The "AI Bull Market Narrative" is once again causing a huge wave! Jensen Huang unveils a trillion-dollar AI blueprint. NVIDIA (NVDA.US) sets sail for a $6 trillion market cap.

NVIDIA CEO Jensen Huang showcased NVIDIA’s “unprecedented AI compute revenue-generation super blueprint” in the AI compute infrastructure sector at the GTC conference in the early hours of March 17 Beijing time. He told global investors that, driven by strong demand for Blackwell-architecture GPU compute power and even more explosive demand for the Vera Rubin architecture-based AI compute system, which is set to enter mass production, his company’s future revenue scale in the AI chip field could reach at least $1 trillion by 2027—far above the $500 billion AI compute infrastructure blueprint for 2026 that was outlined at the previous GTC.

To analysts at firms such as Goldman Sachs, Wedbush, and Morgan Stanley who are bullish on NVIDIA’s stock price prospects, underpinned by an even stronger-than-expected outlook for revenue growth, NVIDIA’s market value is poised to break the $5 trillion threshold again after last October, and it is also very likely to surge toward historical peak levels that are far higher than current ones.

For NVIDIA’s stock, it may soon set a new all-time high again and lift the global AI compute industry chain onto a new upward trajectory. And NVIDIA’s multitrillion-dollar AI compute super blueprint is doing everything it can to prop up the “AI bull market narrative” as the main storyline in capital markets. Based on Wall Street analysts’ average target price, this implies NVIDIA’s market cap will exceed $6 trillion within the next 12 months. Wall Street’s most optimistic view is as high as $8.8 trillion in total market capitalization.

When model scale, inference pipelines, and multimodal/agentic AI workloads drive compute consumption to expand exponentially, the main thrust of tech giants’ capital expenditures is even more inclined to concentrate on AI compute infrastructure. Global investors, too, continue to anchor the “AI bull market narrative,” centered on NVIDIA, Google TPU clusters, and AMD’s new product iterations and expected delivery of AI compute clusters, as one of the most certain cyclical investment narratives in global stock markets. At the same time, it also means investment themes closely tied to AI training/inference—such as power, liquid-cooling heat dissipation systems, and optical interconnect supply chains—will, as uncertainty arises in Middle East geopolitical situations, still remain among the hottest investment camps in the stock market, alongside AI compute leaders such as NVIDIA, AMD, Broadcom, TSMC, and Micron.

At the annual GTC Developer Conference held in San Jose, California, CEO Jensen Huang unveiled a new central processing unit (a data-center server-class CPU), as well as an LPU AI inference compute infrastructure system built on proprietary AI inference architecture technologies from Groq. Groq is an AI chip startup; NVIDIA licensed its technology from Groq in December last year for $17 billion.

These moves are part of Huang’s efforts to consolidate the company’s position in the so-called “inference computing” space. So-called inference computing refers to the entire large-scale compute process involved in responding to query requests from global B-side and C-side users. In this area, NVIDIA’s AI GPU compute stack is facing increasingly intense competition from CPU-based solutions and custom AI ASIC processors developed by companies such as Google (i.e., Google’s TPU-leading AI ASIC technology route). In recent years, NVIDIA chips have dominated the AI large-model training stage, which has also been a major focus for the market.

On the AI training side, where NVIDIA’s AI GPU is nearly monopolized, the need is for stronger general-purpose AI compute cluster capability and rapid iteration across the entire compute stack. On the AI inference side, after frontier AI technologies are scaled and deployed, unit token cost, latency, and energy efficiency matter more.

“The age of AI inference is already here,” Huang said at the GTC conference. “And inference demand continues to rise,” he added.

Dressed in his signature black leather jacket, Huang delivered a speech inside an ice hockey arena that can seat more than 18,000 people. This four-day tech conference has already become one of the largest platforms globally for showcasing AI technology. “I just want to remind you that this is a highly anticipated technical conference,” he told the audience.

The AI inference wave is coming; NVIDIA’s “AI compute blueprint” jumps to a trillion-dollar scale

If you compress Huang’s GTC remarks into a single sentence, the core message is: NVIDIA is completely restructuring itself from “a company that sells AI GPUs” into a “chip powerhouse that sells AI factories.” The official keynote opened with the idea that tokens are the basic unit of modern AI. Huang pushed the industry’s main storyline from “training” to “inference + agentic AI,” and raised the AI infrastructure revenue opportunity for 2025–2027 from the previous $500 billion to at least $1 trillion. This isn’t just a simple increase in demand—it’s telling the capital markets that in the future, compute competition won’t only look at peak training FLOPS; it will come down to who can keep producing tokens at the lowest cost, with the highest level of data throughput, and with the best (lowest) latency on a continuous basis.

Under the narrative of expanding AI compute demand, Huang’s underlying commercial logic is very clear: data centers are no longer “storage centers,” but “AI factories.” Under a fixed power budget, the most critical metrics aren’t single-chip peak performance, but “tokens per watt, cost per token, time to first production.” This is also why he repeatedly emphasized “extreme codesign”—optimizing compute, networking, storage, software, power delivery, and cooling as an integrated whole. According to the official figures, compared with the Blackwell platform, the Vera Rubin NVL72 can achieve up to 10x inference throughput per watt and reduce single-token cost to just one-tenth. For training large-scale MoE models, the required GPU count can also be cut to one-quarter of the original. This isn’t just “chip iteration”—it’s a rewrite of AI infrastructure economics.

At the latest hardware level, the most important change at this GTC is that NVIDIA has formally integrated the CPU, GPU, LPU, DPU, SuperNIC, switch extreme chips, and storage architecture into a platform-level system. The official definition of the Vera Rubin platform includes Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and the newest integrated NVIDIA Groq 3 LPU. The Vera Rubin NVL72 rack consists of 72 Rubin GPUs + 36 Vera CPUs, while the Groq 3 LPX rack is designed specifically to supplement low-latency inference. In a breakthrough way, Huang breaks AI inference into two stages: prefill is handled by Vera Rubin, while decode is handled by Groq AI chips. This means NVIDIA’s answer for the inference era is no longer “have GPUs do everything,” but rather to separate high-throughput and ultra-low-latency processing through heterogeneous computing.

On the software and ecosystem front, Huang’s stance in his speech was equally aggressive. Dynamo 1.0 is defined by NVIDIA as the inference operating system for an AI factory; the official claim is that it can deliver up to a 7x inference performance improvement for Blackwell. In the agent direction, NVIDIA has rolled out Agent Toolkit, OpenShell, and NemoClaw, raised OpenClaw into an “operating system”-style platform for personal AI, and added enterprise deployment for policy control, privacy routing, and security boundaries. At the same time, NVIDIA is also expanding open large model families such as Nemotron, Cosmos, Isaac GR00T, Alpaymayo, BioNeMo, and Earth-2, and previewed the roadmap for the Feynman architecture: the next-generation platform will introduce Rosa CPU, LP40 LPU, BlueField-5, CX10, and Kyber, and continue pushing next-generation AI-factory progress with copper interconnects and co-packaged optical solutions.

Looking further outward, GTC 2026 isn’t just about the data center. NVIDIA is also bringing “physical AI” and “spatial computing” onto the main stage: IGX Thor has entered general availability, targeting industrial, medical, robotics, and edge computing along the rail frontier; the Open Physical AI Data Factory Blueprint is intended to accelerate robots, vision AI agents, and autonomous driving data generation, augmentation, and evaluation; and the Space-1 Vera Rubin Module extends the Vera Rubin architecture to orbital data centers—officially claiming that, relative to H100, it can bring up to 25x AI compute for space inference. This shows that NVIDIA has expanded the “AI factory” paradigm beyond cloud data centers into a unified infrastructure spanning cloud, edge, devices, vehicles, robots, and even space.

The true theme of GTC 2026 isn’t, in fact, a single new product launch like in the past. It’s NVIDIA putting GeForce, data-center compute infrastructure, networking, storage, inference compute systems, agent platforms, robotics, and space computing all into one unified narrative—“upgrading from a single GPU supplier into a full-solution AI infrastructure provider.” That’s why the most worth watching at this conference isn’t a single AI chip parameter, but the fact that NVIDIA is using system-level products to lock in, in advance, the token economics for the coming years, the process of monetizing inference, and even infrastructure pricing power.

AI compute infrastructure dominance is being further consolidated; does NVIDIA’s stock price point to new all-time highs?

“Investors previously had widespread concerns that big-tech companies’ massive AI infrastructure spending would be hard to sustain, but as Huang laid out a $1 trillion revenue-generation opportunity by 2027, investors began to believe NVIDIA’s AI infrastructure demand still has long-term staying power.” Emarketer analyst Jacob Bourne said. “As the entire AI industry moves from the early experimentation stage to large-scale deployment, NVIDIA continues to maintain its leading position in the AI compute market.”

When Huang at GTC lifted the size of NVIDIA’s AI chip and infrastructure opportunity through 2027 to at least $1 trillion in one shot, what the market saw wasn’t just a chip company continuing to sell stronger GPUs, but an infrastructure empire trying to define the production function of the next generation “AI factory.” From the training era to the inference era, from single-chip competition to system-level dominance of full cabinets, full networks, and full software stacks; from Blackwell and Vera Rubin to technology collaboration with Groq for low-latency decoding—NVIDIA is writing token throughput, revenue per watt, and inference monetization capability into a new valuation language.

At GTC, Huang simultaneously proved that demand is still actively expanding with the $1 trillion opportunity size, while also explaining that NVIDIA’s competitive unit is no longer a single AU chip, but the entire AI factory—supported by a complete platform of CPUs, GPUs, LPUs, high-performance network components, a software ecosystem, and an agent toolchain.

Huang’s view that “the inference inflection point has arrived” in essence tells the capital markets: AI capital expenditures have not peaked yet, and truly large-scale deployments are only just beginning. And when NVIDIA packages CPUs, GPUs, LPUs, networks, agent software, and data center economics into the same narrative, what it raises isn’t only a new product cycle—it’s a super ship that is once again heading toward the imagination space of a $5 trillion market value. According to the average target price compiled by TIPRANKS from Wall Street analysts, analysts broadly expect NVIDIA’s stock to rise to $273, which implies upside potential of a striking 51% over the next 12 months in their view. The most optimistic target price is as high as $360. A $273 target price corresponds to NVIDIA’s market cap of roughly $6.6 trillion. By the close of trading on Monday, NVIDIA’s stock price closed at $183.220, with a market cap of about $4.45 trillion.

At the conference, Huang raised the revenue opportunity for AI chips/AI compute infrastructure to at least $1 trillion by 2027, clearly higher than the previous $500 billion by 2026 framing centered on the Blackwell and Rubin architectures. After the GTC conference, Wall Street financial heavyweight Goldman Sachs said that the trillion-dollar revenue prospect in the latest GTC presentation gives the market a longer-cycle demand endorsement, which is enough to ease investors’ worries that “AI capital expenditures might peak in 2026.” In other words, Goldman’s analyst team believes this presentation isn’t just a product showcase, but a re-anchoring of NVIDIA’s future 2–3 year order ceilings and earnings sustainability.

Goldman emphasized that NVIDIA isn’t only releasing another single AI GPU with extremely strong performance; it has formally commercialized inference (inference) in a proprietary way, and upgraded NVIDIA’s AI compute infrastructure across the board into the most core equipment for the next stage of the global AI arms race. As described above, Huang split inference into two stages: prefill and decode. The former is handled by Vera Rubin, while the latter is taken over by Groq 3 LPX/LPU, meaning NVIDIA is further expanding from a “training leader” into an “AI compute inference infrastructure total-solution provider.” Goldman emphasized that the official figures exceed what the market expected: Vera Rubin + LPX can achieve up to 35x inference throughput per megawatt, and provide up to a 10x revenue opportunity for trillion-parameter models.

Goldman said NVIDIA is not just defending the training market; in the inference era constrained by power and sensitive to latency, it has put forward a stronger monetization framework and a more complete heterogeneous compute answer. Goldman is more aggressively bullish mainly because this GTC simultaneously satisfied the two things investors care about most: first, whether demand will peak; second, whether in the inference era NVIDIA’s moat will be diluted by CPUs, in-house ASICs, or other custom chips.

Goldman said this forward-looking $1 trillion guidance far exceeds market expectations, confirming that demand from the cloud computing hyperscalers remains strong and durable. Based on an optimistic assessment of potential catalysts over the coming months, Goldman reiterated its “Buy” rating for NVIDIA and maintained a 12-month target price of $250, emphasizing that the capital expenditure plans of the super cloud service providers and the new model built on the Blackwell and Rubin architectures will continue to reinforce the company’s performance leadership.