I keep buying NVIDIA, and the reason has almost nothing to do with the chips. It’s the software giveaway underneath them. Most investors file NVIDIA (NASDAQ:NVDA | NVDA Price Prediction) under “hardware,” and that framing is exactly why I’m still adding at $207.29. What I actually own is a freemium platform that happens to sell the world’s most expensive accelerators on the back end.
The Free Tier Is the Trap
CUDA is free. Triton is free. NIMs, Dynamo 1.0, Nemotron, BioNeMo, Isaac, Omniverse. All free. Every graduate student, startup, and hyperscaler research team writes code against these libraries. Then the code only runs at full speed on NVIDIA silicon. “NVIDIA has the largest suite of acceleration libraries in the world,” Jensen Huang told analysts on the May call, and that’s the moat in plain language.
The paywall shows up at scale. When a customer moves from prototype to production, they build an AI factory rather than purchasing a single GPU. CFO Colette Kress framed it plainly: “Customers do not buy GPUs; they build AI factories. The right economic metric is not the purchase price of the GPU; it is the lifetime cost of an AI factory producing intelligence.” Switching costs at that level are brutal. You’d rewrite years of CUDA-optimized code, retrain teams, and lose performance. Almost nobody does it.
The Receipts
Three numbers keep the buy button warm. First, growth that shouldn’t be possible at this size. Q1 FY2027 revenue hit $81.615 billion, up 85.23% year over year, with Data Center alone at $75.246 billion (+92%). Networking inside that number grew 199% YoY. Management guided Q2 to $91.0 billion.
Second, the margin structure the software stack enables. Gross margin 71.07%, operating margin 60.38%, ROE 101.5%, ROIC 92.2%. Free cash flow of $48.554 billion in a single quarter. Those are software-company economics attached to a hardware volume business.
Third, capital returns finally showing up. The board authorized an additional $80 billion buyback on top of $38.5 billion remaining, and lifted the quarterly dividend from $0.01 to $0.25. At a forward P/E of 23, I’m paying a market multiple for a compounder returning tens of billions to owners.
Why Not the Obvious Alternatives
NVIDIA’s Data Center segment posted $75.25 billion in a single quarter, which is larger than AMD’s entire company revenue base, and AMD does not run CUDA. Broadcom is the other name people cite for AI silicon, but its custom ASIC business lacks the CUDA software ecosystem lock-in that keeps developers on NVIDIA year after year. I’m paying for a developer base that would need to be pried loose one library at a time.
The Risk I Actually Watch
China. H20 shipments went to zero this quarter, and Q2 guidance excludes China Data Center compute revenue entirely. That’s real. What keeps the thesis intact is that demand outside China is absorbing every wafer TSMC can produce. The $91 billion Q2 guide assumes zero China contribution. Blackwell and Rubin combined carry $1 trillion in revenue visibility through calendar 2027. I don’t need China to make the math work.
What Keeps the Buy Button Active
Reddit is skeptical, insiders are trimming, and the crowd on Polymarket sees limited near-term upside above $210. I’m buying for the long arc, because every free download of CUDA is a future paying customer, and there are roughly 250,000 enterprises that haven’t shown up yet. The freemium hook is set. I keep loading.
Contact [email protected] for any questions or corrections.