SMTM Hosting
Home Page
Services Approach Reviews About
Spend Forecasting Contact

A planning reference, not a price list

What custom AI and infrastructure hosting actually costs to run.

We don't sell fixed packages, which means we also don't publish fixed prices. What we can do is show you the shape of the number: what compute, data, bandwidth, and management actually cost at current market rates, broken down by the variables that move your bill the most.

This page reflects blended, current market rates across cloud GPU providers, managed vector database platforms, and bandwidth vendors as of 2026. It is built to give you a realistic planning range before we scope anything specific to your system.

What actually moves the number

Eight variables account for nearly all of the spread.

Two companies running what looks like "the same" AI product can land 10x apart on monthly spend. These are the variables that explain why.

Traffic & inquiry volume

Concurrent users and API calls per month set your baseline compute and bandwidth floor before anything else is decided.

Model size & GPU class

A 7B parameter model and a 70B+ model are different hosting problems entirely, and the GPU class you need follows directly from that choice.

Inference pattern

Always-on dedicated capacity costs more per hour than autoscaled or serverless inference, but serverless costs more per request at sustained volume.

Vector data volume

RAG and semantic search costs scale with the number of stored vectors and their dimensionality, not just with traffic.

Bandwidth & egress

Media-heavy products move far more data out of the cloud than text-only products, and egress pricing is tiered and provider-specific.

Redundancy & uptime

Single-region hosting and multi-region failover are not the same line item twice. Failover roughly doubles the infrastructure it protects.

Compliance & data residency

Formal compliance programs and data residency requirements add dedicated infrastructure and audit overhead most teams underestimate.

Management depth

Business-hours monitoring and 24/7 SRE-level management are priced very differently, and most outages happen outside business hours.

Compute reference

GPU cloud rates, current market range.

Inference and fine-tuning workloads are priced by the GPU-hour. Rates vary by provider tier, from budget neo-clouds to major hyperscalers, and by commitment level.

GPU ClassTypical UseHourly RangeFull-Time Monthly (single GPU)
A100 (80GB)Mid-scale training, fine-tuning, moderate inference$1.25 to $2.50$900 to $1,800
H100 (80GB)Production inference, 30B to 70B class models$1.50 to $7.00$1,100 to $5,000
H200 (141GB)Memory-bound inference, larger context windows$0.50 to $4.00$400 to $2,900
B200 (latest gen)Frontier-scale training and high-throughput inference$5.00 to $18.00$3,600 to $13,000

Multi-GPU clusters (4 to 8 GPUs, the realistic floor for serious training or high-availability inference) typically run $6,000 to $40,000 per month depending on GPU class, provider, and reservation term.

Data layer reference

Vector and RAG storage, by scale.

Retrieval-augmented systems price by stored vectors, query volume, and embedding dimensionality. Cost is rarely linear. It steps up hard at certain scale thresholds.

Index SizeTypical WorkloadMonthly Range
Under 1M vectorsPilot, early product, small knowledge base$0 to $75
1M to 10M vectorsProduction RAG for a single product line$65 to $400
10M to 100M vectorsMulti-tenant RAG, growing document corpus$400 to $2,000
100M+ vectorsEnterprise-scale search, agent memory at scale$2,000 to $10,000+

Self-hosting the vector layer can undercut managed platforms by 3x to 10x at high query volume, at the cost of taking on the operational overhead directly. That trade-off is one of the first things we model in an actual engagement.

Delivery reference

Bandwidth and egress, by provider type.

Egress, data leaving the cloud toward your users, is billed per gigabyte and tiered by volume. A small number of providers charge nothing for it at all.

Provider TypeEgress Rate10TB/Month Estimate
Major hyperscaler (standard tier)$0.07 to $0.12/GB$700 to $1,200
Major hyperscaler (high volume, negotiated)$0.02 to $0.05/GB$200 to $500
CDN edge delivery (cache hit)$0.004 to $0.02/GB$40 to $200
Zero-egress object storage$0.00/GB$0, storage billed separately

Which of these actually applies depends on how much of your traffic is cache hits versus origin pulls, and whether your architecture can take advantage of zero-egress storage at all.

Illustrative spend tiers

Four shapes of the same question: what will this actually cost.

These are not packages we sell. They are composite ranges built from the reference tables above, illustrating how total monthly spend moves as traffic, redundancy, and compliance requirements scale up. Your actual number depends on your actual architecture.

Pilot / Early-Stage
Up to roughly 10,000 inquiries per month
$1,200 to $3,500 / mo
Compute
Shared or serverless inference, part-time single GPU class capacity
$300 to $900
Vector & data layer
Under 1M vectors, serverless managed tier
$0 to $75
Bandwidth & delivery
Under 1TB monthly egress
$50 to $150
Security & monitoring
Baseline hardening, business-hours monitoring
$400 to $900
Management overhead
Light-touch, monthly check-ins
$500 to $1,200
Composite monthly range
$1,200 to $3,500
Growth
Roughly 10,000 to 250,000 inquiries per month
$3,500 to $10,000 / mo
Compute
Dedicated single H100-class GPU, reserved capacity
$1,500 to $3,500
Vector & data layer
1M to 10M vectors, managed platform
$70 to $400
Bandwidth & delivery
1 to 10TB monthly egress
$150 to $900
Security & monitoring
Extended-hours monitoring, quarterly audits
$900 to $2,000
Management overhead
Proactive management, monthly optimization pass
$1,200 to $2,800
Composite monthly range
$3,500 to $10,000
Scale
Roughly 250,000 to 2,000,000 inquiries per month
$12,500 to $50,000 / mo
Compute
Multi-GPU cluster, 4 to 8x H100 or H200-class, reserved
$6,000 to $25,000
Vector & data layer
10M to 100M vectors, multi-region replication
$400 to $2,000
Bandwidth & delivery
10 to 100TB monthly egress
$900 to $7,500
Security & monitoring
24/7 monitoring, formal compliance audits
$2,500 to $6,000
Management overhead
Dedicated management, weekly optimization
$3,000 to $7,500
Composite monthly range
$12,500 to $50,000
Enterprise / Mission-Critical
2,000,000+ inquiries per month
$50,000 to $200,000+ / mo
Compute
Multi-node reserved clusters, latest-generation GPUs, multi-region
$25,000 to $100,000+
Vector & data layer
100M+ vectors, full redundancy
$2,000 to $10,000+
Bandwidth & delivery
100TB+ monthly egress, global CDN footprint
$7,500 to $30,000+
Security & monitoring
24/7 SRE coverage, formal compliance program
$8,000 to $20,000+
Management overhead
Embedded management team, on-call SLA
$10,000 to $25,000+
Composite monthly range
$50,000 to $200,000+
How to read this page. These figures are a blended planning reference built from current, publicly observable market rates across cloud GPU providers, managed vector database platforms, and bandwidth vendors as of 2026. They are not a quote, a rate card, or a menu of packages. Actual cost depends heavily on architecture choices, provider selection, utilization patterns, and decisions we would only make after a real assessment of what you're running. Every SMTM Hosting engagement is scoped and priced individually. Use this page to walk into that conversation informed, not to skip it.

Let's talk about what you're building

Ready to find your real number?

These ranges are a starting point. Whether you already have a system running or you're starting from nothing, we'll scope the actual figure together.