Businesses Turn to Open-Weight AI Models as Proprietary AI Costs Spiral Out of Control

Corporate Shift to Open-Weight AI Models Accelerates

Corporate America is experiencing a dramatic shift to open-weight AI models as businesses grapple with spiraling AI costs that threaten profitability. Executive mentions of open-weight AI models or open source alternatives in earnings calls and investor conferences surged sixfold in August and September compared to the same period last year, according to data from AlphaSense1

1

. The shift to open-weight models represents a strategic alternative to expensive proprietary systems from OpenAI and Anthropic, as companies seek cost savings and greater control over their AI infrastructure.

While technology companies lead this transition, businesses embrace open-weight AI across diverse sectors including PNC Financial Services, logistics firm CH Robinson, and industrial giant Siemens1

1

3

3

. These open vs closed AI systems represent fundamentally different economic models, with open-weight options allowing companies to customize and run models on their own hardware rather than paying per-token fees to proprietary providers.

Rising Costs of Proprietary AI Models Drive Change

The economics behind this shift are stark. Vinay Kuruvila, chief technology officer at Tinder, revealed the dating app's AI costs exploded from $1 million annually in January to $10 million by July1

1

3

3

. Tinder began routing queries from non-technical users to cheaper open AI models to prevent another tenfold increase in spending. Kuruvila noted that frontier models like OpenAI's Astra and Claude Fable already provide sufficient intelligence for 90% of tasks, making the premium pricing increasingly difficult to justify.

Legal startup Harvey experienced even more dramatic impacts, with gross margin plummeting from 50% at the start of the year to minus 50% by June as customer usage spiked2

2

. The company's margins turned positive again only after releasing its own model built on Moonshot's Kimi K3 in August. Harvey saw a twentyfold rise in token usage costs this year, driven partly by OpenAI and Anthropic shifting from flat subscriptions to token-based billing on top of base subscription fees2

2

4

4

.

Scale Players Demonstrate Dramatic Cost Reductions

AT&T now runs approximately 40% of its AI workloads on open-weight AI models and aims to reach 70% within a year, according to Andy Markus, the telecom's chief data and AI officer1

1

. Processing 45 billion tokens daily, AT&T found it could move traffic from expensive closed-source models to open alternatives at a fraction of the cost while maintaining accuracy. The company achieved cost reductions of up to 56% for coding and advanced AI tasks, with only a 2% decline in performance quality4

4

.

Usage data from AI infrastructure providers confirms this isn't just talk. In August, open-weight models accounted for 56% of all tokens processed through Vercel's AI Gateway, up dramatically from just 7% in December1

1

. OpenRouter shows a similar pattern, with open-weight models from Chinese labs dominating the top 10 by token usage. These metrics suggest the shift represents a fundamental change in AI model economics rather than temporary experimentation.

Data Sovereignty and Vendor Lock-In Concerns Mount

Beyond pure economics, data sovereignty and security concerns are driving adoption of open-weight AI models. Scott Wallace, senior global director of solutions architecture at Digital Realty, cited sovereignty and security as equally important drivers alongside cost in his company's embrace of open models hosted on private infrastructure1

1

. Digital Realty built its own internal chat interface running on open-weight models while using a mix of open and proprietary AI depending on task sensitivity and complexity.

Mistral CEO Arthur Mensch spent July warning European buyers about vendor lock-in risks with closed models, pointing to Anthropic cutting off Windsurf while building Claude Code2

2

. OpenAI's decision to end its Cursor contract after SpaceX acquired the company, citing past terms of service breaches by Musk businesses, demonstrated how proprietary providers can leverage immense control over customers. While Europe's Data Act provides some switching rights, these cover customers leaving suppliers, not suppliers abandoning customers.

Startups Build Custom Models to Escape Cost Trap

Startups are increasingly building their own models on open-weight foundations to escape unsustainable economics. Abridge is developing a clinical model on Nvidia's open weights, while Decagon now routes 80% of customer queries through proprietary models2

2

. Ramp raised $750 million in June and is weighing model training for the first time, with co-CEO Karim Atiyeh stating it "made absolutely no sense a year ago" but is "starting to make a lot more sense now."

However, self-hosting open-weight AI models requires substantial upfront investment in computing capacity, storage, cybersecurity controls, monitoring tools and skilled employees4

4

. Some companies have found this strategy doesn't work due to higher upfront costs and the need for specialized talent, computer infrastructure and proprietary data. Companies without sufficient data of their own face particular challenges in fine-tuning open models to match proprietary AI performance.

Chinese Labs Dominate Open-Weight Market

The open-weight market has become dominated by Chinese companies offering models at significantly lower prices than U.S. competitors, though capable alternatives come from France's Mistral and U.S.-based Nvidia, Reflection AI and Thinking Machines Lab1

1

. Chinese labs can charge less due to more efficient models and China's lower energy costs4

4

. DeepSeek and Zhipu, two leading Chinese open-weight labs, generate revenue by selling tokens directly to users at far lower prices than OpenAI and Anthropic.

Some companies' clients express concerns about data privacy when using Chinese open-weight models4

4

. This creates a complex calculus for businesses weighing cost savings against potential security and compliance risks. AT&T addresses this by taking open models and tuning them with proprietary data so they perform as well as or better than closed alternatives for specific tasks, giving the company greater optionality as open models continue improving.

Implications for OpenAI and Anthropic Valuations

Public market investors are closely tracking how corporate customers allocate AI spending between frontier models and open-weight competitors1

1

. Rapid improvements in cheaper AI systems this year could threaten revenue growth for Anthropic, preparing for an IPO expected to value the company at $2 trillion or more, and OpenAI, in early talks for private funding at a $1.2 trillion valuation after delaying its own IPO until 2027.

Both companies released cheaper versions of their flagship models last week as the price war with Chinese rivals intensifies1

1

. Anthropic showed investors a slide on startups building their own models, reminding them that Harvey still needs Opus for its hardest tasks2

2

. For now, the two frontier AI labs capture the bulk of AI spending, and their demand for computing power has underpinned much of the data center construction boom. However, many firms continue using proprietary AI when tasks require the best available capabilities, suggesting a hybrid approach may emerge rather than complete replacement.