Jensen Huang explains why Nvidia will grow an astounding 70% next year

Every time Wall Street prepares to write Nvidia's victory lap eulogy—claiming we've finally reached "peak GPU demand"—Jensen Huang steps up to a microphone and politely incinerates the bear thesis. Predicting a staggering 70% growth rate on top of a revenue base that already looks like a typo isn't just confident; it's a direct challenge to every analyst who thinks AI hardware is a cyclical commodity business. In a recent candid conversation reported by TechCrunch, Huang addressed the elephant in the room: allegations that Nvidia's growth is being artificially propped up by "circular deals" where they invest in AI startups who turn around and buy Nvidia silicon. Huang dismissed the charge, insisting that Nvidia’s strategy isn't financial engineering—it's market creation. Having tracked silicon cycles for years at WhatIsFuture.com, I can tell you that discounting Jensen’s playbook as a shell game fundamentally misunderstands the structural transformation happening across global compute infrastructure.

Key Takeaways

• Circular deal allegations miss the strategic point: Nvidia’s investments in specialized cloud providers and AI startups aren't round-tripping cash; they are building alternative distribution channels outside traditional hyperscalers.
• Inference is consuming training's lunch: The shift toward reasoning-heavy models (like OpenAI’s o1 series) requires exponential test-time compute, driving an entirely new wave of hardware demand.
• The energy bottleneck is the real threat: Nvidia's biggest growth limit isn't customer demand or TSMC capacity—it's power grid infrastructure and data center thermals.
• CUDA's moat remains unassailable: Competitors pushing custom ASICs or alternative chips are still fighting uphill against Nvidia's unified software and networking stack.

The Myth of the "Circular Deals" Revenue Engine

Let’s talk about the circular financing narrative, because short sellers and skeptics have been beating this drum for months. The argument goes like this: Nvidia puts venture capital into companies like CoreWeave, Lambda Labs, or enterprise AI model builders. Those companies then take that cash and immediately wire it back to Santa Clara to buy H100s and Blackwell racks. On paper, to a traditional auditor, that looks suspiciously like revenue inflation through venture round-tripped capital.

Here’s why that view is dead wrong on the technical merits. What Nvidia is actually doing is aggressive ecosystem building, reminiscent of Intel Capital in the 1990s, but executed at hyperscale speed. Traditional cloud behemoths—AWS, Microsoft Azure, Google Cloud—are all racing to build their own custom ASICs to break free from Nvidia’s margin-heavy grip. If Nvidia relied solely on Big Tech hyperscalers for distribution, they’d eventually get squeezed out as AWS pushes Trainium or Google pushes TPUs.

By seeding a new generation of dedicated "GPU-first" cloud providers and high-growth AI startups, Nvidia guarantees a buyer base that is natively loyal to CUDA. These aren't fake shell companies pushing paper; they are infrastructure providers suffering from an insatiable customer appetite for FLOPs. Jensen isn't buying fake sales; he is financing the construction of a non-hyperscaler distribution pipeline that guarantees Nvidia retains pricing power for the next decade.

Beyond Training: The Massive Compute Pivot to Inference and Physical AI

The core reason Nvidia can credibly project 70% growth isn't just because models are getting bigger—it's because the way we compute AI outputs is fundamentally changing. The industry is rapidly transitioning from pure pre-training (dumping trillions of tokens into a cluster once) to post-training and test-time compute. When models spend extra seconds "thinking" through step-by-step reasoning problems before generating an answer, they consume raw inference compute at rates we’ve never seen before.

At the same time, we are witnessing the migration of intelligence from cloud data centers into physical systems. Autonomous vehicles, industrial automation, and humanoid robotics are moving from research labs into production deployments. Whether it's complex embodied intelligence projects or accessible hardware experiments like when Hugging Face sold a cute $399 open source duck robot, Microduck, the underlying lesson is clear: robotics requires massive real-time simulation compute (like Nvidia Omniverse) before silicon ever touches physical metal.

This physical AI wave creates a double-dip growth engine for Nvidia. Developers use massive Nvidia clusters in the cloud to simulate physics and train reinforcement learning agents, and then deploy edge-optimized Nvidia chips directly into the physical systems. The enterprise market isn't just buying chips to generate text anymore; they are buying systems to model the physical world.

The Real Wall: Energy Constraints and Data Center Realities

If there is a legitimate threat to Nvidia achieving its astounding projection, it isn't competitive silicon from AMD or Intel. It’s thermal dynamics and local electrical grids. Jensen can design the Blackwell and Rubin architectures to be insanely performant, but if a customer can’t pull 120 kilowatts into a single server rack without melting their municipal power substation, those chips will sit in a warehouse.

We are already seeing this friction manifest around the world. Regulators and utility providers are putting their foot down as data center power demands explode. Look no further than municipal policy shifts, such as when Massachusetts hit data centers with new clean power rules, signaling that grid capacity and carbon offsets are becoming firm regulatory barriers. The era of indiscriminately dropping a 100-megawatt facility next to an easy fiber line is over.

This reality is forcing hardware architects to innovate radically on power management and cooling. We need breakthrough approaches, including efforts where developers and hardware teams are pushing boundaries—much like how this founder is teaching chips how to recycle their energy to maximize thermodynamic efficiency. Nvidia knows this, which is why they stopped selling standalone GPUs and started selling full liquid-cooled data center racks complete with proprietary NVLink networking switches. They aren't selling components anymore; they are selling power-optimized compute appliances.

What This Means for Developers, Founders, and the Ecosystem

For founders building in the AI ecosystem today, Nvidia's trajectory offers a clear strategic playbook. First, stop waiting for GPU prices to collapse due to oversupply. Jensen's forward guidance confirms that high-performance compute will remain a premium, constrained resource for the foreseeable future. If your business model relies on cheap, commodity-grade FLOPS to make its unit economics work, you are building on quicksand.

Second, software engineers need to double down on optimizing for the hardware-software co-design layer. Nvidia’s true monopoly isn't just the silicon wafer; it's the CUDA layer, Triton compilers, and NIM (Nvidia Inference Microservices) ecosystem. While open-weight models and alternative runtimes are making progress, the enterprise path of least resistance remains stubbornly locked into Nvidia's software stack. Trying to build custom inference engines on non-CUDA hardware might save you 20% on server costs, but it will cost you double in engineering overhead and time-to-market.

Ultimately, Nvidia’s predicted 70% growth is a reflection of a broader truth: we are in the middle of a multi-trillion-dollar global compute refresh. The old x86 architecture that powered the internet era is stepping aside for accelerated, parallel processing infrastructure built specifically for neural networks. Jensen isn't just selling chips into a bubble; he is billing the entire tech sector for the hardware upgrade of the century.

"The market keeps treating Nvidia like a semiconductor company cycling through consumer demand spikes. In reality, they are operating as the sole utility provider for the next compute era—and energy, not capital, is their only real competitor."

Frequently Asked Questions

Is Nvidia's projected 70% growth rate actually sustainable?

While maintaining a 70% growth rate indefinitely is mathematically impossible as revenue scales, Nvidia's medium-term growth is anchored by a structural shift from CPU-based data centers to GPU-accelerated computing. As enterprise workloads transition from basic cloud hosting to reasoning-heavy inference, physical AI, and industrial simulation, capital expenditures are permanently reallocating toward accelerated computing systems.

What is "circular financing" in AI, and should investors be concerned?

Circular financing refers to accusations that Nvidia invests capital into AI startups (like specialized cloud providers) who then use those funds to purchase Nvidia GPUs, creating an artificial loop of revenue. While the optics attract short-seller criticism, this is a standard ecosystem-seeding strategy. These startups are filling genuine, unmet market demand for high-performance compute that legacy cloud hyperscalers cannot instantly satisfy.

How does the energy grid crisis affect Nvidia's long-term business?

Power availability is the single largest operational bottleneck for Nvidia's growth. High-density architectures like the Blackwell B200 pull unprecedented wattage per rack, forcing data centers to adopt liquid cooling and navigate tight municipal clean-energy regulations. If utility grids cannot scale power delivery, customers may delay chip installations, making power efficiency innovation critical for Nvidia's roadmap.