# A $700 Billion Sovereign Fund Just Made the Chinese AI Cost Argument Impossible to Ignore > Source: > Published: 2026-07-24 01:12:30+00:00 *GIC's public nod to cheaper Chinese open-weight models tells you the AI cost argument has left the developer forum and entered the boardroom.* When a sovereign wealth fund managing more than $700 billion says Chinese AI models can cut adoption costs, you should pay attention. The Financial Times reported that GIC, Singapore's government investment vehicle, has welcomed the rise of open-weight Chinese models including Moonshot and DeepSeek because they could make AI cheaper to use across industries. That's not a lab demo. That's long-horizon capital looking at the bill. The cost gap is the whole story. BasedAI's recent analysis of OpenRouter telemetry, which it says was reported through CNBC, found Chinese open models running 60% to 90% cheaper than leading closed offerings from OpenAI and Anthropic. OpenRouter data also showed Chinese-origin models carrying at least 30% of U.S. company token traffic every week from February 8, 2026, with a peak of 46%. A year ago, that would have sounded like an edge case. It doesn't now. You can see why companies are tempted. Agentic workflows burn through tokens at a rate that makes small per-token differences look large by the end of the month. If your product calls a model thousands of times inside coding agents, support tools, research workflows, or internal copilots, a 60% price cut isn't a nice discount. It's the difference between shipping the feature and rationing it. ## The Price War Is Real For most of the past year, DeepSeek carried the argument. Its January 2025 R1 release showed that a Chinese lab working under tighter chip constraints could still jolt U.S. assumptions about model cost and capability. Now the pattern is broader. Z.ai released GLM-5.2 on June 16, 2026, calling it a long-horizon model with a 1 million-token context window. Vercel's July production index said GLM 5.2's daily token volume grew about 50 times from API availability through the end of June, and that it reached number 11 on Vercel AI Gateway by tokens in the final week. That is fast adoption. It also tells you something awkward for the U.S. frontier labs: users don't always need the most famous model. They need the model that is good enough, cheap enough, and easy enough to run inside the workflow they already have. GIC's view matters because it takes the debate out of benchmark theater. Sovereign funds don't build investment theses around leaderboard screenshots. They look at adoption curves, margins, switching costs, and what their portfolio companies can actually afford. When GIC's chief investment officer Bryan Yeo tells the Financial Times that cheaper open-weight models should help broaden AI adoption, the signal is plain: model pricing is becoming a financial planning input, not just an engineering preference. ## The Compliance Bill Still Arrives None of this removes the security question. It sharpens it. South Korea's Personal Information Protection Commission said in April 2025 that DeepSeek transferred device information, network data, app information, and user prompt inputs to Beijing Volcano Engine Technology without proper consent or disclosure when the service launched in Korea. The commission also said DeepSeek later blocked prompt transfers to Volcano on April 10, 2025. Keep that detail in view. Cheap can become expensive very quickly when data handling is wrong. U.S. officials have already moved in that direction. Congress.gov lists the No Adversarial AI Act, introduced in the House on June 25, 2025, as a bill aimed at restricting federal use of AI systems linked to foreign adversaries. Separate state actions came earlier: Texas banned DeepSeek on state devices on January 31, 2025, New York followed on February 10, and Virginia is listed among states that moved against the app on government systems. For banks, healthcare groups, defense contractors, and anyone selling into government, those facts matter more than the token price. Here is the harder part. Open-weight models can be downloaded, modified, and run locally. That can improve privacy when the deployment is governed properly. It can also leave security teams blind when a developer pulls a model into an internal tool with no procurement review, no central logging, and no legal assessment. The 90% cost delta is easy to model. The unmanaged model on a laptop is not. For startups and mid-market software companies, the trade-off is immediate. If you're not handling regulated data and you're already squeezing gross margins because AI features are expensive to serve, Chinese open-weight models are no longer theoretical competition. Fortune reported in April that Chinese models had surpassed U.S. models in call volume on OpenRouter, and OpenRouter's own June analysis said DeepSeek nearly doubled its token share from January to June 2026. Customers are already moving workloads. OpenAI and Anthropic won't sit still, and they shouldn't be written off as if price is the only thing buyers care about. Enterprise contracts, safety systems, uptime, support, tools, connectors, and product integration all still matter. But pricing pressure has a way of stripping away comfortable assumptions. If a cheaper open model handles the ordinary work, the expensive closed model has to justify itself on the hard work. That's the real issue GIC has put on the table. Chinese open-weight models don't have to beat every U.S. frontier model on every task to change the market. They only have to be strong enough that CFOs start asking why the company is paying frontier-model prices for work a cheaper model can already do. **Also read:** [Meta and BlackRock's $12 billion El Paso bond deal shows the AI arms race is finally moving the credit market](https://startupfortune.com/meta-and-blackrocks-12-billion-el-paso-bond-deal-shows-the-ai-arms-race-is-finally-moving-the-credit-market/), [China is winning the Global South's AI future by giving models away for free](https://startupfortune.com/china-is-winning-the-global-souths-ai-future-by-giving-models-away-for-free/), and [Elon Musk wants rival AI labs to vet each other's models before release](https://startupfortune.com/elon-musk-wants-rival-ai-labs-to-vet-each-others-models-before-release/)