WHY WE OWN IT

The builder of the AI economy

We think of NVIDIA less as a chip company and more as the general contractor of the AI economy. Jensen Huang has spent three decades on one bet, accelerated computing. Today more than 11 million developers build on NVIDIA’s platform (NVIDIA Blog, GTC Berlin keynote live updates, 21 Oct 2026). Data Center revenue reached $89.0 billion in the quarter to 26 July 2026, up 117% from a year earlier (NVIDIA Form 10-Q, quarter ended 26 Jul 2026). Physical AI alone brought in $10 billion in the first half of 2026, Huang says (Published transcript of the BG2 episode (Druckfin)).

“Building the future” is our framing, not his. But his own description of the job is just as broad: NVIDIA is “actually in warehouse and logistics, healthcare, automotive, telecom, every industry on the planet” (Published transcript of the BG2 episode (Druckfin)).

LAYER 01 · THE FUTURE

Jensen’s five-layer cake

Huang’s mental model is that AI is not an app but industrial infrastructure, built in five layers. Every application on top pulls on everything below, all the way down to the power plant (Jensen Huang, “AI Is a 5-Layer Cake”, NVIDIA Blog, 10 Mar 2026).

  1. Applications Where the economic value is created: drug discovery, robotics, legal copilots, self-driving cars.
  2. Models Not just language: biology, chemistry, physics, robotics and the physical world.
  3. Infrastructure Land, power delivery, cooling, construction, networking and the systems that make thousands of processors act as one machine. He calls these AI factories.
  4. Chips Processors that turn energy into computation: parallelism, high-bandwidth memory, fast interconnects.
  5. Energy Power generated in real time to produce intelligence in real time. Huang calls it the binding constraint.

Shown top to bottom: applications sit on models, models on infrastructure, and so on down to energy.

Wording varies by talk. At Davos he split layer two as “chips and computing infrastructure” and layer three as “cloud data centers” (NVIDIA Blog on Huang at Davos, Jan 2026). In Berlin he gave the list as energy, compute, AI factories, models and applications (Stocktwits via Yahoo Finance, 21 Oct 2026). The definitive written version is his 10 March 2026 blog post (Jensen Huang, “AI Is a 5-Layer Cake”, NVIDIA Blog, 10 Mar 2026). NVIDIA sells into the middle three layers and increasingly finances, or owns stakes in, the other two.

LAYER 02 · THE BUILDER

More than a chip company

Half the business is no longer the big four clouds

In the July quarter, Hyperscale revenue was $48.7 billion. The AI Clouds, Industrial & Enterprise segment (AI clouds, sovereign customers, enterprises and AI natives) was $40.3 billion, up 138% year on year (NVIDIA Form 10-Q, quarter ended 26 Jul 2026). On BG2, Huang put on-prem and neocloud customers at about half of NVIDIA’s business (Published transcript of the BG2 episode (Druckfin)). That matters for the bear case: even if the largest clouds shift more work to their own chips, NVIDIA is widening its customer base.

From selling GPUs to selling whole AI factories

NVIDIA now sells the entire plant: GPUs, CPUs, NVLink, networking, rack-scale systems, CUDA-X libraries, models and reference designs. In Berlin it added DSX POD, a pre-built Vera Rubin AI factory in modular containers that can cut on-site deployment from six months to 14 days once utilities are in place (NVIDIA Blog, GTC Berlin keynote live updates, 21 Oct 2026). Each sale is now a factory, not a part.

Making compute financeable

In August 2026 NVIDIA signed memorandums of understanding with six large investors (KKR, Goldman Sachs, Apollo, Brookfield, BlackRock and Blackstone). The aim is independent financing platforms that, over time, could channel upward of $500 billion from outside investors into AI infrastructure (NVIDIA press release, 10 Aug 2026). These are MOUs, not commitments, and they may not lead to definitive agreements (NVIDIA Form 10-Q, quarter ended 26 Jul 2026). The idea: if AI compute can be underwritten like a toll road or a power plant, governments, universities and regional clouds can afford it too.

Cheaper tokens, more demand

Falling costs per unit of AI can grow the market rather than shrink it, as happened with storage and bandwidth. Huang expects the cost of intelligence to keep falling every year while usage compounds, because agents that do useful work get given more work (Published transcript of the BG2 episode (Druckfin)). NVIDIA doesn’t need today’s token prices to last. It needs consumption to grow faster than unit costs fall.

Moving up into models

NVIDIA ships open model families: Nemotron for language and agents, Cosmos for world models, GR00T for humanoids, Alpamayo for vehicles, and more (NVIDIA Blog, GTC Berlin keynote live updates, 21 Oct 2026). It has agreed to buy Hugging Face for about $12.93 billion and promised to keep the platform open to other chips and clouds (NVIDIA Form 8-K, 2 Sep 2026) (Jensen Huang, NVIDIA Blog, 3 Sep 2026). The goal isn’t to beat the frontier labs. It’s to make NVIDIA the easiest place to build, tune and deploy any model.

Physical AI is the next market

Robots, vehicles, factories and medical devices need three computers: one to train, one to simulate (a digital twin) and one inside the machine. NVIDIA sells all three (Published transcript of the BG2 episode (Druckfin)).

LAYER 03 · THE PORTFOLIO

What NVIDIA owns

NVIDIA has quietly become one of the largest corporate investors in tech. As of 26 July 2026 it held $99 billion of equity investments, plus $25 billion more in commitments (NVIDIA Form 10-Q, quarter ended 26 Jul 2026). Two years earlier the figure was about $2.2 billion (Business Insider, 4 Sep 2026) (CNBC, 4 Sep 2026). Roughly half is listed stock and half private holdings, plus about $3.3 billion in infrastructure financiers (NVIDIA Form 10-Q, quarter ended 26 Jul 2026) (The Motley Fool, 13 Sep 2026). Its June 2026 13F listed eight public positions worth about $63.4 billion (NVIDIA Form 13F, holdings at 30 Jun 2026).

NVIDIA says it invests “to enhance our growth opportunities, cultivate our ecosystem, and strengthen our competitive position” (NVIDIA Form 10-Q, quarter ended 26 Jul 2026). Its CFO told analysts it has put “nearly $50 billion” into frontier AI labs (CNBC, 4 Sep 2026). The pattern: back the companies that create AI demand, and invest directly in the bottlenecks (optics, networking, chip design, foundry capacity).

CompanyWhatWhenTerms disclosedValue at 30 Jun 2026Sources
Intel (INTC)Common stock + joint CPU/PC productsSep 2025$5.0B, 214.8M shares at $23.28≈$30.0BIntel 8-K; NVIDIA press release; NVIDIA Form 13F
SpaceX (SPCX)Class A stock; stake traces back to NVIDIA’s xAI investment—Cost basis not disclosed≈$21.0B (122.8M shares)NVIDIA Form 13F; The Motley Fool
CoreWeave (CRWV)AI cloud; latest tranche Jan 2026Jan 2026$2.0B tranche, 22.9M shares at $87.20; earlier cost not fully disclosed≈$4.7B (47.2M shares, total)CoreWeave 8-K; NVIDIA Form 13F
Coherent (COHR)Optics for AI data centersMar 2026$2.0B, 7.79M shares at $256.80≈$3.1BCoherent 8-K; NVIDIA Form 13F
Nokia (NOK)AI-RAN and data-center networking partnershipOct 2025$1.0B, 166.4M shares at $6.01 (2.9% stake)≈$2.2BNokia stock exchange release; NVIDIA Form 13F
Synopsys (SNPS)Chip-design and simulation software partnershipDec 2025$2.0B, 4.82M shares at $414.79≈$2.15BSynopsys 8-K; NVIDIA Form 13F
Nebius (NBIS)AI cloudMar 2026$2.0B announced13F lists 1.19M common shares (≈$0.33B)NVIDIA press release; NVIDIA Form 13F
Lumentum (LITE)Photonics for AI networkingMar 2026$2.0B announcedNot in the 13F listReuters
Generate BiomedicinesAI drug discovery—Not disclosed≈$14MNVIDIA Form 13F
OpenAIFrontier lab; part of a $110B roundFeb 2026$30BPrivateOpenAI
AnthropicFrontier lab; compute partnershipNov 2025Up to $10BPrivateAnthropic
Safe SuperintelligenceFrontier lab; long-term partnershipJul 2026$5B reported (NVIDIA didn’t state the amount)PrivateNVIDIA press release; Reuters
Hugging FaceOpen-model platform: an acquisition, not a stakeAgreed 2 Sep 2026≈$11.9B plus up to ≈$1.0B retention (≈$12.93B); expected to close H1 2027PendingNVIDIA Form 8-K; Jensen Huang
Doctolib, Pigment, Genesis AI, EntalpicFrench health, enterprise software, robotics and AI-for-chemistry startupsOct 2026Minority stakes; amounts not disclosedPrivateTech.eu

Values are taken from the 13F at 30 June 2026 (filed 14 Aug 2026). They are paper marks that change daily and are not realized gains. We show cost only where a filing or press release discloses it. Private stakes have no public mark. Some figures, such as SSI’s $5B, come from news reports rather than NVIDIA.

THE FLYWHEEL

Selling the factory and owning the tenants

CoreWeave shows the loop at work. NVIDIA supplies the GPUs, invests in the company, CoreWeave raises more capital, builds more AI factories and buys more NVIDIA systems. If those factories rent to real outside customers, NVIDIA earns twice: on the hardware and on the equity (CoreWeave 8-K, 23 Jan 2026). The Nokia, Coherent and Synopsys stakes follow the same logic, each sitting in a different layer of the cake.

The fair objection is circularity. If NVIDIA’s money simply comes back as NVIDIA revenue with no end customer, the growth is hollow. The test we apply is whether real external demand sits at the end of the loop: paying users, enterprise contracts and utilization. We watch this every quarter.

LAYER 04 · SIGNALS

FedEx, AT&T and the telcos: what Jensen actually said

FedEx and AT&T (BG2 Pod, 15 Oct 2026). Asked about physical AI, Huang said: “I am super excited for FedEx; they are obviously going to be an AI company, a technology company.” On telecom: “The entire telecommunications industry will be reinvented again into a technology industry… I am delighted to work with AT&T and Verizon and partners around the world as they reinvent themselves.” (BG2 Pod with Jensen Huang, 15 Oct 2026 (video)) (Published transcript of the BG2 episode (Druckfin))

Telcos and edge AI (GTC Berlin, 21 Oct 2026). Six days later he made the edge argument explicitly. He cited about 100,000 telco sites and about 100 GW of distributed base stations and data centers, said NVIDIA and Nokia had made the first commercial phone call on T-Mobile’s network, and joked “Now CUDA makes phone calls” (NVIDIA Blog, GTC Berlin keynote live updates, 21 Oct 2026). Stocktwits quotes him as predicting that “every single telco in the future… will in the future evolve into a distributed AI network” (Stocktwits via Yahoo Finance, 21 Oct 2026).

What that does and doesn’t say. Huang named FedEx as a future AI company and AT&T (with Verizon) as a partner reinventing telecom. He didn’t say AT&T has pivoted into an AI company, and on BG2 he didn’t tie it to edge compute. The edge case came separately, in Berlin, and the operator he showcased there was T-Mobile, which is piloting NVIDIA AI-RAN with Nokia (DCD, 17 Mar 2026). AT&T’s own recent AI moves have mostly been open telco models built with GSMA, AMD, Microsoft and others (GSMA Open Telco AI announcement, 2 Mar 2026) (Computer Weekly, 27 Jul 2026). The direction is right; the details matter. More in The Future: Edge compute.

RISKS

Risk register

  • Custom silicon. The largest customers are building their own accelerators. We see this as the biggest long-term pressure on margins.
  • Circular financing and credit exposure. NVIDIA invests in customers and backs financing structures. In August 2026 it agreed guarantees capped at $105 billion to support data-center leases for an OpenAI affiliate at SB Energy’s Ohio campus (NVIDIA Form 10-Q, quarter ended 26 Jul 2026).
  • Portfolio volatility. A $99 billion book can swing earnings. About $36.9 billion of the listed holdings were under short-term lock-ups at 26 July (NVIDIA Form 10-Q, quarter ended 26 Jul 2026). Some bets will lose money.
  • Overbuild. If AI spending runs ahead of paying demand, orders can slow sharply. History rhymes: see Canals and railroads.
  • Power and physical limits. Huang names land, power and shell as the main constraints for 2027 (Published transcript of the BG2 episode (Druckfin)).
  • Regulation and geopolitics. Export controls (Data Center Hopper shipments to China were under 1% of Data Center revenue last quarter (NVIDIA Form 10-Q, quarter ended 26 Jul 2026)), antitrust scrutiny of a company present in every layer, and the Hugging Face deal still needing approvals (NVIDIA Form 8-K, 2 Sep 2026).
CONVICTION CHECK

What would change our mind

  • Hyperscaler custom chips taking a clearly rising share of new AI capex, at falling NVIDIA gross margins.
  • AI cloud and lab customers whose growth turns out to depend on NVIDIA money rather than outside paying demand.
  • Large write-downs in the investment book, or guarantees being drawn.
  • A durable stall in Data Center revenue growth outside the hyperscalers.
SOURCES

Sources

Last reviewed: 22 Oct 2026. Quotes are from the linked sources; the BG2 quotes are taken from a published transcript of the episode.

Not investment advice. This page is opinion and education, researched and presented with the help of AI analyst personas, and may contain errors. The EarlyConviction team holds a position in NVDA (and in SPCX, which NVIDIA also holds) and may buy or sell at any time. We never disclose position sizes. Read the disclaimer.