Transcription
Hi, welcome back to Be Rich. Today I want to talk about what's happening in AI. AI chip demand is still red hot, and 2026 is shaping up to be another blockbuster year for the semiconductor ecosystem, but the nature of the race is changing.
The first phase of AI was about training models. That phase was capital intensive but relatively concentrated, and it strongly favored Nvidia. The next phase is about inference, meaning serving answers cheaply, quickly, and at scale. This phase is far more fragmented, more competitive, and far more constrained by memory, power, and infrastructure. That shift creates new winners, exposes new bottlenecks, and raises the uncomfortable question about how sustainable the financing behind this entire buildout really is.
Nvidia still dominates, mind you, but the battlefield is shifting. Nvidia remains the primary pick-and-shovel supplier, with revenues more than doubling year-over-year and massive demand for its H200 and B200 GPUs. Goldman Sachs estimates that Nvidia alone could sell $383 billion worth of hardware by 2026. But the motor is no longer static. Competition is intensifying, especially in inference, where performance per dollar matters more than raw power. And we have seen this happening in China, and we have seen this happening even in the US. Hyperscalers are quietly becoming competitors. They're increasingly chip designers themselves. Alphabet is scaling its TPUs. Amazon is pushing Tranium and Inferentia, and Microsoft is doubling its data center footprint over the next two years. This matters because internal chips reduce long-term dependency on Nvidia and the pressure on industry margins.
Inference changes everything. Inference workloads are more memory-bound and compute-bound. That shifts the constraints away from GPUs and towards high-bandwidth memory and system design. This is why memory suppliers are suddenly a critical choking point, and we have seen this happening with memory prices in the market. I'm sure a lot of you have seen it on Twitter. The memes of what the memory price was six months ago to what it is today, and the memory is a real bottleneck. High-bandwidth memory is in acute shortage, and suppliers are pricing this accordingly. Micron Technologies, Samsung, SK Hynix are benefiting massively, with Micron explicitly admitting it cannot meet customer demands, and the shortages will persist. This is not a short-cycle problem. Capacity expansion takes years. This is even true with, you know, metals such as copper, silver, gold. So, the supply chain is being physically hit by limits.
The AI boom is colliding with old-world constraints: power availability, electrical transformers, gas turbines, clean rooms, fab construction timelines. This is not software; it's physics, logistics, and energy policy. And all of that has to come together to make a huge difference. So, in the middle of all this, financing risk is no longer theoretical. A large chunk of demand is indirectly tied to OpenAI, which has a multi-billion dollar computing commitment with Microsoft, Amazon, and of course, Oracle. Analysts are openly saying that if OpenAI does not raise eyewatering amounts of capital soon, markets may start questioning whether 2026 represents peak demand rather than a stepping stone. Valuations assume perfection. Markets have been conditioned to expect relentless quarter-after-quarter growth. Even minor slowdowns are triggering violent sell-offs, as we recently saw across AI stocks. This is not mania yet, but expectations are brittle, and there are some blind spots and risks investors may be underestimating.
Inference commoditizes faster than training. Inference favors cost, efficiency, and specialization that invites competition and margin compression. Nvidia's dominance in training does not automatically transfer one-to-one into inference. Internal chips are structurally deflationary. Custom chips are designed by hyperscalers and do not need to be profit-maximizing the same way merchant silicon does. Their goal is cost control, not external margins. This caps long-term pricing power for independent chip designers.
Capital markets are the real fuel. The entire buildout rests on abundant, patient capital. If rates stay higher for longer, or if mega funding rounds slow, infrastructure demands can fall sharply without warning. This could spook markets earlier in the fall. And the risks have not disappeared. Memory profits invite overcapacity later. Today's shortages are boosting memory margins, but history suggests that aggressive capex cycles are often overshot. When supply catches up, pricing power can evaporate very quickly. This cycle has not been repealed; it's just been delayed.
Energy constraints do not scale linearly. Compute demand can grow exponentially. Power generation grid upgrades cannot. This mismatch is underappreciated and could cap AI development in unexpected ways. The problem is, markets tend to call peaks before they arrive. Even if 2026 turns out to be strong, the fear that it is the peak could dominate price action well before the data turns.
So, what is the bottom line? What should you keep in mind? This is not a classic bubble, or not yet. Demand is real, revenues are real, and use cases are expanding. But the AI boom is moving from a clean, GPU-driven story to a messy, infrastructure-heavy, and capital-constrained one. The transition favors selective winners, punishing complacency and making valuation discipline matter again.
So, this is what I wanted to share with you today, based on what we spoke about yesterday. This further expands what to look forward to into year 2026. So, companies like Google, Amazon, Microsoft, Nvidia, and of course, all our memory chip makers like Micron, Samsung, SK Hynix, all of them are in play. Do pay attention to what's happening. Try and understand what is going on, and like I said, there is a shift happening from what was AI to what AI is today and where the market is driving it towards. We will finally have a world where AI is ubiquitous and it's everywhere, and everyone will be using it. But which will be the companies that will be the standout winners is still very foggy and hard to see.
So, let me know what you think about 2026. Which companies or chip manufacturers or memory makers you think are going to be the winners of 2026? I'm curious to know. Thanks for being with me on Be Rich today. I wish you a wonderful day. See you again tomorrow on the last day of the year. Bye.