Infrastructure Pricing

Section 1

Three deployment options. One software stack.

Every client gets the same stack. The choice is where the hardware lives. Swift's default recommendation is colocation: Avalon owns the hardware and gets real redundancy without needing a server-ready building. Cloud is always available as a starting point; a move to owned hardware requires no software rebuild.

Option 1

On-Premises

Hardware in Avalon's own building. Highest sovereignty. Lowest monthly cost. Requires a stable office and IT-ready team. Client and session data never leaves the building.

Option 2 (Recommended)

Colocation

Avalon-owned hardware in a data center rack. Full ownership without building risk. Real redundancy built in. Best default for most clients who want sovereignty without managing a server room.

Option 3

Cloud

Rented servers on DigitalOcean. No upfront hardware cost. Highest monthly cost at scale. Identical software stack; can move to owned hardware at any point with no rebuild required.

	On-premises	Colocation	Cloud
Hardware owner	Avalon	Avalon	Provider
Cost pattern	Upfront, low monthly	Upfront + rack fee	No upfront, highest monthly
Local AI inference	Yes, owned GPUs	Yes, owned GPUs	Rented or via Swift
Data sovereignty	Highest	High	Lowest

Section 2

Six server roles behind one gateway.

Same names on-premises, in colo, and in the cloud. Runbooks and monitoring work identically everywhere. Avalon holds the ArcVault root key from day one in every option. Swift's access is scoped, logged, and removable at any time.

Role	What it does
ArcGate	Firewall, VPN, network zones, DNS
ArcHelm	Kubernetes, Avalon web and API, Surface
ArcData	Postgres (client and stylist profiles), Neo4j (knowledge graph), ClickHouse (session analytics), object storage (hair assets, renders)
ArcVault	Keys, secrets, Ledger. Avalon holds root key from day one in every option.
ArcRig	GPU inference: hair visualization rendering, texture transfer models, style recommendation, fine-tuning
ArcBuild	CI runners, preview environments, agent build lanes

Section 3

Cloud infrastructure costs, billed direct, no markup.

These costs apply during the build phase for all options, and as the full ongoing cost for Option 3 (cloud-first). All accounts registered to Avalon. Figures verified against DigitalOcean and Cloudflare current rates (Sept 2026); re-confirm at proposal time.

Dev AI Tooling (during build)

Tier 1: 1 developer~$1,000 / mo

Tier 2: 2 developers~$2,000 / mo

Tier 3: 3 developers~$3,000 / mo

Rate per developer$1,000 / mo

Includes frontier model API access, Claude Code, GitHub Copilot, and AI-powered CI. All accounts registered to Avalon.

Cloud Infrastructure (DigitalOcean + Cloudflare)

App + API (2 nodes, 4 vCPU / 16 GB each)$252

Managed Postgres HA (client and stylist profiles)$122

Neo4j (knowledge graph, style patterns)$168

ClickHouse (session analytics + telemetry)$63

Object storage R2 (hair assets, renders)$15

CDN, DNS, monitoring, CI/CD runners$75

Est. during build~$695 / mo

Post-launch base: ~$1,300/mo. Scales with active users, style catalog size, and rendering load. All cloud accounts registered to Avalon; zero platform take-rate.

Cost item	Phase	Monthly	Who pays
Dev AI toolingAPIs, Claude Code, Copilot, CI, per developer on the engagement	During build	$1,000 / dev	Avalon
Cloud infrastructure (build phase)App/API, Postgres, Neo4j, ClickHouse, storage, CDN	During build	~$695	Avalon
Cloud infrastructure (post-launch)Scales with user base and rendering load	Post-launch	~$1,300+	Avalon

Ownership: Every cloud account (DigitalOcean, Cloudflare) is registered under Avalon credentials from day one. Swift has scoped access for deployment and monitoring; access is logged and removable. No vendor lock-in: the full stack runs on standard infrastructure any engineer can take over.

Section 4

On-premises and colocation hardware (Options 1 and 2).

Hardware is identical across on-prem and colo. Colocation adds a monthly rack fee; the client's building costs are removed. Get vendor quotes before any number goes into a proposal. GPU hardware is priced separately in Section 5.

Tier	Users / Agent lanes	Server hardware	Colo monthly (half rack, 5 kW)
Small	Under 25 users, 1-2 lanes	~60% of Standard	Market rate, 2-3 local quotes
Standard	25-75 users, 3 lanes	$30,000-$91,000	Market rate, 2-3 local quotes
Large	75+ users, 6+ lanes	~175% of Standard	Market rate, 2-3 local quotes

Hardware sourcing: The $30,000 floor assumes certified refurbished servers (Techbuyer, ServerMonkey, eBay Business). The $91,000 ceiling is all-new. Compute nodes (ArcHelm, ArcBuild) and switches are good candidates for refurbished. UPS units, PDUs, and ArcVault nodes should be new or under current warranty; power protection and key storage are not places to accept unknown service history.

Section 5

GPU options for the Hair Intelligence Engine.

GPU hardware powers Avalon's Hair Intelligence Engine: real-time visualization rendering, texture transfer model inference, style recommendation, and fine-tuning on Avalon's style catalog. GPU prices must be re-verified the week each proposal is sent.

Starter Build

~$13,000

2x RTX 5090 (32 GB each)

Fits 1-2 agent lanes, launch scale. Runs 70B-class models and real-time visualization. Cards can be upgraded later without replacing server hardware.

Standard Build (Recommended)

~$39,000

2x RTX PRO 6000 Blackwell Max-Q (96 GB each)

Fits 3-6 agent lanes. 192 GB total VRAM. Runs texture transfer and recommendation models simultaneously. Covers the full Hair Intelligence stack at scale.

Heavy Build

~$78,000-$83,000

4x RTX PRO 6000 Server Edition (96 GB each)

384 GB total VRAM. Multiple large models serving many agent lanes, high-concurrency rendering, and fine-tuning in parallel. Server Edition is passive-cooled for rack use.

GPU memory	Runs well (4-bit quantized)
64 GB (2x RTX 5090)	70B-class models: style recommendation, cataloging, structured drafting
96-192 GB (1-2x RTX PRO 6000)	Texture transfer models, 70B at higher quality, several models simultaneously
384 GB (4x RTX PRO 6000)	Multiple large models serving many agent lanes, concurrent rendering, and fine-tuning in parallel

Re-quote rule: GPU prices have roughly doubled since launch as AI data centers absorbed the world's memory supply. The September 2026 numbers above are a floor, not a ceiling. Every GPU number must be re-verified the week the proposal is sent.

Section 6

Frontier API baseline: what the platform costs today.

What Avalon pays for AI before any owned inference hardware. Frontier API and SaaS tooling scale linearly with team and agent count. They do not decrease as usage grows; unlike owned hardware, they carry no fixed ceiling.

AI-Assisted Development (full stack)

Raw API usage per developer$100-$300 / mo

Claude Code + GitHub Copilot~$400 / mo

AI-powered CI tooling~$300 / mo

All-in per developer~$1,000 / mo

Autonomous Agent Lanes

Business hours only (~8 hr/day)$500-$1,500 / lane

24/7 lane$1,500-$3,000 / lane

3 lanes 24/7 (example)$6,000 / mo

Frontier API and SaaS tooling costs scale linearly with every new developer and agent lane added. Owned hardware costs are fixed once purchased; every additional agent lane after payback is near-zero marginal cost.

Section 7

The ownership argument.

The core case for buying hardware rather than renting cloud GPUs indefinitely. Scenario: Standard ArcRig build vs. renting equivalent H100 compute from DigitalOcean through Swift.

Frontier API monthly

$6,000

3 agent lanes at ~$2,000/lane/mo. Scales with every new lane added.

Owned inference monthly

~$1,100

Standard ArcRig amortized over 3 years (~$54k total hardware). Near-zero marginal cost after payback period.

Annual savings

~$59,000

Hardware pays itself back in 8-10 months. Everything after is inference at near-zero marginal cost.

Comparison	DO base rate	2x GPU 24/7 (base)	Cost to Avalon
H100 on-demand (DO)	$4.41/GPU/hr	$6,440/mo	$8,370/mo
H100 reserved 12-month (DO)	$3.26/GPU/hr	$4,760/mo	$6,190/mo
Standard ArcRig owned (2x RTX PRO 6000)	$39,000 one-time hardware		~$1,100/mo amortized

Standard ArcRig vs. reserved DO H100 through Swift ($6,190/mo): payback in 6-7 months. Every month after that is ~$6,190 kept. Over 36 months, owned hardware saves approximately $183,000 vs reserved H100 rental, before power costs.

Section 8

Own or borrow from Swift: start anywhere, move anytime.

Avalon can borrow any component from Swift's Grid until usage justifies owning it. The software stack is identical regardless of which option is chosen. Moving from borrowed to owned requires no rebuild at any point.

Component	Avalon owns	Borrowed from Swift
Brain (ArcRouter + models)	Runs on Avalon ArcRig; all client data stays in Avalon infrastructure	Avalon apps call Swift's Grid through ArcRouter; no GPUs needed
Vault (keys, secrets, Ledger)	Runs on Avalon ArcVault nodes	Swift hosts it; Avalon receives SDK and scoped keys
Surface workspace	Runs on Avalon ArcHelm	Workspace on Swift's Surface instance
Build pipeline	Runs on Avalon ArcBuild nodes	Builds run on Swift's runners, billed per use

Section 9

Proposal and re-quote rules.

Hardware costs are planning estimates; get vendor quotes before any number goes into a final proposal. Labor retainers are quoted separately.

Item	Rule
GPU prices	Re-verify the week the proposal is sent. RTX PRO 6000 prices have been volatile; September 2026 figures are a floor, not a ceiling.
Server hardware	Re-confirm within 30 days. Get vendor quotes before any number goes into a final proposal.
Colocation rack fees	Get 2-3 local quotes per proposal. Rates vary by market and facility tier.
Cloud compute	Confirm ArcBuild node count against Avalon's actual agent lanes before finalizing. Section 3 figures are based on verified DO rates (Sept 2026).
Markup basis	Cloud accounts billed at vendor rate with no markup. Labor retainers are a separate line item in every proposal.

Three paths to ownyour AI stack.