NVIDIA is the only AI company

NVIDIA just wrapped GTC 2026 and Jensen Huang reminded everyone that while the rest of the industry argues about model benchmarks, NVIDIA sells the infrastructure they all depend on. The picks-and-shovels play isn't a cliché when you're the only credible supplier. With $1 trillion in projected purchase orders for Blackwell and Vera Rubin through 2027, 86% of data center GPU revenue, and a software moat that has been compounding for two decades, NVIDIA isn't competing with OpenAI or Anthropic or Google. It's taxing them. And that distinction matters more than any model leaderboard.

The GTC 2026 signal

Jensen Huang's keynote at GTC 2026 ran nearly three hours. The headline was a staggering projection: purchase orders for Blackwell and Vera Rubin chips could reach $1 trillion through 2027, doubling last year's $500 billion estimate. Vera Rubin ships in Q3 2026 with HBM4 support. Rubin Ultra follows in the second half of 2027. But the more revealing moment was subtler. NVIDIA unveiled a new inference-focused chip built on technology licensed from Groq for a reported $20 billion. As AI computing matures and shifts from training to always-on inference, NVIDIA isn't waiting for the market to move. It's buying the pieces it doesn't already own and folding them in. Huang also walked through "every single layer" of AI, from accelerated computing and networking to storage, software, and models. The message was clear: NVIDIA doesn't just sell chips. It wants to define the entire physical plant of the AI economy.

The platform company thesis

There's a useful distinction between companies that build AI products and companies that sell the means to build AI products. OpenAI, Anthropic, and Google DeepMind are in the first category. NVIDIA is squarely in the second, and it's a far better business. In fiscal Q3 2026, NVIDIA posted $51.2 billion in data center revenue alone, representing 90% of the company's total $57 billion quarter. Every major AI lab, every hyperscaler, every startup training a model is writing checks to NVIDIA. When Elon Musk's xAI needed to build the world's most powerful training system, Colossus 2, it used NVIDIA GB200 chips. There was no meaningful alternative. This is the platform company thesis in its purest form. NVIDIA doesn't need to win the model race. It doesn't need to figure out product-market fit for chatbots or agents. It just needs demand for compute to keep growing, and every trend in AI points in that direction. As Huang put it at GTC: "If they could just get more capacity, they could generate more tokens, their revenues would go up." The AI companies are NVIDIA's distribution channel.

Why CUDA is more durable than any model lead

The conversation about NVIDIA's dominance usually focuses on hardware, but the real moat is software. CUDA, NVIDIA's parallel computing platform, has been building ecosystem depth since 2006. Two decades of optimization, documentation, libraries, and developer muscle memory have created a lock-in that is qualitatively different from anything in the model layer. Model leads are fleeting. GPT-4 was state of the art for months, not years. Claude, Gemini, and open-source alternatives keep narrowing the gap. A model moat depends on staying ahead in a race where the track is getting shorter. CUDA's moat works differently. It's embedded in university curricula, research workflows, production infrastructure, and millions of lines of existing code. Teams spend the majority of their time on DevOps and optimization rather than model development. Switching costs aren't just financial, they're cognitive. An entire generation of ML engineers thinks in CUDA. Forbes reported in March 2026 that NVIDIA holds 86% of data center GPU revenue, down slightly from 90% in 2024. That gentle decline looks like progress for competitors until you realize the overall market is expanding so fast that NVIDIA's absolute revenue keeps climbing. The pie is growing faster than the slice is shrinking.

The counter-narrative: AMD and custom silicon

AMD is the most credible GPU challenger. At CES 2026, CEO Lisa Su unveiled the full MI400 lineup, with the MI450 shipping in Q3 2026 and promising up to 10x performance gains over the MI300X in some applications. AMD claims comparable compute to NVIDIA's Vera Rubin with 1.5x higher memory capacity and scale-out bandwidth. The MI500, built on a 2nm process with HBM4E memory, is already slated for 2027. The commitments are real. OpenAI signed on for 6 gigawatts of AMD Instinct GPUs over multiple years. Oracle plans to deploy 50,000 AMD AI chips. These are not token gestures. Then there's custom silicon. Google's TPU v6e offers up to 4x better performance per dollar compared to NVIDIA's H100 for large language model training. Anthropic committed to hundreds of thousands of Google TPUs in late 2025. AWS deployed over 500,000 Trainium2 chips for Anthropic's model training, the largest non-NVIDIA AI cluster in production. The New York Times reported in January 2026 that Amazon and Google are eating into NVIDIA's chip supremacy, with Broadcom's CEO disclosing $10 billion in Google chip sales to Anthropic alone. For high-volume inference workloads where you're running the same model repeatedly at massive scale, TPUs and Trainium can offer meaningfully better cost-per-token economics. Some companies are quietly making the switch. But here's the catch: training new frontier models still requires NVIDIA. Inference is being commoditized. Training is not. And NVIDIA keeps moving the goalpost with annual hardware refreshes. Huang calls himself "Chief Revenue Destroyer" because each generation intentionally obsoletes the last. Competitors aren't just chasing a moving target, they're chasing one that accelerates.

Intel's mobile moment, and whether NVIDIA could fumble

The most instructive analogy isn't AMD. It's Intel. Intel dominated the PC era so completely that it couldn't imagine a world where the next compute platform wouldn't run on x86. Then mobile happened, ARM ate the market, and Intel spent a decade trying to catch up. The lesson: monopolies don't fall because a competitor builds a better version of the same thing. They fall when the game changes entirely. Could NVIDIA have an Intel moment? The scenario would look something like this: inference becomes the dominant workload, specialized ASICs optimized for specific model architectures outperform general-purpose GPUs on cost and latency, and the CUDA ecosystem advantage erodes as higher-level abstractions (like PyTorch's hardware-agnostic backends) reduce switching costs. There are early signals. Groq's inference chips were compelling enough for NVIDIA to license the technology for $20 billion rather than compete against it. Hyperscalers collectively plan to invest $660 billion in AI infrastructure in 2026, and some fraction of that is going to proprietary silicon specifically to reduce NVIDIA dependency. But the timeline matters. NVIDIA's annual product cycle, its willingness to cannibalize its own hardware, and the sheer depth of CUDA integration mean competitors need to be not just better but dramatically better to justify the switching cost. Intel's fumble was a decade of denial. NVIDIA, so far, shows no signs of complacency.

What this means for indie builders

If NVIDIA controls supply, it controls the pace of AI progress, and that has real consequences for anyone building outside the hyperscaler bubble. GPU scarcity is already biting. NVIDIA confirmed ongoing GeForce RTX 50 Series shortages through fiscal 2027 due to a global memory crisis. Gaming GPU production has been cut 30-40%, and prices are climbing. Data center GPUs are even harder to come by for small teams. When NVIDIA prioritizes trillion-dollar hyperscaler contracts, indie developers and small startups are at the back of the queue. This creates a two-tier AI economy. Well-funded companies with multi-year NVIDIA contracts can train and iterate freely. Everyone else is constrained by whatever compute they can rent at market rates, rates that NVIDIA's pricing power keeps elevated. The practical response for small teams is increasingly to avoid the training game entirely. Fine-tune open-source models on rented compute. Use inference-optimized providers where AMD and custom silicon competition is actually driving prices down. Build at the application layer where the marginal cost of compute matters less than taste, speed, and product intuition. The irony is that NVIDIA's dominance might be the strongest argument for the open-source model ecosystem. If you can't afford to train your own frontier model, you need someone else to release theirs. And you need inference to be cheap, which means you need NVIDIA's competitors to succeed, at least in that segment.

The bottom line

NVIDIA is not an AI company in the way most people use the term. It doesn't ship chatbots or agents or AI-powered products. It ships the substrate on which all of those things run. That's a stronger position than any of its customers hold, and it's likely to remain so for the foreseeable future. The real question isn't whether NVIDIA will be dethroned. It's whether the AI economy that NVIDIA enables will be accessible enough for more than just the largest players to participate in. Right now, the answer is uncertain, and that should concern everyone building in this space.

References