Beyond Training: The Inference Economy
One of the prevailing concerns around NVIDIA has been whether the company's dominance could erode as the AI industry shifts from training large language models to inference — the process of actually running those models to generate outputs. While it is inevitable that NVIDIA will cede some market share over time, the sheer size and growth rate of the AI compute market make this a manageable concession. The company is not simply riding on the strength of its GPUs; it has built a far broader ecosystem of hardware, software, and networking that is deeply entrenched across the industry.
The introduction of the Grace Blackwell 3 LPX system at NVIDIA's 2025 GTC conference is a case in point. This architecture is purpose-built for the emerging inference economy, offering separate systems optimized for outputting tokens from AI models, distinct from those that ingest and process data during training. This is a strategically important move: by producing inference-specific chips on different manufacturing nodes, NVIDIA sidesteps the bottleneck of relying solely on the most advanced — and capacity-constrained — fabrication processes. Inference chips can be manufactured at volume alongside training hardware, ensuring supply keeps pace with exploding demand.
A Trillion Dollars in Sight
At GTC, NVIDIA signaled that it expects at least a trillion dollars in orders over the next two fiscal years through 2027 — a figure that exceeded prior estimates by a couple of hundred billion dollars. With the stock trading near the low end of its range since mid-2024, this kind of pipeline arguably presents a compelling entry point. The gap between the company's execution and its stock price reflects not a micro-level concern about NVIDIA's competitiveness, but rather a macro-level anxiety about the broader AI spending cycle.
The central question investors are wrestling with is straightforward: will the massive capital expenditures flowing into AI infrastructure ever generate adequate returns? This is not something NVIDIA can control through product launches or keynote presentations alone. It is a macro call — a bet on whether AI spending will plateau or continue accelerating. The prevailing view among those bullish on the company is that as 2026 gives way to 2027 without a dramatic slowdown in hyperscaler and enterprise spending, the market will be forced to rerate the stock upward.
Agentic Computing and the Personal GPU
Beyond the data center, NVIDIA is laying the groundwork for an entirely new computing paradigm. The Nemo Claw platform envisions a world of agentic AI systems — always-on, autonomous agents that operate continuously on behalf of users. In this vision, every knowledge worker has a GPU at their desk running models on an ongoing basis, transforming the personal computer into something closer to a personal AI engine.
This is a growth vector that has scarcely been modeled into current revenue projections. Today, the AI hardware conversation is dominated by hyperscaler deployments — massive GPU clusters installed by the likes of Microsoft, Google, and Amazon. But the agentic computing paradigm suggests a much larger addressable market that extends to enterprises, small businesses, and eventually individual users. If even a fraction of this vision materializes, the total demand for NVIDIA's silicon could dwarf current forecasts.
The Final Frontier: AI in Space
Perhaps the most striking signal from GTC was NVIDIA's move into space-based computing. The company has begun hiring engineers and partnering with emerging satellite companies like StarCloud to deploy GPUs in orbit — pushing compute closer to what might truly be called the final frontier. While this may sound like science fiction, it reflects a pragmatic strategy: overcoming terrestrial bottlenecks in energy and construction by moving some compute infrastructure off-planet entirely.
The launch of the DSX reference architecture for data centers further illustrates how NVIDIA is thinking beyond individual racks and clusters. The company is now addressing the fundamental substrate of computing — where data centers are physically located, how they consume energy, and how they scale to meet demand. By broadening its scope from chip maker to infrastructure architect, NVIDIA positions itself to capture value across the entire AI supply chain, not just at the silicon level.
The Macro Puzzle
NVIDIA finds itself in an unusual position: a company delivering consistently exceptional results that nonetheless struggles to break out of a trading range it has occupied for months. The analogy of a star student who keeps turning in perfect grades captures the dynamic well — the market has simply come to expect excellence, and excellence alone is no longer enough to move the stock.
What will ultimately drive a rerating is not another product announcement or another record quarter. It is the resolution of macro uncertainty. When the broader market gains confidence that AI infrastructure spending is not a bubble but a durable, multi-year buildout — one that generates real returns for the companies making the investments — NVIDIA's valuation will reflect the extraordinary business it has become. Until then, the company will continue to execute, expand its addressable market, and wait for the market to catch up with reality.