The AI world is buzzing, and it’s not just about bigger models. Last week, news broke: Anthropic is talking with Samsung to build its own AI chips. Days later, DeepSeek announced it’s doing the same. What gives? For ambitious founders, this isn't just a tech headline; it’s a masterclass in strategic vertical integration—and a stark debate over what drives the most audacious moves in tech.

Key Takeaways

  • Large language model (LLM) providers like Anthropic and DeepSeek are actively moving to design or build their own custom AI chips.
  • One core argument for this shift is a strategic imperative: to "own the compute." This offers greater control, security, and reduces dependency on external suppliers like Nvidia.
  • Another major driver is the potential for significant efficiency gains. Optimizing silicon specifically for their unique models could deliver far better performance than general-purpose GPUs.
  • A counter-argument suggests this isn't about unmet technical needs; major chip manufacturers like Nvidia would readily customize chips for high-volume customers. Instead, it’s a play to capture the high margins currently held by chipmakers.
  • Founders should evaluate their own "compute" — the core, irreplaceable component of their business — and question when outsourcing creates unacceptable risk versus when vertical integration is simply chasing margin.

The Disagreement

When LLM providers start designing their own silicon, it forces a question: Are they solving an unmet technical need, or are they trying to capture value?

Podcast guest Rory lays out the case for strategic necessity. He highlights two key arguments that changed his own initial skepticism. First, there's the existential need to control the underlying infrastructure. Rory states, “You got to own the compute if you don't own the compute you're screwed.” In a world where access to cutting-edge GPUs can make or break an AI company, relying entirely on an external supplier, no matter how dominant, presents a strategic vulnerability. Think of it like a car company needing to secure its engine supply.

The second argument for owning the silicon centers on optimization. General-purpose GPUs, while powerful, aren't perfectly tuned for every LLM architecture. Rory explains, “If you build your own silicon, you can optimize the silicon for your model and probably get way you significantly more efficient than you might do buying a general purpose computing platform from Nvidia and adopting it to your specific model.” This isn't just about speed; it's about energy efficiency and cost reduction at massive scale, which could translate into a competitive edge in model training and inference.

But Jason, the co-host, calls foul on the "specialized needs" argument. Having worked in the semiconductor industry, he knows how the game is played. Jason argues that if a customer like Anthropic or DeepSeek is driving enough volume to Nvidia, the chip giant would absolutely build custom silicon for them. He says, "The only thing that makes zero sense to me is the argument that hey, open a we need to build our own chips because we have very specialized needs that Nvidia can't meet. Yes, I I have a little bit of experience in the semiconductor industry. If if you're driving that much volume to them and you need a special version of a chip, they'll build it for you." His conclusion: the real driver isn't a technical shortfall from Nvidia, but a desire to take a bigger slice of the profit pie. "This is just responding to believing that the margins are so high. Nvidia to survive, we have to recapture that margin. I just think the idea that it's customized for us is just soft language."

Who's Right (and When They're Wrong)

Both Rory and Jason hit on crucial truths. Rory is right that owning critical infrastructure provides unparalleled strategic control and the potential for deep, specialized optimization. For companies betting their entire existence on AI performance and availability, a certain degree of vertical integration makes sense. It’s a long-term play for strategic independence and efficiency.

However, Jason is likely more correct about the immediate, economic motivation. Nvidia's margins on AI chips are astronomical. It’s hard to ignore the allure of capturing a portion of that value, especially when LLM companies are burning billions on compute. The "customized for us" argument might be a convenient narrative, but the underlying ambition to control a larger share of the value chain is real. It’s not just about needing a custom chip; it’s about needing a custom chip that they profit from building, rather than Nvidia.

The truth probably sits squarely in the middle: these companies are pursuing vertical integration for both strategic control and margin capture. The strategic necessity provides the justification to investors for such a capital-intensive, high-risk endeavor, while the promise of higher margins fuels the ambition.

What to Do With This

Founders in their 20s and 30s should stop and define their own version of "the compute." What's the single most critical, expensive, or bottlenecked component in your business that, if outsourced, leaves you vulnerable? Then, evaluate that component: could bringing it in-house or designing a bespoke solution give you a strategic moat and materially improve your economics? Don't fall for "soft language" about specialized needs if the real driver is just chasing margins. Instead, calculate the true cost of outsourcing vs. the investment and long-term gain of owning that critical piece, both for control and potential efficiency.