Inception and Jais: The Arabic LLM the UAE Built With Cerebras

Inception is G42's AI research and product subsidiary. Its most public output is Jais — an open-weight Arabic-language large language model built in partnership with Cerebras Systems and the Institute of Foundation Models at MBZUAI, the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi.

Jais launched in August 2023 with 13 billion parameters. A 30-billion-parameter version followed in November 2023. In December 2025, the team released Jais 2 — 70 billion parameters, trained on the largest Arabic-first dataset ever assembled, running on Cerebras Wafer Scale Engine infrastructure. Positioned as the sovereign alternative to running Arabic workloads through American models.

The name is a reference to Jebel Jais, the highest mountain in the UAE. The training corpus for the original release was 116 billion Arabic tokens and 279 billion English tokens. The compute ran on Condor Galaxy 1, the G42–Cerebras multi-exaFLOP supercomputer. The target audience is the 400+ million Arabic speakers worldwide, and the strategic point of the project has nothing to do with the size of that market.

What it actually is

The strategic point of Jais is not that it out-benchmarks GPT-4 in Arabic. It doesn't, and it isn't trying to. The strategic point is that it exists at all, in a format the UAE government and UAE enterprises can deploy without licensing anything from OpenAI or Anthropic. Sovereign inference. Local data residency. Arabic-first performance. Open weights the state can host on domestic infrastructure.

This is a national capability play, not a market share play. The value is optionality — the UAE now has a domestic model it can point at any workload where American licensing terms become inconvenient. That is the entire game. Every national procurement conversation from here forward includes the phrase 'or we can run Jais internally,' and that phrase is worth billions in leverage.

The Israeli mirror

Israel is running the exact same play with Hebrew. AI21 Labs has been building sovereign-grade Hebrew-capable models for years — the Jurassic and Jamba families. DICTA is building open academic Hebrew models out of Bar-Ilan University. The Israeli government is quietly funding sovereign LLM work through the Innovation Authority. The Hebrew-speaking user base is 9 million people; the strategic logic still holds, because sovereignty is not a function of market size.

The parallel is exact. Two small, technically dense, geopolitically exposed countries. Each building a domestic model in a non-English language, backed by state capital, positioned as sovereign optionality against the American stack. Neither one competing with OpenAI on the open leaderboard. Both winning the game that actually matters — the one where the state can operate its own AI without asking permission from a Silicon Valley board.

Why the pairing matters commercially

The UAE and Israel are, quietly, converging on the same technical stack for sovereign AI. Both need chips. Both need training capacity. Both need model expertise. Neither can produce all of it domestically. The commercial logic of pooling — through third-jurisdiction vehicles, through joint research, through licensed IP, through discreet compute-sharing arrangements — is obvious to everyone involved and unspeakable in public.

Jais and the Hebrew models are the visible edge of that convergence. The invisible edge is the talent, the compute contracts, the licensing deals, and the shared vendor base that will get built in the next three years, all of it moving through Cyprus, Singapore, and Switzerland to keep the paperwork clean.

The English press will notice this in 2028. The people building it are moving now.