Click to follow
The column "AI Future Guide" was launched by Tencent News. It invites industry experts, entrepreneurs, and investors from around the world to discuss technological development, business models, application scenarios, and governance challenges in the AI field. This issue focuses on the recent status of AI investment and financing in Silicon Valley.
Text丨Hao Xin
Editors丨Su Yang, Liu Yuqi, Wang Yisu
Sweeping away the haze of last year's layoffs and plummeting stock prices, Silicon Valley is re-emerging under the spotlight with a "Gatsby" attitude by virtue of AI.
According to the incomplete statistics of Light Cone Intelligence, in 180 days, Silicon Valley has completed 42 financings in the field of artificial intelligence, and won 55% of the global fundraising amount. Among them, 8 artificial intelligence star unicorn companies have emerged, with an average round financing amount of 330 million US dollars.
Hot money poured down from the sky of Silicon Valley, and dollars piled up new stories.
"For the AGI era, this year is the best year in the past ten years, but it will be the worst year in the next ten years." The atmosphere of entrepreneurship is full of every corner, which has almost become the consensus of Silicon Valley.
"Participating in five or six gatherings a week, from technology seminars to application opportunities, Silicon Valley has been in the excitement of AI for the past six months." Kong Xianglai, a Chinese who graduated from Stanford and is resident in Silicon Valley, not only devoted himself to the entrepreneurship of AI e-commerce shopping guides , and even the AGI Advent faction, an AI community that it accidentally launched, also unexpectedly exploded under this wave of enthusiasm.
(Silicon Valley AI Exchange Conference is hot on the scene)
Chinese entrepreneurs on the other side of the ocean have also been infected. Star entrepreneurs such as Wang Xiaochuan and Li Zhifei, as well as many well-known investors such as Dai Yusen, managing partner of ZhenFund, and Zhang Yutong, partner of GSR Ventures, do not want to "look at flowers in the fog". , Watching the Moon in the Water”, flew to Silicon Valley three times in a row in January.
Although the field of AGI in China and the United States is equally hot, the pattern and ecology of the market are completely different.
The most obvious difference is that, compared with the “100-model competition” in China, basic large-scale model entrepreneurship is not very popular in Silicon Valley. "OpenAI is the only one, and only big companies such as Google and Meta challenge it. Startups rarely make basic large-scale models." Cheng Hao, partner of Yuanwang Capital and founder of Xunlei, said to Guangcone Intelligence, but in China, Some people still want to gamble, because it is still uncertain who will be China's OpenAI in the end, and the process of creating gods is dangerous and sexy.
Outside of the big model, Silicon Valley is bustling with flowers blooming in the middle layer and vertical application layer. But in contrast to China, although investors have a keen sense of smell and have long smelled the "meat", they "see more and invest less"; in the field of AI entrepreneurship, there are few companies other than brand-new large-scale models. The voice of medium and large companies.
Whether it is due to the retreat of US dollar funds or the domestic AI entrepreneurial environment, unlike the Internet era, in this round of technological frenzy, Chinese entrepreneurs and VCs are experiencing unprecedented confusion, and the Copy to China strategy seems to be gradually failing up.
"Silicon Valley attaches great importance to technology and focuses on upgrading large-scale model capabilities, while in China, more attention is paid to business models." Kong Xianglai said to Opticone Intelligence. In the AGI era, their business logic remains unchanged.
Combining in-depth interviews with investors and entrepreneurs, Lightcone Smart reviewed the financing and development of AI start-ups in Silicon Valley in the first half of this year, hoping to find opportunities and warnings in the wave of AI 2.0 entrepreneurship, with a view to giving domestic investors and A little inspiration for startups.
AI gold digging wave, where is the hot money going?
According to data from PitchBook, a foreign venture capital data analysis company, in the first half of the year, a total of 1,387 financings occurred in the global AI field, raising a financing amount of 25.5 billion U.S. dollars, with an average financing amount of 26.05 million U.S. dollars. According to data from financial service agency Carta, in Q1 of 2023, US round A AI start-up financing will increase by 58.4% month-on-month, and the valuation of seed round companies will increase by 19%.
Half of the hot money went to Silicon Valley. According to the incomplete statistics of Lightcone Intelligence, there were 42 financings in Silicon Valley in the first half of the year, with a total amount of about 14 billion US dollars, accounting for 55% of the world's total financing amount. The average round of financing is $330 million, which is nearly 13 times the average financing level.
The most comprehensive Silicon Valley AI start-up company combing and tabulation: Light Cone Intelligence
The AI upstarts led by OpenAI have become well-deserved protagonists. Among the 40 companies that received investment, nearly 60% of the companies were established within one year. Its financing rounds are also at an early stage. Among the 42 financing events, seed rounds accounted for 40%, and B rounds (including B rounds) accounted for 86%.
Kong Xianglai told Guangcone Intelligence that most of the start-ups in Silicon Valley adopt a small and sophisticated approach. The number of early-stage entrepreneurial project teams is 3-5 people, and the size of the mid-term project team is also controlled between a dozen to dozens of people. The effect of Midjourney is astounding all over the world. With a team of 11 people in the early stage, it achieved a revenue of 100 million US dollars. "The AGI field pays more attention to technological innovation, and it is useless to pile up people." Kong Xianglai said bluntly.
This is very different from the way of starting a business in China. The number of domestic entrepreneurial teams is generally large. For example, the media reported that Wang Huiwen’s Light Years Beyond had 70 people before it was disbanded. Wang Xiaochuan’s open letter revealed that Baichuan Intelligent initially formed 50 people.
However, although the scale of AI entrepreneurial teams in Silicon Valley is generally small, their ability to attract money is amazing.
So far, the rankings of AIGC startups by the amount of financing are: OpenAI ($11.3 billion), Inflection ($1.525 billion), Cohere ($445 million), Adept ($415 million), Runway ($195.5 million) , Character.AI ($150 million) and Stability AI (about $100 million).
Standing behind them are still well-known companies and bigwigs in the technology circle. Lightcone Intelligence compiled statistics and found that in the first half of the year, Microsoft participated in 5 AI financing projects, Google 4 times, Nvidia 6 times, and OpenAI 3 times. Together, these giants participated in about 43% of AI financing.
To focus more, Silicon Valley is currently focusing on three directions of AI: the first is the basic large model layer; the second is the middle layer where development tools and databases are located; the third is the vertical application layer.
Cheng Hao introduced to Guangcone Intelligence that except for the two companies OpenAI and Anthropic for the basic model, other entrepreneurs are doing open source models; entrepreneurs at the tool level are mobilizing all talents and elites to build an open source community, and the core barrier is to create a developer ecosystem ; There are two types of start-up companies gathered in the application layer, one is companies in vertical fields such as legal and HR recruitment, and the other is general-purpose companies related to copywriting, Wensheng pictures, and Wensheng videos. The application layer generally gathers B-end products, and the C-end products are much less, which belongs to the state of seeking explosive models in the cracks of giants.
From the perspective of different opportunities, in the first half of the year, there were 8 financings at the basic large model layer in Silicon Valley, 12 financings at the middle layer, and 23 financings at the vertical application layer. However, the financing amount is inversely proportional, with financing amounts of US$11.08 billion, US$350 million, and US$2.52 billion.
On the surface, the basic large model layer seems to be the hottest investment field, but in fact it is completely supported by OpenAI. If OpenAI’s huge financing of 10.3 billion US dollars is excluded, the financing ratio of the entire vertical large model layer will directly increase from 79% to 79%. % plummeted to 21%.
As can be seen from the figure above, the vertical application level is currently the hottest investment field in Silicon Valley, with many rounds of financing, but the amount of single financing is not high; the basic large model layer OpenAI occupies an absolute leading position, while other large model companies have a single round Its financing is relatively high, but it is difficult to compete with OpenAI, and its business scope is also making up for the shortcomings of OpenAI; the middle layer is a new continent recently discovered by VCs. At present, it has invested in Pincone, a vector database company with a valuation of 750 million US dollars. One foot into the ranks of unicorns.
Kong Xianglai said, “Silicon Valley investors are divided into two factions, one is only optimistic about OpenAI, and believes that OpenAI will dominate 2C-end applications in the future, so there is no need to invest too much time in the field of C-end applications, and instead invest in B-end and AI companies that are deeply integrated in the industry; the other faction holds the opposite attitude, actively embraces the open source community, and is also optimistic about vertical applications on the 2C side, believing that unicorn companies can also emerge from this field.”
Overall, in the first half of this year, the following directions set off wave after wave of gold nuggets in Silicon Valley:
Basic large model layer: small parameter basic model, general large model.
Middle layer: vector database, AI tool chain, model deployment tool.
Vertical application layer: Wensheng video, Wensheng voice, AI dialogue robot, generative AI search, legal vertical field application, humanoid robot.
Kong Xianglai added and concluded that investors in Silicon Valley are focusing on several directions at present, namely AI Agents (intelligent body), multi-modality (Wensheng map, Wensheng video), solving industry illusions (law, medicine), personalized direction ( Character AI and other dialogue robots), middleware for large language models, and industry scenario applications that are deeply empowered by AI.
(Source: CB Insights)
Standing at the cusp of the gold rush, the value of these upstarts has risen with the tide. In 2022, the ChatGPT and AIGC fields will attract more than 2.6 billion US dollars in gold, and a total of 6 unicorns will be born. As of May 8 this year, unicorns The membership of the club has risen to 14 (Midjourney has no valuation yet).
1,000 VCs raised their placards to bid at the same time, and what they brought was Baby, which was established less than four months ago, and raised two rounds of financing worth hundreds of millions of dollars. On June 29th, Inflection AI completed a new round of financing of 1.3 billion US dollars, becoming the second most funded generative artificial intelligence startup company, and the capital once again succeeded in making gods.
At the same time that the miracle came to Silicon Valley, the knockout round officially began.
Data, the only barrier in the AIGC era
In just half a year, Silicon Valley has already contributed a wave of real gold and silver lessons to AI entrepreneurship.
One is that companies like Jasper that have grown up by accessing the OpenAI API interface have been hit.
As the first batch of AIGC unicorn companies, Jasper seized this wave of AI upgrade opportunities, and its valuation soared to US$1.5 billion. But its problem is also very fatal. The product barriers of shell companies are very thin. Its user experience and brand are good, but not the best. It is easy to be replaced by differentiated products built in high-value segments. Improperness is its biggest problem.
Cheng Hao believes that Jasper's biggest competitors are giant competitors such as ChatGPT, Microsoft Copilot, and NotionAI. The problem is that the extra value created is not thick enough. For companies like Jasper, the core is to find ways to work hard on data storage, multi-person collaboration, and workflow integration to increase user stickiness.
The other is the VC-backed chatbot team, which raised a lot of money in last year's financing boom and expects to sell it to enterprises this year. But at the beginning of the year, there were so many chatbots in the market, and the technical barriers were not high, they were easy to be copied, and once again entered the vicious circle of homogenization.
(Source: Network)
In addition, Neeva, an AI search platform for the C-end, was eventually acquired by a large company due to difficulties in commercial implementation. With this lesson learned, almost all companies in Silicon Valley are now frantically attacking the enterprise market.
"Start-up companies must first choose the right route, whether it is '+AIGC' or 'AIGC+'." Cheng Hao believes that choice is more important than hard work.
The criterion for choosing "+AIGC" and "AIGC+" is the proportion of AI in the entire business value chain. If a company's AI component accounts for 10% and business logic accounts for 90%, it is more suitable to take the "+AIGC" entrepreneurial path; if its AI component accounts for more than 50%, it is more suitable to take the "AIGC+" route .
The dominoes have begun to fall, and it is unclear who will be the next Jasper, but it is becoming clearer that startups in the AIGC era must also defend their moats.
Investor Chamath Palihapitiya believes that either you are at the absolute bottom and master the data scene; or you are at the absolute top and have core computing power resources.
“For middle-stage companies, they may be worth a lot today, but they may be worth nothing tomorrow,” Chamath Palihapitiya said.
"Data is the only barrier in the AIGC era." Kong Xianglai said.
Kong Xianglai believes that neither the model nor the computing power is the moat of the AIGC era. No matter how high the score of the model training is, it will eventually fall into practical application. The user distribution data accumulated in the scene will be included in the fine-tuning model data set of the enterprise. After continuous iteration, it will be formed on the enterprise side. The data flywheel, on this basis, the large model that is fine-tuned and trained will become more and more accurate, forming a positive feedback effect.
With the open source of LLaMA2, the technical barriers of large models have been further broken. As Fu Sheng said, "AI start-up companies wake up in the middle of the night with a smile", the pace of technology completion will be accelerated, and the competition for data will become more intense.
Chinese-style AI, don't copy Silicon Valley, and don't copy your old self
The opportunities and lessons of real money are in front of us. Where will Chinese-style AI go?
"Large models will lead to a revolution in interaction and productivity." Cheng Hao told Guangcone Intelligence that the industries driven by natural language will benefit first. Therefore, intelligent customer service, pre-sales consulting, writing, translation, legal in vertical fields, and HR recruitment These "low-hanging fruits" will naturally be targeted by entrepreneurs early on.
However, "the application of law and psychological counseling in the United States is hot. The fundamental reason is that the labor costs of lawyers and psychological counselors are very high, and the economic model of large-scale model applications can run smoothly. In China, there is no such environment, and blind copying will not work." .” As early as two months ago, some investors made Chinese-style comments on the investment boom in the United States.
In addition to Wen Shengtu and Digital Man, like Silicon Valley, many Chinese entrepreneurs are also aiming at the new generation of intelligent customer service. But at the same time, worries and anxieties about "homogenization" are quietly spreading.
"Customer service marketing accounts for nearly half of the 8 AI application entrepreneurship projects," Yang Ji (pseudonym), who is also engaged in AI customer service marketing entrepreneurship, exclaimed in a low voice, looking at the competitors on the road show. As the roadshow progressed, his expression became more and more tense.
Yang Ji told Lightcone Intelligence that the technology is relatively mature and the demand is clear. Customer service marketing has become the fastest-running scene, and it has now moved from the domestic market to Southeast Asia. Yang Ji's entrepreneurial experience reflects the common problem of a group of entrepreneurs from the Internet to the current AIGC. They don't want to spend energy on the hard bones of technology, but just want to take shortcuts by finding scenarios and making applications.
I am afraid that Chinese entrepreneurs will have to step on the pits that Silicon Valley has stepped on. Hegel's words come true again, "The only lesson that mankind can learn from history is that man does not learn from history."
There is no revenue forecast, no number of users, and PPT-style roadshows are staged one after another, making it impossible for Chinese VCs to start. "ChatGPT is new, and the confused thing is not knowing what to invest in; at this stage, there is nothing to invest in."
It is easy to flock to do simple applications, but there is no shortcut to the rise of the AI industry.
Large-scale model giant companies, including OpenAI, have now come to their own development bottleneck period. Yin Yifeng, technical engineer of Hugging Face, a foreign open source model community, confirmed to Lightcone Intelligence, "It may be difficult to make new breakthroughs in technology for at least half a year, which will greatly limit the large-scale implementation of applications."
Silicon Valley's technology has hit the ceiling, and it has to go back and strengthen the technical capabilities of the middle layer. For China, this is also an excellent opportunity to complete the industrial chain.
If the development of a large model is compared to "building a house", then the AI Infra (AI infrastructure) in the middle layer is a "toolbox". Referring to the development trend of Silicon Valley, the creation of data tools, model deployment of enterprise mobile terminals, and the practice and application of AI Agents will all be the next development direction.
(Source of AI Infra Industry Chain: CB Insights)
Christensen mentioned in "The Innovator's Answer" that there is a cornucopia that can retain wealth in the industrial chain. TMT VC investor Na Liu mentioned, "Currently, the cornucopia in the field of AI Infra is changing in the value chain, from AutoML, a platform-based solution with an integrated structure (focusing on performance) to modularization (flexibility, speed, convenience) Mainly).” Behind this is that enterprises want to open the “black box” process, and hope to be able to flexibly adjust each component in the model and construction workflow to obtain the system and analysis results that best suit their specific needs .
The most valuable link is also the most difficult bone. China is now short of tools and raw material manufacturing factories. This also explains the root cause of why China lacks competitive large models: the bottom layer is weak, and the upper layer is weak.
Taking the data of the three elements of AI as an example, China's data-related industrial chains are almost all "one-stop all-inclusive" by big cloud companies, lacking in-depth cultivation in a certain vertical field, and for start-up companies, every link will be It's an opportunity to do fine work. "Data preparation" is an opportunity with Chinese characteristics, which includes data quality, data labeling, data synthesis, and application malls and projects.
At present, synthetic data companies have gradually gained the favor of capital. A generative AI company with the ability to synthesize image data, “Kuawei Intelligence” completed the Angel and Pre-A rounds of 100 million-level financing within one year last year; this year, Guanglun Intelligence, which was just established this year, completed three rounds of financing within half a year. After the Angel + round , and its cumulative financing amounted to tens of millions of RMB.
After the technology worship in the early stage, more and more entrepreneurs have realized that the OpenAI road is not the only solution in the era of large-scale models.
In addition to the large model, the middle layer is a seemingly niche, but more secure path; while the application layer, which seems to be the most "low-hanging fruit", is a single-plank bridge with thousands of troops and horses, and it is easier for the winner to take all. Big factories go first.
But for Chinese-style AI, it is easier to know that it cannot blindly copy Silicon Valley; what is more difficult to do is not to copy the "old self" and go to the old road of re-delivery where people grab projects.
Welcome to pay attention to Light Cone Intelligence and get more cutting-edge scientific and technological knowledge!