Simply now, on the NVIDIA GTC convention in Taipei, China, NVIDIA CEO Jensen Huang as soon as once more targeted the subject on the event course of the AI trade.
Totally different from the concentrate on the generative AI wave two years in the past, this time Jensen Huang gave a brand new judgment:
“Generative AI has arrived, and sensible AI has arrived.”
The period of sensible AI has arrived
In his view, the most important change within the AI trade up to now few years will not be the continual development of mannequin parameter scale, however that AI has begun to change into an actual manufacturing instrument and immediately have an effect on financial actions.
For example this modification, Jensen Huang first confirmed a set of knowledge from the code internet hosting platform GitHub. He identified that software program growth is among the earliest fields the place generative AI has been carried out, and it’s also one of many largest teams of information staff on this planet. Presently, there are about 30 to 40 million skilled software program engineers all over the world who depend on programming work, and there are additionally a whole bunch of tens of millions of scholars and newbie builders concerned.
In his speech, the variety of code submissions on GitHub was used as an essential indicator to measure the change in AI productiveness:
- In 2023, the variety of code submissions was about 300 million;
- In 2024, it elevated to 400 million;
- In 2025, it reached 500 million;
- And the info within the first few months of 2026 has already proven a a number of – fold enhance in comparison with the earlier stage.
Jensen Huang believes that these numbers mirror that AI – assisted programming instruments are considerably bettering the effectivity of software program growth.
“Software program engineers all over the world create a wage worth of about $3 trillion.” He stated, “And these software program additional assist international financial actions value practically $100 trillion.”
In accordance with his calculation, if AI can enhance the productiveness of software program builders a number of occasions, the financial worth launched will far exceed the software program trade itself.
Lately, with the speedy growth of code – technology instruments, “whether or not programmers can be changed by AI” has all the time been the main focus of trade debate. In response, Jensen Huang gave a transparent reply in his speech.
He believes that the event of AI is not going to scale back the variety of software program engineers, however will stimulate enterprises to recruit extra builders. The logic is easy: if an engineer can create greater output with the help of AI, enterprises usually tend to enhance R & D funding moderately than scale back the scale of the R & D crew.
“Some folks say that AI will scale back employment. That is full nonsense.” Jensen Huang stated.
In his view, what actually determines the size of employment will not be the unit labor price, however the capacity of unit labor to create worth. When software program engineers can full extra work with the assistance of AI, the market demand for software program and digital capabilities will additional increase.
Jensen Huang then turned the subject to AI infrastructure. He identified that as AI strikes from the laboratory to the precise manufacturing setting, the trade’s focus has shifted from mannequin capabilities to Token output capabilities.
Prior to now, Token was only a technical indicator within the mannequin operation course of; now, Token has change into a unit that may immediately generate income. In different phrases, AI firms will not be producing software program merchandise within the conventional sense, however constantly producing Tokens.
Whoever can generate extra Tokens at a decrease price and better effectivity can have stronger industrial competitiveness.
“As a result of Token has now change into a revenue – making unit – Token is now a revenue – making unit that may usher in income. Simply because it could actually now make a revenue, AI firms need to construct extra Tokens, generate extra Tokens, and construct extra AI factories. That is why the demand for computing energy in Taiwan, China has skyrocketed. And that is why all of you’re so busy and your companies are doing so effectively. The truth is, simply take a look at a few of your inventory costs.” Jensen Huang stated.
That is additionally an essential cause why the development of knowledge facilities all over the world is heating up constantly and the demand for AI computing energy in Taiwan, China is rising quickly.
In his description, the AI manufacturing facility (AI Manufacturing unit) is regularly changing the standard knowledge heart and turning into the core of the brand new spherical of computing infrastructure building.
From the appliance period to the agent period
Nevertheless, in Jensen Huang’s view, the larger change isn’t just the advance of mannequin efficiency, however the change within the computing paradigm itself.
Prior to now few many years, computer systems have adopted the sample of: software → code → working system. Customers full duties by clicking on the interface and getting into instructions.
Within the AI period, a brand new structure is taking form: Agent → massive – language mannequin → instrument system.
Jensen Huang confirmed a typical structure diagram of an Agent system.
On this structure, the massive – language mannequin is answerable for understanding issues, reasoning, and planning; the peripheral framework is answerable for managing context, invoking instruments, coordinating process execution, and managing lengthy – time period and brief – time period reminiscence. To finish duties, the agent can invoke browsers, databases, spreadsheet instruments, knowledge evaluation engines, CAD design software program, and numerous enterprise techniques.
The entire course of is extra like a digital worker moderately than conventional software program. “Prior to now, we began purposes, clicked buttons, and entered content material.” Jensen Huang stated, “Sooner or later, we solely want to clarify our intentions to AI.” Then AI will routinely write code, invoke instruments, and full duties.
The rise of brokers has additionally sparked one other controversy: if AI can full work, will software program firms be eradicated?
Jensen Huang’s reply is the alternative.
He believes that the Agent period will give rise to much more software program techniques than at this time. The reason being that the variety of digital brokers is now not restricted by the inhabitants measurement. Sooner or later, each enterprise course of, each enterprise hyperlink, and even each private process might have its personal unique agent. And these brokers must name a lot of exterior instruments and providers to finish their work.
Due to this fact, software program is not going to disappear, however might want to exist in a type that may be referred to as by AI.
“This is among the finest occasions for the software program trade.” Jensen Huang stated.
On this context, NVIDIA’s lengthy – collected CUDA ecosystem can even welcome new alternatives.
Prior to now, the CUDA library was primarily for builders; now these capabilities might be immediately referred to as by brokers and change into a toolset when brokers carry out duties. In a way, the message Jensen Huang is making an attempt to convey may be very clear: within the generative AI period, we talk about what the mannequin can do, whereas within the sensible AI period, we talk about what work the mannequin can full.
When AI begins to generate income, drive GDP development, and may carry out advanced duties by invoking instruments by means of brokers, it’s now not only a chatbot, however is turning into a brand new computing platform.
“NVIDIA is initially a software program firm”
After speaking concerning the computing paradigm change caused by brokers, Jensen Huang as soon as once more emphasised a view he has repeatedly talked about in recent times:
NVIDIA is basically a software program firm.
Subsequently, Jensen Huang defined the core structure and working logic of AI brokers.
He stated that the agent is the last word decoupled and distributed computing mannequin, which must mobilize a lot of completely different computing models to run collaboratively. A whole AI agent consists of 5 core elements: mannequin, framework, instruments, abilities, and runtime. Every element runs dispersedly on completely different nodes within the knowledge heart. He vividly in contrast it to a working particular person: the mannequin is the “mind” of the agent, answerable for considering and resolution – making; the framework is the “physique”, carrying the general operation; the runtime is like an unique studio, supporting the implementation of varied instruments. The entire system completes computing energy scheduling and process execution in an extremely – massive – scale mode.
In accordance with his introduction, every work means of the agent is cut up into completely different modules of the pc and accomplished step-by-step. Amongst them, the massive – language mannequin undertakes core clever duties akin to considering, context processing, environmental notion, logical reasoning, plan planning, and motion execution. This course of will activate the Grace Blackwell NVLink 72 computing energy cluster in batches. Within the means of the agent invoking instruments, the CPU undertakes the computing work and might be tailored to C compilers, Python, JavaScript, and numerous accelerated computing instruments.
Jensen Huang believes that the instrument software capacity of present AI brokers continues to be in its main stage, and will probably be upgraded to be extra skilled and proficient sooner or later. For that reason, NVIDIA’s CUDA X library has undergone an essential improve. Your complete collection of library merchandise can be accompanied by unique AI ability manuals, which can be utilized by AI brokers to study independently and grasp the use strategies of instruments, tremendously bettering the power of brokers to unravel numerous core trade issues. Sooner or later, the computing energy worth and software potential of brokers calling CUDA X instruments can be tremendously launched.
In all the agent computing energy system, numerous {hardware} and practical modules have clear divisions of labor. Device computing duties are accomplished collaboratively by the CPU, GPU, and huge – mannequin; the safety safety framework is deployed on the CPU and NVIDIA BlueField DPU safety processor to make sure the general operation security; the general process scheduling and orchestration work is uniformly led by the CPU, forming a heterogeneous computing system with clear ranges and clear divisions of labor.
In his speech, Jensen Huang particularly talked about the core ache level of AI computing – the reminiscence system. He stated that the working reminiscence of the agent is realized by means of the KV cache, masking advanced operations akin to reminiscence retention, knowledge compression, info retrieval, matching of structured and unstructured knowledge, and finding out the logical relationships and ontological associations of varied knowledge. The general processing course of is extraordinarily troublesome and sophisticated. He predicted that the iterative improve of the AI – particular reminiscence system will drive a revolutionary change within the international storage system.
Evaluating with the standard software program operation mode, Jensen Huang emphasised that the brand new computing paradigm represented by AI brokers has important variations. Prior to now, most software program was a centralized operation mode the place a single binary file was tailored to a single working system. In distinction, brokers undertake a brand new computing logic of decoupling, distribution, and heterogeneity. That is additionally the core motivation for NVIDIA to dedicate itself to the analysis and growth of the subsequent – technology Vera Rubin platform.
Concerning the brand new Vera Rubin platform, Jensen Huang emphasised that it’s in no way a single chip or an bizarre GPU product, however a whole finish – to – finish revolutionary system. The platform takes the GPU because the core place to begin, integrates core {hardware} akin to GPUs, Vera, and NVLink 72, depends on a number of CPUs to finish international process orchestration, and is provided with an iteratively upgraded revolutionary storage system to construct a full – hyperlink computing energy base. On the identical time, the platform integrates CX – 9 {hardware}, the DOCA software program stack, and a constructed – in safety processor, which may encrypt knowledge all through the method of static state, transmission, and use, and comprehensively shield the safety of excessive – worth AI mannequin knowledge primarily based on the confidential computing structure.
Jensen Huang stated bluntly that Vera Rubin is NVIDIA’s most formidable R & D mission in its growth historical past. All 40,000 engineers of the corporate are concerned within the mission, and on the identical time, it brings collectively the energy of trade companions to implement it. It’s a particularly advanced system that has been comprehensively polished and reconstructed from scratch. He admitted that NVIDIA has lengthy accomplished the strategic transformation from a single GPU producer to a full – stack system producer. The presently launched Vera Rubin system is essentially the most advanced and full AI computing energy system designed within the trade up to now.
When speaking concerning the final industrial wants and the enterprise transformation course, Jensen Huang stated that the core calls for of consumers and companions will not be merely to acquire pc {hardware}, however to construct a mature and environment friendly AI manufacturing facility. Based mostly on this trade pattern, NVIDIA is beginning a brand new spherical of strategic transformation. Presently, NVIDIA’s core applied sciences have been totally carried out in infrastructure – stage software situations, and on the identical time, it hyperlinks with numerous industrial ecological companions akin to energy crops, cooling techniques, and energy grid suppliers to construct a whole AI industrial ecosystem.
Sooner or later, NVIDIA will proceed to construct a full – stack computing energy system to supply core assist for international clients to construct massive – scale, excessive – efficiency AI infrastructure.
It’s value noting that on this speech, Jensen Huang elaborated intimately on NVIDIA’s new industrial positioning and formally proposed the “new paradigm of the AI manufacturing facility ecosystem”, clearly stating that NVIDIA’s growth focus has been comprehensively upgraded from the standard computing ecosystem to a manufacturing facility – primarily based ecological system serving the trillion – stage AI infrastructure.
Jensen Huang distinguished between NVIDIA’s previous and new ecological types. Prior to now, NVIDIA took the computing ecosystem because the core and deeply built-in its computing layer, software program, and computing stack into numerous enterprise platforms and third – celebration libraries, broadly empowering the digital computing energy wants of all industries.
Now, the newly constructed AI manufacturing facility ecosystem has fashioned a transparent upstream – downstream industrial closed – loop: trade companions are the upstream basis assist for NVIDIA, and NVIDIA depends by itself full – stack technical capabilities to output a whole AI manufacturing facility ecosystem to the downstream. The core aim is now not merely to output GPU chips or computing energy techniques, however to assist clients construct extremely – advanced and extremely – massive – scale AI manufacturing facility infrastructure.
He stated bluntly that the AI manufacturing facility has entered the stage of enormous – scale implementation with extraordinarily excessive funding and excessive thresholds. Presently, the development price of a single 1 – gigawatt (GW) – stage AI manufacturing facility continues to rise, from the preliminary $20 – 40 billion to the present $50 – 60 billion, and it’ll quickly exceed $80 billion and even $100 billion sooner or later. The funding of a whole bunch of billions of {dollars} in a single mission signifies that the AI manufacturing facility has extraordinarily excessive necessities for implementation stability and operation reliability. It should be constructed as soon as and put into regular manufacturing instantly. Its capital funding price and system building complexity have reached an unprecedented stage within the trade.
To unravel the issue of constructing an extremely – advanced AI manufacturing facility, NVIDIA depends on Omniverse’s digital simulation capacity to attain a full – course of innovation. Totally different from the standard pc R & D mode – first designing the chip after which simulating the system operation within the system – now all of the AI manufacturing facility infrastructure of NVIDIA might be constructed, simulated, examined, and optimized prematurely on the Omniverse digital platform. By the empowerment of the digital simulator and digital structure, the trade can full the total – course of deduction of the extremely – massive – scale AI system earlier than breaking floor and investing large quantities of cash, utterly avoiding implementation dangers and realizing the trade’s lengthy – standing know-how implementation imaginative and prescient.
Jensen Huang particularly launched the core system DSX that helps the implementation of the AI manufacturing facility ecosystem, forming a whole infrastructure format akin to NVIDIA’s current product matrix. Amongst them, the RTX collection corresponds to GPU {hardware}, DGX corresponds to the built-in computing energy system, and the brand new DSX platform is exactly focused in any respect situations of AI infrastructure. Counting on the core capabilities masking the system, software program, and full technical stack, NVIDIA can empower small and medium – sized enterprises to rapidly construct world – class AI cloud service capabilities.