Liwa / Insights / Introducing DesertGrid

From kilowatts to tokens: introducing DesertGrid

Liwa Insights·26 May 2026·8 min read

Every piece on this blog circles the same claim: in AI, cheap power is the only durable moat. It is easy to assert and harder to prove. So we ran the experiment. If you genuinely had the cheapest reliable power in the region, what is the first thing you would build on it? Our answer has a name now. It is called DesertGrid, and it turns Liwa kilowatts into tokens.

What DesertGrid is

DesertGrid is a Gulf-based AI API. You point any OpenAI-compatible client at it, swap your key, and call efficient flash-class language models at prices built on a low power base, from about $0.15 per million tokens blended. It is privacy-first, PDPL-native, has no hard rate limits, and offers white-label options. Later it opens into a marketplace where anyone can list and monetize their own hosted model on the same cheap-power infrastructure (desertgrid.ai).

The models it serves are real and already cheap: DeepSeek V4 Flash runs near $0.14 per million input tokens, Qwen3 Turbo near $0.04, because mixture-of-experts designs fire only a fraction of their parameters per token (DeepSeek, CloudZero).

$0.15Blended / 1M tokens

$0.10/kWh Liwa power underneath

OpenAICompatible, drop-in

Why an API is what cheap power builds first

Walk the logic. The cost of running a model is dominated by electricity, paid on every token, forever. So the provider with the cheapest power has the lowest cost floor, and in a market where token prices keep falling, the lowest cost floor is the last one standing. An API business is simply the most direct way to convert a power advantage into revenue: every query is a kilowatt-hour wearing a price tag.

There is a second tailwind. As inference gets cheaper, total demand for it explodes rather than shrinks, the Jevons pattern we keep returning to. Cheaper tokens invite a hundred new uses, and all of them have to run somewhere, continuously, sensitive to every cent. The cheapest place to run them wins the volume (Stanford HAI, AI Index).

If cheap power is the moat, an API is the product that monetizes it most directly, one kilowatt-hour, one token, at a time. So the question for anyone sitting on cheap power, or thinking about reserving some, is this: are you going to sell the kilowatts wholesale, or turn them into tokens and capture the margin yourself?

The same thesis, two layers

Liwa and DesertGrid are the same bet expressed at two layers of the stack. Liwa is the infrastructure: secured power at $0.10/kWh, a liquid-cooled hall rated to 150 kW/rack, white-label, in a UAE free zone. DesertGrid is a tenant on that infrastructure, the proof that the power advantage flows all the way up into a product the market can buy with two lines of code.

That is also the invitation. You do not have to choose between them. Reserve space on Liwa and build your own DesertGrid. Or use DesertGrid as a customer and let the power advantage reach you as a cheaper token. The grid underneath is the same either way.

Where this meets Liwa

DesertGrid runs on the exact economics Liwa sells: power secured at $0.10/kWh and a liquid-cooled hall rated to 150 kW/rack, in a UAE free zone. If you want the infrastructure, reserve space and run your own brand on it. If you just want the cheap tokens, DesertGrid is live in private beta at desertgrid.ai. Same grid, two doors in.

Questions we're sitting with

If the power advantage flows all the way up into the token price, who captures the most margin, the landlord or the operator?
When tokens keep getting cheaper, is the cheapest place to run inference the only safe place to stand?
Would you rather sell your kilowatts wholesale, or turn them into tokens and sell those?

Own the grid, or use it.

Reserve liquid-cooled, 150 kW-ready capacity at $0.10/kWh under your own brand, or try the cheap tokens DesertGrid builds on top of it.

Reserve space →Visit DesertGrid

Sources

DesertGrid is a Segments venture in private beta; model availability and pricing track the open-model market and may change.