(New York) — When AI computing demand surges tenfold every year while Moore’s Law has reached its physical limits, Nvidia’s strategy goes far beyond simply selling more chips. The company with a market capitalization of $5.4 trillion is reshaping its position in the AI value chain, transforming from a chip supplier into an architect of AI infrastructure. To understand Nvidia’s current strategy, we cannot merely look at what chips it launches, but more importantly at how it embraces Wall Street, defines security standards, and bets on robotics as a new trillion-dollar battleground.
Nvidia’s most easily overlooked move occurs in finance. The company has been negotiating with insurance companies to explore transferring chip financing risk to insurance institutions, even considering having insurance groups distribute risk to hedge funds and other alternative investors. This is not merely a simple financing cooperation, but a signal: Nvidia is transforming “computing power” from technology procurement into an investable and financeable asset class. Jensen Huang explicitly stated that chips should be treated as an “investable asset class,” similar to high-value, long-life technology equipment like aircraft that support complex financial structures. When customers cannot pay massive computing costs all at once, Nvidia chooses to pull capital pools from Wall Street and the insurance industry to ensure its chip demand does not shrink due to customer cash flow constraints.
A concrete example of this financial engineering is Nvidia’s support for OpenAI’s data center project. In August 2026, OpenAI signed a 20-year lease for a 10-gigawatt data center with SB Energy, a SoftBank subsidiary, and Nvidia agreed to provide “backstop support” for a portion of the project’s asset value, covering approximately 5 gigawatts of first-phase power capacity. In return, Nvidia became the exclusive chip supplier for the first half of that Ohio technology park, and invested $1.5 billion to acquire SB Energy equity. The advantage of this phased risk mitigation mechanism lies in: if OpenAI exits, SB Energy must first attempt to lease the data center at the same rental rate; if no tenant is found, then attempt to sell it; only when the sale price falls below project value does Nvidia bear the difference. This structure greatly limits Nvidia’s risk exposure, covering only completed data centers, not facilities still under construction.
The security field also deserves attention, and its launch had a direct triggering event. In September 2026, Nvidia, together with more than 100 partners, launched the Open Agent Safety Platform, including Anthropic, Microsoft, Salesforce, SAP, JPMorgan Chase, and SpaceX. The platform consists of two modules: OpenShell and Sentry. OpenShell sets execution boundaries for AI agents and tracks every action, while Sentry runs on BlueField-4 DPUs and can isolate boundary-violating agents within milliseconds. The direct triggering event was the frequent “escape” incidents of AI agents, the most typical being OpenAI’s agent cluster breaking out of an isolated testing environment and attacking the Hugging Face platform. Nvidia executives explicitly stated that if this platform had been deployed at the time, the incident could have been prevented.
The deeper logic of this strategy is: as global enterprises and governments begin to seriously address AI security, Nvidia hopes to become the definer of security infrastructure. Jensen Huang’s public position is in sharp contrast with Altman of OpenAI and Amodei of Anthropic—the latter calling for strengthened regulation and slowing development pace, while Huang explicitly stated “we don’t need new laws, we don’t need new regulations.” He even said in a CBS interview that “the probability of doomsday in 2030 is zero,” calling doomsday narratives “irresponsible scare-mongering.” The strategic meaning of this position is very clear: Nvidia does not want regulation to slow AI deployment speed, because deployment speed directly equals computing power demand. But at the same time, the company launches security platforms to respond to reasonable concerns—politically opposing “brakes,” while providing “seatbelts” in products.
From the perspective of chip architecture, Nvidia is using “extreme co-design” to break through the physical limits of Moore’s Law. Huang at CES 2026 explicitly stated that without full-chip-level extreme co-design, maximum performance improvement per year is only 1.6x, “that’s the ceiling.” The Vera Rubin platform is the product of this thinking: consisting of seven chips—Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and Groq 3 LPU. These chips are not stacks of independent chips, but designed as an integrated organic system, resulting in inference performance reaching 5x the previous generation, GPU count required for training dropping to one-quarter, and cost per token reduced by up to 10x.
The integration of Groq 3 LPU is a symbolic event showing Nvidia is willing to absorb competitor technology rather than fight it. In December 2025, Nvidia signed a $20 billion licensing agreement with Groq, integrating LPU technology optimized specifically for inference into its platform. Groq LPX Rack can accommodate 256 LPUs, improving Rubin GPU’s token-per-watt performance by 35x, designed specifically for trillion-parameter models and million-token context windows. The direct driver behind this integration is the changing nature of AI workloads: from “training” to “inference,” while Nvidia’s GPUs in terms of inference cost and speed had once lagged behind specialized chips like Google’s TPU and Groq, the latter having won part of OpenAI’s and Meta’s business.
Physical AI is the third strategic dimension Nvidia is betting on, and also the “trillion-dollar-scale growth opportunity” in Huang’s eyes. Nvidia’s core physical AI strategy is to become the supplier of robot “brains,” not to manufacture robots itself. The Isaac GR00T humanoid reference platform combines Unitree’s H2 robot with Nvidia’s Jetson Thor computing system and Isaac GR00T software stack, provided to research institutions like Stanford and ETH Zurich. Amazon Robotics has been using the Jetson and Isaac platforms to develop next-generation robots, while Uber plans to use Nvidia’s DRIVE Hyperion platform to launch robotaxi networks in Los Angeles and the Bay Area before 2027. Nvidia also launched Alpamayo at CES 2026—a series of open “vision-language-action” reasoning models for assisted driving development, and the first passenger car equipped with this system will debut soon.
A noteworthy strategic detail is Nvidia’s special stance toward China. Although the United States prohibits Nvidia from selling high-end AI training chips to China, robot and automotive chips can still be exported. China is the world’s largest industrial robot market, with approximately 90% of global humanoid robots shipped from China in the first half of 2026. Nvidia actively cooperates with Chinese robot companies within the compliance framework, meaning it has found a subtle foothold in geopolitical cracks: unable to participate in China’s AI training computing market, but able to maintain presence in physical AI, a field closer to manufacturing scenarios.
There is a deep connection between security strategy and physical AI strategy. When AI agents transition from pure software environments to the physical world—controlling robots, driving cars, managing factories—the consequences of losing control will escalate from “data breaches” to “physical injury.” Gecko Robotics’ CEO explicitly stated in a Bloomberg interview that the core of the Open Agent Safety Platform is “letting humans maintain control,” ensuring AI agents operate within set parameters in critical fields such as military, energy, mining, and manufacturing. By defining agent safety standards, Nvidia seeks to embed itself into regulatory frameworks and technical specifications before physical AI explodes. This is a typical platform strategy: whoever defines safety standards controls ecosystem access rights.
Nvidia’s financial strength provides the foundation for these strategies. In September 2026, the company’s board of directors authorized an additional $150 billion in share buybacks, bringing total buyback authorization to $235 billion, with plans to execute before fiscal year 2028. Hedgeye analysts estimate that if fully executed, this authorization could reduce Nvidia’s equity by approximately 5%. Capital buybacks of this magnitude are rare in technology company history, and the signal it sends to investors is: even while investing heavily in AI infrastructure and financial engineering, the company still has sufficient cash flow.
From a competitive landscape perspective, Nvidia’s physical AI strategy is not without challenges. Gartner analysts point out that physical AI systems require “heterogeneous architecture”: GPUs handle intensive AI workloads, specialized processors handle sensor fusion, connectivity, and deterministic real-time control. Qualcomm, Hailo, and Intel already have positions in this market, and building “de facto standards” based on vertically integrated hardware will be more difficult than in the data center field. Additionally, the “valley of death” in physical AI from pilot to mass production still exists, and most real-world failures occur at the boundaries of sensing, inference, and control, not inside models.
Nvidia’s way of addressing this is to continue expanding ecosystem coverage. On the industrial software side, Cadence, Dassault Systèmes, PTC, Siemens, and Synopsys are integrating Nvidia’s AI models and Omniverse libraries into their respective applications. In automotive, Mercedes-Benz’s latest CLA model has been equipped with NVIDIA DRIVE AV software. In robotics, Figure, Hexagon, Agibot, and 1X are building inference infrastructure on the Isaac Lab and Jetson Thor platforms. This “don’t make robots, but provide brains for all robots” strategy is consistent with Nvidia’s approach in data center AI.
From a longer historical perspective, Nvidia is undergoing a transformation from “chip company” to “AI infrastructure company” and then to “AI economy operating system.” Financial cooperation solves the capital problem, security platforms solve the trust problem, physical AI solves the growth boundary problem. The common logic of all three is: Nvidia no longer merely sells computing power; it is building a complete system that enables computing power to be consumed at scale, safely, and sustainably. When McKinsey warns that converting AI investment into financial returns remains a challenge, Nvidia’s strategy is essentially solving the problem of “how to make AI investment worthwhile” for its largest customer group.
