NVIDIA's Third Growth Curve: Three Businesses Not Yet Visible in Its Financials

Chapter 1: The Same Playbook—Platforms, Open Models, and Joint Laboratories

Consider three seemingly unrelated businesses side by side: autonomous vehicles, AI drug discovery, and humanoid robots. On the surface, they are three separate industries. In practice, they are the same playbook executed three times. NVIDIA does exactly the same thing in each market—it first establishes a development platform (DRIVE, BioNeMo, or Isaac), then opens a set of pretrained models (Alpamayo, Clara, or GR00T), and finally establishes a joint laboratory or flagship partnership with a leading player in the relevant field (Uber, Eli Lilly, or Hon Hai (Foxconn)). This is not a chip-selling business. It is a direct extension of the strategy CUDA originally used in data centers into the physical and scientific domains.

The narrative can therefore be drawn as a single line: data centers → inference-dedicated systems (Groq 3 LPX and Vera CPU) → physical AI and scientific AI. The first stage generates today's cash flow, the second is a transition now scaling up, and the third—the three businesses discussed in this article—represents a future that remains invisible in the financial statements.

The investment assessment must therefore begin with a qualitative distinction: these three businesses represent option value, not cash flow. Their contribution to EPS in FY2027 and FY2028 will be close to zero. Investors buying NVIDIA will not see any upside surprise from these businesses in next quarter's results. What they determine is something else: the multiple the market will be willing to assign NVIDIA after FY2030. This distinction underpins the entire article. We are not evaluating “how much these three businesses earn today,” but rather “how much each of these three options is worth and the probability that each will be exercised.”

Chapter 2: Autonomous Driving—NVIDIA Is Not Selling Autonomous-Driving Chips, but a Complete Branded-Vehicle Autonomy and Robotaxi Launch Package

The market's standard narrative for autonomous driving is “NVIDIA versus Tesla in a battle over chips.” That framework misidentifies the product. NVIDIA is not selling a single chip; it is selling a complete three-layer launch package. The same stack is sold through two different channels. For automakers, it is an end-to-end “branded-vehicle autonomy” solution that allows Mercedes, BYD, and others to offer capabilities ranging from L2+ to L4 under their own brands. For ride-hailing platforms, it is a “robotaxi launch package” that allows companies such as Uber to operate fleets without maintaining their own autonomous-driving teams. One R&D effort supports two monetization paths:

NVIDIA positions Alpamayo as “the world's first thinking autonomous-driving model”—using reasoning to address edge cases rather than exhaustively enumerating rules. This is the key point of direct competition with Tesla FSD. The release cadence of three versions in six months is itself a signal of investment intensity.

The customer list clearly divides into two groups. On the branded-vehicle side, BYD, Geely, Isuzu, Nissan, Hyundai/Kia, and VinFast are all using the stack to develop L4 systems. Mercedes-Benz is furthest ahead in production adoption, with the CLA becoming the first mass-production vehicle equipped with the complete stack. On the robotaxi side, customers include Uber, Bolt, Grab, Lyft, and TIER IV. The largest single opportunity is Uber, which is using Hyperion 10 and the end-to-end DRIVE AV software stack to expand its L4 fleet. NVIDIA states that the platform could support Uber's gradual expansion to 100,000 vehicles, beginning in Los Angeles and San Francisco in the first half of 2027 and expanding to 28 markets by 2028.

Chapter 3: AI Drug Discovery—Understand How a Drug Reaches the Market Before Assessing What AI Has Changed

Before discussing NVIDIA, this chapter must establish a foundation that most technology investors lack: what exactly must a drug go through between the laboratory and the pharmacy? Without this foundation, it is impossible to price any news about “AI-accelerated drug development” correctly.

Essential Background: The Drug-Development Funnel

The pharmaceutical industry's harsh economics begin with a funnel: of 5,000–10,000 compounds screened, approximately 250 reach preclinical testing, roughly five enter human clinical trials, and only one ultimately reaches the market. The entire process takes 10–15 years.

Once the funnel is understood, the proper interpretation of AI drug-development news becomes clear: an “AI-designed molecule entering Phase 1” still faces a 92.1% probability of failure before reaching the market. The true graveyard is Phase 2, where the question is not whether the molecule was designed elegantly, but whether the right target was selected.

What Has AI Changed? Let the Data Speak

AI primarily compresses the earliest stage of the funnel. Insilico Medicine's rentosertib, a treatment for pulmonary fibrosis, took less than 18 months and approximately US$2.6 million to progress from a target hypothesis to a preclinical candidate—a stage that traditionally requires four to five years. The drug's Phase 2a results were published in Nature Medicine in 2025: after 12 weeks, FVC improved by 98.4 mL in the 60 mg dose group, compared with a decline of 20.3 mL in the placebo group. It entered Phase 3 in July 2026, making it the world's most advanced AI-designed drug to date.

However, aggregate statistics provide a necessary reality check. An analysis published by a BCG team in Drug Discovery Today in 2024 showed that AI-discovered molecules achieved Phase 1 success rates of 80%–90%, far above the historical average of 52%. AI is demonstrably effective at creating “drug-like molecules.” However, their Phase 2 success rate is approximately 40%, comparable with traditional pharmaceutical companies. In other words, AI has passed the chemistry test but not yet the biology test. AI has not demonstrated that it is better than humans at selecting the right target—the most expensive problem in drug development. Of the 6,147 drugs in clinical development globally, only 67, or approximately 1%, were discovered using AI. DSP-1181, the first AI-designed molecule to enter Phase 1, developed by Exscientia in 2020, was discontinued after Phase 1. As of August 2026, no de novo AI-designed drug has been approved anywhere in the world.

NVIDIA's Position in This Market: Sell Picks and Shovels, Do Not Bet on the Gold Mine

Once the probability structure above is understood, NVIDIA's strategic choice becomes exceptionally clear. At the JP Morgan Healthcare Conference in January 2026, NVIDIA and Eli Lilly announced the establishment of a joint AI innovation laboratory in the San Francisco Bay Area, with combined investment of up to US$1 billion over five years. The laboratory is being built on the BioNeMo platform and Vera Rubin architecture. Before this announcement, Eli Lilly had already built a system containing 1,016 Blackwell Ultra GPUs—described as the pharmaceutical industry's most powerful AI factory. Amgen's Freyja uses 248 H100 GPUs, while Novo Nordisk's Gefion supercomputer contains 1,528 H100 GPUs. Other ecosystem participants include Genentech, Thermo Fisher, which is jointly developing autonomous laboratories, Illumina, Schrödinger, Chai Discovery, Basecamp Research, and Boltz.

How does BioNeMo itself generate revenue? The answer is: almost no direct revenue. The BioNeMo framework has been fully open source under Apache 2.0 since November 2024 and is free to use. The only publicly priced direct software revenue comes from the NVIDIA AI Enterprise license required to deploy NIM microservices in production environments—US$4,500 per GPU per year. Even this is a platform-wide license, and NVIDIA does not disclose BioNeMo's share. The real money is in hardware: every partnership above involves purchases of hundreds or thousands of GPUs. NVIDIA's vice president of healthcare has publicly stated that healthcare is already a business generating more than US$1 billion in annual revenue, with an internally estimated total addressable market of US$100 billion. However, NVIDIA does not report a healthcare segment, making these figures impossible to verify externally.

The clearest evidence of NVIDIA's actual position is how it has deployed its own capital. In July 2023, NVIDIA made a US$50 million PIPE investment in AI drug-discovery flagship Recursion, causing Recursion's share price to rise by approximately 80% that day. Two and a half years later, a 13F filing disclosed in February 2026 showed that NVIDIA had sold its entire position of 7.71 million shares, exiting with an estimated loss of approximately 50%. Meanwhile, Recursion's BioHive-2 supercomputer, built with 504 H100 GPUs, continues to operate, and its BioNeMo partnership remains unchanged. Recursion's clinical pipeline also validates the statistical reality of the Phase 2 graveyard: its two most advanced drug candidates were both discontinued in May 2025, and its best current data come from a 12-participant open-label trial. Exiting the equity investment at a loss while continuing the hardware and platform business is the clearest example of “selling picks and shovels without betting on the gold mine”—and this choice was made by the player that understands the market best.

Chapter 4: Robotics—The Earliest-Stage Business, but Deployment Has Already Begun on Taiwan-Linked Production Lines

Robotics is the earliest-stage of the three businesses. NVIDIA is again following the same playbook: Isaac GR00T Reference Humanoid is the first open humanoid-robot reference design, built on Jetson Thor and the Isaac GR00T development platform. In FY27Q2, NVIDIA extended its Halos autonomous-driving safety system into Halos for Robotics, adding an end-to-end safety layer for physical AI. Platforms, open models, and reference designs—the strategy is identical to its approach in autonomous driving and drug discovery.

There are currently no credible revenue figures for this business, so the discussion should remain appropriately brief. However, one deployment signal is worth noting: Hon Hai's (Foxconn's) AI server plant in Houston, Texas, plans to introduce humanoid robots based on NVIDIA Isaac GR00T N into its production lines. The first genuine commercial deployment scenario for robots is not the household assistant depicted in science fiction, but the production lines of the Taiwanese assembly companies discussed in the previous article in this series, “NVIDIA's Golden Decade: Is Taiwan Capturing Revenue or Profit?” The first customer for physical AI is the factory that manufactures physical AI hardware itself.

Why Are These Three Businesses Invisible in the Financial Statements?

This is the foundation of the article's analytical integrity and must be answered with figures. Autonomous driving and robotics, together with PCs, game consoles, workstations, and AI-RAN base stations, are jointly reported within the Edge Computing segment. The segment generated US$7.2 billion in FY27Q2 revenue, up 27% YoY and 13% QoQ, accounting for 7.5% of total revenue. In the previous quarter, FY27Q1, revenue was US$6.4 billion, up 29% YoY. Compared with the data center segment's 117% YoY growth over the same period, Edge Computing is NVIDIA's slowest-growing segment. The company does not separately disclose the shares attributable to autonomous driving and robotics, making external disaggregation impossible. AI drug discovery is even more indirect: its revenue takes the form of GPU purchases and cloud-computing consumption, which are blended into the data center segment without even an identifiable reporting segment.

Accordingly, the “third growth curve” is a line that can barely be drawn on today's income statement. This is precisely why the article characterizes it as an option from the outset: investors are not buying an existing growth curve, but the probability that such a curve will emerge in the future.

Conclusion: How Should These Three Options Be Valued?

Place this article back within the framework of the series. The initiating-coverage report established a three-scenario valuation framework for NVIDIA's core business, assessing the quality and durability of data center cash flow. The three businesses discussed here are the actual substance behind the “bull-case scenario” in that framework. If the market remains willing to assign NVIDIA a growth-stock multiple after FY2030, the reason will not be a seventh consecutive year of high data center growth, but that at least one of these three options has begun to be exercised. Conversely, if investors believe only in the base-case scenario, the value of these three businesses can simply be set at zero. This article provides the full analytical basis needed to determine whether that “zero” is overly conservative.