AI transactions have entered the “settlement stage”! Can the Nvidia (NVDA.US) GTC Berlin conference continue to write new logic for the computing power supercycle?

The Zhitong Finance App learned that from October 20 to 22 local time, Nvidia (NVDA.US) will hold the GTC conference in Berlin, Germany. CEO Huang Renxun will take the stage to deliver a keynote speech at 11:00 local time on October 21 (17:00 Beijing time). According to Nvidia's official website, the conference topics covered various topics such as accelerated computing, agentic AI (intelligent AI), robotics, open models, physical AI (physical AI), sovereign AI, and AI infrastructure.

If you put it in the past few years, this is probably just another high-profile technology event for the artificial intelligence (AI) industry. However, at this point in time, AI transactions are moving from simply expecting increased demand for transaction computing power to a “settlement stage” where more attention is paid to return on investment.

As the scale of global AI infrastructure investment continues to expand, the capital market's examination of AI transactions is also gradually shifting from “will demand for computing power grow” to a more realistic question: can huge AI capital expenses eventually be converted into sufficient revenue, cash flow, and productivity?

Therefore, the GTC Berlin conference may become an important “stress test” for the current AI industry. What the market really wants to hear from Hwang In-hoon is not only how much the performance of next-generation computing platforms will improve, but three more core questions: How long can AI computing power demand grow? Where does the return on investment in AI infrastructure come from? And can this round of capital expenditure ultimately translate into real economic value?

Highlight 1: Where will the $2 billion European AI deal go?

Before the official opening of the conference, Nvidia had already left GTC Berlin with a lot of suspense.

On October 13, Hwang In-hoon revealed at an event in Paris that Nvidia will announce about 2 billion US dollars of additional AI investment or transactions in Europe within the next week. The year before, Nvidia had invested around $8 billion in Europe. If the two parts are added together, Nvidia's investment and commitment in Europe would be close to $10 billion.

While in Paris, Hwang In-hoon announced that Nvidia provided financing to four French startups, including medical technology company Doctolib, robotics AI company Genesis AI, software company Pigment, and AI materials company Entalpic. Later, at the London event, he named British AI companies Boltz, Prima Mente, Hiverge, Riverlane, and Foresight.

This means that what the market needs to pay attention to during the GTC Berlin period is not only what technology Nvidia will release, but also what areas the additional $2 billion European deal will eventually flow to, and whether these investments can be transformed into larger AI infrastructure requirements.

Germany is probably one of the key landing points. According to reports, Germany is preparing to advance an AI collaboration with Nvidia, focusing on physical AI tools for robot and autonomous driving training, as well as AI models for industrial applications. If the relevant cooperation is further implemented during the GTC period, it will further strengthen Nvidia's layout in the European industrial AI field.

Therefore, Europe is not a simple “venue” for this conference, but may be becoming an important market for the next phase of Nvidia's AI ecosystem expansion.

Highlight 2: Can Vera Rubin and the AI factory prove that the computing power cycle continues?

For Nvidia, hardware is still one of the core highlights of the GTC conference, but the way the market understands Nvidia products has changed.

In the past, when investors discussed Nvidia, they often focused on the performance, video memory, and manufacturing process of a single GPU; today, Nvidia is increasingly emphasizing the concept of a complete “AI factory” — the GPU, CPU, network, storage, software, and data center infrastructure are integrated into one whole.

At the GTC conference in San Jose in March of this year, Nvidia revealed the Vera Rubin platform and positioned it as a full-stack computing platform for Agentic AI, covering various aspects such as Vera CPU, GPU, network, and storage. At the same time, Nvidia also announced the DSX AI Factory reference design and DSX Air to perform digital simulations prior to physical data center construction.

Therefore, what is more noteworthy about the GTC Berlin conference market is probably not “whether there are any new chips”, but rather how much Vera Rubin's actual deployment speed, customer order intensity, and unit token calculation cost can actually be reduced.

If Nvidia can show that large cloud vendors have entered the large-scale deployment stage and prove that next-generation platforms can significantly reduce AI inference costs, then market concerns about excessive investment in AI computing power may be mitigated to a certain extent.

Conversely, if the conference still releases mainly forward orders, data center plans, and capital expenditure commitments, and there is no simultaneous verification of actual deployment and terminal revenue growth, then it may reinforce investors' concerns about the “self-circulation” of AI infrastructure investment.

Therefore, what the market now needs is not just a technical roadmap, but commercial fulfillment capabilities.

Highlight 3: Can Agentic AI become the new engine for the next round of computing power requirements?

If AI infrastructure requirements over the past two years were mainly driven by big model training, then the next stage of the story Nvidia is now trying to tell is probably Agentic AI.

Traditional chatbots are more “one question and one answer,” and AI agents that can complete tasks independently require continuous calls to models, search databases, use tools, read files, run code, and perform multiple rounds of reasoning.

This means that AI computation may change from a one-time “request—response” to a continuously running computational process.

From Nvidia's perspective, this will form a new demand curve: in the past, AI computing power demand was mainly driven by model training scale, but in the future, it may be driven by pre-training, post-training, inference, and intelligent workload.

This is also an important reason why Vera Rubin was positioned by Nvidia as an “Agentic AI Platform.”

For investors, a key question is whether AI agents can actually form large-scale commercial applications, so that the demand for inference computing power can take over the training requirements.

If AI moves further from “training a model” to “digital employees that operate around the clock” in the future, then data centers will no longer just be fixed assets built once and used for many years, but may continue to expand as the number and frequency of use of agents increases.

This will directly relate to one of the market's biggest concerns: is AI capital expenditure a one-time wave or can it form a long-term continuous infrastructure cycle.

Highlight 4: Can Physical AI open up new computing power markets such as automobiles and robots?

Compared to generative AI, physical AI is still at an earlier stage of development, but Nvidia is significantly improving its strategic position.

The topics at GTC Berlin this time covered robotics, autonomous driving, digital twins, and industrial AI, and the lineup of participants also showed a very obvious industrial color.

Participating companies/organizations	Belonging field	Involved direction
mercedes-benz	cars	Physical AI, autonomous driving
Siemens energy	Energy and grids	AI data center, power and grid
Nokia	tele	AI infrastructure, communication networks
Wandelbots	robots	Physical AI, robotics
London Stock Exchange Group (LSEG)	financial data	Enterprise AI, financial data
Revolut	Fintech	Enterprise AI, Fintech
Black Forest Labs	AI model	Generative AI
ElevenLabs	AI voice	Generative AI, voice AI
n8n	AI workflows	Intelligent AI, AI workflow
European Sovereign AI Center (CESAI)	Sovereign AI	Sovereign AI, AI infrastructure
Stuttgart High Performance Computing Center	High performance computing	High-performance computing, AI infrastructure

Judging from the lineup of participants, GTC Berlin's industrial coverage has clearly surpassed the traditional AI chip field. The participation of companies such as Mercedes-Benz, Siemens Energy, Nokia, and Wandelbots made industries such as automobiles, energy, power grids, communications and robotics an important part of the conference.

It is particularly noteworthy that Mercedes-Benz and Wandelbots are compatible with physical AI applications such as autonomous driving and robotics, and the participation of Siemens Energy's digital grid business leaders echoes the electricity demand brought about by the expansion of AI data centers.

This means that Nvidia is promoting AI further from cloud and internet companies into traditional European industrial systems.

If robots, autonomous driving, and industrial AI begin to enter the large-scale deployment stage, then the AI market space will further expand from software and data centers to the global manufacturing and automation industries.

For Nvidia, this may mean a new source of demand for computing power; for the market, it means that the AI industry's valuation logic has an opportunity to further expand from the “big model” to a real-world productivity revolution.

Highlight 5: Will the reduction in AI inference costs lead to more demand?

“Lowest Cost per Token” is also a key word worth paying attention to in GTC this time.

After AI entered the era of inference, the indicators that investors are concerned about are shifting from simple peak computing power to unit calculation costs.

As intelligent AI continuously calls models, a complex task may require a large number of tokens. If the computational cost of each token is too high, it is difficult to commercialize AI applications on a large scale; conversely, if computing power efficiency continues to improve and token costs continue to decline, the threshold for using AI will also be lowered.

There is an economic logic behind this that deserves the market's attention:

Increased efficiency does not necessarily mean lower aggregate demand.

If technological progress drastically reduces AI computing costs, enterprises and consumers may instead increase the amount of AI used, thereby driving overall computing demand to continue to grow.

Similar logic has appeared in the Internet, cloud computing, and mobile communications industries.

Therefore, what needs to be observed at this GTC Berlin conference is not whether Nvidia can reduce the cost of a single token, but whether AI usage can grow more rapidly after the cost is reduced.

If the answer is yes, then “falling computing power prices” may instead be a catalyst for the next round of demand growth.

Highlight 6: Will sovereign AI become a new source of demand for computing power in Europe?

Sovereign AI is another important main line that distinguishes GTC Berlin from ordinary AI technology conferences.

Over the past few years, the global AI infrastructure has been highly concentrated in large US tech companies. But as AI is increasingly seen as part of national competitiveness and industrial security, many economies in Europe, the Middle East, and Asia are seeking to build their own AI infrastructure.

For Europe, sovereign AI means not only having a local AI model, but also requiring local computing power, data centers, networks, and data infrastructure.

This is exactly what has given Nvidia a new market space.

Topics at the GTC Berlin conference include sovereign AI. Executives from European companies and AI companies such as Deutsche Telekom, Dassault Systèmes, and Mistral AI were also invited to participate in discussions at related events before Hwang In-hoon's keynote speech.

At the same time, judging from the list of participants, European sovereign AI centers, high-performance computing institutions, and large industrial enterprises all participated, indicating that Nvidia is trying to connect AI infrastructure with the local European industrial ecosystem.

As far as Nvidia is concerned, if European countries, companies, and industrial groups all want to have their own AI infrastructure in the future, then the source of AI capital expenditure growth will no longer be entirely dependent on a few US tech giants.

But on the other hand, the European market also means more complex regulation, data sovereignty, export controls, and competition for local AI chips. Therefore, GTC Berlin is also an important window for observing Nvidia's global strategy.

Final question: Can AI capital expenditure complete the “closed loop of economic value”?

At the end of the day, the most important question at the GTC Berlin conference is not whether Nvidia can once again present a more powerful computing product.

Over the past few years, the AI industry has formed a very strong growth chain:

Increased model capacity → demand for computing power increased → cloud vendors increased capital expenses → Nvidia revenue increased → market expectations raised for AI → more capital entered into AI infrastructure.

However, by 2026, the market began to question whether this chain could actually close the loop.

Especially as AI infrastructure is increasingly dependent on debt financing, the problem has spread further from stock valuations to the credit market.

The recent financing and profit pressure faced by AI infrastructure companies is a microcosm of the market beginning to re-examine the return on AI capital expenditure. Investors are now focusing not only on “strong demand for AI,” but on revenue, free cash flow, return on capital, quality of customer contracts, utilization of computing power, financing costs, and how much the end user is actually willing to pay for AI.

In other words, the AI industry is moving from a “technology cycle” to an “economic cycle.”

Therefore, the signals that the GTC Berlin conference may release can be boiled down to five questions:

First, are Vera Rubin's orders and deployments strong enough to prove that the AI infrastructure cycle continues?

Second, can Agentic AI drive the continuous growth of inference demand and shift AI computing power requirements from a training cycle to a more stable usage cycle?

Third, can AI factories and next-generation computing platforms continue to reduce the cost per token and create more demand for AI usage through improved efficiency?

Fourth, can robotics, autonomous driving, and industrial AI create real large-scale commercial orders to open up a new computing power market for Nvidia other than cloud computing?

Fifth, and most importantly — can AI capital expenses of big tech companies ultimately translate into revenue, cash flow, and productivity improvements?

If Hwang In-hoon can give a strong enough answer in Berlin, then recent market concerns about the AI bubble, excessive capital expenditure, and peak demand for computing power may be mitigated to a certain extent.

However, if the conference still mainly releases technology roadmaps, forward orders, and capital expenditure plans, and the actual commercial returns of AI applications are still lacking verification, then GTC may instead become an important point for the market to further examine AI valuations.

Therefore, what really needs to be verified at this GTC Berlin conference is not “whether AI can become stronger”, but rather whether AI can actually create enough economic value after becoming stronger to support the ever-expanding capital investment.