OpenAI and Anthropic face revenue threats: Under cost pressures, more AI startups are turning to open-source models.

Harvey, a legal AI unicorn, saw its gross margin plummet to -50% this year, becoming a microcosm of the growing trend as API fees from OpenAI and Anthropic continue to rise. More and more AI application startups are embracing open-source weighted models to regain control over costs.

This trend is accelerating. Startups such as Harvey, Abridge, Decagon, and Ramp, spanning the legal, medical, financial, and customer service sectors, have successively announced plans to develop or customize their own models, with some companies already switching 80% of their traffic to their own models.

Leading investment firms such as Sequoia Capital and General Catalyst also provided support behind the scenes.

This move puts direct pressure on OpenAI and Anthropic, both of which are preparing for their highly anticipated IPOs, and the outflow of application-layer customers could erode their most important revenue streams.

The sharp drop in gross margin forced Harvey to restructure its model strategy.

Harvey's situation best illustrates the point.

This legal AI startup, valued at $15.6 billion, has long used OpenAI's GPT-4 as the core of its products. According to Bloomberg, after a major update to its AI agent in March of this year, user adoption surged, but this was followed by a precipitous drop in gross margin—from about 50% at the beginning of the year to -50% in June.

Harvey's AI Token usage has increased twentyfold this year.

Harvey co-founder and president Gabe Pereyra stated that previously, the performance of AI applications heavily relied on the capabilities of the underlying models, making the purchase of the best models from OpenAI and Anthropic a necessary choice, while the significance of self-trained models was relatively limited. However, this logic has begun to loosen under the heavy pressure of costs.

In August of this year, Harvey released its own model, powered by Moonshot AI's Kimi K3. This model boasts performance close to Anthropic's best product, but at a fraction of the cost. Combined with adjustments to other AI usage strategies, Harvey's gross margin has returned to positive territory.

The wave of using open-source models has swept across multiple vertical sectors.

Harvey's transformation is not an isolated case; this trend has spread to multiple industries.

• Medical technology startup Abridge recently announced that it will build its own basic models for clinical scenarios based on NVIDIA's open-source models.
• AI customer service startup Decagon says that 80% of query requests now flow through its proprietary model.
• In the fintech field, Ramp and Rogo are the first to explore the possibility of training their own models.
• In the field of programming tools, Cursor, which has been incorporated into SpaceX, and Cognition, valued at $48 billion, are among the earliest AI application companies to release customized models.

Karim Atiyeh, co-CEO of Ramp, stated that training proprietary models was previously economically unfeasible, but this logic is reversing with the completion of a $750 million funding round in June and a significant leap in the performance of the open-source weight system.

Dr. Lan Xuezhao, founder and managing partner of San Francisco-based venture capital firm Basis Set, took a more assertive stance:

"If you don't optimize costs and refine your proprietary models, you're inherently inefficient. A company that doesn't consider building its own models may not be able to secure funding at all."

Dual pressures from closed-source giants: price increases and market penetration.

This shift is driven not only by costs themselves, but also by the increasingly aggressive business strategies of OpenAI and Anthropic.

Both companies have recently shifted to charging enterprise users extra for model usage, which, combined with the previous base subscription fee, directly penalizes the "token maximization" usage model. Reportedly, Uber exhausted its entire year's AI budget by April after encouraging engineers to maximize the use of Anthropic's Claude Code.

At the same time, both OpenAI and Anthropic have been actively recruiting in core sectors of startups such as law, finance, and healthcare this year, launching plugins and piloting industry applications, directly competing with their customer base.

Startups' reliance on model vendors also carries another risk: the potential loss of access. Less than a week after SpaceX completed its acquisition of Cursor, OpenAI announced it was suspending model access to the programming tools startup, citing past violations of its terms of service by Musk's company.

It was against this backdrop that, at an investor forum held at Anthropic's offices, attendees specifically discussed the trend of startups building their own models. Anthropic stated that Harvey still relies on its most powerful Claude Opus model to handle the most complex tasks.

Real-world challenges of the open-source path

Switching to open-source weight models is not without its costs; the path is also fraught with obstacles.

Talent is the primary bottleneck. Matt Kraning, a partner at Menlo Ventures, an investor in Anthropic, points out that professional engineers capable of fine-tuning models can earn millions of dollars and are easily poached by large organizations such as OpenAI and Anthropic.

Data is another hurdle. Enterprises need a lot of proprietary data to train their own models, and Harvey had to purchase training data from AI data provider Mercor because it could not access its clients' sensitive legal documents.

The infrastructure costs of open-source models are also significant. Andrew Dai, CEO of visual AI startup Elorian, points out that the expense of downloading open-source weight models and managing the computing infrastructure oneself is considerable. For early-stage companies with low traffic, paying for closed-source models on a pay-as-you-go basis may be more economical.

Furthermore, some companies have indeed failed after attempting this approach. The startup Salespeak previously announced plans to build its own large language model, but abandoned the project after several months of exploration. Its co-founder and CEO, Omer Gotlieb, stated that they did not see any "significant advantages" compared to existing models from Anthropic and OpenAI.

Dependence remains, the game continues.

Despite the surge in self-developed technologies, most startups do not expect to completely break free from OpenAI and Anthropic.

Logan Bartlett, managing director at Redpoint Ventures , which also invests in Anthropic, Abridge, and Ramp, expects the industry’s motivation to reduce its reliance on the Anthropic model to continue to grow, but acknowledges that startups will still pay for optimal performance when necessary. “They won’t cut their own limbs out of spite,” he said.

Anthropic's demonstration materials also confirm this reality—Harvey still needs Claude Opus to handle the most complex tasks, indicating that at this stage, a hybrid strategy of using open source and closed source may be the optimal solution for most startups.

This struggle over AI costs and control is essentially a process of application-layer startups seeking to renegotiate with model-layer giants. As the performance of open-source models continues to catch up with closed-source products, this tension will continue to test the boundaries of interests for both sides as OpenAI and Anthropic move towards their IPOs.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.