Sriram Krishnan (White House AI Adviser) breaks down the US-UAE AI acceleration partnership and American AI dominance strategy
May 19, 2025 with Sriram Krishnan
Key Points
- The US-UAE AI partnership requires Gulf nations to match every megawatt of data center capacity built abroad with equivalent US infrastructure investment, channeling foreign capital into American AI buildout.
- American companies must operate and maintain GPUs deployed under the deal, with physical verification and remote-access controls blocking access from countries of concern, locking allies into the US AI stack.
- The White House is scrapping Biden-era GPU export restrictions and plans a formal AI Action Plan within six months covering infrastructure promotion and competitive protection against China.
Summary
Sriram Krishnan, Senior Policy Adviser for Artificial Intelligence at the White House, lays out the strategic logic behind the US-UAE AI acceleration partnership signed during President Trump's first Middle East visit of the new administration.
Krishnan came to the role after a career spanning large tech companies and Andreessen Horowitz. He joined following a conversation with AI and Crypto Czar David Sacks, motivated by the view that the previous administration was taking American AI policy in the wrong direction. The mandate he's been working under since day one is explicit: American AI dominance.
The deal structure
The partnership has three components. The first, and the one Krishnan says got lost in press coverage, is that Gulf nations investing in AI infrastructure abroad must make matching investments in US data centers. If a country builds X megawatts of capacity in the UAE, X megawatts of equivalent capacity gets built in the United States. That's net-new capital, not a repackaging of existing commitments.
The second component is operational control. The vast majority of GPUs deployed in the UAE under this deal will be run, hosted, and maintained by American companies. Krishnan frames this as a direct market-share gain for US hyperscalers against non-American competition.
The third is security. Every GPU shipped carries physical verification protocols making diversion impractical, combined with remote-access controls ensuring that entities from countries of concern cannot reach the hardware.
The strategic frame
Krishnan's analogy is the Microsoft Windows and Office ecosystem: lock in the stack early, and switching costs compound over years and decades. He argues that once a country has spent heavily on American hardware, optimized models to run on that hardware, and built applications tuned to those models, the practical cost of switching to a Chinese alternative becomes prohibitive — even if an API theoretically exists. The goal is to replicate that dynamic at a geopolitical scale, getting resource-rich allies permanently tied to American AI infrastructure before Chinese alternatives mature.
The 5G/Huawei parallel comes up explicitly. Krishnan says the difference this time is that America starts in the lead — Nvidia, AMD, and the frontier model labs are all domestic — but acknowledges the lead is narrower than it looked before DeepSeek. He describes R1 as having been, for a period, the only competitive reasoning model outside of OpenAI's o1, which he treats as a genuine wake-up call about the margin the US actually has.
The Biden-era diffusion rule, which divided the world into GPU-haves and GPU-have-nots and effectively shut allies out of American AI infrastructure, has been scrapped. The UAE deal is described as a template for similar partnerships to follow.
Who qualifies
On whether neo-clouds like CoreWeave, Crusoe, and Lambda can compete with hyperscalers for these contracts, Krishnan says the primary criteria from Commerce Secretary Howard Lutnik is that companies are American and can sign on to the security protocols. Scale is secondary. Any American company that clears that bar should be eligible.
What the region actually wants
Demand from Gulf nations is concentrated on inference at scale, not foundation model training. The use cases are healthcare, education, predictive infrastructure management, and hydrocarbon processing optimization. Krishnan describes the conversation as roughly 80-90% large language models with some image models layered in for industrial applications. Robotics is generating interest but is not yet a primary commercial conversation. Agentic applications are early-stage, with some prototypes visible but nothing at deployment scale.
The domestic plan
A formal American AI Action Plan is due within the first six months of the administration, which Krishnan estimates means roughly four to six weeks out. It will cover two tracks: a promote side addressing energy, open source, and regulatory runway for US companies across the full stack from semiconductor tooling to agents and robotics; and a protect side focused on counterintelligence and maintaining the competitive lead against China.
On open-source models specifically, Krishnan says he wants an American open-source model that decisively outcompetes alternatives built on PRC values and censorship constraints. He names Llama and a forthcoming model from Sam Altman as candidates he's rooting for, and notes that Polymarket currently prices an OpenAI open-source release in 2025 at 85%.
The core bet is simple: get allies locked into the American AI stack now, fund US infrastructure build-out with their capital, and let compounding switching costs do the strategic work over the following decade.