Jensen Huang: With a home power consumption of 240W, this is the "first AI" delivered to Elon Musk.

On October 15, 2025, Jensen Huang, the CEO of NVIDIA, personally handed a device as small as a paperback book to Elon Musk.

The location was the Starship launch base in Texas.

He said, "Imagine having the smallest supercomputer working next to the biggest rocket."

This was not an ordinary device delivery but a grand ceremony. Amid the applause of the engineers lining up on both sides, Musk solemnly received the machine named DGX Spark.

Meanwhile, on the other side of the world, an unprecedented acquisition had just been completed: BlackRock, Microsoft, and NVIDIA jointly formed a consortium to acquire Aligned, one of the world's largest data - center operators, for $40 billion. Behind this deal lies the underlying consensus of the crazy expansion of the AI industry: Computing power is the core resource.

But while capital was heavily betting on the "cloud battlefield" at the 5 - gigawatt level, Jensen Huang quietly opened another door.

DGX Spark is neither a larger GPU nor a host with the most powerful performance.

It can run large models with 200 billion parameters locally, connect to the desktop systems of Dell, Lenovo, and HP, and support the operation of private models on Ollama, Roboflow, and LM Studio.

It represents that AI is moving from the cloud center to the personal realm. It is no longer just an infrastructure built far away but is truly placed on your desktop for the first time.

The significance of this 1.2 - kilogram supercomputer goes far beyond a product launch.

Because what really matters is not that he handed an AI supercomputer to Musk.

It is that he handed the future direction of the entire AI to each of us.

Section 1

The First AI Device Closest to You

"For the first time, we've brought an AI supercomputer to everyone's desk." When Jensen Huang said this, he wasn't describing a future vision but holding a real machine that could be picked up.

This machine is called DGX Spark.

It's only the size of a book, weighs 1.2 kilograms, and has a power consumption of only 240 watts. It can run on an ordinary power socket. However, what it can do used to require a large - scale data center and even hundreds of kilowatts of electricity in the past.

It can run large models with 200 billion parameters locally without connecting to the cloud at all. You can train, fine - tune, and deploy AI applications right at your office desk.

Jensen Huang used several "you"s during an earlier press conference. This was something he specifically wanted to emphasize: AI shouldn't be the privilege of only a few companies. It should be as accessible to everyone as a mobile phone or a laptop.

In the past, AI was usually remotely invoked:

To run a large model, you had to connect to the cloud of OpenAI or Anthropic.

To have AI view images, listen to sounds, or help you write, you had to upload data and have others process it.

The "intelligence" of AI was essentially accomplished on "someone else's computer."

And DGX Spark changes this: For the first time, we've packaged real AI capabilities into a personal device that can be used at any time.

Inside Spark is NVIDIA's latest - generation GB10 Grace Blackwell chip, equipped with 128GB of unified memory. It can support complex tasks such as image generation, voice recognition, local Q&A, search, writing, programming, and reasoning. This is not a "smaller GPU" but a complete toolkit of AI capabilities.

Why is this machine special?

Jensen Huang didn't talk about parameters but painted a picture: Imagine an artist, a designer, or a programmer being able to do what they want with this machine in the office or at their desk.

This also explains why traditional PC manufacturers like HP, Dell, Lenovo, and ASUS have fully integrated the Spark architecture right away. Because this is no longer "AI for enterprises" but "AI for everyone."

This is not just a machine; it marks the beginning of something:

For the first time, AI is so close to us.

Section 2

Not "Delivering a Model," but "Handing Over the Torch"

Many people think that DGX Spark is a new computing device. But Jensen Huang himself doesn't introduce it as a "new product."

He called this delivery:

It's like handing the "torch" to everyone who needs it.

The scene of this delivery was indeed like a "lighting ceremony."

In Texas, at the SpaceX Starship launch base.

Jensen Huang carried Spark through the rocket factory, walking among a group of engineers in work uniforms. Musk appeared in the cafeteria, unpacking desserts for employees and children. After meeting, he personally led Jensen Huang to visit the rocket factory and then smilingly received the machine handed over by the other party. The applause of the people present witnessed this real - life in - person delivery.

This is not the first time.

Jensen Huang still remembers the scene in 2016 when he handed the first DGX - 1 supercomputer to OpenAI:

"At that time, I delivered the machine to San Francisco like a delivery guy. The customer was a non - profit organization called OpenAI. That machine became the starting point for them to train GPT."

This is the second time. But the meaning is completely different.

The first time, AI was just starting. The second time, AI is starting to enter everyone's daily life.

He didn't talk about the chip frequency or mention the computing - power benchmark. He emphasized a scenario:

"This machine is like an assistant beside you, waiting for your first question."

NVIDIA has been making such a transformation in the past two years.

They are no longer just delivering chips but delivering immediately usable capabilities. When you turn on the machine, you can run the image - generation model FLUX.1, use it as a visual - search agent, or deploy your own Q&A robot, voice assistant, or writing tool.

"It's not a toolkit but a torch. In the future, every developer, creator, and company employee can ignite their own AI."

We often say that AI is entering all industries. But in the past, in most cases, you had to wait for others to provide services, open APIs, and release models for you to use. You were just a "user."

But from the moment when someone personally gets DGX Spark, puts it on the desk, plugs it in, turns it on, and trains the model by themselves, the role changes.

You change from a user to a "torch - lighter."

What matters is not the reduction in size but that AI has finally moved from the cloud into personal hands.

Section 3

From 1GW to 240W: Achieved by Three Key Factors

In the past, using a large model might require invoking an entire data center. From power supply, cooling, to maintenance and scheduling, the cost of AI was incredibly high.

A top - level training server could have a power consumption of 100,000 watts. A super factory could easily exceed 1 gigawatt (1 billion watts), which is equivalent to millions of households turning on lights and air - conditioners simultaneously, about the electricity consumption of a small city.

But now, a 240 - watt desktop device can run a large model. Moreover, there is no need to wait in line, no cloud - service fees, and no need to hand over data.

This transformation didn't happen suddenly. There are three key factors behind it, which Jensen Huang explained very clearly.

Factor 1: Integrate the Entire AI Process into One Device

Jensen Huang mentioned:

"We don't just provide chips. Instead, we package the entire set, from chips, programming languages, to pre - trained models. Customers can use it right after plugging in the power, without having to assemble it themselves."

This is like in the early days of computers when you had to assemble the motherboard, memory, and hard drive by yourself, while now you buy a laptop that's ready to use out of the box.

The biggest change in Spark is that it pre - integrates all the components required to run AI, including chips, memory, models, software tools, and microservices. You can use it as soon as you get it, just like you can take photos, take a taxi, or chat as soon as you turn on your phone.

This greatly reduces the cost of building an AI system.

Factor 2: High Efficiency Means Real Affordability

When investing in AI, you can't just look at performance but at the return on investment per unit of energy consumption. If you save 3 times more electricity than others, the customer's profit will be 3 times more.

What does this mean?

It's not about whose chip is faster but about who can do more work with the same amount of electricity.

NVIDIA has optimized everything from chips to networking technology, enabling DGX Spark to achieve far - greater data efficiency with 240 watts. Moreover, since the device is small, it hardly needs an additional cooling system, and even the radiator can be simplified.

This means:

• Individuals can afford it.
• Companies can deploy it quickly.
• It no longer depends on a professional computer room.
• The cost is reduced from the "tens of millions level" to the "tens of thousands level" or even lower.

Factor 3: Everyone Can Access the Ecosystem

In the past, deploying AI required a very complex environment: a cloud - service account, a large amount of remote computing power, security - permission settings, multi - team collaboration, and approval processes.

But now, almost all major PC manufacturers, from Dell and Lenovo to HP, have integrated the Spark architecture. As long as you use these machines, you can directly run models compatible with Spark.

NVIDIA has also pre - installed a complete AI software stack, including common training tools, model interfaces, and deployment environments. This is like in the past when you had to get a driver's license to drive, while now you can take a taxi with just a click.

For the first time, the threshold of AI has been lowered to the level where you can use it with just a click.

Jensen Huang summarized it in one sentence:

"It's not that customers are waiting for AI to become cheaper. Instead, we need to make AI readily available."

From centralized deployment to terminal usability, from kilowatt - level devices to hundred - watt - level desktops, the significance of DGX Spark is not just that it can be used but that anyone can use it.

This is the real key to reducing the cost of AI.

Section 4

AI Sovereignty Is Not Only for Nations but Also for Individuals

In previous AI layouts, discussing "sovereignty" often meant competition at the national level.

For example, NVIDIA's chip - export restrictions, countries' rush to build large AI models, and governments' support for local computing - power platforms... The core of all this is that AI cannot rely entirely on imports and that one must have autonomous capabilities.

Jensen Huang also said that any country should not completely outsource its national data and then import intelligence. Even if it can buy technology from the outside, it should retain its own AI - training and deployment capabilities.

But he also added:

Not only countries but also every company and individual need sovereignty.

The underlying issue is: Whoever owns the data owns the intelligence.

AI models need to learn continuously, and the choice of training data determines their final output capabilities. In the past, large models like GPT, Claude, and Gemini were powerful, but they were all trained by others. When you use them, you are actually invoking someone else's intelligent system.

And the emergence of Spark changes this: You are no longer just borrowing but can train, deploy, and even customize models by yourself.

This is why Jensen Huang keeps emphasizing:

"Enterprises cannot rely solely on external AI services, as this will make proprietary data vulnerable. In the future, every company should have its own AI employees. Just as the HR department is responsible for human employees, the IT department should be responsible for recruiting, training, and managing digital employees."

This is not just a metaphor, as there is a complete toolchain to support it:

The AFX architecture, a collaboration between NVIDIA and the intelligent - data - infrastructure enterprise NetApp, enables enterprises to directly convert their own data (such as PDF contracts, design drawings, and experimental data) into semantic materials that AI can understand. The entire process of accessing the model for answering, generating, and analyzing is completed within the company, without leaving the company, uploading, or leaking data.

More importantly, this private - AI capability is no longer the patent of large enterprises.

Individual developers can also deploy local models on Spark without uploading data, opening API permissions, or being restricted by platform interfaces. All you need is a machine, some materials, and a little patience, and you can have your own exclusive AI assistant.

In the past, people regarded AI as a remote tool: connect to the cloud, enter instructions, and wait for a reply. But in the future, AI will become your partner beside you, understanding your language, familiar with your needs, and protecting your privacy.

This transformation gives new meaning to AI sovereignty: It is not only the bottom line for countries but also the choice for each individual.

Section 5

After the Desktop Revolution, the AI Application Ecosystem Is Completely Rearranged

In the past decade, almost all AI applications were deployed in the cloud:

Chatbots were embedded in web pages or apps.

Code assistants had to be installed as browser extensions.

Video generation required waiting in line for server rendering.

Although these services are powerful, the underlying logic is the same: users connect to the server → ask questions → remote computing → receive results. Users are always "users" and not "participants."

But starting from DGX Spark, this relationship is being rewritten.

AI is no longer far away in the cloud but runs directly on your desktop.

When NVIDIA released Spark, it specifically listed a number of practical application scenarios:

• Ollama: It can run open - source large models (such as Mistral and Gemma) locally and have conversations without an internet connection.
• Roboflow: You can train visual - recognition models by simply dragging in images for quick fine - tuning.
• LM Studio: A platform for building local Q&A robots, supporting long - document processing and targeted knowledge upload.
• NYU Laboratory: It uses Spark to train privacy - sensitive AI locally on campus without relying on cloud services.
• Zipline Drones: They perform edge computing and task reasoning on Spark.

This type of application is rapidly emerging. Their common feature is that they don't require remote connection, cloud - based invocation, or large - model authorization. Users can customize, modify, and retrain them.

That is to say, for the first time, AI has become your own thing like an app.

This brings about a major change: The entry points are being rearranged.

In the past, model manufacturers defined the capabilities, users queued up to invoke them, and software developers could only use others' APIs. Now, whoever seizes the local experience first becomes the new entry point.

Generative computing is becoming a new form of entry point, just like search engines, operating systems, and browsers. Take Perplexity as an example. You don't click on a website to retrieve existing information but directly generate answers. In the future, it won't be you asking and it answering; instead, it will understand you first and actively offer help.

What does this mean?

It's not