Silicon, Sovereignty, and the $1 Trillion Compute Race

'The point of a company like ours is to make an insane amount of compute available. Compute is no longer a component. It is the new oil'

• Jensen Huang, Founder & CEO, NVIDIA

❓ What You'll Learn

• Why the market for AI chips is heading toward $418.8 billion by 2030.
• How one factory step in Taiwan, called advanced packaging, sets the speed limit for all AI progress.
• Why are the biggest cloud companies spending over $725 billion in 2026 to build their own chips and stop paying Nvidia.
• Why the US and China are fighting over machines built by a single Dutch company.
• How AI data centers are quietly bringing nuclear power back to life.

💎 Why It Matters

Spending on AI infrastructure is projected to pass $3.5 trillion by 2030.

Nvidia alone made $193.7 billion from data centers in FY2026, which now makes one chip company the backbone of the global tech economy.

Power has shifted. Countries used to fight over borders and oil. Now, they fight over who controls the chips and the supercomputers running them.

🔍 Problem

A few companies own everything. Nvidia holds roughly 90% of the AI chip market, and TSMC is years ahead of everyone in the packaging step that finishes those chips.

Export controls are splitting the world in two. New restrictions in January 2026 are pushing the US and China into completely separate technology stacks.

Furthermore, AI is hitting a wall that money cannot bypass. Data centers are expected to pull 1,300 TWh of electricity by 2035, more than most grids were ever built to handle.

💡 Solution

Free software tools like OpenAI's Triton and AMD's ROCm now let AI models run on different brands of chips, which dismantles Nvidia's biggest advantage.

The Ultra Ethernet Consortium is building an open way to connect thousands of chips, replacing Nvidia's proprietary networking.

New startups are trying stranger ideas. Chips the size of a whole dinner plate, chips built only for speed, and open designs anyone can license. All of it is aimed at making AI cheaper to run.

🏁 Players

The Core Ecosystem & Challengers

• Nvidia • Still the king, with its H200 and upcoming Vera Rubin chips, and a 75% gross margin almost no hardware company has ever achieved.
• TSMC • Builds nearly every advanced AI chip on Earth. It just raised its Arizona investment to $265 billion to move some of that work onto American soil.
• Broadcom and Marvell • The quiet winners. They design the custom chips for the cloud giants and posted record AI revenues of $10.8 billion and $2.4 billion in early 2026.
• ASML • The most important company most people have never heard of. It is the only maker of the machines that print advanced chips, with 65 shipments planned for 2026.
• Tenstorrent • Run by legendary chip designer Jim Keller, now shipping open-design servers that hit 308 tokens per second on DeepSeek models.
• Intel Foundry • Attempting the comeback of the decade with its 18A manufacturing process, reportedly reaching 85% yields and winning back its own chip production.
• Huawei and SMIC • China's answer to the ban. Their Ascend 910C chip has pushed production yields to 40% despite having no access to Western equipment.
• Cerebras Systems • Valued at $23 billion for one strange idea, building a single chip the size of a dinner plate so data never has to travel between chips at all.

🔮 Predictions by Aybars D.

• Custom chips will trap companies, not free them • Google, Amazon, and Microsoft are building their own silicon to escape Nvidia. But once your AI runs only on Amazon's chips, you have not escaped anything. You have just changed landlords.
• Intel becomes a national utility • With improving yields and the US government now holding equity, Intel is ceasing to be a normal company. It becomes too important to fail.
• Sovereign AI is the new space race • National AI infrastructure will grow into a $148 billion market by 2032, led by Gulf states that refuse to let their data leave their borders.

☁️ Opportunities

• Advanced packaging. TSMC's capacity for this final assembly step is fully booked through 2027. Anyone with spare capacity can name their price.
• Open networking hardware. Switches built for the new open standard can take billions in market share from Nvidia's proprietary system.
• Power for compute. Microsoft restarted Three Mile Island and Amazon bought into a nuclear plant. Small reactors and uranium are now AI investments in disguise.

🏔️ Risks

• The memory bottleneck • A new generation of AI memory arrives in late 2026. If SK Hynix or Samsung stumbles, global server shipments could halt overnight.
• One company holds the whole roadmap • Everything depends on ASML's supply chain in the Netherlands. One serious disruption there freezes TSMC, Intel, and Samsung at the same time.
• Rules as weapons • Proposals like the US AI Overwatch Act would let Congress cancel export licenses after they are granted, making long-term planning nearly impossible.

🔑 Key Lessons

• Software decides the winner, not silicon. Free tools now let AMD chips run standard AI workloads at 90 to 95% of Nvidia's speed without rewriting code. That is how monopolies actually crack.
• Electricity is the real ceiling. The limit in 2027 is not how many chips can be manufactured; it is how many gigawatts the winners can lock down.

🔥 Hot Takes

• China's Nvidia ban is protectionism, not patriotism. Blocking chips that its own companies legally bought is not revenge on Washington. It is a forced subsidy to guarantee a market for Huawei's weaker products.
• Most 'sovereign AI' is theater. A data center in the Gulf running American chips and American software is not independent. It is American technology with a local address.

😠 Haters

'AMD will never catch CUDA.'

They do not need to win everywhere. AMD's software now reaches 90 to 95% of Nvidia's performance on the workloads most companies actually run. Good enough, at a lower price, is how moats fall.

'Intel Foundry is finished'

Tell that to their 18A process. Yields have reportedly hit 85%, ahead of Samsung, and real chips are shipping. The comeback is already happening.

🔗 Links

• Nvidia FY2026 Earnings • Data center revenue reaches $193.7 billion.
• TSMC's US Mega-Expansion • Arizona investment lifted to $265 billion.
• Broadcom and Marvell • Record AI chip revenues in fiscal 2026.
• Huawei Ascend 910C • Chinese production yields climb to 40% under sanctions.
• Tenstorrent's Galaxy Blackhole • Open-design processors expanding fast inference globally.
• Constellation and Microsoft • Restarting Three Mile Island to power data centers.
• Ultra Ethernet Consortium • The open standard challenging Nvidia's networking.

See you in the next one. To be continued..

Written by Aybars Dorman | July 19, 2026

#AI #Technology #Innovation #Semiconductors #ArtificialIntelligence #DeepTech #Geopolitics #SupplyChain #AIChips #Nvidia #TSMC #ASML #SovereignAI #ChipWars #ExportControls #DataCenters #ComputeInfrastructure #DormanReview