---
title: "The limits of \"extracting\" silicon: how to keep GPUs from being \"idle\"?"
type: "News"
locale: "en"
url: "https://longbridge.com/en/news/294576875.md"
description: "The AI industry is focusing on improving GPU utilization, triggering a \"chip-making war\" and an explosion in the AI Infra market. Giants like Google and OpenAI are investing in self-developed chips, while platforms like Baseten, Fireworks, vLLM, and SGLang are seeing skyrocketing valuations, attracting massive financing. As the focus of large models shifts to inference, how to maximize the potential of silicon chips becomes crucial, driving the formation of a new market worth hundreds of billions"
datetime: "2026-08-01T08:19:49.000Z"
locales:
  - [zh-CN](https://longbridge.com/zh-CN/news/294576875.md)
  - [en](https://longbridge.com/en/news/294576875.md)
  - [zh-HK](https://longbridge.com/zh-HK/news/294576875.md)
---

# The limits of "extracting" silicon: how to keep GPUs from being "idle"?

The AI industry has suddenly entered a "chip-making war."

On July 20, The Information reported that Google is developing a server chip codenamed "Frozen v2," planning to directly embed parts of the Gemini model architecture into the silicon.

A month earlier, on June 24, OpenAI unveiled its first inference chip Jalapeño, co-developed with Broadcom, which took only nine months from design to tape-out.

Additionally, Anthropic, Zhipu AI, and DeepSeek have also been reported to be making relevant arrangements.

However, not every company has the funds or capability to recreate a chip. Therefore, a market is emerging that is attracting the attention of capital and giants: AI infra. The underlying logic is that there is still room for optimization in the already deployed data centers and GPUs.

The AI inference platform Baseten saw its annual revenue grow 20 times, with its valuation skyrocketing from $2.1 billion to $13 billion. In the same sector, Fireworks quadrupled its valuation to $17.5 billion within seven months, with annual revenue exceeding $1 billion.

Recently, this sector has also welcomed a strong force. The open-source inference engines "Gemini" vLLM and SGLang announced their commercialization in succession, with seed round financing exceeding $100 million, gathering almost all the names of giants and top venture capitalists in the AI industry, marking the beginning of a strong competition.

Thus, "GPU utilization" has become the latest keyword in Silicon Valley, leading to a new trillion-dollar market.

In this article, we invited the co-founder and core member of RadixArk, incubated by SGLang, as well as data center experts, to discuss the four-layer architecture of AI Infra, how the industry can maximize the potential of "silicon," and the key technologies behind it.

The focus of large models is shifting from training to inference, which will pose completely different requirements for computing power.

Training is extremely costly, but once completed, the task comes to a halt. In contrast, inference is different; whether it's a chatbot or an agent, the more AI applications there are, the more requests the GPU needs to handle, and these requests are continuous, leading to an explosive growth in inference demand.

> **Zhu Banghua**
> 
> **Co-founder of RadixArk**
> 
> As new applications gradually increase, the total amount of inference will be much larger than before. As everyone continues to expand applications and the intelligent boundaries of models, the demand for inference will grow increasingly. It has now reached a point where it is very difficult to find cards on the market, and every company is short on cards.

The current situation is that every company on the market is short on cards, and for wealthy giants, the most direct method is, of course, to continue building data centers and purchasing more GPUs.

Tech giants like Meta, Google, and Microsoft continue to increase capital expenditures, with a total expected to exceed $1 trillion this year. Jensen Huang predicts that by 2030, global annual investment in AI infrastructure will reach $4 trillion.

But Wall Street is worried.

> **Ethan Xu**
> 
> **Former Microsoft Energy Strategy Manager, Director of Breakthrough Energy Research**
> 
> Wall Street's concerns are not unfounded; we have past experiences to learn from. The bursting of the internet bubble, for example, saw a massive influx of funds into fiber optic construction and internet companies. Although they realized their value in the long run, significant problems arose in the short term.

So, if we cannot significantly increase computing power in the short term, what other methods can satisfy the rapid development needs of AI? The answer is: **Increase GPU utilization**. The underlying industry behind this is called AI Infra (Artificial Intelligence Infrastructure Layer).

First, from the perspective of the overall architecture of AI Infra, let's clarify a misconception about GPUs: **You think GPUs are busy, but in fact, most of the time, GPUs are quite "idle."**

GPUs are just one part of the entire AI system. From the moment a user request enters the system, it needs to go through network transmission, resource scheduling, model loading, memory management, and finally complete inference and return results. Throughout this process, if any link encounters a bottleneck, the GPU can only stop and wait.

A study released by Carnegie Mellon University (CMU) in April of this year conducted fine-grained telemetry on 756 GPUs in a large academic AI cluster for 31 days, covering six models from A6000, L40S, A100, H100 to the latest B200. The results found that many times GPUs were in a state called "execution-idle." How serious is this "seemingly busy yet actually idling" situation? In the clusters they observed, **nearly 20% of execution time and about 11% of energy consumption were wasted on waiting**. More concerning is the inference scenario, where Azure Code workloads consume up to 65% of their energy on this idling, and OpenAI's Chat requests also reach 52%.

So, why are GPUs not fully utilized? First, we need to understand the four-layer architecture of AI infrastructure, where each layer affects GPU utilization.

**Layer 1: Power Infrastructure**, which refers to electricity. This layer determines whether the GPU can operate continuously and stably.

**Layer 2: Hardware Infrastructure**, which refers to the hardware. In addition to GPUs, this now includes high-bandwidth memory (HBM), high-speed interconnects (NVLink, InfiniBand), and CPUs. This layer determines the theoretical computational limits that can be provided.

**Layer 3: System Software**, which includes CUDA, compilers, communication libraries, memory management, and kernel libraries, determines whether each computation can be executed to the physical limits of the hardware.

**Layer 4: Service Orchestration**, this layer is responsible for coordinating GPU resources, deciding which tasks to prioritize, and dynamically scheduling between different services. Inference engines, training frameworks, request scheduling, and resource management all belong to this layer, with representative projects like vLLM and SGLang, which are also the areas of most innovation in recent years.

Having clarified the architecture of AI infrastructure, let's return to our question: why are so many GPUs underutilized, and what are the main bottlenecks?

Theoretically, any issues in any of the four layers can affect GPU utilization: insufficient power and cooling will force the GPU to throttle; if hardware interconnects lag, GPUs will spend a lot of time waiting for data rather than processing it; inadequate optimization in system software will prevent each computation from reaching the physical limits of the hardware; and improper request scheduling will cause computations that could be merged to be split apart, and reusable results to be recalculated, leading to waste.

However, the nature of the problems in these four layers is not the same. The power infrastructure and hardware infrastructure layers are "hard" problems, where power, cooling, and chip modification cycles are measured in years, and the optimization potential is rapidly reaching its peak. The system software and service orchestration layers are "soft" problems, essentially engineering and algorithmic issues, where optimizations can lead to several times performance improvements 
Against this backdrop, the industry's attention is rapidly shifting towards these two "soft" issues.

> **Chen Zhenlin (Richard)**
> 
> **RadixArk Member of Technical Staff**
> 
> Although costs mainly occur at the physical layer, most of the work needs to be focused on the software layer and service orchestration. By investing human resources in this area, you can achieve the greatest potential for optimization.

To give a specific example, a NVIDIA GB200 NVL72 cabinet has a procurement cost of about 4 million dollars. However, if the software stack is not well scheduled and optimized, the actual utilization rate of the GPU may only be 50%, which means only 2 million is effectively used, while the remaining 2 million dollars are wasted in the form of electricity costs, depreciation, and opportunity costs. Conversely, if the utilization rate can be increased from 50% to over 90%, the effect is equivalent to having an additional cabinet for free.

> **Chen Zhenlin (Richard)**
> 
> **RadixArk Member of Technical Staff**
> 
> The better the software layer performs, the better the service and token consumption will be. By rewriting and optimizing this layer, we can help improve the interaction between the inference engine and training framework, including the scheduling of post-training processes like reinforcement learning, which is particularly popular now, to increase GPU utilization.

This is why, when hardware costs reach this level, every optimization at the software layer becomes a direct business issue.

Not long ago, netizens discovered that the largest proportion of employees at Anthropic are now engineers related to AI infrastructure, which is closely related to this logic.

> **Chen Zhenlin (Richard)**
> 
> **RadixArk Member of Technical Staff**
> 
> Anthropic's inference team now has about 200 people. Over the past three years, they have gone from extreme instability in infrastructure to now being extremely stable, and have strengthened cost and optimization, which has directly impacted their revenue, allowing them to achieve profitability in the second quarter OpenAI, Anthropic, and Google will place such AI Infra teams internally within the company, as this has become one of the competitive barriers for major cutting-edge laboratories. However, for smaller model teams or startups without the funding to optimize, two open-source frameworks have become their pillars: SGLang and vLLM.

> **Chen Zhenlin (Richard)**
> 
> **RadixArk Member of Technical Staff**
> 
> This is why open-source communities like SGLang emerged two years ago. You need such new frameworks to redefine this wave of new hardware system combinations to better adapt to it.

Having understood the four-layer architecture and the two "soft" problems that need optimization, what should be done specifically?

Currently, all optimization work can be summarized as answering four questions: **Which computations do not need to be redone? Which waits can be eliminated? Which computing power is underutilized? Which resources are not coordinated?**

The SGLang and vLLM mentioned earlier revolve around these questions, although their design philosophies are slightly different. vLLM has gained widespread attention for proposing PagedAttention, emphasizing conventional and general inference deployment scenarios. SGLang, on the other hand, starts from the inference execution process, placing greater emphasis on computational reuse and system-level optimization.

Let's take a look at how the RadixArk team, incubated by SGLang, is currently thinking about these four core issues of AI Infra optimization. First, let's address the computation issue.

▍Extreme Reuse of K/V Cache

There is a "very silly" fact in large model inference: when the same piece of text is repeatedly fed to the model, it has to be computed from scratch each time.

This is a significant flaw in practical business, especially in agent workflows, where the same tool description may be called dozens of times, starting from zero each time. The industry now refers to this as "inference tax," which is somewhat akin to an IQ tax; it's essentially money wasted.

Can this tax be avoided? Of course, it can.

Anthropic previously implemented an optimization that directly cut costs by 90%; something that originally cost ten dollars now only costs one dollar after optimization. The core idea behind this is called KV Cache reuse, which simply means storing what has been computed once and using it directly next time, without recalculating from scratch 
During decoding, each layer of self-attention in the Transformer maps historical tokens into Key/Value tensors, referred to as KV, and the Cache is the storage. With KV Cache, during the first request's prefill phase (the process from when the user finishes inputting the prompt to generating the first token), the system can store each token and the K/V of each layer from the prompt. When decoding each new token, only the KV of the new token is calculated and added to the cache, while the K/V of historical tokens is reused directly.

However, KV Cache has a drawback. While it can reduce the computational load during decoding and lower the first token latency (TTFT), it comes at the cost of occupying GPU memory. When the context is long and there are many user requests simultaneously, KV Cache can consume a large amount of memory. Once the memory is insufficient, the model struggles to improve concurrency and speed, often becoming the main bottleneck for inference services.

SGLang's solution to this issue is a core technology called **RadixAttention**.

The core idea is to use a data structure called a "Radix Tree" to organize the massive requests' KV Cache. You can think of it as a family tree with shared prefixes, where requests with the same beginning share the same branch, and only when the content starts to diverge do they grow new branches.

> **Zhu Banghua**
> 
> **Co-founder of RadixArk**
> 
> In today's world of intelligent agents, there are often many shared system prompts, followed by different user prompts. Under the same user prompt, everyone continues to ask questions, so the tree structure keeps extending downward. This is also a visionary aspect of SGLang's paper, which anticipated this pattern around 2023. As the number of users programming intelligent agents increases, the benefits brought by the Radix Tree, including SGLang's own prefix cache system, become more significant.

In simple terms, **one of the core optimizations achieved by SGLang is the sharing and reuse of KV Cache across requests and machines.** Ordinary KV Cache serves only one request, but SGLang transforms KV Cache into a globally reusable resource However, while this may sound simple, it is quite complex in engineering. For example, where to store the cache, how to quickly match it, how to share multiple requests, and who to evict when there is insufficient video memory—all these steps require meticulous design. Having a good cache structure alone is not enough; a "scheduler" is also needed to actively arrange requests with shareable prefixes to be processed together. Otherwise, even if the cache is available, if the scheduler distributes requests to different times and different GPUs, the cache is essentially wasted. In this regard, SGLang has implemented many optimization strategies.

For instance, Cache-aware scheduling involves processing similar requests together. You can think of the operation of a large model like a restaurant kitchen. If the kitchen is currently processing three consecutive orders of spicy hot pot, many seasonings and preparations can be reused, leading to high efficiency. Even if there are differences, they are just small variations like mild, medium, and spicy.

But if the order changes to spicy hot pot, sushi, steak, spicy hot pot, sushi, pizza, spicy hot pot, the chef has to switch back and forth, and wash pots in between, reducing the opportunities for reuse. Therefore, SGLang uses Cache-aware scheduling to group similar prefix requests together, rearranging the order to three spicy hot pots, two sushi, steak, and pizza, making it easier for the cache to hit and further improving efficiency.

Another mechanism called Eviction is used to discard the least likely caches when there is insufficient video memory, similar to managing memory on your phone. Photos and documents from ten years ago may not be as valuable as data from the last few months. Therefore, commonly used system prompts, high-frequency RAG documents, and recently accessed conversation prefixes are more likely to be retained, while older, low-reuse, and unlikely-to-be-accessed KV Cache will be evicted when memory is insufficient.

Another technology that makes AI Infra appear very intelligent is called distributed cache-aware load balancer. Large-scale services typically have many GPUs. If a long prompt's KV Cache is on GPU 1, but the next request with the same prefix is sent to GPU 3, GPU 3 does not have that cache and must recalculate. Therefore, SGLang also tries to send requests to the GPU that "already has notes," which can significantly reduce the amount of repeated calculations 
There are other small technical details that we will not elaborate on for now. Overall, although SGLang has set up the architecture, how each team deploys the infrastructure still involves considerable engineering considerations. Next, let's look at the second major challenge in optimization work: waiting.

▍Waiting: Keep the GPU Busy

In the current market, memory prices can rise by 90% in a quarter, high-end graphics memory is out of stock until 2027, and even with money, you might not be able to grab a GPU. However, in the end, half of the time, these expensive cards are just "waiting idly."

**The first is the waiting between requests.** The earliest inference engines operated on a "queue system," where requests came one by one, and one was calculated before the next. Later, it evolved to "batch processing," where a batch was gathered for processing, but gathering a batch also requires waiting, which does not solve the waste problem. Moreover, if there are short and long requests in a batch, the short ones have to wait for the long ones to finish before they can all be processed together, leading to a poor user experience for short requests.

The current mainstream approach is called Continuous Batching, which no longer waits for a full batch to be ready before processing. Instead, it operates like a bus that can pick up and drop off passengers at any time, allowing new requests to board and completed requests to disembark at any time, keeping the GPU fully loaded. This modification can increase the actual throughput of the GPU by 2 to 4 times.

In addition to request waiting, **there is also waiting during the computation phase.** The inference of large models actually consists of two stages: "reading in" and "generating." The reading stage (Prefill) requires quickly processing a large amount of text and has high computational demands. The generating stage (Decode) outputs one character at a time, which requires less computation, but each character must reference the entire memory, making it extremely sensitive to memory bandwidth. In the past, these two parts were mixed together in the same GPU, resulting in Prefill waiting while Decode was computing, and Decode waiting while Prefill was computing, dragging each other down.

The current approach is called Prefill/Decode separation, which splits these two stages onto different GPUs, each using the most suitable hardware configuration, and transmits intermediate results via high-speed networks. For example, DeepSeek adopts this solution, where 32 GPUs form a minimal computing unit during the Prefill "reading" stage, and when it truly begins to "answer word by word," the task is handed over to another batch of GPUs, allowing both sets of cards to work independently without interference 
Today, PD separation has become one of the most important architectural evolutions for inference engines. In the distributed inference framework Dynamo launched by NVIDIA, PD separation is placed at the core of the entire architectural design, and SGLang is also one of the earliest inference engines in the industry to support large-scale PD separation deployment.

Next, let's talk about the third major issue: the problem of fully utilizing computing power.

▍Computing Power: "Release One More Time"

When NVIDIA's H100 was released, the Tensor Core computing power increased several times compared to the A100, and the entire industry expected AI inference to grow in sync and achieve a performance leap, but the reality was not so. The main reason behind this is that GPUs spend a lot of time on data movement and are limited by serial decoding, preventing computing power from being fully utilized.

There are two main directions to solve this problem.

**One is low-precision computing.** When large models are running, the GPU memory mainly contains three things: model weights, intermediate activation values, and KV Cache. The precision of these three parts can be set independently. Using smaller data formats to store or compute any of these parts can reduce memory usage and leverage faster low-precision computing power on the GPU. In suitable scenarios, it can even release double the computing power space.

**The other is speculative decoding,** which means using a small model (draft model) to "guess" the next few characters, and then having the large model perform a one-time verification. If the guess is correct, it generates multiple characters in one computation.

To make an analogy, the large model is a very capable but busy teacher, while the small model is a fast assistant. Previously, every time a student wrote a character, they had to ask the teacher if it was correct, and only after the teacher confirmed could they write the next character. But with speculative decoding, the assistant first helps the student write a small draft, and the teacher checks this small section at once. If everything is correct, it is all approved; if there is an error in the middle, they start over from the point of error.

In the best-case scenario, this method can increase the generation speed by 2 to 3 times. The inference engine plays a crucial role here by integrating these new capabilities into the system as quickly as possible, bridging software and hardware, and optimizing collaboration.

> **Zhu Banghua**
> 
> **Co-founder of RadixArk**
> 
> SGLang itself is a comprehensive solution; it is an inference engine. It needs to ensure, for example, that we perform kernel optimization, scheduler optimization, and optimizations such as CUDA Graph execution optimization and continuous batching Even further up, you will have routing (request routing), including how prefix cache is routed to the corresponding inference engine instances. Therefore, SGLang actually combines a lot of technologies together to achieve the ultimate and best inference solution.

Our guest also mentioned that inference engines like SGLang are now in a deeply bound joint optimization with chip manufacturers. When chip manufacturers design new hardware, they are already collaborating with the SGLang team for joint tuning. By the time the hardware is officially released, SGLang has almost simultaneously completed day-0 support, and both parties have compressed the silicon's performance to the extreme.

> **Zhu Banghua**
> 
> **Co-founder of RadixArk**
> 
> Regardless of which hardware manufacturer produces a new GPU or new hardware, they must prove that this hardware is very high performance. Basically, it is now the default choice for many people to run SGLang, and SGLang has even become a hardware evaluation tool. For example, you can see that AMD's latest hardware and NVIDIA's latest hardware are now the first launch partners of SGLang, and even on the day of hardware release, they will announce how well it performs on SGLang. We are now collaborating with all these hardware manufacturers, and SGLang itself and our team can compress the performance of their hardware to the extreme, allowing for better inference performance, training performance, and so on.

This also explains why, when RadixArk was incubated from SGLang for financing, institutional investors such as NVIDIA, AMD, MediaTek, Databricks, and other major players in AI hardware and system layers were almost all present. The list of individual investors is even more impressive, including Intel CEO Pat Gelsinger, Broadcom CEO Hock Tan, OpenAI co-founder John Schulman, and PyTorch creator Soumith Chintala, among others, making it a very luxurious lineup.

For them, investing in an inference engine like RadixArk is a strategic bet. They need an infrastructure entry point that connects computing power and applications, gathering more models, enterprise applications, and agents into their computing power ecosystem Finally, let's talk about an important optimization of SGLang in reinforcement learning: Miles.

▍Resource Coordination: Adapting Miles for Post-Training in RL

Reinforcement learning has become the core training method for the new generation of models, including OpenAI o1 and DeepSeek R1, which has significantly changed the form of AI training.

Traditional model training is a "single-attribute" task: feed data, update parameters, feed data again, and update parameters again, with the GPU performing the same type of work throughout the process. In this case, hardware configuration and scheduling strategies only need to be optimized for the task of "training." However, reinforcement learning requires the integration of inference, evaluation, and parameter updates—three completely different computations running in a loop, each with vastly different hardware requirements.

> **Zhu Banghua**
> 
> **Co-founder of RadixArk**
> 
> To be precise, reinforcement learning must perform online inference. Its load is that I first let the model generate a lot of content to be answered, then score this content, update the model, answer some new questions, and then score it again. So it is inherently an alternating process of training and inference.

If traditional training frameworks are used to run reinforcement learning, they will compete for resources within the same batch of GPUs, with generation waiting during training and training waiting during generation, leading to a significant waste of computing power during phase switching and mutual waiting.

Thus, resource coordination has become the key to improving efficiency. This is also why RadixArk has launched the training framework Miles in addition to SGLang.

> **Chen Zhenlin (Richard)**
> 
> **Member of Technical Staff at RadixArk**
> 
> If we purely ignore training and only focus on inference, we may not be able to bet on the future. Because ultimately, many people will still want to train their own models and then perform inference. Therefore, if we want to build an overall AI software infrastructure, we believe both sides need attention.

Miles is an open-source training framework aimed at post-training for large models. Post-training refers to the process of refining the model after it has acquired basic capabilities through methods such as SFT (Supervised Fine-Tuning) and reinforcement learning, making it better at following instructions and thinking.

**Compared to traditional training frameworks, the biggest difference of Miles is that it considers "training" and "inference" within the same system.** This is very important in reinforcement learning, as mentioned earlier, where inference and training are an alternating process; if not well-coordinated, a lot of computing power will be wasted After Miles and SGLang work together, SGLang continuously produces new samples on the inference side, while Miles continuously updates model weights on the training side. The updated model can then be used for the next round of generation. This optimization allows for smoother integration between inference and training, reducing wasted time on waiting and switching, thereby forming a more efficient post-training closed loop.

> **Chen Zhenlin (Richard)**
> 
> **RadixArk Member of Technical Staff**
> 
> Mismatches are still common in today's engineering field, which can lead to situations where you need to share some common model weights and kernels in inference and training models to maintain alignment. Many people use SGLang, and then using Miles brings stronger training capabilities, improving stability for inference tasks, long-range tasks, and agent tasks. MoE (Mixture of Experts) is an architecture that everyone is currently using, but MoE itself has some routing, which can lead to different parallel strategies in training and inference, as well as varying final accuracy. Therefore, you need stronger engineering capabilities to make this work. Miles itself brings optimization in this area and better integrates training and inference.

Currently, Miles has undergone extensive engineering optimizations in areas such as new hardware support, MoE training, and Training-Inference Alignment, allowing the entire post-training process to run more stably and efficiently. Since the project launched in November 2025, Miles has been adopted by several cutting-edge laboratories.

We have discussed so much technical theory because this layer is truly important for the development of large AI models. It aims to squeeze out the limits of "silicon" and systematically release the computational efficiency of the entire AI system.

As we mentioned earlier, both SGLang and vLLM are open-source architectures, primarily maintained by a group of researchers and engineers in their spare time, driven by passion. However, as the world faces a shortage of computing power and needs to "squeeze" the limits of silicon to release more computational power, relying solely on the collaboration of the open-source community is no longer sufficient to meet the enormous demand. Therefore, the founding teams of SGLang and vLLM have chosen to formally start businesses to push the technology forward more quickly and with greater investment.

 Since being officially incubated from SGLang, RadixArk has greater ambitions. They believe that just as cloud computing transformed servers, future AI infrastructure may evolve into a foundation that the entire AI industry can share.

Recently, Jensen Huang led the establishment of an open-source AI alliance, which includes model companies, chip companies, and AI infrastructure projects like SGLang.

Today, the speed of iteration for open-source models is increasing rapidly, with new SOTA models emerging every few weeks, from DeepSeek V4, to Zhipu GLM-5.2, and the recently released Kimi K3 on July 17. However, open-source models only provide weights, and these weights do not automatically become tokens. Anyone who downloads these models still needs to solve a series of engineering problems such as deployment, memory usage, throughput, and scheduling, which SGLang can now support from day-0.

This means that once a model is released, any developer can immediately deploy it using engines like SGLang, and the same inference framework can run on NVIDIA, AMD, and more emerging AI chips. **The inference framework is increasingly resembling the "operating system" of the AI era, becoming a crucial layer connecting models and chips.**

> **Zhu Banghua**
> 
> **Co-founder of RadixArk**
> 
> Our larger purpose in creating RadixArk is actually to hope that in the future, the term frontier lab will no longer exist, because everyone can be a frontier lab. I believe that the frontier itself is also a very dynamic concept. Two years ago, everyone thought that Anthropic was not a frontier company, and even before that, people might have thought that Google was too far from OpenAI to catch up. So in this regard, I think AI is changing too quickly now.

> We believe that there will be more and more challengers, more and more people with their own ideas, who will create better AI models and AI products. Our goal with RadixArk is to empower these more innovative individuals, enabling them to do their best work, and we provide them with the strongest support in infrastructure. From the moment they start thinking about what they want to do, they will have the capability to compete with the current frontier labs. Ultimately, perhaps everyone in the world will be able to train frontier-level models themselves and achieve the best inference, so we actually hope to play the role of a supporter Therefore, **when AI Infra capabilities transition from the exclusive resources of a few leading model giants to a public good that any startup can directly access, the threshold for engaging in AI work will be significantly lowered.** This, perhaps, is the most important significance of "extracting the limits of silicon."

We also hope to see that after the full release of infra dividends, more and more teams with new ideas, new architectures, and new application scenarios will no longer be slowed down by infra, bringing AGI closer to us

### Related Stocks

- [DTCR.US](https://longbridge.com/en/quote/DTCR.US.md)
- [CLOU.US](https://longbridge.com/en/quote/CLOU.US.md)
- [SOXX.US](https://longbridge.com/en/quote/SOXX.US.md)
- [SOXL.US](https://longbridge.com/en/quote/SOXL.US.md)
- [IGV.US](https://longbridge.com/en/quote/IGV.US.md)
- [XDAT.US](https://longbridge.com/en/quote/XDAT.US.md)
- [SMH.US](https://longbridge.com/en/quote/SMH.US.md)
- [DAT.US](https://longbridge.com/en/quote/DAT.US.md)
- [XSW.US](https://longbridge.com/en/quote/XSW.US.md)
- [IDGT.US](https://longbridge.com/en/quote/IDGT.US.md)
- [USD.US](https://longbridge.com/en/quote/USD.US.md)
- [SSG.US](https://longbridge.com/en/quote/SSG.US.md)
- [SOXS.US](https://longbridge.com/en/quote/SOXS.US.md)
- [SOXQ.US](https://longbridge.com/en/quote/SOXQ.US.md)
- [XSD.US](https://longbridge.com/en/quote/XSD.US.md)
- [PSI.US](https://longbridge.com/en/quote/PSI.US.md)
- [FTXL.US](https://longbridge.com/en/quote/FTXL.US.md)
- [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md)
- [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md)
- [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md)
- [AVGO.US](https://longbridge.com/en/quote/AVGO.US.md)
- [META.US](https://longbridge.com/en/quote/META.US.md)
- [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md)
- [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md)
- [NVD.DE](https://longbridge.com/en/quote/NVD.DE.md)

## Related News & Research

- [LIVE MARKETS-AI's not that fun and easy. Just look at the SOX](https://longbridge.com/en/news/294098193.md)
- [Alphabet, Amazon, and Meta Will Spend Over $500 Billion on AI in 2026. Nvidia Collects a Huge Share.](https://longbridge.com/en/news/293878735.md)
- [MediaTek eyes bigger share of AI market](https://longbridge.com/en/news/294536060.md)
- [Here’s How Much You Would Have Made Owning Invesco PHLX Semiconductor ETF Stock In The Last 5 Years](https://longbridge.com/en/news/294562109.md)
- [Nvidia's Potential $250 Billion OpenAI Financing Deal Is Reviving a 1990s Tech Bubble Habit, Analyst Warns: 'Usually Don't End Well'](https://longbridge.com/en/news/294077412.md)