Preview of Nvidia GTC Conference: Integrating Groq technology to aggressively attack inference chips, Samsung’s first OEM production, OpenAI may become the first batch of customers

# Ye Zhen

Source: Wall Street Insights

As NVIDIA’s GTC conference is about to kick off, three key signals may reshape the AI industry landscape: a major push into the AI inference market by integrating Groq technology; the possible first-time introduction of Samsung as a chipmaker to break TSMC’s monopoly, with OpenAI expected to be among the first major customers; and potential further expansion of robotics physical AI and open-source model ecosystems.

Local time on Monday, the highly anticipated annual NVIDIA GTC developer conference will open in San Jose, California, USA. CEO Jensen Huang’s keynote speech has long been regarded as a key bellwether for the AI industry.

For investors, the most noteworthy aspects of this year’s conference are NVIDIA’s strategic shift from training to inference and adjustments to its supply chain layout.

Based on reports from media including The Information and The Deep View, three critical signals may emerge at GTC:

First, NVIDIA may make a major foray into the AI inference market by integrating Groq technology.

Second, chip manufacturing may partially shift from TSMC to Samsung in the short term.

Third, the robotics physical AI and open-source model ecosystems are expected to expand further.

## Pushing into the Inference Market, with Groq Chips as a Key Tool

The AI industry is gradually shifting from “training-first” to “inference-driven”. In the training sector, NVIDIA has built strong technological and ecological advantages with GPUs. But in the inference market, competitors such as Cerebras are gaining share with faster and lower-cost solutions.

Against this backdrop, the market is paying close attention to NVIDIA’s response. According to The Information, Jensen Huang is expected to announce a new chip system combining NVIDIA and Groq technologies at the conference. This follows NVIDIA securing a roughly $20 billion technology license from Groq in late 2024.

Groq-developed chips, known as LPUs (Language Processing Units), are optimized specifically for inference workloads. This will also mark the first time NVIDIA has directly integrated another company’s AI processor into its server rack architecture.

## Supply Chain Restructuring and First Major Customers Confirmed

The manufacturing and commercial deployment of the new system are also in the spotlight for capital markets.

According to The Information, the Groq LPU is expected to be manufactured by Samsung in the second half of 2025. This arrangement is highly significant, as it would be the first time NVIDIA server chips are produced by a foundry other than TSMC, breaking its long-term reliance on a single supplier.

However, sources cited by the above media noted that this shift may be mostly phased. As the next-generation LPU requires tighter integration with NVIDIA’s future AI chips, production may later return to TSMC.

On the demand side, NVIDIA is expected to announce OpenAI as one of the first customers for the system. The chip system could be used to power AI agents performing tasks such as coding.

## Underlying Architecture Changes and Future Technology Roadmap

For investors focused on chip technology, the architecture of the NVIDIA-Groq system also reveals potential integration challenges and opportunities.

According to The Information, the new rack structure will differ significantly from existing systems: each rack will house 256 Groq chips. Meanwhile, Intel processors will handle communication management within the system. This design suggests that NVIDIA’s current architecture has not yet fully merged with the LPU.

NVIDIA clearly has longer-term plans, however. According to two sources involved in development, cited by The Information, the company is internally exploring deeper integration of the LPU into its future product roadmap. One proposal is to merge the Groq processor with the next-generation Feynman GPU (the successor to the Rubin architecture) into a single chip to improve performance and lower overall costs.

## Physical AI and Expansion of Open-Source Model Ecosystem

Beyond computing infrastructure, the AI application ecosystem is also a major highlight of the GTC conference.

As “AI Moore’s Law” continues — with computing efficiency roughly doubling every four months — NVIDIA’s布局 in robotics and physical AI is drawing intense attention. Especially amid the accelerated development of the humanoid robot industry in China, whether NVIDIA and its partners can provide more cost-effective solutions for scenarios such as autonomous driving will be a key market focus.

At the same time, NVIDIA is moving rapidly in open-source models. The company previously released the Nemotron 3 Super model with 120 billion parameters and announced plans to launch Nemotron 4 Ultra with four times the parameter count. Improved model capabilities are expected to further reduce enterprise AI inference costs and enhance overall return on investment.

The signals released at this GTC may largely shape the AI industry landscape in 2026.

---

## Risk Warning and Disclaimer

The market is subject to risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this article is at your own risk.