NVIDIA’s Road To $10T

This piece for paid subscribers is focused on the impact the Groq acquisition could have for NVIDIA. As a company, NVIDIA has been better at M&A than most. They have the best view of any company in the world on what is happening in the AI ecosystem, and so their decisions should be analyzed through that lens. As a reminder, we are AI people not financial people so please do your own research before investing. Disclosure: We are long NVIDIA and have been for years.

NVIDIA has come a long way from being a specialized gaming hardware vendor to the high-priest of the generative era. Its GPU technology is absolutely central to the first act of the AI revolution: The Great Build. For three years, the world has been obsessed with the “Training” phase—a brute-force arms race to forge intelligence inside massive, power-hungry clusters. This, along with the company’s perfectly timed “triple threat” of hardware dominance (through its GPU technology), software locking (through the CUDA platform), and a complete monopoly on the infrastructure of the AI era, has led to NVIDIA surpassing the $4 trillion market cap.

While this massive growth past the $4 trillion mark has been fueled by the capital-intensive “construction” of AI, investors are now understandably questioning how much more growth potential remains in a world where the biggest models have already been “built.”

An assessment of the company’s recent strategic activity seems to suggest that Nvidia is clearly looking to progress beyond supplying the picks and shovels, and may already be fixated on the next leg of value creation.

The $20B acquisition of Groq—a company defined by its “unhinged” but architecturally pure focus on deterministic execution—signals that Jensen Huang is no longer interested in just winning the training war. He is moving to monopolize the marginal cost of the token.

In this newsletter, we explore Nvidia’s potential path to a $10 trillion market cap, and how the Groq acquisition may possibly prove to be the missing piece of the puzzle.

The Groq Acquisition

In December 2025, NVIDIA executed what analysts are calling a “Strategic Capture”—a $20 billion “reverse acqui-hire” of Groq. The deal, which is Nvidia’s largest purchase to date, was not a traditional merger but a surgical extraction of intellectual property and human capital designed to bypass the 18-month regulatory “purgatory” typical of such mega-deals.

By licensing Groq’s entire patent portfolio and bringing 80% of its engineering staff—including founder Jonathan Ross and former Chief Architect Dennis Abts—under the NVIDIA Research umbrella, Jensen Huang effectively “defanged” his most dangerous inference rival while simultaneously absorbing the cure for the GPU’s greatest weakness.

The “Abomination” and the Solution

As outlined by industry skeptics, the standalone Groq architecture was a “violently imbalanced” machine: a 144-wide VLIW (Very Long Instruction Word) monster that placed an immense burden on the compiler. While the hardware was simple, its “cycle-accurate” requirements meant that any slight synchronization hiccup across a server rack would cause the entire execution to “burst into flames.”

NVIDIA didn’t buy Groq for the hardware it had; they bought it for the hardware only NVIDIA could build. They are “fixing” the Groq abomination using three core IP synergies:

● Optical Clock-Forwarding: In early 2026, NVIDIA significantly accelerated its transition to optical die-to-die (D2D) and chiplet interconnects to solve the “power wall” and bandwidth bottlenecks of traditional copper. This technology solves the “sync nightmare” that plagued Groq. By forwarding a master clock over optics with sub-16-picosecond accuracy, NVIDIA can force hundreds of Groq-style chips to act as a single, coherent, deterministic brain—eliminating the “jitter and drift” that previously ground Groq’s execution to a halt.

● Hybrid Bonding & HBM4: Groq was trapped in a “SRAM-only” prison on ancient 14nm nodes. NVIDIA is migrating this architecture to the Vera Rubin platform, utilizing TSMC’s hybrid bonding to stack massive amounts of SRAM directly onto the compute die, while backing it with the immense bandwidth of HBM4.

● The “Hellish” Compiler Integration: The real prize was Groq’s compiler—refined over six years of “money-incinerating” cloud operations. By integrating this deterministic scheduling logic into CUDA 13, NVIDIA has moved from dynamic hardware scheduling (where the chip “guesses” what to do next) to a “pre-computed script” model.

From “LPU” to “Vera LPU”

The result is the rumored Vera LPU (Language Processing Unit). By stripping away the “legacy” overhead of a general-purpose GPU—like branch predictors and complex caches—and replacing them with Groq’s deterministic dataflow logic, NVIDIA is targeting a 10x reduction in cost-per-token.

How Groq Could Enable The Inference Inflection Point

To understand the path NVIDIA would need to take to 2x (or possibly even 3x) from here, we need to examine the fundamental shift in where the money is being spent...