According to an article published by the US Wall Street Journal on August 1st, in order to develop AI models that can replace those from China and offer better value for money, several startups in Silicon Valley are competing to develop their own open-source models. However, some projects face little interest from venture capital investors, and they have to operate with limited budgets.
By the end of 2025, Arcee AI, an unknown startup in Silicon Valley, has invested most of its remaining funds into a ambitious goal: to create open-source large-scale models with the highest possible capabilities.
According to reports, Alcy AI completed pre-training in 33 days and achieved its goals. The investment required was much lower than that of leading industry laboratories. The final product is a high-cost-performance AI system that users can download and customize themselves. It has become a domestic competitor to China’s emerging models that are shaking the global AI competition.
Currently, Silicon Valley and Washington are increasingly concerned that Chinese open-source models such as Kimi, Tongyi Qianwen, and DeepSeek have approached the performance of top American products. These models could potentially challenge the profitability of American AI companies in the coming years.
Therefore, a few domestic manufacturers in the United States are competing to develop competitive models of similar products. American media claim that some users have concerns about so-called data security risks associated with Chinese AI, and therefore prefer to use American alternatives. Chinese large-model manufacturers deny that their products pose any security risks.
Although there is a desire to compete with Chinese AI in terms of cost-effectiveness, American open-source AI companies generally face a major bottleneck: difficulties in financing. Although there is ongoing domestic demand for developing more powerful open-source models to compete with Chinese products, investment institutions are reluctant to fund such start-up projects.
Alsi team photo. The Wall Street Journal
Alceste Company CEO Mark McGuire said directly, "Almost all of the leading venture capitalists have rejected us outright."
Alsi, Reflection AI, Poolside and other new-generation open-source model companies are betting on the growing market demand: The market needs high-performance open-source large models that combine the efficient and low-cost advantages of Chinese models with the ability to avoid "geopolitical risks".
Pوسيد's co-founder and co-CEO, جيسون وورنر, said: "The market is extremely eager for American companies to launch open-source artificial intelligence products with top-notch performance." The company released a new generation of open-source weight model series called Laguna S 2.1 in July.
Open-source weight models allow users to download the weight values corresponding to billions of parameters of the model. These models can be run locally on specialized hardware, and their weights can be fine-tuned using incremental data for further development. Such models are often referred to as open-source models in the industry. These models offer greater transparency, as all core materials such as weight information and training code are fully available to users.
The Wall Street Journal states that the United States was originally a leader in open-source AI. The entire open-source AI community largely developed based on Meta’s Llama series of models. However, China quickly caught up and surpassed them.
In the past year, many companies have seen the cost of AI computing power skyrocket, so they have switched to using open-source models from China, which offer much lower pricing.
This shift in market focus has brought open-source technologies back to the forefront of public attention, and it also highlights the growing gap between China and the United States in this field. To enhance the competitiveness of high-end products, American companies including OpenAI have recently lowered the pricing of some models.
Microsoft's venture capital fund M12's managing partner, Michael Stewart, commented: "The industry has finally begun to realize that in the future, open-source models will be the default choice. This trend is already beginning to emerge."
NVIDIA CEO Jensen Huang is one of the most vocal supporters of open-source AI. In July, NVIDIA led the release of a letter calling for increased support for open-source models, and also urging policymakers to avoid introducing "premature restrictive provisions."
Although executives from both Anthropic and OpenAI expressed overall support for open-source models, they also pointed out the need to balance the pros and cons. Dario Amodei, CEO of Anthropic, mentioned in a recent blog post that open-source models pose potential risks of being abused, used in cyberattacks or biological weapon attacks.
NVIDIA has released several open-source models from the Nemotron series. Earlier this year, it joined with several AI laboratories to form an open-source technology alliance. NVIDIA has also become one of the largest investors in open-source startups, making significant investments in Reflect AI (which has raised over $2 billion and plans to launch its first open-source model later this year), Bosideng, and the Thinking Machine Lab, which released its first open-source model, Inkling, in July.
Apart from NVIDIA and its invested companies, the overall size of the open-source AI industry in the United States remains small. However, Chinese open-source models still maintain a leading performance.
Venture capital institutions are showing little willingness to invest in open-source projects, further hindering the development of this industry. Many investors question how companies that use open-source models can generate stable revenues through free use; they also worry that such technologies could disrupt the core businesses of OpenAI and Anthropic, damaging their own investments.
In the United States, closed-source large model laboratories still hold a dominant position in the industry, being tied to almost all of the top investors in Silicon Valley. According to statistics from data firm PitchBook, in just the first quarter of this year, total funding for AI startups in the US reached $255.5 billion, with nearly two-thirds of this funding going to companies like OpenAI, Anthropic, and xAI during their fundraising rounds.
Early investor Joe Freud, who once helped Alcy with financing and worked at Emergent Capital, said, "I've heard investors make blunt excuses over and over again: 'I don't want this project to succeed; investing in it will drag down my investments in Anthropic and OpenAI.'"
Alsi McGuire originally thought that venture capital would show greater interest, but instead, there was widespread skepticism and denial. Recently, on social media platform X, there has been an increasing number of calls for the development of self-developed open-source weight models in the United States that can compete with Chinese products.
"I support each of the local open-source AI foundations, and I have almost individually contacted them all. McQuinn said his goal is to make the United States catch up and surpass China."
McQueade has worked for the AI platform Hugging Face, and previously served in the Amazon Web Services team. In 2023, he co-founded Alsy with Jacob Solavitz and Brian Benedict. Initially, they did not plan to develop a large-scale foundational model. After Meta significantly reduced its resources invested in the Llama model, he saw an opportunity in the market.
Alsi has accumulated a total financing of $50 million, with a post-investment valuation of $240 million. The company consists of about 30 employees. Initially, the company launched a small model with 4.5 billion parameters, and subsequently iterated to produce the "Trinity Large" version. The team completed training using approximately 2,000 NVIDIA B300 chips.
The “Three One Large Model” still has a much smaller scale compared to industry-leading models. It lags behind Anthropic and OpenAI products in many authoritative evaluation metrics. However, Alsi aims to rely on a new round of financing to develop a next-generation model that is larger in scale and more powerful.
According to the report released by the world's largest open-source AI model platform in spring 2026, Chinese-developed open-source models account for 41% of the global total downloads. China ranks first worldwide, surpassing the United States. On the global list of mainstream large models, all of the top six models are from Chinese teams. Over the past 12 months, Chinese models maintained a size limit among global open-source models for 9 months, with their iteration pace continuing to lead the world. The cumulative downloads of Chinese open-source models have exceeded 10 billion times, making China the country with the highest number of downloads globally.
In mid-July, Moon's Dark Side released the Kimi K3, which achieved performance levels that matched or even surpassed those of closed-source systems in the United States. The model weights were made fully public, allowing developers to download and deploy it on their own infrastructure, without incurring API fees.
Kimi K3 Brand Visual Image. The Dark Side of the Moon
In May, DeepSeek announced a permanent discount of 75% on the V4-Pro API. The cache hit price dropped to 0.025 yuan per million tokens, setting a new global low price. DeepSeek's low-price strategy is not a loss-making subsidy, but rather a way to compress computing power consumption by 27% compared to previous products, using sparse attention and mixed expert architectures. On July 31, DeepSeek announced in the API documentation update that the official version of DeepSeek-V4-Flash API has started public testing. The model version was updated to DeepSeek-V4-Flash-0731 (hereinafter referred to as V4-Flash-0731).
Beijing Zhongguancun College's dean, Liu Tieyan, said that the vitality of technology lies in its open and fluid nature. A number of Chinese open-source large models have progressed from individual breakthroughs to collective achievements, providing new solutions and pathways for the development of artificial intelligence worldwide. This will accelerate the positive and beneficial development of artificial intelligence technologies.
