Silicon Valley VC Zhang Lu: Silicon Valley's Big Model Market Divided into Three Categories, Three Major Application Areas Iterating Quickly

In the global hub of tech innovation, Silicon Valley, how advanced has the AI industry ecosystem become? Where are investors pouring their money into?

Zhang Lu, a well-known investor in Silicon Valley and the founder of Fusion Fund, shared these key insights at the China AIGC Industry Summit:

• Currently, startups can basically adopt a "cocktail" model, leveraging cutting-edge big model APIs, using open-source models as a foundation, and making modifications for model optimization.
• In Silicon Valley, the model market is relatively clear, mainly falling into three categories, with a personal preference for open platforms.
• Artificial intelligence is a super tool, and our opportunities may be ten times greater than the internet era, but only one-third of these opportunities are left for startups. Zhang Lu graduated from Stanford University and currently manages nearly $400 million in capital, focusing on investments in the medical, AI, and deep technology sectors.

To fully capture Zhang Lu's thoughts, Quantum Bit collaborated with a large model to edit and organize the speech content without altering the original intent, hoping to provide you with more inspiration.

The China AIGC Industry Summit, organized by Quantum Bit, gathered 20 industry representatives for discussions. With nearly a thousand in-person attendees and 3 million online viewers, the event garnered widespread attention and coverage from mainstream media.

Key Points:

• Artificial Intelligence is an efficient super tool, representing the trend of overall industry digital transformation, known as Digital Transformation.
• At the infrastructure level, we need to address several major challenges facing AI: high computing costs, high power consumption, data privacy issues, and latency problems.
• In Silicon Valley, the theme of AI is empowerment rather than destruction or revolution, meaning not only startups but also large tech companies can be empowered. For startups, finding the right entry point is crucial.
• Startups can adopt a cocktail model, utilizing APIs from major platforms, optimizing with their own open-source models, and building industry-specific models.
• Compared to last year, expectations for AI have become more realistic this year. Last year focused on model and algorithm performance, while this year focuses on large-scale industrial applications and cost control.
• In terms of application opportunities, healthcare, financial insurance, and robotics are areas with rapid iteration speeds.
• Infrastructure is crucial as it determines the cost of AI applications. Moreover, the cost of large-scale applications is the primary business consideration.

In Silicon Valley, the theme of AI is about empowerment rather than destruction or revolution.

Hello everyone, I'm Zhang Lu. Speaking of today's theme - artificial intelligence, although it is not a new topic in Silicon Valley, since 2016 and 2017, it has become the base work. However, new themes have gradually emerged, and by 2022 and 2023, generative AI has not only sparked widespread attention in the tech industry but also across all industries towards artificial intelligence.

I have always emphasized that AI is a tool, representing the trend of overall industry digital transformation, known as Digital Transformation. I also proposed a few years ago that AI is an efficient super tool this time, capable of driving the digital transformation of entire industries. Not only is AI widely used in the tech industry, but other industries are also generating a large amount of high-quality data, providing a stage for showcasing its capabilities.

I often stress that besides technology, we should also focus on data.

Many industries lacked high-quality data to train models a decade ago, despite having AI. However, in the past two years, driven by industrial automation, various industries have generated a large amount of high-quality data, becoming the cornerstone for the application and development of artificial intelligence. Everyone can see that artificial intelligence is a super tool, offering opportunities potentially ten times greater than the internet era, yet only one-third of these opportunities are left for startups.

Especially in Silicon Valley, the theme of artificial intelligence is empowerment rather than destruction or transformation.

Empowerment means not only startups but also big tech companies can benefit.

In Silicon Valley, the characteristics of consumer-facing applications lie in the vast amount of high-quality user data held by big tech companies. Therefore, startups and companies need to clearly identify where the innovation opportunities lie at the application level, which application scenarios and industries can obtain a large amount of high-quality data. At the same time, big tech companies, especially those in Silicon Valley, are building their own ecosystem platforms, such as NVIDIA, Meta, Microsoft, etc., which have been constructing ecosystems for years, aiming to support the development of startups.

Hence, as entrepreneurs and innovators, finding the right entry point is crucial.

The model market in Silicon Valley is relatively clear.

In terms of future model investments and types, the model market in Silicon Valley is already quite clear, divided into three categories:

• The first category is represented by OpenAI and Anthropic, backed by major cloud service companies, which provide models as services and products to third parties.
• The second category includes companies like Apple, Salesforce, NVIDIA, whose language models are excellent but primarily intended to support their own product iterations and upgrades, not as third-party products.
• The third category consists of open-source platforms, which I personally have high hopes for, and we have invested in several companies in this field.

The open-source ecosystem in Silicon Valley is thriving, with Meta as a leading company, and its LLama 2 model has helped many enterprises grow rapidly.

In the coming months, we may see the release of LLama 3 and contributions to open source from Mistral.

There are also small open-source models, such as a company we recently invested in - NEXA AI, which released a small model running at the edge, with the smallest possibly being 1 to 2 billion tokens. The rapid development of these open-source models provides another option for startups and large enterprises, enabling more efficient optimization of energy consumption and infrastructure costs.

Startups can adopt a cocktail model.

Our artificial intelligence investments are divided into two dimensions: the application end and the AI infrastructure end.

• At the application end, the focus is on Healthcare Enterprise, AI, and Industrial Automation.
• At the infrastructure level, from the chip layer to the Cloud layer and then to the security layer, all are critical technological nodes.

The application side is relatively clear, with different application scenarios in each industry. However, at the infrastructure level, we need to address several major challenges faced by artificial intelligence:

The infrastructure layer is not a single-layer solution but an ecosystem investment, spanning from hardware to software, from the cloud to the data end, all requiring strategic positioning to better drive the development of the AI ecosystem.

In Silicon Valley, the development direction of artificial intelligence applications mainly focuses on the B2B sector, and we do not invest in B2C companies.

I have always emphasized the importance of data, how to use massive high-quality data to optimize models, making artificial intelligence solutions more practical and commercial.

Currently, startups can adopt a cocktail model, leveraging APIs from major platforms, combining them with their own open-source models for optimization, and building industry-specific models.

In this process, we have observed two characteristics:

First, data quality is more important than quantity. High-quality data can optimize models at lower costs and faster speeds.

Second, we do not need one model to solve all problems. In specific application scenarios and industries, training industry-specific small models can perform on par with general large models, and even faster in terms of cost, efficiency, and response speed.

Three major application areas iterate quickly.

In terms of application opportunities, healthcare, finance insurance, and robotics are fast-paced fields -

The healthcare sector in the US market is huge, accounting for 20% of GDP. The finance insurance industry has high data quality, standardization, and diverse application scenarios. Robotics encompass not only humanoid robots but also various automation and mechanical arms, serving as new data interfaces.

Space data and space technology are also crucial fields, with space data currently being both high-quality and high in quantity, offering substantial value. The potential trend we often discuss is the regulation of the entire tech industry, with data regulation at its core. These industries share a significant characteristic: they are highly regulated, making it easier for them to adapt to regulations.

Another significant opportunity lies in AI Infrastructure, which is a problem we must address.

Despite the immense prospects for artificial intelligence development, challenges may hinder large-scale deployment. We need to not only pursue better algorithms and models but also focus on costs, especially the costs of generative AI. GPU costs are high, as are the costs of computing power and energy. We may not have enough energy to support large-scale generative AI applications, which is another reason for pursuing smaller models. Additionally, there are issues such as latency, how to address them through edge computing technology, and data privacy and security concerns.

In terms of computing power, GPUs and various chip designs are hot topics of discussion. From software to hardware, everyone is exploring new architectures and designs. I am studying Materials Science and Engineering at Stanford University, and I am also involved in the research and development of new materials. Currently, opportunities can be divided into two main categories:

• One is based on silicon-based new designs, such as AMD's APU, Google's TPU, and LPU.
• The other involves the exploration of new materials, such as optical computing, 3D design, etc. In terms of energy consumption, there have been some innovations emerging, such as inference optimization and integration of new algorithms, with some companies having the potential to reduce energy consumption by over a hundredfold.

Edge computing is also a promising technology that can address energy consumption and latency issues.

We have been investing in edge computing since 2018, releasing industry reports covering everything from chips to serverless edge, and then to cloud edge solutions, aiding in the deployment of edge applications in the era of artificial intelligence.

There are various entry points for edge solutions, such as NEXA AI's open-source model Octopus V3, a generative AI language model running at the edge, which helps broaden the direction of rapid AI development. Additionally, the flexibility of edge computing at the open-source level is also an advantage.

Data privacy concerns include model security, Cloud Network security, and data privacy. Innovations like model security protection at the chip level, network security solutions, as well as federated learning and data encryption, present opportunities for advancement.

In summary, expectations for artificial intelligence have become more realistic this year compared to last year. While last year focused on the performance of models and algorithms, this year's focus is on large-scale industrial applications and cost control.

Therefore, I have listed several points to share:

First, as a startup, it is crucial to understand the market outlook for competition and cooperation and find good ecosystem support for growth.

Second, finding the right application scenarios is crucial. As mentioned earlier, 80% to 90% of opportunities in Silicon Valley's consumer market are dominated by large companies. Overall, in the field of AI innovation, about 70% of advantageous opportunities are also held by large companies. However, the remaining 30% of opportunities represent a market potential that is ten to twenty times larger than the internet era, indicating significant market potential.

Third, data quality and diversity are also crucial. Especially now that models can be made smaller and tailored for specific tasks, the quality of data is more important than quantity. Defining data quality and ensuring data diversity are both critical.

Lastly, infrastructure is key as it determines the cost of AI applications, with the cost of large-scale applications being the primary business consideration.

The good news is that we are just getting started. The challenges in both models and infrastructure present opportunities for growth. I am currently in Silicon Valley. I often joke with my partners that it feels like there is a little surprise candy waiting for us every day. I am delighted to collaborate with outstanding entrepreneurs at the forefront of innovation in Silicon Valley. I hope to promote more exchanges through sharing and look forward to meeting more entrepreneurs from home and abroad in Silicon Valley in the future.

— End —