Artificial intelligence (AI) chip makers in the United States have complained in recent years about Washington’s chip export ban against China, but they are less grumbly now as they see rising opportunities in the Middle East.
On a trip to the Middle East on May 13, US President Donald Trump was welcomed by Saudi Arabia’s Crown Prince and Prime Minister Mohammed bin Salman. Both sides agreed to boost their cooperation in AI development.
Trump said Saudi Arabia’s DataVolt plans to invest $20 billion in AI data centers and energy infrastructure in the US. At the same time, Google, DataVolt, Oracle, Salesforce, AMD and Uber are committing to invest $80 billion in cutting-edge transformative technologies in both countries.
The same day, the US Commerce Department said its Bureau of Industry and Security (BIS) will not enforce the AI diffusion rule. That rule dates from January, when the Biden administration announced US companies, such as Nvidia and AMD, were required to apply for export licenses if they ship high-end AI chips to foreign countries other than 18 US allies.
“The good news about the US is that the administration has changed,” Mohsen Moazami, President of Groq International and an Iranian-American entrepreneur, who joined Trump’s recent tour to Saudi Arabia, told Asia Times in an interview on the sidelines of a London event called AI Rush on May 16. “The administration has changed, and therefore the mind has changed.”
“What the former administration thought,” he said, “could be different from what the current administration thinks. Right now, the current administration is very keen on making American technology available to all the legitimate players on the planet to ensure that they benefit from the usage, and that America benefits from them using American technology,” Moazami said.
Commenting on the ongoing US-China chip war, he said the market would decide who would win.
He noted that Jensen Huang, Nvidia’s chief executive, “was on record 10 days ago saying 50% of the world’s AI researchers are Chinese. A country with that much wealth and human resources should be respected, as long as everybody abides by the set of rules in a fair competition.”
Moazami added: “We don’t steal from you, and you don’t steal from us. Then we can compete and let the market decide who is the winner.”
Nvidia and Groq are both headquartered in California. Nvidia makes graphic processing units (GPUs) for AI model training, while Groq makes language processing units (LPUs) for AI inference, which uses trained AI models to reason and make predictions for machine learning. Groq is unrelated to technology guru Elon Musk’s AI model Grok.
In October 2022, the Biden administration unveiled a package of export controls, forbidding Nvidia from shipping A100 and H100 AI chips to China. A year later, it also banned the exports of A800 and H800 chips to China. Nvidia then tailor-made H20 chips for the Chinese markets, but it still faced shrinking revenue and margins in China and rising challenges from Huawei Technologies’ Ascend AI chips.
In his final week at the White House in January, US President Joe Biden announced the AI diffusion rule, which would create heavy compliance costs for American AI chipmakers.
However, on May 13, the BIS rescinded the AI diffusion rule and replaced it with three guidelines to forbid companies from
- using Ascend chips,
- deploying US chips to help Chinese firms train their AI models or
- re-exporting US high-end chips to China.
The new rules shifted compliance burdens from chipmakers to the end users.
The Trump administration also helped American AI firms open new markets in the Middle East.
Also on May 13, Humain, an AI development unit of Saudi Arabia’s state-owned Public Investment Fund (PIF), announced a strategic partnership with Nvidia to build AI data centers in the kingdom with a projected capacity of up to 500 megawatts powered by several hundred thousand of Nvidia’s most advanced GPUs over the next five years.
In the first phase, Nvidia will deploy an 18,000-unit Nvidia GB300 Grace Blackwell AI supercomputer with NVIDIA InfiniBand networking to Saudi Arabia.
On May 15, Humain named Groq as its official inference chip provider. In February, Groq opened a data center in Dammam, Saudi Arabia.
“Groq is fortunate to be chosen to power the largest cluster of AI inference in the Middle East,” Moazami said. “We have an extension of GroqCloud coming out of Dammam. Saudi Arabia is now serving tokens [providing AI and machine-learning services] not only domestically but also internationally.”
“Low energy cost is the most critical factor,” he said. “At the end of the day, things not only have to be fast – they have to be cheap.”
He added that the power cost is only three US cents per kilowatt hour in Saudi Arabia, compared with more than 30 cents per kWh in the United Kingdom.
Chinese media said Huawei’s Ascend chips currently have a 30% market share in Saudi Arabia, mainly due to an AI project in the Neom planned city.
China’s AI development plan
On April 9, the US government informed Nvidia that it would need to obtain licenses to export H20 chips to China.
Huang told Reuters on May 17 that his company is evaluating how to address the Chinese market but will not downgrade the H20 further as “it’s not possible to modify Hopper anymore.”
According to a Financial Times report, Nvidia will open a research and development center in Shanghai to support Chinese clients while complying with US export rules amid growing AI chip trade.
Lin Jian, a spokesperson of the Chinese Foreign Ministry, said on May 16 that the US overstretches the concept of national security, abuses export controls and long-arm jurisdiction, and groundlessly and maliciously blocks and suppresses China’s chips and AI industry.
“China firmly opposes this and absolutely does not accept it. China urges the US to drop its protectionist acts and unilateral bullying, and stop its egregious suppression of China’s technology businesses and AI industry,” Lin said. “China will take firm measures to defend its right to development and Chinese businesses’ legitimate rights and interests.”
On Monday, Huawei published a technical paper, saying that its Ascend AI server using the DeepSeek AI model can deploy inferences faster than Nvidia’s Hopper.
Jin Lei, a columnist at QbitAI.com, said Huawei successfully used advantages in mathematical theories and algorithms to offset the physical limitations of its hardware, maximizing its servers’ performance.
At the AI Rush event, some cloud experts said combining different chips and models is the upcoming trend in the AI sector.
Christian Reilly, Cloudflare’s Field Chief Technology Officer for EMEA, said DeepSeek’s success showed the importance of using the right technology at the right time.
“As technologists and engineers, we always build for scale and peak. That is a great idea, until you can’t do it,” Reilly said. “DeepSeek engineers were constrained by their limited capacity. So they engineered this differently.”
“Originally, there was a defined way that we kept scaling out the number of parameters … but if you look at DeepSeek, it was genius.”
In late January, Wall Street investors were shocked by DeepSeek’s rise. The company used the distillation method and 2,000 H800 chips to train its AI models.
Alexandre Pereira, co-founder and chief executive of 2501.ai, said the AI sector has made enough GPUs to sustain itself for several years, and it’s time for AI companies to consider what types of GPUs and language models they need instead of choosing the most powerful ones.
Read: US uses ‘poison pills’ to isolate China from supply chains
