Deploy global computing power nodes: GoodVision AI builds a full-domain dynamically scheduled AI Compute Grid

Jensen Huang posted a piece about AI factories, which gained over a million views in less than 12 hours.

The post did not discuss more powerful models, but focused on land, power supply, computer rooms and complete computing power systems. He judged that infrastructure, rather than algorithms, is becoming the main constraint for AI growth. Because these things that sound far away from ordinary users are determining an increasingly realistic question: why are model calls getting cheaper, while enterprises' AI bills may grow faster than their business?

The answer is hidden in a seemingly ordinary task:

A software company asked Agent to assist in troubleshooting online faults. It needs to read alerts and operation logs, locate abnormal codes, retrieve internal documents, analyze upstream and downstream dependencies, generate repair plans, then call tools to run tests and check results. What the engineer sends may only be one instruction, but to complete it, Agent has to perform dozens of steps continuously in the background. Each step may call the model, and each call consumes computing power.

These calls are not the same type. Sorting out logs does not require the most powerful model, but judging the cause of a fault may require complex reasoning; some codes can be handed over to cloud models, while logs containing customer data must remain inside the enterprise; ordinary analysis can wait for a few more seconds, but alerts from production systems cannot afford long network round trips. If all tasks are sent to the same model and the same data center, what the enterprise gets may not be a better result, but it will definitely bear higher costs and greater latency.

This is the true meaning of AI entering the infrastructure stage. Competition no longer only occurs between models, but also in every allocation behind the models: what model should be used for what task, at which node it runs, and at what cost and speed it is completed. When such choices occur millions of times every day, what determines the enterprise's AI cost is not only how much computing power it purchases, but also whether the computing power is used in the right place.

This is also the field that GoodVision AI is laying out. It undertakes local and regional inference demands with a modular AI Factory, and a single factory can deploy up to 1.5MW of inference computing power in a 200-square-meter space. Relevant project plans have covered Japan, South Korea and the United States. But what is more noteworthy than how much computing power is built is that GoodVision AI tries to answer another question on top of these infrastructures:

When a task is split into more and more calls, who decides where each request should go?

The money spent on training is often one-time, but the money spent on inference is permanent

In the past few years, the industry's attention has been almost entirely focused on training — parameter scale, training clusters, and ranking performance. This makes sense, as that is the stage that determines the upper limit of model capability.

But training and inference are two completely different ways of spending money.

Training is like building a power plant: huge investment, concentrated cycle, once it is completed, it is done, and a model is mostly trained only once. Inference is like power transmission and consumption: every call, every step executed by Agent, and the operation of every terminal device, consumes computing power in real time. A model is trained once, but inference will occur billions or trillions of times.

GoodVision AI saw this difference very early. David Wang, founder and CEO, founded the company in 2019, focusing on the infrastructure gap that emerged when AI capabilities moved from "demonstrable" to "large-scale delivery". In his statement, AI infrastructure is gradually shifting from focusing on "where the computing power is" to focusing on "how each Token is efficiently produced, routed and delivered". While the industry focuses on the upper limit of model capabilities, GoodVision AI focuses on who manages the operating costs below that upper limit.

Today, this judgment is becoming the industry consensus.

The figure Jensen Huang presented at GTC 2026 is that the computing demand for AI inference has increased by about a million times in the past two years. Correspondingly, the global capital expenditure on AI infrastructure has exceeded 400 billion US dollars in 2025, and the forecasted scale in 2026 will further exceed 600 billion US dollars — AI infrastructure has not cooled down with the end of the training competition, but has entered a longer construction cycle due to the outbreak of inference demand.

Agent also complicates inference costs.

What Agent changes is not only the consumption of Tokens, but also the consumption pattern. In the past, the user asked one question and the model answered one; now, to complete a task, an Agent needs to read materials, call tools, check results, and continue to act based on feedback. One task is thus split into dozens or even hundreds of model calls.

Tokens are shifting from "conversation consumption" to "system consumption". When calls enter the main process of an enterprise, the cost is no longer an occasional API fee, but an operating expense that continues to grow with the business.

The real meaning of this for enterprises is that the cost structure of AI has changed from one-time procurement to a bill that keeps jumping. As long as the bill is continuous, efficiency is no longer an engineering problem that can be optimized later — it is an operational problem from the very first day.

Thus, a seemingly contradictory phenomenon begins to appear: the same enterprise may face both surplus and shortage of computing power at the same time. High-end computing power is queued up for competition, while a large number of simple tasks that could have been completed by lightweight models locally in dozens of milliseconds occupy the most expensive resources and travel around the earth before the results are sent back.

What really needs to be solved is how to allocate computing power: which requests should use flagship models, which requests can be handed over to lightweight models, and which requests should remain local.

Every call is a multiple-choice question

This is exactly the layer that GoodVision AI wants to solve. David Wang has deep experience in the cloud computing industry, with resumes spanning IBM, AWS, Alibaba Cloud and Tencent Cloud. He used to be a partner of IBM and a member of AWS senior management, participated in the early team building of Alibaba Cloud, and later served as the head of Tencent Cloud North America.

These experiences make him pay more attention to the practical application of AI in enterprises. The technologies used by enterprises will continue to change, from cloud servers and GPUs to model APIs and Agents, but the demand has always been consistent: obtain truly useful AI capabilities at lower costs and in a more stable way.

GoodVision AI first established its foothold in cloud services, which is also its first layer of entry connecting enterprises with global computing resources. Facing AWS, Microsoft Azure, Google Cloud, as well as enterprises' own private clouds and local computer rooms, it helps customers complete cloud migration, architecture design, multi-cloud governance, operation and maintenance, and continuous optimization. The combination method depends on business scale, cost, data security and global deployment requirements, without being bound by a single cloud vendor.

The value of this layer of business is first reflected in "helping enterprises move to the cloud", and at the same time allows GoodVision AI to go deep along the real workload — figuring out how customers use computing power, where Tokens are spent, where latency is stuck, and where the compliance boundary is drawn. This is information that external suppliers can hardly get from PPT. When enterprises start to introduce large models and Agents, this capability naturally extends to model calls, cross-cloud scheduling and private inference.

This judgment is completed by the Smart Routing Engine.

After an AI request enters the platform, the system first identifies the request intent, then comprehensively considers task complexity, Token cost, response speed, data sensitivity, regional compliance and node availability, and allocates resources in real time between different models and computing resources.

In specific scenarios, complex reasoning and general knowledge tasks can be routed to flagship models of cloud services such as ChatGPT, Claude and Gemini; lightweight tasks such as translation, summarization, customer service response and repetitive processing are handed over to low-cost small models; requests involving sensitive data or extremely high latency requirements are sent to enterprise private models or edge nodes to reduce data transmission between external clouds.

The Smart Routing Engine also supports Agent to Agent (A2A) calls. When an Agent processes a task, it often needs another Agent to retrieve, analyze, review or generate content. The system will arrange appropriate models and operation nodes for this call according to task objectives, capability matching, cost, latency and data permissions. In this way, multiple Agents can collaborate more stably to complete a job.

Multi-model routing answers "which model should be called", and GoodVision AI wants to go one step further: where this model should run. The same model, deployed in public clouds, enterprise computer rooms or edge nodes, has completely different costs, latency and data boundaries. Therefore, what the Smart Routing Engine schedules is not only the model, but also the computing power location where the model is located — it not only provides a unified API entry, but also is responsible for resource scheduling of the enterprise AI system, that is: simple tasks use lightweight resources, and complex tasks retain necessary model capabilities; when cloud platforms, model services or computing nodes fluctuate, requests can be automatically switched to other available environments.

This logic has been verified by actual business.

GoodVision AI once served an image and video generation platform. At that time, the customer's monthly AI expenditure had exceeded 500,000 US dollars, and Token consumption was still growing at a rate of about 50% per month. After migrating the relevant models to GoodVision AI Factory for private deployment, the customer's overall cost dropped by about 60%, and network latency dropped by about 50%.

Its way of "saving money" comes from the simultaneous rearrangement of models, computing power and deployment locations, rather than simply replacing a cheaper model.

Computing power is starting to move to places with large populations

After the request is selected, it will eventually fall to a certain physical location. This brings out the heavier half of GoodVision AI's business.

David Wang has a judgment: the current AI infrastructure is still in a stage similar to "before the emergence of CDN". Centralized large clouds are still most suitable for training and complex reasoning, but as Agents begin to process high-frequency tasks such as translation, customer service, coding, image and video processing, more and more requests should actually be completed by nodes close to users, close to data, and close to the business site.

There are clear business reasons behind this. AI glasses, robots, autonomous driving, and smart factories are extremely sensitive to latency, and many scenarios simply do not allow requests to be sent thousands of kilometers away before waiting for results; data from industries such as finance, healthcare and government affairs often cannot even take the step of "crossing the border".

GoodVision AI's AI Factory is a modular inference center for local and regional demands. According to company materials, it uses immersion cooling to control PUE below 1.2, can deploy up to 1.5MW of inference computing power in a 200-square-meter space, and is connected to the Smart Routing Engine to become a local node in the global scheduling network. At present, relevant projects are being promoted in Japan, South Korea and the United States. The nodes in Japan are planned to be expanded in stages in the next three years, with a total scale of 100MW. The company's expansion path is to first build flagship nodes with local partners, then deploy edge nodes to surrounding areas, and replicate this infrastructure delivery model to more regions.

It is worth noting that building a data center is notoriously a heavy and slow task. Why can a company founded in 2019 advance projects in three countries at the same time?

The answer is hidden in the previous computing power cycle.

From land acquisition, construction approval to power-on and operation, traditional data centers are often calculated in years, while enterprises only need a few months to purchase GPU servers — the time gap in between is the most realistic gap in current AI infrastructure. Relying on existing mine and energy infrastructure resources, GoodVision AI cooperates with North American infrastructure partners in Texas, Wisconsin and other places to upgrade existing facilities that already have power supply and sites into AI-ready inference infrastructure. According to company materials, the first phase of the Fukushima project in Japan is expected to be completed and put into operation within three months.

This is a quite meaningful industrial succession: the power and sites left over from the previous computing power cycle have become the rarest admission tickets for this round of AI inference implementation. What it brings is not so much a cost advantage as a time advantage — in a track where everyone is rushing for construction periods, this may be the more valuable one.

Just like the early oil industry competing for important oil fields, whoever can occupy energy and space resources in key regions in advance will have the ticket to enter the next round of AI infrastructure competition. From David's perspective, the competition in AI infrastructure is presenting a logic similar to the energy industry.

Based on this judgment, GoodVision AI plans to connect AI Factory nodes distributed in different countries and cities, and eventually form a dynamically schedulable global "AI Compute Grid", covering more than 100 cities and more than 100 AI Factories in the long run, so that computing power, like internet content, can be distributed to users more efficiently according to demand.

In the future, such power indicators, land resources and infrastructure nodes that can carry AI computing power will become scarce assets.

From selling resources to selling judgments

Looking back, the three businesses of cloud services, Smart Routing Engine and AI Factory are connected to each other and develop along the same link.

Cloud services allow GoodVision AI to access enterprise systems and connect to scattered multi-cloud resources; the Smart Routing Engine is responsible for understanding each request and completing real-time allocation of models and computing power; AI Factory provides controllable, low-latency, locally deployable underlying computing power. The three together form a complete link from enterprise demand, to Token scheduling, to computing power delivery.

This link is translating into rapid growth. According to the latest disclosed data of the company, GoodVision AI achieved revenue of about 24 million US dollars in the first nine months of fiscal year 2026 ending June 30, 2026, which is nearly 5 times of the about 4.82 million US dollars in the same period of the previous year, with a year-on-year increase of 397.9%. It is more noteworthy that the company achieved revenue of about 13.45 million US dollars only in the three months from April to June 2026, not only a 544% year-on-year surge, but even exceeding the total revenue of about 10.55 million US dollars in the previous six months, showing that its business growth is still accelerating significantly. According to the latest management forecast, GoodVision AI's full-year revenue for fiscal year 2026 is expected to reach about 38.87 million US dollars, about 5 times of the 7.74 million US dollars in the previous fiscal year.

For an infrastructure company still in the expansion stage, the significance of this set of figures does not lie in the absolute scale, but in proving that the business model has been verified, and the company is moving from the verification stage to the large-scale stage.

It is worth noting that what GoodVision AI delivers is changing. Early customers purchased cloud resources, and their core demand was to obtain more suitable computing power. As the Smart Routing Engine undertakes more enterprise requests and AI Factories are put into operation one after another, the company needs to further judge which model each request should be handed over to, at which node it runs, and at what cost and latency it is completed.

The core that enterprises purchase has extended from the computing power itself to the way of using computing power.

Affordable to buy, and affordable to use

Back to the online fault troubleshooting scenario at the beginning.

Such cost problems usually come from the mismatched model selection and computing power scheduling. Millions of requests with different difficulty levels are processed in the same way. As the industry shifts from training to inference, how to adapt models and computing power to specific tasks has become a required lesson for enterprises.

The power of electricity truly changed the world after the power grid was spread out.

This AI grid is just starting to be laid now.