As AI model costs continue to climb, a growing number of AI startups are reassessing their reliance on major model providers like OpenAI and Anthropic, pivoting toward open-weight models that can be downloaded, modified, and trained in-house. This shift not only promises to lower operating costs for AI application companies but could also pressure the revenue growth and business models of the two leading model developers.
Legal tech startup Harvey exemplifies this transformation. The company, valued at $15.6 billion, had primarily relied on OpenAI’s GPT-4 and other models to deliver specialized AI services for lawyers. But after Harvey updated its AI agent tools in March, customer usage surged dramatically, driving up model inference costs. Gross margin plummeted from roughly 50% at the start of the year to negative 50% by June—demonstrating that rapid growth in AI service demand doesn’t necessarily translate into higher profitability.
Harvey reports that AI token usage has increased 20-fold this year. Sustaining such growth would be untenable if the company remained dependent on OpenAI and Anthropic’s high-cost models. As a result, Harvey began adopting open-weight models and launched its own AI model in August, built on the foundation of Moonshot AI’s Kimi K3 from China. According to people familiar with the matter, the new model’s performance approaches that of Anthropic’s most advanced models but costs only a fraction of the price. Combined with other adjustments to its AI usage strategy, Harvey’s gross margin has returned to positive territory.
Gabe Pereyra, Harvey’s co-founder and president, said that AI application performance previously depended heavily on the underlying model’s capabilities, making it necessary to pay for OpenAI and Anthropic’s best models—with limited value in training proprietary models. But that calculus has begun to shift under cost pressure.
Open-weight models make model parameters publicly available for external developers to download and modify, allowing companies to conduct post-training and customization using their own data without relying entirely on large AI companies’ closed models. For AI application businesses, this not only reduces per-call model costs but also increases control over their technical architecture.
The Self-Built Model Wave Spreads Across Industries
This shift is spreading across multiple sectors. Medical tech startup Abridge has announced plans to build a customized foundation model for clinical scenarios, based on Nvidia’s (NVDA) open models. AI customer service startup Decagon now handles 80% of queries through its proprietary model; fintech company Ramp and financial AI startup Rogo are also evaluating in-house model training.
In the software development space, Cursor and Cognition—valued at $48 billion—were early adopters of customized models among AI application companies. Cursor has since been acquired by SpaceX. This signals that the competitive focus for AI application companies is shifting from simply accessing the most powerful foundation models to building models tailored to their business needs at lower cost.
Karim Atiyeh, Ramp’s co-CEO, said the company previously believed building its own models made no economic sense, but that has changed as open-weight AI models have improved dramatically. Ramp, which completed a $750 million funding round in June, is now evaluating training its first proprietary model. He noted that a year ago, building models in-house “made absolutely no sense,” but with open model advancements, the economics have become more compelling.
Venture capital firms are also backing this trend. Sequoia Capital and General Catalyst are both supporting companies pursuing this path, partly because owning model capabilities reduces AI spending while avoiding complete outsourcing of critical technology to OpenAI or Anthropic.
Lan Xuezhao, founder and managing partner of San Francisco-based venture firm Basis Set, takes an even harder line: “If you’re not optimizing costs and fine-tuning your own models, you’re by definition inefficient. If a company isn’t considering building its own models, it probably can’t even get funded.”
Cost Pressure and Supplier Risk
The focus on AI model costs is also tied to rapidly increasing usage. Some AI companies now pay not just model subscription fees but also additional charges based on model usage volume. OpenAI and Anthropic have both recently introduced lower-priced models in response to customer cost concerns, but enterprises still face substantial token usage fees.
For example, Uber exhausted its entire annual AI budget before April this year. One reason: engineers were encouraged to maximize use of Anthropic’s Claude Code, causing AI usage to spike rapidly. Such cases highlight how AI agent tools, once deployed across enterprise workflows, can drive explosive growth in model usage while rapidly escalating costs.
For AI startups, another appeal of open-weight models is reduced supplier dependency. After SpaceX completed its acquisition of Cursor, OpenAI announced it was suspending Cursor’s access to its models, citing Cursor’s past involvement in Elon Musk-affiliated companies violating OpenAI’s terms of service. Such incidents have prompted some companies to consider the risk of building core products heavily dependent on a single model provider—if that provider changes its business strategy or terminates service, it could directly impact product operations.
The shift toward proprietary models is particularly noteworthy for OpenAI and Anthropic because both companies are themselves actively expanding into the AI application market. Both have accelerated hiring this year and launched plugins and pilot programs across industries including legal, financial, and healthcare—putting them in more direct competition with the AI startups that use their models.
Anthropic recently presented investors with data on the trend of AI startups building their own models, including Rogo and medical AI company OpenEvidence in related briefing materials. Anthropic also acknowledged Harvey’s proprietary model development but noted that Harvey still requires advanced models like Claude Opus for its most complex workloads. This reflects the reality that building proprietary models doesn’t mean abandoning OpenAI or Anthropic entirely—rather, a hybrid model is emerging where “in-house models handle high-volume general tasks, while top-tier closed models handle complex tasks.”
The Real-World Obstacles to Building Your Own Model
However, building or fine-tuning models comes with significant barriers. First is talent cost. AI engineers with model training and post-training capabilities are in short supply, with top talent commanding annual compensation in the millions of dollars—while also facing poaching risk from large AI companies like OpenAI and Anthropic. Matt Kraning, partner at Anthropic investor Menlo Ventures, noted that engineers capable of model fine-tuning can earn millions of dollars annually and are highly susceptible to being recruited away by major institutions.
Second is the data problem. To build a truly competitive specialized model, companies must obtain large volumes of high-quality, domain-specific data. Harvey cannot directly use clients’ sensitive legal data for training, so it has turned to AI data company Mercor to acquire relevant specialized data and capabilities.
Additionally, running open models isn’t necessarily cheaper than using closed models. After downloading a model, companies still bear the costs of compute infrastructure, model hosting, and maintenance. Andrew Dai, CEO of visual AI startup Elorian, pointed out that the expense of downloading open-weight models and managing compute infrastructure independently is substantial—for early-stage companies with low traffic, paying per-use fees to closed models may actually be more economical.
Some companies have also discovered that building their own models isn’t worth the effort. Salespeak previously planned to build its own large language model but abandoned the effort after months of testing, concluding that the in-house model offered insufficient advantages over readily available models from Anthropic and OpenAI. Omer Gotlieb, Salespeak’s co-founder and CEO, said the company failed to see a “significant advantage” over off-the-shelf models.
As a result, the AI industry is more likely to see model mixing rather than a wholesale shift from closed to open models. Logan Bartlett, managing director at Redpoint Ventures, said the trend of AI startups reducing dependence on Anthropic’s models may continue, but when companies need the most powerful AI capabilities, they will still use the best models available—even at higher prices. “They’re not going to cut off their nose to spite their face,” he said.
Revenue Impact on OpenAI and Anthropic
This trend is unfolding as OpenAI and Anthropic prepare to enter public capital markets. Anthropic’s annualized revenue surpassed $65 billion as of end-July, up sharply from approximately $9 billion at the end of 2025; OpenAI’s annualized revenue crossed $40 billion in July this year. Anthropic projects revenue of approximately $190 billion to $200 billion by 2028.
But as enterprise customers begin seeking cheaper models, both companies’ revenue growth sustainability will depend on whether their premium models’ performance advantage over open-weight alternatives is sufficient to justify the price gap. On the other hand, the AI application market itself is still growing rapidly, so even as some companies increase their use of proprietary models, they may continue to employ OpenAI and Anthropic for their most complex tasks.
For the broader AI industry, this signals that the value distribution between model providers and application companies may be shifting. As open-weight model performance gradually approaches that of top-tier closed models, AI startups gain more options—switching between different models based on task requirements, cost considerations, and data control needs. This means OpenAI and Anthropic face not only competition from each other but also from an increasingly mature open model ecosystem.