AI Under Fire: Middle East AI, Sovereign Models, and the Future of Voice

This is the third episode of Stolen Chat. Wesley had just flown in from Abu Dhabi. A month earlier, he'd lived through Iran's missile strikes on the UAE — air raid sirens every thirty minutes, the sound of intercepted missiles landing like earthquakes. He does voice AI research at MBZUAI, the Middle East's dedicated AI university, and is incubating an Arabic-first Voice AI project. After this conversation, my understanding of Middle East AI was completely rewritten.

What it's like to work under missile fire

On February 28th, a missile alert woke Wesley from his afternoon nap. The first week was the most intense: triple-digit drones and double-digit missiles daily. His apartment was 10 kilometers from a military base in a straight line. Interceptor debris fell at random — the Burj Al Arab caught fire and made the news, but residential areas were hit too. The closest debris landed two kilometers away, at the university.

He and a few classmates drove east to the Oman border that night. The second week, he went back to Abu Dhabi and returned to work. The school was humane — hybrid work, immediate survey to check on everyone.

What surprised him: business barely stopped. During the second week of the war, he had a meeting with BCG. He emailed asking whether to postpone. Their reply: no need, proceed as normal.

"Short-term, operations keep running. Long-term, what gets hit is confidence — and confidence affects the flow of capital and talent."

What the Middle East's AI university actually is

MBZUAI — Mohammed bin Zayed University of Artificial Intelligence — is named after the UAE's president. Founded in 2020, first students arrived in 2021 due to COVID.

The president is Eric Xing, former head of CMU's Machine Learning Department. Most early faculty came from CMU. Yann LeCun has visited; Sam Altman came last summer and received an honorary degree.

Growth has been staggering: Wesley started with 3 departments, graduated with 7 or 8. Faculty doubled in two years. Undergrad admissions have begun.

"AI investment in the Middle East is concentrated in the UAE and Saudi Arabia, proportional to economic power. They've shown genuine commitment to the tech transformation."

Why build Arabic Voice AI

Wesley's project, AudarAI, builds multilingual Voice AI centered on Arabic — speech-to-text, text-to-speech, Voice Agents.

Why not just use OpenAI or ElevenLabs? Because the differences between Arabic dialects aren't like British vs. American English. They're like Mandarin vs. Cantonese — or even bigger. North African dialects diverge even more dramatically.

Major model providers only handle Modern Standard Arabic (MSA), which is literary language — nobody actually speaks it daily. It's like building an AI that only speaks classical Chinese. The emotional delivery might be perfect, but locals will immediately feel it's fake.

"Emotionally expressive speech can simultaneously have terrible accent and pronunciation. The two aren't mutually exclusive."

AudarAI's strategy: perfect UAE dialect first, then Saudi, then expand to GCC countries, then mainstream languages. Niche market, expanding outward step by step.

Voice is the next era's AI interface

This is the single most memorable judgment from the episode.

Wesley believes voice will become the universal interface for human-AI interaction in the next era. The logic is simple: over 70-80% of human-to-human information transfer happens through voice — online chats, conference calls, in-person coffee chats. If AI is going to work closely with humans, the way humans communicate with AI must mirror the way humans communicate with each other.

"You can't be having a conversation with someone, then stop to transcribe everything into text and upload it for AI processing. The efficiency gap is too big. We need AI with native audio reasoning and audio generation capabilities."

The current problem: voice modality lags text-based LLMs by one to two years. The biggest technical bottleneck is latency. Two approaches: end-to-end (voice to voice) — low latency but a black box, hard to customize; or cascaded pipeline (speech → text → LM → speech) — can build agents and industry customization, but the LM inference step creates a latency bottleneck.

"You can tolerate a model thinking for minutes. But in face-to-face conversation, more than two seconds of silence already feels abnormal."

Sovereign AI is a real need in the Middle East

I asked Wesley: is sovereign AI in the Middle East a genuine need or a political slogan?

Real need. He gave a vivid example: in UAE supermarkets, locally produced products carry the label "Proudly Made in UAE."

The logic: if you use third-party models, you can't guarantee they conform to local culture, social customs, and religious factors. Chinese and American models won't adapt to Arab culture and Islam.

The deeper concern is education. If we believe the next generation is AI-native, and AI plays an increasingly important role in their education — if you don't control the AI in that process, how do you ensure citizens raised under that education still identify with your nation?

"Rather than imposing all kinds of restrictions and regulations, it's better to have the capability to train your own local models."

G42: more like a Chinese state enterprise than Silicon Valley

G42 is one of the UAE's most important tech conglomerates, deeply affected by US-China tensions — previously extensive cooperation with Huawei, later restricted by the US.

Wesley's observation: G42 resembles a large Chinese state-owned enterprise more than Silicon Valley. It has solid subsidiaries (Inception, Precise), combining internal R&D with government and local enterprise service contracts.

"American influence is still stronger. You need compute, you need NVIDIA GPUs, so you play by their rules."

It's essentially a B2B service contractor: regardless of whether the technology comes from America or China, it deploys locally and serves local clients.

Where the opportunities are in Middle East AI

Context-dependent, but several clear directions: Energy is the undisputed top priority — ADNOC has its own AI division, partnering with G42 to form AIQ for oil exploration, reserve estimation, and process optimization. Fintech, especially in Dubai. Healthcare, with G42's M42 subsidiary investing heavily.

"Different countries focus on different priorities. China focuses on smart manufacturing and embodied intelligence because manufacturing is strongest. The Middle East doesn't have strong manufacturing — it relies on imports and re-exports — so the focus is energy, finance, and health."

For external founders: if you have energy industry background plus AI capabilities, the Middle East might be one of the best markets globally.