There is a resource at the centre of every AI breakthrough, every chatbot response, every autonomous agent running in the background of a business. It is not data. It is not talent. It is silicon specifically, the specialized chips that power AI workloads at scale.
The world does not have enough of them. And that scarcity is reshaping the global economy faster than most people realise.
The new infrastructure race
When OpenAI closed its $122 billion funding round in March, the capital was not earmarked for research alone. A significant portion is going toward compute chips, data centres, and the infrastructure required to train and run increasingly powerful models. Amazon committed $50 billion. Nvidia put in $30 billion. These are not passive investments. They are bets on who controls the compute layer.
The same pattern plays out everywhere. Anthropic, xAI, Meta, Google, and Microsoft are all building or leasing data centre capacity at a pace that strains global chip supply. Nvidia’s H100 and H200 GPUs have had wait times measured in months. Blackwell chips are already oversubscribed before they reach full production. The demand is structural it does not stop between product cycles.
Why chips are not like other tech
Software can be copied instantly. Data can be replicated. Chips cannot. Fabricating a leading-edge AI accelerator requires a supply chain spanning dozens of countries, years of R&D, and manufacturing plants that cost tens of billions to build. TSMC in Taiwan handles the majority of the world’s most advanced chip production. A single disruption geopolitical, natural, or logistical would send shockwaves through every AI company on earth simultaneously.
That concentration is the hidden fragility underneath the AI boom. The models get the headlines. The chips get the funding.
The race to build alternatives
Nvidia currently holds an estimated 80% of the AI accelerator market. That dominance is why every major tech company is now investing in custom silicon. Google has its TPUs. Amazon built Trainium and Inferentia. Microsoft is developing Maia. Meta revealed four generations of its MTIA chip family this month. Apple, Qualcomm, and MediaTek are optimising for on-device inference.
The startup layer is moving just as fast. Cerebras built a wafer-scale chip designed to eliminate the bottlenecks of traditional GPU clusters. Rebellions in South Korea raised $400 million this year ahead of an IPO. Groq has built chips purpose-built for inference speed. Each of these bets is a wager that the compute market is too important and too concentrated to leave to one supplier.
The geopolitics of compute
The US government has already moved to restrict advanced chip exports to China. In response, Chinese labs are racing to develop domestic alternatives. Huawei’s Ascend chips are powering DeepSeek and other Chinese frontier models despite being cut off from Nvidia hardware. What began as a trade restriction has accelerated China’s domestic chip industry by years.
Every major economy now treats AI compute as strategic infrastructure. The EU is funding sovereign AI initiatives. India is building data centre capacity. Japan signed a $10 billion Microsoft partnership this month partly to ensure domestic AI processing stays within national borders. The chip is the new oil well and nations are drawing borders around it.
What comes next
The demand is not slowing. Every new AI capability longer context windows, multimodal reasoning, agentic workflows requires more compute, not less. The models are getting larger. The use cases are multiplying. The infrastructure required to support them is years behind the ambition.
The companies that secure reliable compute access will set the pace of AI development. The ones that cannot will be left optimising for efficiency while their rivals scale. That is the real race underneath the AI race and most of the world is only just beginning to understand it.
Previous
