Sell NVIDIA Stock — The Competition Is Coming for Jensen's Throne

The Intel Precedent#

In 2015, Intel controlled roughly 80% of the client and server CPU market.¹ They had dominated for two decades. Their manufacturing process was years ahead of competitors. The “Intel Inside” sticker was synonymous with computing itself.

By 2024, that market share had plummeted to around 60%.¹ In some segments—desktop CPUs, data center revenue—Intel no longer outsells AMD by 9:1. The ratio is closer to 2:1.² In Q3 2024, AMD’s data center division actually outearned Intel’s for the first time ever: $3.549 billion versus $3.3 billion.²

Between 2021 and 2024, Intel’s revenue declined by more than 30%.¹ That’s not a correction. That’s a structural collapse. (If you think rapid market cap losses only happen to smaller companies, check out how Apple lost $112 billion in two days after a disappointing product launch—even tech giants aren’t immune to investor sentiment shifts.)

What happened? AMD happened. ARM happened. Apple’s decision to build its own silicon happened. Intel got complacent, missed manufacturing transitions, and competitors caught up.

The Communications of the ACM put it bluntly: “Intel’s commitment to manufacture its own designs turned an Intel strength into a glaring disadvantage.”³

Here’s the thing about monopolies: they’re not defended by being good. They’re defended by being so far ahead that catching up is impossible. The moment catching up becomes merely difficult instead of impossible, the lead starts eroding.

Athena Character @ openart.ai

NVIDIA’s Moat Is Real. But It’s Not Forever.#

NVIDIA bulls will correctly point out that NVIDIA’s position is different from Intel’s. And they’re right—in some ways.

NVIDIA’s CUDA ecosystem represents an 18-year head start in GPU software development.³ Millions of developers are trained in CUDA programming. PyTorch, TensorFlow, and every major AI framework are optimized first for CUDA. Libraries like cuDNN and cuBLAS have been refined through countless iterations. The ecosystem is genuinely impressive.

But CUDA isn’t magic. It’s a software moat—and software moats can be bridged.

AMD’s ROCm software stack has narrowed the performance gap considerably. PyTorch now officially supports ROCm. AMD has funded development of ZLUDA, a drop-in CUDA compatibility layer that allows many CUDA applications to run on AMD hardware without modification.⁴ Microsoft is reportedly developing toolkits to convert CUDA models to run on AMD chips.⁵

The CUDA lock-in is real, but it’s weaker than the market believes. And every year, the alternatives get better.

The Smart Money Isn’t Just Reducing—It’s Diversifying#

Remember those institutional investors from Part 1 who are reducing their NVIDIA positions? They’re not moving to cash. They’re moving to alternatives.

Microsoft is deploying AMD Instinct MI300X accelerators to power Azure OpenAI Service workloads, achieving what it calls “leading price/performance” on GPT inference.⁶ Meta allocated 43% of its 2024 GPU shipments to AMD—173,000 MI300X units versus 224,000 from NVIDIA.⁷ Oracle is building out MI300X capacity for its cloud offerings.⁷

The hyperscalers aren’t stupid. They saw what happened when everyone became dependent on Intel for CPUs. They’re not going to let NVIDIA achieve permanent lock-in without building alternatives.

The Power Wall#

Data centers are not ready for this. According to a Deloitte survey of 120 US-based data center and power company executives, 72% consider power grid and capacity to be “very or extremely challenging” for AI infrastructure buildout.¹⁰

Fewer than 5% of the world’s data centers can support even 50kW per rack.¹¹ Blackwell configurations require 60-120kW per rack. The infrastructure gap is enormous. And when infrastructure fails—as AWS’s 16-hour us-east-1 outage demonstrated—the cascading effects can cost hundreds of millions in losses across the entire cloud ecosystem.

Goldman Sachs Research forecasts global power demand from data centers will increase 165% by 2030.¹² Grid connection requests in key regions like Virginia now face wait times of four to seven years.¹³ The AI buildout is running headlong into physical infrastructure constraints.

What does this mean for NVIDIA? It means there’s a ceiling on how many Blackwell GPUs can actually be deployed, regardless of demand. And it creates massive pressure to develop more power-efficient alternatives.

The Efficiency Arms Race#

Google’s new Ironwood TPU—their 7th generation custom AI chip—delivers 4,614 TFLOPs of peak performance while claiming twice the performance per watt of its predecessor.¹⁴ Each Ironwood pod consumes approximately 10 megawatts while delivering 42.5 exaFLOPS of compute.¹⁴

Intel’s Gaudi 3 accelerator costs roughly half the price of an NVIDIA H100 ($15,625 vs $30,678 per chip) and consumes 600W versus 700W.¹⁵ Intel claims 2.3x better performance-per-dollar for inference workloads.¹⁶

These alternatives may not match NVIDIA’s raw performance in every benchmark. But for many inference workloads—which account for a growing share of AI compute—they’re “good enough” at significantly lower cost and power consumption.

Athena Character @ openart.ai

The Inference Shift#

Here’s the structural change that NVIDIA bulls are underestimating: the AI industry is shifting from training to inference.

Training is where NVIDIA dominates most completely. Training runs consume massive amounts of compute over days or weeks, and NVIDIA’s hardware and software are optimized for exactly this use case.

But inference—actually running trained models to generate outputs—is becoming the dominant share of AI spending. Inference workloads will make up 75% of global AI compute demand by 2030.¹⁷ One analysis suggests the training-inference crossover will occur by 2027.¹⁸

Inference is a fundamentally different workload. It’s more latency-sensitive, more cost-sensitive, and more amenable to specialized hardware. Companies like Google, Amazon, and now Qualcomm are building chips specifically optimized for inference rather than training.

NVIDIA can compete in inference. But they’ll face far more competition there than in training. And the customers who are panic-buying H100s and B200s for training today will be shopping for alternatives when it comes time to scale inference.

The Hyperscaler Insourcing Threat#

The biggest risk to NVIDIA isn’t AMD or Intel. It’s NVIDIA’s own customers building their own chips.

Google, Amazon, Meta, and Microsoft are all investing billions in custom silicon. They have the resources, the engineering talent, and—critically—the scale to justify the investment. A hyperscaler running millions of inference queries per day can amortize chip development costs across an enormous volume.

The logic is simple: why pay NVIDIA’s 73% gross margins when you can build your own chips at cost?

Google has been on this path the longest—TPUs have been in development for a decade. But Amazon’s Trainium program is catching up fast. And the deal between OpenAI and Broadcom signals that even the AI labs themselves are looking to reduce dependence on NVIDIA.

This isn’t speculation. This is happening right now. Every hyperscaler is diversifying away from NVIDIA dependence because they’ve seen what single-vendor reliance did to them with Intel.

Want contrarian tech analysis that challenges Wall Street groupthink? Subscribe for investment insights that follow the data, not the hype.

Part 3 drops soon with the valuation math Wall Street doesn’t want you to see.

The Rise of Smaller, Specialized Models#

For the past three years, the AI industry has been obsessed with scaling. Bigger models. More parameters. More training compute. NVIDIA’s GPUs were perfectly positioned for this arms race.

But the industry is shifting. Small language models (SLMs) are gaining traction for enterprise use cases. Anthropic’s Claude Haiku 4.5 matches much of the accuracy of larger models while running twice as fast and costing one-third as much.²² IBM’s Granite 4.0 “Nano” and “Tiny” models can run directly on local devices.²²

NVIDIA itself acknowledges that SLMs could perform 70-80% of enterprise AI tasks, with large-scale systems reserved for the most complex reasoning.²²

What does this mean? It means a growing share of AI workloads can run on cheaper, less power-hungry hardware. Edge devices. CPUs with integrated AI accelerators. Purpose-built inference chips like Google’s TPUs or Amazon’s Inferentia. (Meanwhile, the hype around “autonomous” AI agents continues despite 70% failure rates on real-world multi-step tasks—suggesting the industry may be overestimating near-term AI capabilities while underestimating infrastructure requirements.)

The “bigger is better” era that drove NVIDIA’s revenue explosion may be giving way to an efficiency era where power consumption and total cost of ownership matter more than raw FLOPS.

The Power Crunch Favors Alternatives#

Data center power consumption from AI is projected to reach 68 gigawatts globally by 2027—almost a doubling from 2022 levels.¹³ Individual AI training runs could require up to 8 gigawatts by 2030.¹³

This is creating intense pressure to develop more power-efficient solutions. Every watt saved at the chip level translates to massive savings at the data center level—in power costs, cooling costs, and infrastructure costs.

NVIDIA’s Blackwell chips are more efficient per operation than prior generations. But they’re still power-hungry compared to purpose-built inference accelerators. Google claims Ironwood delivers 30x better power efficiency than its first Cloud TPU.¹⁴ Intel positions Gaudi 3 as a TCO leader partly because of its lower power consumption.¹⁵

When the industry hits power constraints—and it’s hitting them now—efficiency becomes more valuable than raw performance. NVIDIA’s advantage in peak performance may become less relevant than competitors’ advantages in performance-per-watt.

The CUDA Moat Is Eroding#

AMD’s ROCm 7 software stack, expected soon, claims up to 3.5x better inference performance compared to ROCm 6. PyTorch officially supports ROCm. HIP (Heterogeneous-compute Interface for Portability) allows developers to write code that runs on both AMD and NVIDIA GPUs with minimal changes.⁴

Microsoft is reportedly building toolkits to convert CUDA models to ROCm.⁵ OpenAI has added support for AMD Instinct accelerators in Triton 3.0.¹⁹ The open-source community is actively working on bridging the software gap.

CUDA’s dominance isn’t going to disappear overnight. But the “lock-in” is weaker than the market believes. And every year, the friction of switching decreases.

Q: Doesn’t NVIDIA’s software ecosystem make switching costs prohibitive?#

A: Switching costs are real but declining. AMD’s ROCm now officially supports PyTorch. Microsoft is reportedly building CUDA-to-ROCm conversion tools.⁵ AMD has funded ZLUDA, a drop-in CUDA compatibility layer.⁴ The friction of switching decreases every year. And for new workloads, customers increasingly evaluate total cost of ownership rather than defaulting to NVIDIA.

Q: What about NVIDIA’s lead in training? Isn’t that where the money is?#

A: Training is where NVIDIA dominates most completely today. But inference workloads are growing faster and will represent 75% of AI compute demand by 2030.¹⁷ Inference is more price-sensitive and more amenable to specialized hardware. NVIDIA faces stiffer competition in inference than in training.

Q: Don’t hyperscalers still need NVIDIA alongside their custom chips?#

A: Yes—for now. Google uses both TPUs and NVIDIA GPUs. Amazon uses both Trainium and NVIDIA. But the mix is shifting. Every custom chip deployed is one fewer NVIDIA GPU purchased. And as custom silicon matures, the mix will shift further away from NVIDIA.

Q: What about Blackwell’s performance leadership?#

A: Blackwell is genuinely impressive. But it also consumes 1,200W per chip.⁸ Fewer than 5% of data centers can handle Blackwell power densities.¹¹ Performance leadership matters less when you can’t actually deploy the hardware due to infrastructure constraints.

Q: Isn’t Qualcomm too late to the market?#

A: Qualcomm’s AI200 doesn’t ship until 2026. They’re late. But they’re not entering a stable market—they’re entering a market that’s still being defined. Their mobile chip expertise in power efficiency could translate well to inference workloads where TCO matters. And they already have a major customer (Humain) signed up for 200MW of capacity.²¹

NOTE

Take away this: NVIDIA’s $4.4T valuation prices in a monopoly that history says won’t last. Seven major competitors are shipping alternatives, hyperscalers are building custom silicon, and the AI workload mix is shifting from training (where NVIDIA dominates) to inference (where competition thrives). You can own a great company at the wrong price—and this is the wrong price.