Joe
NVIDIA: Selling AI Factories, Not Just Chips
If you've never studied NVIDIA before, the first thing to understand is that it stopped being a graphics card company years ago. In the most recent quarter (Q2 FY2027, ended July 26, 2026), NVIDIA reported $96.2 billion of revenue: $89.0 billion from Data Center and $7.2 billion from Edge Computing. The old "Graphics" segment is now the smaller, slower-growing piece inside Edge. The company's core is supplying the compute, networking, systems, and software that large AI data centers depend on. It calls itself a "data center scale AI infrastructure company," which is accurate: its customers are hyperscalers, AI model makers (AI labs), AI cloud providers (neoclouds), enterprises, and sovereign buyers, while consumer gaming is secondary.
My core judgment: NVIDIA's economic engine is not just selling GPUs; it sells the most productive, most fungible, and most durable AI computing platform, allowing customers to build and finance AI factories with confidence. The company captures value through high-margin hardware (~75% gross margin) and an expanding full-stack pull-through (networking, CPUs, software). Growth is currently driven by a global AI infrastructure buildout so broad that demand would double next year, but supply constraints limit actual growth to about 70%. The model's biggest open questions are whether custom silicon from major customers erodes its share, and whether its increasing use of balance-sheet support to finance customers creates hidden risk.
How NVIDIA Actually Makes Money
Reportable segments are Compute & Networking and Graphics, but the more useful split is by market platform. Data Center revenue grew 117% year over year to $89.0 billion, split between Hyperscale ($48.7B, +102% y/y) and AI Clouds, Industrial & Enterprise ("ACIE", $40.3B, +138% y/y). Edge Computing was $7.2B, +27% y/y. The Data Center business sells GPU systems (Blackwell Ultra, soon Vera Rubin), CPUs (Grace/Vera), networking (InfiniBand, Spectrum-X Ethernet), and related software. Gamers and professional visualization now sit in Edge.
This is not a one-product company. A full buildout includes a mix of Compute (GPUs/CPUs) and Networking; software like CUDA and AI Enterprise; and increasingly system-level offerings. The economic logic: customers are buying an "AI factory" that turns electricity, chips, and software into tokens/intelligence. NVIDIA earns revenue on each element of that factory, and gross margin is high because the product is a complete, integrated system—not a commodity component. Q2 GAAP gross margin was 75.0%, up 2.6 points year over year.
Customer concentration is real but diversifying. Hyperscalers still dominate, but ACIE (neoclouds, enterprises, sovereigns) grew faster and now accounts for nearly half of Data Center revenue. Management says non-hyperscaler growth (ACIE) will represent roughly half of Data Center business, and sovereign AI revenue tripled year over year in Q2.
Why Customers Choose NVIDIA: The Three Pillars
Ian Buck, NVIDIA's VP of hyperscale/HPC and original CUDA architect, gave the clearest statement in an October 2026 interview: NVIDIA wins on productivity, fungibility, and durability.
- Productivity: Each generation delivers more tokens per dollar and per watt. Vera Rubin is guided to be 30x higher throughput per megawatt and 35x lower token cost than Grace Blackwell Ultra. Software optimizations (CUDA, Dynamo, PyTorch, etc.) mean customers get immediate performance. This is not just chip specs; it is the accumulated 20-year software stack.
- Fungibility: The same GPU can do training, inference, reinforcement learning, and accelerated computing. When a customer spends billions on a data center, they can redeploy the compute across workloads. Buck says older GPUs (Volta from 2019) are still sold out at CoreWeave. That fungibility makes NVIDIA infrastructure easier to finance.
- Durability: The platform keeps generating value after NVIDIA moves to next generation. Customers view the purchase as a long-lived, redeployable asset. That durability is why financing institutions are willing to underwrite AI builds—and why regional neoclouds are exploding.
To those pillars I would add a fourth: full-stack system integration. Because NVIDIA co-designs GPU, CPU, networking, and software, it can capture more of the data center TAM. Revenue opportunity per gigawatt has expanded from ~$18B (Hopper) to ~$25B (Blackwell) to ~$40B (Vera Rubin). That is the company monetizing not just the chip but the surrounding system.
The software ecosystem (CUDA) is not just a moat—it is the distribution channel. Hundreds of domain-specific libraries and frameworks run on NVIDIA GPUs; AI researchers and startups default to it. That makes NVIDIA the default neutral platform for model builders, including closed-model labs (OpenAI, Anthropic, Grok, Meta) and open-model providers (Mistral, Qwen, etc.). The goal is to be the infrastructure that all AI models run on.
Growth Engine: The Broadening of AI Demand
NVIDIA's growth is not one hyperscaler or one model. It is a layered expansion:
- Hyperscalers: Top five hyperscalers are expected to spend nearly $800B capex in 2026 and $1.3T in 2027; NVIDIA supplies them and their internal workloads. AWS alone is deploying an additional 2 million GPUs through Q2 FY2029.
- Neoclouds/AI clouds: Regional and specialized providers are bringing capacity online fast—CoreWeave, Nebius, Nscale, and about 100 others. Neocloud installed capacity is expected to go from ~3GW at end-2025 to ~8GW at end-2026, using NVIDIA DSX reference designs.
- Enterprises and sovereigns: Automotive, financial services, healthcare, and governments are building on-prem AI factories; sovereigns often prefer regional neoclouds over foreign hyperscalers. AI-native startups are funded by $400B VC in H1 2026, roughly 70% spent on compute.
- Edge: Gaming and workstations grow modestly, plus autonomous driving and robotics, but this is not the main story.
The company's own growth guide says it all: it expects ~70% revenue growth in fiscal 2028, but customer forecasts point to demand doubling. The difference is supply constraint. That constraint—memory, CoWoS packaging, power, data center shell—is the main brake. NVIDIA has committed $279B to supply procurement, primarily memory, and has secured capacity through partnerships.
What's Changing: From Hardware Sales to Ecosystem Financing
The most important emerging mechanism is not a new chip but a change in how NVIDIA gets customers to buy. Because AI labs and neoclouds are growing faster than their balance sheets can support, NVIDIA has started:
- Revenue-sharing agreements: It provides take-or-pay capacity commitments to neoclouds, enabling lenders to underwrite projects, in exchange for a share of revenue above a floor. NVIDIA "gets paid twice"—once on hardware sale and again on rental revenue.
- Balance-sheet support: Nearly $50B in equity investments in frontier AI labs, plus partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to raise over $500B third-party capital. It signed guarantees up to $105B for OpenAI's PORTS-Pike site (4.25GW), with total max gross exposure $108.5B including other guarantees.
This is a double-edged sword. It accelerates the AI buildout and locks NVIDIA into future generations. But it turns NVIDIA from pure supplier into a quasi-financier, with credit exposure and potential circularity: NVIDIA's investments help customers buy NVIDIA products. Management sees the risk as limited because NVIDIA compute is fungible and can be redeployed. I would flag this as the biggest watch item for anyone studying the model: if customer demand disappoints, these guarantees and revenue shares become a new kind of economic engine—one that could burn cash and compress margins.
Limitations and Threats
- Supply constraint: Growth is capped by memory, packaging, and power. If supply loosens faster than expected, growth can surprise upward; if not, the 70% guide is the ceiling.
- Custom silicon: Microsoft Maia, Amazon Trainium, OpenAI's Jalapeno, and others. NVIDIA's response is that its performance per token and software ecosystem remain superior, but if hyperscalers internalize more of their own inference/training, NVIDIA loses growth from those large customers. Currently, hyperscalers still buy NVIDIA in addition to their own chips, but the mix matters.
- China: Essentially shut out of data center compute (less than 1% of Data Center revenue in Q2, no China compute in future outlook). Not a huge near-term revenue loss but a long-term TAM cap.
- Customer concentration in financing: About a quarter of next year's business may be tied to AI labs backed by NVIDIA's balance sheet. If any of those labs fails, NVIDIA not only loses revenue but also credit exposure.
- Margin mix: New products and lower-margin Hopper China shipments may dilute gross margin; Q3 guidance is 74.0%, down 100 basis points.
What I'm Watching Most
If I had to track only a few things, they would be:
- Data Center revenue growth vs. hyperscaler capex forecasts. If hyperscaler spending decelerates and non-hyperscaler (ACIE) growth cannot fill the gap, NVIDIA's growth engine will stall. Watch the split between Hyperscale and ACIE.
- Vera Rubin ramp and token-cost trend. The entire customer-value story rests on NVIDIA delivering a large per-generation improvement in token cost. Any delay or underperformance would undermine the "more compute drives more revenue" loop.
- The financing/guarantee structure. The amount of balance-sheet support, revenue share income, and guarantees will reveal whether NVIDIA has found a durable growth accelerant or is underwriting its own demand.
- China/export policy and supply constraints—these dictate the ceiling and margin.
The moment I would change my view: if I see non-hyperscaler customers failing to renew or expand capacity without NVIDIA financial support, or if custom-silicon benchmarks start matching NVIDIA on total cost of ownership for inference at scale, then the "AI factory toll booth" assumption would need to be revisited. Until then, NVIDIA's core model—selling the most productive, fungible, durable compute platform—is intact and still expanding, even if supply constraints are the binding limit.






