Sovereign AI vs Public Cloud LLM: Total Cost and Risk Comparison for Oman Government

An Omani permanent secretary asks a finance director a simple question. "If we standardise on a public-cloud language model for the ministry, what does it cost over three years, and what does it expose us to?" The honest answer is not a single figure. It is two figures, one priced in Omani rials and one priced in jurisdictional risk, and a clear-eyed view of which dominates at each query volume. This article works the comparison the way a sovereign buyer should work it, with public 2026 vendor pricing on one side, the real cost of on-premise hardware and operators on the other, and the line items that most TCO spreadsheets quietly omit.

The TCO question, framed for sovereign buyers

Sovereign buyers do not buy AI the way a startup does. The horizon is three years, not three months. The relevant baseline is not "one engineer testing prompts" but "the ministry's full caseload, every working day, for the duration of the budget cycle." For Omani institutions, the credible volume range is 100,000 to 500,000 queries per day, depending on whether the AI sits behind a single directorate, a full ministry, or a citizen-facing channel.

A useful query is not a one-token autocomplete. The realistic profile for a ministerial workload is an average input of 4,000 tokens (a briefing paragraph plus retrieved context from a document store) and an average output of 600 tokens (a structured summary, a draft paragraph, a bilingual reformulation). At 200,000 queries per day, that is 800 million input tokens and 120 million output tokens daily, or roughly 24 billion input tokens and 3.6 billion output tokens per month. Holding this profile fixed lets us compare cloud and on-premise on the same workload rather than on rhetorical assumptions.

The cost question has two answers. The first is the bill. The second is what the bill does not include: the legal exposure, the lock-in, the egress cost of leaving, the deprecation risk, and the operational overhead of governing a foreign supply chain. A complete TCO is the sum of both.

The cloud-LLM cost model

Public-cloud LLM pricing in 2026 is well documented. Anthropic publishes Claude Opus 4.7 at 5 US dollars per million input tokens and 25 US dollars per million output tokens, with Claude Sonnet 4.6 at 3 and 15, and Claude Haiku 4.5 at 1 and 5. AWS Bedrock mirrors these rates almost exactly for the Anthropic family, with cross-region inference adding around 10 percent. OpenAI's frontier-tier pricing sits in a similar band, with mid-tier variants priced lower and reasoning-heavy variants priced higher. Provisioned throughput on Bedrock is hourly, in the range of 40 to 200 US dollars per hour per model unit.

Plug the 200,000-queries-per-day workload into Sonnet-class pricing. Twenty-four billion input tokens at 3 US dollars per million is 72,000 US dollars. Three point six billion output tokens at 15 US dollars per million is 54,000 US dollars. That is 126,000 US dollars per month on inference alone, before any of the other line items the bill carries.

Those other line items are not small. Egress is the most often forgotten. AWS charges 0.09 US dollars per gigabyte for the first 10 TB of internet egress, with NAT Gateway processing adding 0.045 US dollars per gigabyte for traffic from a private subnet. A retrieval-augmented architecture that ships embeddings, document chunks, and tool outputs back and forth across regions easily produces 5 to 20 TB of monthly egress for a 200k-per-day workload, or 450 to 1,800 US dollars per month. Cross-region transfer at 0.02 US dollars per gigabyte for redundancy adds a similar layer.

Then comes the redundancy multiplier. A serious ministry will not run a single-region cloud deployment. Active-active across two regions multiplies steady-state spend by roughly 1.6x to 1.8x once data replication, idempotency tooling, and cross-region inference markup are included. Governance overhead, identity federation, audit-log shipping, FinOps tooling, contract management with the hyperscaler's local representative, and the legal review needed for Article 23 cross-border posture under Royal Decree 6/2022, runs another 5,000 to 15,000 US dollars per month at this volume.

Aggregating these honestly puts a 200k-per-day public-cloud LLM workload at roughly 220,000 to 280,000 US dollars per month at Sonnet-class quality, or 2.6 to 3.4 million US dollars per year, before any volume discount and before any growth.

The on-premise cost model

The on-premise side is dominated by capex and the unit economics of inference hardware. The right reference point in 2026 is a 4U or 8U rack-mounted appliance with two to eight NVIDIA H100 or H200 accelerators. A complete 8x H100 server sits in the 200,000 to 450,000 US dollars range fully built, including chassis, NVMe storage in the tens of terabytes, redundant power, and the network appliance. Single H100 cards are roughly 25,000 to 40,000 US dollars depending on form factor, and an H200-based server is 10 to 25 percent more.

What does that hardware deliver? vLLM benchmarks on a single H100 show steady-state throughput in the range of 2,400 to 2,780 tokens per second for 70B-class dense models under realistic concurrent load, climbing to 12,000 tokens per second and beyond on smaller 8B models with FlashInfer. Eight H100s in a single chassis push that ceiling upward by close to a linear factor for batched workloads. At 920 million tokens per day (the 200k-query workload), an 8x H100 appliance runs comfortably below half-utilisation, leaving headroom for fine-tuning, retrieval re-ranking, and concurrent specialist models.

Operational cost is small. Power for an 8x H100 chassis at sustained load is roughly 6 to 10 kW, or 4,500 to 7,500 kWh per month. At Omani commercial tariffs that is in the order of 250 to 450 OMR per month per appliance. Rack space inside a ministry data hall is effectively free at the margin. Maintenance, including spare parts cover and an annual refresh of NVMe drives, sits at 3 to 5 percent of capex per year. Staffing is the largest soft cost and is discussed below.

For a Workstation-tier appliance, a single Apple M3 Ultra Mac Studio, the picture is different again. Real-world benchmarks show roughly 30 to 41 tokens per second on Gemma-3-27B class models, comfortable for one to four concurrent operators, at a capex below 15,000 US dollars and a power footprint under 200 W. This tier handles a chief of staff or a small intelligence cell, not a 200k-per-day ministry, but it is the cheapest credible entry point for a sovereign deployment.

Break-even by query volume

The break-even calculation is what makes the discussion concrete. The table below uses 2026 published vendor pricing for cloud and a mid-range capex assumption for on-premise, amortised over three years, including power, staffing, and a single hardware refresh.

The break-even point on Sonnet-class pricing for the assumed workload profile sits around 12,000 to 18,000 queries per day for a Tower-tier appliance and around 30,000 to 45,000 queries per day for a Rack-tier appliance. Below the break-even, cloud wins on cash. Above it, the on-premise cost curve flattens while the cloud bill grows linearly. At ministry scale, the gap is an order of magnitude.

Risk dimensions cost cannot price

Cash is only one axis. The dimensions that the cloud bill never shows are the ones that matter to a sovereign buyer.

The first is jurisdictional exposure. The United States CLOUD Act grants US authorities the right to compel disclosure of data held by US-headquartered providers regardless of where the bytes physically reside. A Bedrock or Foundry deployment in any region answers to this regime. China's Data Security Law creates a mirror-image obligation on Chinese providers. Neither pricing page mentions either statute, and yet for a defence ministry or central bank, this is the dominant cost.

The second is vendor lock-in. Cloud LLM contracts increasingly include reserved capacity, custom fine-tunes, and platform-specific tooling that is non-portable. The exit cost, retraining staff, rewriting integration code, repaying remaining reserved-capacity commitments, sits in the 6-to-12-month and several-hundred-thousand-US-dollar range for any non-trivial deployment. A sovereign buyer who has standardised on a single hyperscaler has implicitly priced themselves into a future contract negotiation they cannot walk away from.

The third is model deprecation. Hyperscalers retire models on their schedule, not the customer's. A ministry that has built workflows, prompts, and fine-tunes against one generation can be forced into a costly migration when that generation is sunsetted. On-premise deployments retain old model weights as long as the institution chooses, which is the right default for archival and auditability of past decisions.

The fourth is telemetry leakage. Every cloud LLM request is logged on the operator's side. Even with contractual retention controls, the metadata, who queried, when, with what frequency, against which document store, leaks to a foreign operator and is legible to a foreign regulator under subpoena. Sovereign on-premise systems keep this telemetry inside the perimeter by construction.

The hidden line items most TCO analyses miss

Five line items are routinely absent from cloud-versus-on-premise comparisons.

Compliance audit cost. A cross-border AI deployment under Royal Decree 6/2022 requires periodic legal review of the cross-border transfer regime, vendor compliance attestations, and the institution's own data protection impact assessment. This runs 25,000 to 75,000 US dollars per year for a serious ministry, mostly invisible inside the legal department's existing budget.

Governance overhead. Identity federation between an Omani institution's directory and a hyperscaler's IAM is non-trivial. Audit-log shipping, retention, and cross-jurisdiction defensibility add ongoing engineering. None of this appears on the inference invoice.

Egress at scale. A retrieval-augmented architecture or a multi-modal workload (vision, audio) sends real volume across the wire. The 0.09 US-dollar-per-gigabyte rate looks cheap until a defence directorate ships its case-file index nightly to a cross-region embedding service.

Cloud fine-tuning costs. Fine-tuning on a hyperscaler is multiples more expensive than inference, and the fine-tuned weights live on the hyperscaler's storage. Replicating those weights to a second provider for resilience usually requires re-running the fine-tune from scratch, which doubles the cost.

The exit clause. Most cloud LLM contracts include a 90-day data-purge SLA on cancellation, but no clause for retrieving the fine-tuned model weights. The institution that fine-tuned a billion-parameter model on classified internal correspondence cannot, in practice, take that model with it on the way out.

When public cloud is genuinely the right answer

Public cloud LLM is a credible choice for several Omani sovereign workloads. The shared platform provided by Mu'een already covers cross-government use cases where a national-level deployment is the right primitive. Public-facing chatbots over already-published material (open data, public regulations, ministry FAQs) carry no sovereignty risk because the input set is itself public. Marketing copy generation, internal employee training material, citizen-facing translation of public documents, and developer-side code completion against non-classified codebases are all reasonable cloud workloads. The break-even analysis above also favours cloud below roughly 15,000 queries per day, where the on-premise capex cannot amortise.

The line is bright where it should be bright. If the input or output of a query would be classified, restricted, or commercially sensitive in any other context, it does not belong on a foreign-operated cloud, regardless of region label, regardless of contractual promise, and regardless of price.

The realistic 3-year picture, 200k queries per day

An Omani ministry standardising on a Sonnet-class cloud LLM for a 200,000-queries-per-day workload over three years pays roughly 9 to 10 million US dollars on the bill, plus an estimated 600,000 to 1.2 million US dollars in egress, redundancy, governance, and compliance audit overhead. It carries CLOUD Act exposure on every byte of input and output, accumulates lock-in through reserved capacity and platform-specific tooling, and ships its full operational telemetry to a foreign operator.

The same ministry deploying a sovereign Rack-tier appliance with two H100-class servers, an institutional fine-tuning capability, and a fully on-premise retrieval-augmented stack pays roughly 1.4 to 1.7 million US dollars in three-year all-in cost, including hardware, power, staffing at one full-time-equivalent, and a mid-cycle hardware refresh. It carries no jurisdictional exposure, retains every fine-tune as a sovereign asset, and exits the discussion of foreign legal process entirely.

The cost saving is roughly 7 to 8 million US dollars across three years. The risk reduction is categorical. The procurement question is not "which is cheaper", it is "what is the right tool for which workload", and the answer for any sensitive ministerial work is unambiguously the sovereign appliance. Hosn ships this pattern in three reference tiers, sized to a specific concurrency and classification target, with pricing by quotation against your actual workload.

If your ministry is running this comparison and would like a one-hour briefing tailored to your concurrency, classification, and integration requirements, the next step is simple. Email [email protected] or message +968 9889 9100. We will walk through the architecture, the models, and a credible plan against your timeline.