Saudi Arabia’s AI Compute Strategy: What 480 Megawatts Actually Buys, and Why Your Assistant Does Not Need It
Why power replaced floor space as the binding constraint, the decisive difference between training and inference economics, how to calculate your own API-versus-hardware crossover, and the narrow cases that genuinely need national scale.
Megawatts have become the unit in which national AI ambition is announced. The Hexagon data centre is described at 480 megawatts; HUMAIN carries a mandate around 500. These are genuinely large numbers, and they are quoted far more often than they are translated.
This article translates them. What a megawatt of AI capacity physically contains, why power rather than floor space became the limiting unit, how much of it an ordinary enterprise workload actually consumes, and the specific — narrow — set of circumstances in which a Saudi organisation needs national-scale compute rather than a rack of its own.
Why capacity is measured in power, not square metres
A traditional data centre was constrained by space and network. An AI data centre is constrained by electricity and by the ability to remove the heat that electricity becomes. A rack of general-purpose servers might draw 5 to 10 kilowatts. A rack densely populated with AI accelerators can draw several times that, and the difference is not incremental — it changes the building.
| Conventional hosting | AI compute | ||
|---|---|---|---|
| Binding constraint | Floor space and network | Power delivery and heat removal | |
| Rack density | Roughly 5–10 kW | Several times higher | |
| Cooling | Air, generally sufficient | Frequently liquid at high density | |
| Utilisation pattern | Spiky, follows user traffic | Sustained near-full load during training | |
| What limits growth | Adding cabinets | Grid connection and substation capacity |
This is why capacity is announced in megawatts. Power is the number that cannot be worked around once the building exists.
Training and inference are different businesses
The single most useful distinction for reading any compute announcement is whether the capacity is aimed at training or at inference. They have opposite economic shapes, and a buyer who conflates them will size their own requirement wrongly by orders of magnitude.
| Training | Inference | ||
|---|---|---|---|
| What it does | Produces model weights from data | Uses finished weights to answer a question | |
| Duration | Weeks to months, continuous | Milliseconds, on demand | |
| Hardware appetite | Thousands of accelerators in one fabric | One to a few, per concurrent request | |
| Who realistically does it | National programmes, frontier labs, a few large enterprises | Essentially every organisation deploying AI | |
| Where national capacity matters | Decisively | Rarely |
National megawatt figures are overwhelmingly about the first column. Almost every enterprise buyer lives entirely in the second.
This is the crux. When an organisation reads that the Kingdom now has 480 megawatts of government data-centre capacity and concludes that its customer-service assistant should be built on national compute, it has silently moved from the second column to the first. The assistant is inference. It always was.
What an ordinary enterprise assistant actually needs
It is worth stating plainly, because the gap between the national figures and the enterprise requirement is several orders of magnitude and almost nobody says so out loud.
Illustrative relative magnitudes, not measurements. The point is the gap between the top and bottom rows, which is larger than the chart can honestly draw.
A retrieval assistant answering questions from a company’s own documents, for a few hundred concurrent users, runs on hardware that fits in a rack and often on a single server. That is not a limitation to apologise for. It is the correct sizing, and it is why on-premises deployment is economically reachable for a mid-sized Saudi organisation rather than being a luxury.
- 2019SDAIA establishedThe national authority, and the body that would later set the rules AI deployments run under.
- May 2025HUMAIN launchedA PIF company with a mandate around 500 megawatts and chip-supply arrangements with NVIDIA and AMD.
- 2025$9.1bn committedAcross roughly 70 investment deals into Saudi AI-sector companies in a single year.
- Early 2026Hexagon inauguratedDescribed as the world’s largest government data centre, at 480 megawatts.
- 2026Year of AI declaredThe designation follows the build-out rather than announcing it.
Read in order, the capacity preceded the designation. That sequencing is the reason the figures are worth taking seriously.
The cost shapes are genuinely different, and the crossover is calculable
Per-token API pricing and owned hardware are not two prices for the same thing. They are two shapes. One is proportional to use with no floor; the other is a floor with near-zero marginal cost. Which is cheaper depends entirely on volume, and the crossover point is something you can compute rather than argue about.
- 1Measure real volumeQuestions per day, and average tokens in and out per question — measured on Arabic text, which is more expensive per sentence than English.
- 2Price the API pathVolume times per-token rate, plus egress, plus the engineering time to manage rate limits and retries.
- 3Price the owned pathHardware, plus power, plus cooling, plus a share of an engineer, amortised over three years — not two, and not five.
- 4Add the asymmetriesThe API path grows with success. The owned path does not, until you exceed the box, at which point it steps.
- 5Price the risks separatelyProvider price changes and deprecations on one side; hardware refresh and your own downtime on the other.
- 6Decide on the shape, not the point estimateIf your volume is near the crossover, choose on residency and control instead, because the money is a tie.
Energy, water and the questions worth asking a host
AI capacity at national scale draws attention to power sourcing and to cooling water, both of which matter in a desert climate. The Kingdom’s renewable programme is directly relevant here, and any organisation with sustainability reporting obligations should ask about it rather than assume.
| Question | Why it matters | |
|---|---|---|
| What is the PUE, and measured how | Power usage effectiveness varies with method and season; an annualised figure measured on site is worth more than a design figure | |
| What is the energy mix | Directly feeds your own emissions reporting, and increasingly your customers’ questions | |
| What cooling technology is used | Liquid cooling changes water draw and heat reuse options; in a desert climate this is not a footnote | |
| What happens at grid interruption | Generator runtime and fuel on site determine how long your assistant keeps answering | |
| Is capacity contracted or available | A published megawatt figure is a facility total, not an allocation to you |
The last row causes the most disappointment. Announced national capacity is not a queue you automatically stand in.
The shapes, not the prices. The owned path is flat because the marginal cost of one more answer on hardware you already own is close to zero.
When you genuinely do need national-scale compute
There are real cases. They are narrower than the enthusiasm suggests, and it is worth being able to recognise your own.
If your project is in the bottom three rows, national compute capacity is excellent news for the Kingdom and irrelevant to your architecture.
What the build-out does change for ordinary buyers
It would be wrong to conclude that none of this matters to a normal enterprise. Three things change, and they are worth naming precisely, because they are second-order rather than direct.
- SupplyHardware and expertise become locally availableAccelerators, integrators and people who have done this before are easier to source domestically than they were.
- PriceIn-Kingdom hosting becomes competitiveDomestic capacity puts downward pressure on the premium previously charged for staying in-country.
- SkillsA trained population is the durable changeCompute can be procured in eighteen months. The million-plus people trained through SAMAI is the part that cannot, and it is what makes on-premises deployment operable rather than theoretical.
Honest limits
The megawatt figures cited are published national figures which we have not independently audited, and facility capacity totals are not allocations available to any particular organisation. Verify current numbers and actual availability with SDAIA or the operator.
The relative-magnitude chart is illustrative rather than measured. Real ratios depend on model size, quantisation, batching and traffic pattern, and they move quickly as hardware and inference software improve.
What we can state precisely is our own sizing, because we deploy it: Elbi runs on hardware inside your premises, typically at single-server scale for an enterprise deployment, with indexing on the same machine and no external model call in the answer path.
Common questions
Far less than national capacity figures suggest. A retrieval assistant answering from a company’s own documents, for a few hundred concurrent users, typically runs on a single server, and indexing runs on the same machine. National figures of 480 or 500 megawatts are sized for training foundation models and serving very large models at national scale, which is a different workload by several orders of magnitude.
Because power delivery and heat removal, not floor space, are the binding constraints. A rack of general-purpose servers might draw 5 to 10 kilowatts, while a rack densely populated with AI accelerators draws several times that, often requiring liquid cooling. Power must be committed years ahead through grid connections and substations and cannot be retrofitted cheaply, which makes it the honest unit for announcing capacity.
Training produces model weights from data over weeks or months, continuously, across thousands of accelerators in one fabric — realistically done by national programmes, frontier labs and a few large enterprises. Inference uses finished weights to answer a question in milliseconds on one or a few accelerators, and is what essentially every organisation deploying AI actually does. National megawatt announcements are overwhelmingly about training capacity.
It depends on volume, and the crossover is calculable rather than arguable. Per-token pricing is proportional to use with no floor; owned hardware is a floor with near-zero marginal cost. Measure questions per day and average tokens per question on Arabic text specifically, price both paths including power, cooling and a share of an engineer amortised over three years, then compare. If you land near the crossover, cost is not your deciding factor and you should choose on residency and control instead.
Yes, but indirectly. Hardware, integrators and experienced people become easier to source domestically; domestic capacity puts downward pressure on the premium previously charged for in-Kingdom hosting; and the more than one million citizens trained through SAMAI make on-premises deployment operable rather than theoretical. Compute can be procured in eighteen months, but a trained population cannot, which makes the skills effect the durable one.
Keep reading
HUMAIN, ALLaM and the sovereign AI stack
A Saudi model served over a foreign cloud is still a foreign network call. Here is what HUMAIN provides, where ALLaM actually runs, and the six questions that separate a real residency claim from a hosting address.
Read Article
Saudi Arabia’s Year of AI 2026 explained
Saudi Arabia designated 2026 the Year of Artificial Intelligence. The compute, the capital and the training were already built. Here is which of it changes an enterprise AI decision, and what a national programme still cannot answer for you.
Read Article
How much does an AI chatbot cost in Saudi Arabia
Quotations for the same brief routinely differ by an order of magnitude, and the reason is almost never margin. It is that the word "chatbot" covers four different products. Here is how the cost is actually built.
Read Article