Good morning. Just a few years ago, countries competed for oil.

Now they're competing for GPUs.

South Korea spent the weekend announcing new AI infrastructure partnerships with NVIDIA, while government officials signaled they want priority access to NVIDIA's upcoming Vera Rubin systems. At the same time, SK Hynix signed a multi-year agreement with NVIDIA to develop next-generation memory for AI data centers. The message is becoming harder to ignore: AI compute is starting to look like strategic infrastructure.

It remains to be seen whether governments can build AI factories as quickly as they announce them.

In today's newsletter:

Markets move. Headlines catastrophise. But somewhere inside the noise is the story that matters — the opportunity, not the fear. 

The Daily Upside was built by Wall Street insiders to find it — global business and finance, reported without the alarm.

Join 1M readers →

NVIDIA’s next bottleneck may not be GPUs. It may be everything around them.

The big headline this week was NVIDIA’s South Korea deal spree. But the quieter investor signal is more important: NVIDIA is locking down the infrastructure layer around AI before it becomes a bigger constraint.

NVIDIA and SK Hynix announced a multi-year partnership to develop advanced memory for AI data centers. That matters because high-bandwidth memory is one of the most important ingredients in modern AI systems. GPUs get the attention, but memory helps determine how fast those GPUs can actually move data. If HBM supply gets tight, NVIDIA’s ability to ship next-generation systems could be limited even if GPU demand remains massive.

At the same time, South Korea said it wants priority supply of NVIDIA’s upcoming Vera Rubin GPUs, a sign that governments are now treating AI compute like strategic infrastructure. That shifts NVIDIA from “chip supplier” to geopolitical asset.

The other overlooked piece is power. AI data centers are running into grid and cooling constraints, and NVIDIA has already said a meaningful chunk of data center power is lost before it reaches AI workloads. That means the winners may be the companies that solve full-stack efficiency, not just raw chip performance.

Green Chip Take: The next phase of NVIDIA’s growth may be less about selling more GPUs and more about controlling the whole AI factory stack: chips, memory, networking, power, cooling, and partners.

Today’s overlooked AI economy story is Google reportedly ordering more than 3 million tensor processing units from Intel for 2028 delivery. On the surface, this is an Intel manufacturing win. For NVIDIA investors, it is really a reminder that the largest AI customers do not want to depend on one supplier forever.

Google has been building its own TPUs for years, but this report matters because of scale and timing. A multi-million-unit order suggests Google wants more control over cost, supply, and performance for its AI workloads. If Google can manufacture more of its own AI chips, it may reduce some reliance on NVIDIA GPUs inside its own cloud infrastructure.

That does not mean NVIDIA demand suddenly falls. Training frontier models, serving enterprise customers, and supporting broad AI workloads still require massive compute capacity. NVIDIA also benefits from software, networking, developer adoption, and full-system design advantages that custom silicon does not automatically replace. NVIDIA’s own product portfolio continues expanding from GPUs into full AI factory systems.

Green Chip Take: This is a mild long-term risk to NVIDIA, but not an immediate demand problem. The smartest read is that AI demand is so large that customers are using both NVIDIA GPUs and custom silicon. The pie is growing, but the competitive map is getting more complicated.

For the past two years, investors have worried that NVIDIA couldn't build enough chips to satisfy AI demand. Jensen's latest comments suggest the opposite problem may soon matter more: demand continues to outpace even NVIDIA's rapidly expanding supply chain.

The key takeaway isn't that supply constraints are gone. It's that NVIDIA believes it has secured enough capacity to support another phase of aggressive growth across both GPUs and CPUs. Jensen also emphasized that upcoming products like Vera Rubin and NVIDIA's expanding AI ecosystem are positioned for large-scale deployment.

📄 NVIDIA's latest Blackwell Ultra announcement reveals where the company's next generation of AI infrastructure is headed.

☁️ Meta is reportedly exploring AI superclusters that could consume gigawatts of power, a reminder that AI infrastructure spending remains enormous.

⚔️ AMD says several hyperscalers are expanding deployments of its MI350 platform as it pushes for more AI market share.

🏭 TSMC expects AI-related revenue to double again in 2026, reinforcing strong demand across the semiconductor ecosystem.

🎤 Watch Jensen Huang's Computex keynote where he lays out NVIDIA's vision for AI factories and sovereign AI.

📊 This chart shows how NVIDIA's data center revenue has grown from a side business into the company's dominant profit engine.

🕳️ Why high-bandwidth memory (HBM) may be the most important component in AI infrastructure that investors aren't talking about.