Insource Compute Book a reviewBook a 30-minute technical review

Managed on-premise AI

Your AI bill is someone else's pricing decision.

Insource Compute runs open-weight AI models on infrastructure you already own — inside your perimeter, with no per-token meter. We deploy it, manage it, and keep it current. You stop paying to rent intelligence.

Book a 30-minute technical review Start a 14-day evaluation

For enterprises running proprietary, regulated, or sovereign data.

Built on

Llama · DeepSeek · Qwen · Mistral

— open-weight models you keep.

01 / The problem

Nobody decided to put a business-critical cost line under a single vendor's unilateral control. It happened one API key at a time.

108%

YoY growth in enterprise AI spend

Source: Zylo, 2026

73%

of organizations exceeded their AI budget

Source: Zylo, 2026

7%

of C-suite leaders have established AI ROI measurement

Source: KPMG Global AI Pulse, June 2026, n=2,145

Per-token prices have fallen. Enterprise AI bills went up anyway — usage grows faster than unit price drops, and the models you actually want cost more, not less. OpenAI's own inference spend rose roughly 4x in a single year while list prices were falling.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation
Insource Compute · Technical note Evidence 01 — Pricing Evidence 01 § 02

The rate card is not the bill.

2.1 Reasoning tokens are invisible cost. o3 lists at $2/$8 per million tokens — an 80% cut from its own launch price. But high reasoning effort adds 10,000–30,000 “thinking” tokens to a single response, billed at the output rate. Your cost per query can climb while the rate card falls.1

2.2 Falling prices have not produced falling bills. GPT-4's $30/$60 became GPT-4o's $2.50/$10 in a year. Enterprise AI spend still rose 108% year over year. Usage grows faster than unit price falls — which is why your bill is not forecastable from a rate card.1,3

2.3 You hold no price protection. No lock, no cap, no notice period. Models are deprecated and traffic rerouted at the vendor's discretion.

Unit price down. Total spend up.

Indexed, start = 100 200150100500 208 17 AI spend ↑ 108% Unit price ↓ 83% BeforeAfter Indexed, start = 100 200150100500 100baseline 208enterprise AI spend 17price per 1M output tokens Total spend — rising ↑ 108% Unit price — falling ↓ 83% BeforeAfter
FIG. 2.1 Both series indexed to 100. Unit price is list price per million output tokens, GPT-4 ($60) to GPT-4o ($10). Spend is year-over-year enterprise AI spend growth. Cheaper tokens have not produced cheaper bills.

Precedent · Broadcom / VMware

In December 2023, Broadcom acquired VMware for ~$61B. It then cut roughly 168 product bundles to four, ended perpetual licensing, and raised the minimum purchase from 16 cores to 72. Customers reported increases of 8–15x. Perpetual-to-subscription conversions exceeded 1,000%; AT&T was reportedly quoted 1,050%. Gartner called it a once-in-a-decade disruption to enterprise IT budgets.2

Nothing in your AI contracts prevents this.

1 OpenAI published API list prices, input / output per 1M tokens. o3 reduced from $10/$40 to $2/$8, June 2025.

2 Broadcom / VMware: customer and press reports; Gartner commentary.

3 Zylo, 2026 SaaS Management Index: year-over-year enterprise AI spend growth.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation

03 / What we do

Insource it.

  1. 01

    Assess

    We model your real token volume, workload mix and existing GPU capacity, and tell you honestly whether this pays. If it doesn't, we say so.

  2. 02

    Deploy

    Open-weight models installed on your infrastructure, inside your perimeter. Nothing leaves your network.

  3. 03

    Operate

    We manage, monitor, patch and upgrade it. New open models get evaluated and rolled out as they ship. You hire no one.

You don't become an AI infrastructure company. You just stop renting one.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation

04 / For Finance

From a variable bill you don't control to a fixed cost you do.

Flat annual platform fee plus a monthly managed-service fee. No per-token charges, ever.

  • Forecastable.

    The same number in month 12 as in month 1, whatever usage does.

  • Usage stops being rationed.

    Teams stop being told to go easy on AI because of cost.

  • The asset is yours.

    Compute you own is depreciable and reusable. Tokens you rent are gone.

  • No exposure

    to repricing, deprecation, or new billable token categories.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation

05 / For Security & IT

The model comes to your data. Your data never goes anywhere.

  • Zero egress.

    Inference happens inside your perimeter. Prompts, documents and outputs never transit a third party.

  • No third-party processor

    in the data path — removes an entire category of DPA, residency and cross-border exposure.

  • Air-gap capable.

    Deployable with no outbound internet at all, for OT-adjacent environments.

  • Auditable.

    You control model versions, logging, retention and access. Nothing changes underneath you.

77%

of enterprise AI users paste company data into model prompts

82%

of them from unmanaged personal accounts

10.7% → 34.8%

share classified as sensitive, over two years

Source: LayerX, 2025

In 2023 Samsung banned generative AI on company devices after engineers pasted proprietary source code into ChatGPT.

The alternative to a sanctioned internal model isn't no AI. It's shadow AI.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation
Insource Compute · Technical note Evidence 02 — Capability Evidence 02 § 06

Open-weight models are a fiscal quarter behind the frontier — and ahead of what your workload needs.

6.1 Epoch AI measures the gap between the best open-weight and best closed models at ~3–4 months.1 The UK AI Security Institute independently puts it at 4–7 months.2

6.2 It has been closed faster: DeepSeek R1 shipped roughly seven weeks after OpenAI's o1, matching or beating it on AIME, MATH-500 and SWE-bench Verified.3

6.3 Most enterprise AI work isn't frontier work. Extraction, classification, summarization, retrieval and internal Q&A are solved problems for current open-weight models.

6.4 And where you genuinely need frontier capability, keep it. We're not asking you to cancel your API account. We're asking you to stop routing 90% of your volume through it.

How far behind: measured as a range

Epoch AI ~3–4 mo UK AI Security Institute 4–7 mo DeepSeek R1 after o1 ~7 wks · outlier 012345678 Months behind the best closed model Epoch AIUK AI Security InstituteDeepSeek R1 after o1 ~3–4 months4–7 months~7 weeks · outlier 012345678 Months behind the best closed model
FIG. 6.1 Two independent estimates, plotted as ranges rather than a single number. R1 is the fastest observed case, not the norm.

1 Epoch AI, open vs. closed model capability gap.

2 UK AI Security Institute, independent estimate.

3 DeepSeek R1 release; reported AIME, MATH-500, SWE-bench Verified results.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation

07 / Hardware

You probably already own the hard part.

Seismic imaging and reservoir simulation put serious GPU compute in energy companies years before anyone said “LLM.” You have the clusters, the power and cooling built for them, the vendor relationships, and staff who run them. That's the slow, expensive part — and you're past it.

We'll tell you exactly what your existing capacity supports, and where it doesn't, the smallest sensible addition. Most enterprise workloads run comfortably on models far smaller than the headlines suggest.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation

08 / Questions

What people ask first.

What if we need frontier models for some work?

Keep them. This is about the 90%, not the 10%.

Is open-weight actually good enough?

For your workloads, almost certainly. We'll prove it on your own documents during the evaluation.

What hardware do we need?

We assess before you commit. If you don't have the capacity, we tell you what it costs — before you've spent anything.

Who runs it?

We do. Fully managed: deployment, monitoring, patching, model upgrades.

What if we want out?

The models are open-weight. You keep them and keep running them. No lock-in — that's the entire point of the company.

Does the evaluation require sending you sensitive data?

No. It runs in an isolated single-tenant environment, we never train on anything you send, and no production or regulated data is needed.

The offer

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

14 days. No cost, no commitment. Or talk it through with an engineer first.

Book a 30-minute technical review or start the 14-day evaluation

09 / The evaluation

Start with a 14-day evaluation. Nothing gets installed and nothing sensitive moves.

We stand up an isolated, single-tenant environment running the exact open-weight models we'd deploy for you. Your team tests them against real workload types — not production data — and judges the quality directly.

Start a 14-day evaluation or book a 30-minute technical review
  • 14 days. No cost, no commitment.
  • Single-tenant and isolated. We never train on your inputs.
  • No production or regulated data required.
  • You finish with a written assessment: what your workloads need, what your hardware supports, and what it would cost.

10 / Get started

Find out what you're actually spending — and what you'd stop spending.

Book a 30-minute technical review

Next step: pick a time on the calendar.

Your role routes you to the right engineer. Privacy