Finvest
NVDA Semiconductors · Mega cap · AI infrastructure · Thesis updated June 10, 2026

NVIDIA Is the AI Factory Toll Road

01 Running thesis

The lead keeps widening

NVIDIA is still the clearest way to own the AI infrastructure buildout. The company is not selling a single chip. It sells a full computing system: accelerators, CPUs, networking, and software. That matters because big customers want speed, not parts they must stitch together themselves.

The bull case got stronger after Q1 FY2027. Total revenue was $81.6 billion, up 85 percent year over year. Data Center revenue was $75.2 billion and now makes up over 90 percent of the company. The next Vera Rubin platform is expected to start production shipments in Q3 2026, and the new standalone Vera CPU opens a market management frames at $200 billion.

A second growth path is sovereign AI. These are national AI infrastructure projects, not just cloud company orders. Revenue from this area grew more than 80 percent year over year, with deployments in nearly 40 countries. If that demand keeps broadening, NVIDIA becomes less dependent on only the largest U.S. cloud buyers.

The bear case is still real. Three direct customers recently represented 21 percent, 17 percent, and 16 percent of total revenue. China is also a live risk because export controls have already cut into the data-center opportunity there, and Chinese regulators issued a preliminary finding tied to the Mellanox acquisition approval. The business is excellent, but the thesis depends on AI spending staying high and on customers not moving too much work to their own chips.

Jul 2026▲Incorporated Q1 FY2027 results. Vera Rubin shipments moved to Q3 2026, the standalone Vera CPU widened the market, and sovereign AI became a clearer demand driver.
Jun 2026→Reviewed after GTC and Q1. The thesis held, with focus on inference share and custom chip announcements.
Feb 2026▲Raised conviction after Q4 FY2026. Rubin, the Anthropic investment, and the Groq licensing deal strengthened the platform story, while China risk grew more specific.
Nov 2025▼Trimmed the view after customer concentration rose sharply and China H20 orders failed to materialize. The growth story stayed strong, but the risk profile widened.
02 Business model

Systems first, chips second

NVIDIA designs the chips, but the deeper business is the platform around them. A customer buying NVIDIA for AI usually needs GPUs, CPUs, networking, systems software, and developer tools. CUDA is the software layer that makes it hard to switch, because much of the AI world already writes code for it.

The model is becoming more capital heavy. NVIDIA has invested $10 billion in Anthropic and made $18.6 billion of additional Q1 FY2027 investments in private companies and infrastructure funds. Some of those companies may indirectly buy or use NVIDIA products in the cloud. That can help secure future demand, but it also ties the balance sheet to the health of the AI startup and infrastructure cycle.

NVIDIA is also using intellectual property deals to add technology faster. The Groq licensing agreement gives access to low-latency inference technology, which means technology aimed at making AI answers faster after a model is trained. The open question is whether this improves Rubin's cost and power efficiency enough to matter at scale.

03 Product portfolio

What NVIDIA sells

Growth engine

Data-center GPUs

Hopper, Blackwell, and Vera Rubin accelerators train and run AI models. This is the product line the whole thesis depends on.

Growth engine

Networking

NVLink, InfiniBand, and Spectrum-X connect thousands of chips so they act like one giant computer. Networking is a key reason NVIDIA sells systems, not only chips.

Option

Vera CPUs

The standalone Vera CPU is NVIDIA's push into a data-center CPU market management says is worth $200 billion. It puts NVIDIA closer to x86 and ARM incumbents.

Cash cow

CUDA and AI software

CUDA, NVIDIA AI Enterprise, and NIMs help developers run AI workloads on NVIDIA hardware. This software layer is the moat that slows switching.

Steady

GeForce gaming

GeForce RTX cards serve PC gamers and creators. It is still a strong brand, but it is now much smaller than Data Center.

Steady

Professional visualization

RTX workstation products serve designers, engineers, manufacturers, and AI developers. The RTX PRO 5000 Blackwell workstation extends this line.

Option

Automotive and robotics

NVIDIA DRIVE, AI Cockpit, robotics tools, and models like Alpamayo aim at physical AI. This is a long-term bet that cars and robots need the same kind of AI compute.

04 Business segments

The data-center company

Data Center92%growing fast
Edge Computing8%modest

Q1 FY2027 mix uses NVIDIA's new two-platform reporting structure. Data Center was $75.2 billion of $81.6 billion in total revenue, and three direct customers recently represented 21 percent, 17 percent, and 16 percent of total revenue.

05 Risk factors

What could break it

AI spending digests

High impact · Medium odds

NVIDIA's revenue is tied to other companies' capital spending. If cloud companies or national AI projects decide the return is too slow, orders can pause quickly. Data Center is over 90 percent of revenue, so a slowdown would hit the whole company.

We watchHyperscaler capital spending guidance, AI application revenue disclosures, and any pause in large AI cluster orders.

Customer concentration and custom chips

High impact · Medium odds

Three direct customers recently represented more than half of total revenue. These customers still need NVIDIA for frontier AI, but they are also building their own chips. The risk is not that NVIDIA loses everything. The risk is that more everyday inference work moves to cheaper custom silicon.

We watchCustom chip usage at large cloud providers, especially for inference workloads.

China and export controls

High impact · High odds

U.S. export controls have already damaged NVIDIA's China data-center business, including a $4.5 billion H20-related inventory charge in fiscal 2026. China-based competitors are also getting stronger while NVIDIA is restricted. That could create a separate AI hardware market outside NVIDIA's reach.

We watchU.S. export license rules, H20 shipment updates, and Chinese AI chip adoption by local cloud and model companies.

China antitrust ruling

Medium impact · Medium odds

China's antitrust regulators issued a preliminary finding that NVIDIA's compliance with U.S. export controls violated terms from the Mellanox acquisition approval. A final adverse ruling could bring penalties or business limits. This is separate from the export rule problem and could add more pressure.

We watchFinal determination from China's antitrust regulator on Mellanox acquisition compliance.

Ecosystem investment losses

Medium impact · Medium odds

NVIDIA is investing large sums to support AI partners and infrastructure funds. The Q1 FY2027 filing disclosed $18.6 billion of investments in private companies and infrastructure funds. This can help lock in future workloads, but weak AI startup returns could lead to write-downs.

We watchQuarterly investment disclosures, impairment charges, and workload commitments tied to partner investments.
06 Quick answers

In one breath

What does NVIDIA actually sell?

NVIDIA sells AI computing platforms. That includes GPUs, CPUs, networking, systems software, and tools developers use to train and run AI models.

Why is NVIDIA so important to AI?

Modern AI needs huge amounts of parallel computing. NVIDIA's chips do that work well, and its CUDA software makes it easier for developers to stay on NVIDIA systems.

What is NVIDIA's biggest risk?

The biggest risk is a slowdown or shift in AI infrastructure spending. Data Center is over 90 percent of revenue, so the company is very exposed to one buildout cycle.

What is sovereign AI?

Sovereign AI means national AI infrastructure built or backed by governments. For NVIDIA, it is important because it adds demand beyond the largest cloud companies.