Nvidia Says US$1T by 2027, Here’s the Hardware Shift It’s Banking On

Nvidia’s US$1T Goal Just Raised the Stakes for Inference

Nvidia CEO Jensen Huang has just held the company’s biggest annual conference, GTC, and announced a major new product aimed at the next shift in how AI is being used.

If we look at the AI lifecycle, it feels like we are moving from the first phase, which was focused on training and building AI models, into the next phase, where those models are actually being used inside real products. That next phase is inference, which is the computation that happens when AI generates an answer, completes a task, or responds to a user query. In simple terms, training is what teaches the model, while inference is what makes it useful.

Nvidia has been the clear leader in training, building specialised GPUs and AI racks designed to process huge datasets and learn patterns at scale. But as AI products move into the mainstream, inference is becoming the more important workload because it happens constantly. Every time someone asks ChatGPT a question or uses an AI tool to generate an output, that is inference.

That matters because current GPUs are not perfectly optimised for this shift. They are extremely powerful, but they also consume a lot of energy and are not always the most efficient option for fast, high-frequency inference workloads.

Nvidia is now responding to that change with a new product, the Nvidia Groq 3 LPX rack, which combines 72 Vera Rubin GPUs, Nvidia’s newest chip, with 256 Language Processing Units from Groq, a startup Nvidia has backed with US$20B of capital.

The LPU chip is specifically designed for inference. It is built for speed, memory efficiency, and rapid response to AI queries. Together, the system can generate 700 million tokens per second, which is around 350x faster than Nvidia’s Hopper generation from two generations ago. It also has 500x more high-bandwidth memory than Hopper.

Why Would Hyperscalers and Large Corporations Want This?

Hyperscalers like Microsoft Copilot and Google Gemini handle millions of queries every day, so speed and cost per query are everything. Faster inference means more users can be served for every dollar of hardware deployed, which directly improves unit economics.

At 700 million tokens per second, this system allows operators to run far more AI workloads on the same physical footprint.

So these hyperscalers would benefit from better unit economics per query, but this would be another high upfront capex cost for a long term profitbailty play if its products have high demand.

What’s the upside for Nvidia

Inference is a much larger and more recurring revenue opportunity than training because every AI query becomes a billable compute event. That extends Nvidia’s moat beyond just building the infrastructure to train AI and pushes it deeper into the ongoing usage layer of the market.

The US$1T Blackwell and Rubin sales target by the end of 2027, which is double prior guidance, shows enormous confidence in demand. But it also highlights the scale of what Nvidia is aiming for. With annual 2025 revenue at US$130B, that target implies a massive step up in growth over a very short period.

What was also notable is that the acquisition of Groq looks very strategically accretive. Nvidia has effectively bought the company building the inference chip and is now integrating that capability directly into its own products.

The risks for Nvidia

The US$20B Groq deal now needs to deliver. It is a major bet that LPUs can become the standard chip for inference, which means Nvidia is not just backing a product, it is backing a whole shift in AI hardware architecture.

Competition in this market is also getting more intense. AMD is pushing harder, Google has its own TPUs, Amazon is developing Trainium and Inferentia, and a growing group of inference-focused startups are targeting the same opportunity.

At the same time, customers increasingly want to reduce their dependence on Nvidia. That makes the software layer more important, and the coalition of open-source AI companies Nvidia announced can be read as a move to lock in that layer before competitors do.

The US$1T revenue target by 2027 is enormous. Any macro slowdown, capex pullback, or tighter geopolitical restrictions on chips could make that target much harder to achieve.