What are Chinese AI models, and why should commercial real estate investors care? Chinese AI models are open-weight large language models built by companies such as DeepSeek, Alibaba (Qwen), MiniMax, and Moonshot (Kimi) that now run a rapidly growing share of enterprise AI workloads because they cost 60 to 90 percent less than the leading US models. On July 7, 2026, CNBC reported that Chinese open models have taken more than 30 percent of US enterprise token usage every week since February, spiking as high as 46 percent, based on data from the model-routing platform OpenRouter. For CRE investors deciding how to use Chinese AI models for commercial real estate work without letting AI spend balloon, this shift rewrites the cost math. For a broader framework on picking between providers, see our complete guide on AI model comparison for CRE investors.
Key Takeaways
- Chinese open-weight models like DeepSeek and Qwen now handle 30 to 46 percent of US enterprise AI usage, up from roughly 11 percent a year earlier, driven almost entirely by price.
- The cost gap is stark: DeepSeek V3.2 lists around $0.42 per million output tokens, while frontier US models run into the tens of dollars per million, a 60 to 90 percent discount.
- The smart CRE move is a two-tier AI stack: route high-volume, low-sensitivity tasks to cheap models and reserve premium US models for sensitive, high-stakes analysis.
- Data governance is a live risk: US lawmakers have opened inquiries into Airbnb and Anysphere over Chinese-model use, a warning for firms handling tenant data and LP information.
- Cheaper "good enough" AI lifts the ROI of routine work like lease abstraction and comps, where a Bain study found most enterprises still capture only single-digit cost savings.
Chinese AI Models for Commercial Real Estate, Explained
Chinese AI models matter to CRE because they collapse the cost of the routine language work that fills a real estate shop's day. The core news: on OpenRouter, a platform whose users are about 94 percent non-Chinese developers, four of the top five models by usage are now Chinese, led by DeepSeek V4 Flash. Chinese open models jumped from about 1.2 percent of global usage in late 2024 to nearly 30 percent in a matter of months, and among US companies specifically that share has topped 30 percent every week since February 8, 2026, reaching 46 percent at its peak.
The driver is not ideology, it is arithmetic. As one analyst summarized the trend, when a task does not need the best model, teams route it to the cheapest one that is good enough. Open Chinese models can be 60 to 90 percent cheaper than Anthropic's Claude or OpenAI's GPT line. AI startup Lindy said it moved 100 percent of its traffic from Claude to DeepSeek to save millions, and Airbnb uses Alibaba's Qwen for customer-service chatbots. An Andreessen Horowitz partner estimated that roughly 80 percent of US startups now build on Chinese base models.
Why the Cost Collapse Matters for CRE Budgets
For CRE firms, cheaper tokens turn AI from a line item you ration into infrastructure you can run at scale. A mid-sized multifamily operator that abstracts 500 leases a month, drafts market summaries, and screens hundreds of listings was, a year ago, choosing between a large AI bill or simply doing less. At frontier prices, processing a single 30-page lease might cost a few dollars in tokens; at Chinese-model prices, the same job can cost a few cents. Across thousands of documents a quarter, that is the difference between a rounding error and a real operating expense.
This is the same dynamic we covered in the collapsing cost of frontier AI, now accelerating. Lower unit costs mean a firm can afford to run AI over the whole rent roll rather than a sample, re-underwrite a portfolio monthly instead of quarterly, and let analysts iterate freely. The point is not to shave a few basis points off a cap rate; it is to make AI-assisted analysis cheap enough that using it everywhere becomes the default. For how to structure that spend, see our guide on AI cost management for CRE.
A Two-Tier AI Stack for CRE Firms
The right response is not "switch everything to the cheapest model," it is to route each task by sensitivity and complexity. A practical two-tier framework looks like this:
- Tier 1, cheap and high-volume: Lease abstraction, OCR cleanup, comps summarization, first-pass rent-roll parsing, meeting notes, and marketing copy. These are low-sensitivity, high-repetition tasks where a good-enough open model at a fraction of the cost is the rational default.
- Tier 2, premium and high-stakes: Investment-committee memos, LP communications, complex DSCR and IRR modeling narratives, legal-adjacent review, and anything touching tenant data or competitive strategy. Here the marginal cost of a frontier US model with clear data governance is worth paying.
The tiering echoes the same match-the-model-to-the-job logic behind GPT-5.6's Sol, Terra, and Luna models: pick the model for the task, not the other way around. Tools like OpenRouter make this routing mechanical, letting a firm send each request to the model that fits its cost and risk profile. For CRE teams that want this routing built and governed for them, The AI Consulting Network specializes in exactly this.
The Data Security and Vendor Risk You Cannot Ignore
Before a CRE firm pipes deal data into any model, it has to weigh where that data goes. Chinese open-weight models can often be self-hosted, which can actually improve data control, but many teams call them through third-party APIs whose data-handling terms deserve scrutiny. The political dimension is now concrete: US lawmakers have opened inquiries into Airbnb and Anysphere, the maker of the Cursor coding tool, over their use of Chinese models such as Qwen and Kimi.
For CRE, the sensitive categories are obvious: tenant personal data, borrower financials, LP identities, and pricing strategy. None of that belongs in an unvetted API regardless of the vendor's home country. The defensible posture is a written model policy that classifies data, names approved models for each class, and keeps regulated or confidential information on vetted, contractually governed systems. The AI Consulting Network helps CRE firms build exactly this kind of model-routing and data-governance policy so that cost savings never turn into a compliance problem.
Real-World CRE Applications and the ROI Reality
Where do cheap models actually earn their keep in CRE? The clearest wins are volume tasks with a human check at the end: abstracting a stack of leases into a comparison table, extracting expense lines from a T12, drafting a first-pass offering memo, or flagging any deal with a DSCR below 1.25x for a closer look. These are exactly the jobs where paying frontier prices was never justified, and where NOI-level accuracy still gets confirmed by a person.
The caution comes from the numbers. A Bain and Company survey of 951 companies found cumulative enterprise AI investment has passed $1 trillion, yet only about 4 percent of firms reported cost savings above 30 percent, and the largest group saw 10 percent or less. Cheaper models help, but they do not create return on their own. That still comes from choosing the right workflows, a point JLL research on real estate technology adoption keeps making: most firms have started AI pilots, but only a small share have scaled them into measurable savings. CRE investors looking for hands-on help matching cheaper AI to the workflows that actually move returns can reach out to Avi Hacker, J.D. at The AI Consulting Network.
Frequently Asked Questions
Q: Are Chinese AI models safe for commercial real estate firms to use?
A: It depends on the task and the deployment. For low-sensitivity, high-volume work like lease abstraction or comps, they can be a cost-effective choice. For tenant data, borrower financials, or LP information, use vetted, contractually governed systems and avoid piping confidential data into unvetted third-party APIs.
Q: How much cheaper are Chinese AI models than US frontier models?
A: Roughly 60 to 90 percent cheaper for comparable tasks. DeepSeek V3.2, for example, lists around $0.42 per million output tokens, versus tens of dollars per million for top-tier US models, which is why US enterprise usage of Chinese models has climbed to as much as 46 percent.
Q: What is the best way for a CRE firm to control AI costs?
A: Adopt a two-tier stack. Route routine, low-risk tasks to cheap models and reserve premium models for sensitive, high-stakes analysis. A model-routing layer such as OpenRouter automates the choice, and a written data-classification policy keeps confidential information on approved systems.
Q: Does cheaper AI guarantee a better return for real estate investors?
A: No. A Bain study found only about 4 percent of enterprises achieved cost savings above 30 percent despite more than $1 trillion in AI investment. Savings come from applying AI to the right workflows, not simply from paying less per token.