AI Industry Shifts Focus to "Inference," Can Nvidia's "Trillion-Dollar Expectations" Win Over the Market?

[Global Times Special Correspondent Feng Yaren, Global Times Reporter Yang Shuyu, Special Correspondent Ren Zhong] As the global artificial intelligence (AI) competition heats up, chip giant NVIDIA once again becomes the focus of technological rivalry. On the 16th, local time, NVIDIA’s annual GTC conference kicked off in San Jose, California. This industry event attracted tens of thousands of tech professionals and prompted media to deeply examine the race for computing power: in a context where cost-performance competitors are surrounding it and major players are accelerating the “de-NVIDIA-ization,” how will this computing giant maintain its dominance? Can its estimated valuation of up to $5 trillion be sustained amid the fierce wave of “reasoning” AI? The market is watching NVIDIA and its challengers’ next moves.

Using Technology to Protect the “Moat”

At the conference, NVIDIA founder Jensen Huang unveiled a new type of central processing unit (CPU) and an AI system built on Groq’s technology—a startup specializing in custom inference chips—to improve AI response speed. The new server system combines Groq’s language processing unit (LPU) with Vera Rubin servers to create a new AI inference infrastructure.

According to Bloomberg, the LPU is a dedicated chip optimized for accelerating the “inference” process—generating responses to AI prompts. In this architecture, NVIDIA provides it as a “co-processor” to work alongside the main accelerator, which performs better on more complex, multi-stage tasks. Huang stated that this architecture offers a significant leap in computing performance compared to the previous generation of graphics processing units (GPUs).

In recent years, NVIDIA has accelerated its technological R&D pace significantly. The company is no longer limited to its iconic GPUs. “They are connecting these technologies to protect their moat,” said Daniel Newman, CEO of Futurum Group, a tech analysis firm, when discussing this new product. At GTC, Huang also showcased NVIDIA’s layout in robotics, autonomous driving, and AI agents.

Stock Price Closes Higher

NVIDIA is trying to calm market growth anxiety with grand performance expectations. Previously, NVIDIA projected that chips based on the Blackwell and Rubin architectures could generate a $500 billion revenue opportunity by 2026. At this year’s GTC, Huang predicted that its latest AI processor would help the company reach $1 trillion in sales before 2027.

Buoyed by this forecast, NVIDIA’s stock rose by 4% during the day, then the gains narrowed, and it ultimately closed up 1.2%, easing market concerns about its growth prospects and the “AI bubble.” Bloomberg reported that the company’s stock has stagnated in recent months, with a year-to-date decline of 3.4% before the GTC.

Reuters commented that betting on AI inference chips marks Huang’s effort to solidify NVIDIA’s position in the so-called “inference computing” (the process of AI providing answers). As the global AI industry shifts from “model training” to “commercial deployment (inference),” giants like OpenAI and Anthropic are shifting their strategic focus from purchasing training chips to serving hundreds of millions of end users calling AI systems. Since inference chips are far more energy-efficient during task execution and response generation after the training phase, the global market is increasingly interested in cheaper, more streamlined inference hardware. Huang said, “The inflection point for inference has arrived,” adding, “Demand is still growing.”

Capital Floods into the “Inference” Wave

Although NVIDIA currently holds about 90% of the market share, the surrounding AI hardware landscape is showing signs of encirclement. According to Business Insider, NVIDIA’s high-priced GPUs are prompting some customers to seek to reduce dependence, and a number of competitors are emerging. Companies like Meta are accelerating their in-house chip development. Bloomberg also mentioned that CPUs, which were somewhat weaker in training, are demonstrating strong substitution potential in deployment due to their cost advantages.

The focus of AI is also evolving. Business Insider reports that while GPUs dominate in training, inference is a continuous process highly sensitive to costs. Cloud giants and startups are racing to develop competing AI chips, especially for inference. Amazon has launched the Trainium and Inferentia series chips as lower-cost alternatives to NVIDIA. Microsoft recently announced the Maia 200 inference chip. Additionally, many startups are developing cheaper, more efficient dedicated chips to reshape industry standards. With billions of dollars pouring into this inference wave, the sector has spawned several highly competitive unicorns.

Furthermore, China remains NVIDIA’s biggest geopolitical challenge. Business Insider reports that U.S. government restrictions on security and trade also limit NVIDIA’s potential growth. Huang has repeatedly warned that blocking sales to China will only accelerate China’s domestic industry development. Huawei, a key player in China’s domestic industry, is seen as NVIDIA’s most direct competitor. It manufactures chips, servers, and network equipment, and operates its own cloud platform. Meanwhile, Chinese chip startups like Cambrian are emerging as strong alternatives outside NVIDIA.

Zhongguancun Information Consumption Alliance Chairman Xiang Ligang told the Global Times that NVIDIA still holds an advantage in AI hardware, and it’s unlikely to be shaken in the short term. However, “more products are emerging in the inference track,” and future competition may mainly revolve around price. Many companies hope to challenge NVIDIA’s dominance with lower prices, an area where NVIDIA’s advantage is not prominent.