the on-prem server adds infrastructure cost without meaningfully reducing the cloud bill, and the capital math doesn't close the way it does for general-purpose business AI.

Build it with us

NexFlow sources, specifies, and commissions setups like this for SMBs in the USA, UK, Europe, and Australia. We shortlist the hardware, handle the OS and Docker setup, deploy the Ollama models, wire the first production n8n workflows end to end, and run a 14-day shadow period before go-live. You own the hardware, the model weights, and the workflows from day one, no lock-in, no per-execution metering, no vendor who can take your operations offline with an export control the way Fable 5 was pulled in June 2026. If you'd rather start smaller, our agentic AI guide covers the software-only version of this same architecture.

The 15-minute map call is where to start: we review your team size, usage patterns, and compliance needs, and tell you honestly whether local makes sense right now or whether a cloud-first build is the better first move. If local is the call, we scope and quote the hardware and commissioning cost. If it isn't, we say that too.