NVIDIA GTC global conference, which ends today, is the place where the company unveils its latest AI chips, platforms, and software, making it a bellwether event for the direction of the AI industry.
With the resurgence of AI agents and the changes that AI has started to produce in society, this year is even more critical for NVIDIA, writes eToro analyst for Romania, Bogdan Maioreanu.
Within the AI ecosystem, there is turmoil over new technologies and demand shifts that are also shaping investors’ expectations for the technological giants’ evolution. In this context, NVIDIA highlighted its new AI full‑stack strategy—hardware, CUDA software, and cloud-scale platforms- to reinforce its technological lead and long-term growth narrative.
NVIDIA was at the forefront of AI evolution by providing the hardware needed to train AI models. But with the increasing demand for queries, technology companies are shifting part of their attention to inference, meaning putting the model to work on new, unseen data to produce predictions, decisions, or generate content.
The processors designed for AI training (such as those from NVIDIA or Google) are excellent at AI inference, but they are often not the most cost-effective or energy-efficient choice for that purpose. While AI training requires massive computational throughput to process vast datasets, inference requires lower latency and high efficiency. Both processes demand massive computing power from AI chips, but with different capabilities.
In this context, major clients (Microsoft, Amazon, Google, Meta) are developing their own chips, optimized for inference, precisely to reduce their dependence on (and bills for) NVIDIA’s chips. With the introduction of AI Agents, inference will likely be a much larger market by volume than training—you train a model a few times, but run it billions of times—so the total market is growing, even if NVIDIA’s percentage share is shrinking. However, NVIDIA is evolving too, becoming an integrated ecosystem, a full-stack AI platform provider.
In addition to its hardware, NVIDIA benefits from its software ecosystem (CUDA) and its “end-to-end” solution—complete servers, networking, and software—which continues to make it the preferred supplier for many large customers. The latest news shows that NVIDIA is making decisive steps in inference by acquiring the core assets, IP, and talent of the startup Groq, apparently paying around 20 billion dollars in cash for those. Groq’s LPU architecture demonstrated an order‑of‑magnitude lower latency and higher tokens‑per‑second per user than NVIDIA H100 clusters on large LLMs.
Analysts are considering this move as one to maintain the large moat the company has in AI hardware. It might also signal that NVIDIA expects the “Inference Flip”—where revenue from running models exceeds training—to define the next phase of AI buildout, and wants architectural options tuned to that phase under its own roof. However, this field will be much more competitive than the one in which it is estimated that 90% of the current AI-specialized data centers are based on NVIDIA architecture.
This week, NVIDIA entered the AI agents field too, saying it is collaborating with the creator of OpenClaw to create NVIDIA’s NemoClaw, which combines the ease of OpenClaw’s ability to build agents with the privacy and security controls needed to run these agents within an enterprise. And the interesting news continued, as Reuters reported, with NVIDIA winning China’s approval to sell its second-most powerful AI chip, the H200, in the Asian country.
NVIDIA CEO Jensen Huang disclosed that it has already received purchase orders from Chinese customers. And above all this, Huang has projected $1 trillion in AI infrastructure revenue over the next three years.
Despite these developments, the stock price has lost over 2% over the past 5 days. With the current turmoil in the AI sector, taken by storm by AI agents and the uncertainties about how its new capabilities will influence the industry, the key for investors is to assess whether the overall situation still justifies NVIDIA’s over 4 trillion-dollar valuation. And the interest for the company is really high. At the end of last year, NVIDIA was the most-held stock among both Romanian and global investors on the eToro trading and investment platform.
The largest share of NVIDIA’s revenue—as well as its growth—comes from data centers, and AI cannot exist without those. NVIDIA is currently the dominant force in this space. It is expanding into other areas beyond chips, such as networking and photonic technologies for networking and data transmission, as well as chip architecture. In this area, copper is beginning to face physical speed limitations. But eventually, all is tied to the execution of its current plan to change the company into a full-stack AI platform provider. If NVIDIA executes its roadmap and succeeds in deploying inference with much cheaper AI agents on its own hardware, then this might be the way this technology giant reinvents itself for the next growth cycle.
_
