This website uses cookies

Read our Privacy policy and Terms of use for more information.

Fei-Fei Li, Geoffrey Hinton, a moderator, and Andrew Ng seated across the main stage at Ai4 2026 in Las Vegas. Li sits at far left in a light blue blazer with her hands clasped, while Hinton, dressed in black, sits beside her near a small table with water bottles. The moderator sits in a white armchair holding notes and turns toward Ng, who sits at far right in a dark suit and light blue shirt while speaking. A large illuminated white “Ai4” logo dominates the black stage backdrop, framed by blue and purple lighting. Dark silhouettes of audience members fill the foreground, including an attendee holding up a phone toward the stage.

Fei-Fei Li, Geoffrey Hinton and Andrew Ng onstage at Ai4 2026 in Las Vegas. Photo by Shane Snider [email protected] or on LinkedIn

Consistent, Coordinated Effort to Stifle Open-Source AI: Andrew Ng

Key Takeaways:

  • Andrew Ng said at Ai4 that a “very well-funded effort” is trying to stifle open-source AI, arguing that the rationale for restrictions has shifted from extinction risk to bioweapons, job losses, competition with China, and other concerns.

  • Ng argues that companies that have invested billions of dollars in proprietary AI models have an economic incentive to restrict open-source alternatives that can be released for anyone to use at little or no cost.

  • Jim Harris argues that the large cost difference between proprietary U.S. frontier models and Chinese open-source and open-weight models threatens the economics of investments in expensive proprietary AI.

  • Harris calls the resulting practice “token arbitrage,” with companies routing AI workloads among models based on factors including accuracy, cost, latency, privacy, security, data sovereignty, reliability, capability, scalability, and geopolitical risk.

  • Chinese open-source models have rapidly increased their share of U.S. usage on OpenRouter, although Harris notes that OpenRouter represents only part of the U.S. AI market and should be viewed as an indicator rather than a measure of the entire market.

  • Harris argues that restricting access to lower-cost Chinese open-source models would increase AI workload costs for U.S. companies while reducing choice for businesses, academics, and researchers.

At Ai4 in Las Vegas on August 5th Andrew Ng, founder of DeepLearning.AI and Coursera appeared on main stage along with Geoffrey Hinton, often called the Godfather of AI; Fei-Fei Li, co-director of the Stanford Institute for Human-Centered AI often called the godmother of AI. Ng’s comments regarding open source are important. Here is an edited transcript of his comments:

The real story is, over the last few years, a number of [US] businesses have invested billions of dollars, to train large AI models.

And one of the leading forces of AI technology is open source or open-weight software [models]. That's when someone trains an AI model and releases it free for anyone [in the world] on the internet to use.

It turns out if you spend billions of dollars training something, building something, it's really annoying if someone else gives the same thing away for free. It really degrades the value of your investment.

Over the last three to four years, I've been really surprised at the intensity of the fear-mongering to stoke up fears in our government. They tried to pass laws to stifle [free open-source AI software]. Software that free for anyone to use.

Over these last few years, the same people have said AI will lead to human extinction.

That caused a bunch of us, including I think Fei -Fei and me, to fly to DC to testify, reassure our government, no AI will not lead to human extinction.

After that narrative lost credibility, and our government no longer believes that, thank God, the same people said, “Oh, AI could create bioweapons.” After the RAND and other studies, no, AI is not that dangerous, not going to create bioweapons.

Then the same people are now saying, yeah, it will lead to massive job loss. And most we see is [due to] China, China, China.

So the same people have been repeatedly shifting their narrative over the last three to four years to stifle the ability of others to release software free for you, free for anyone to use . . . We must pay attention to this.

Another [narrative has been] an AI conscious being will attack us. It’s part of this ongoing, very well-funded effort to stifle open source.

Why is This Important?

In August 2026, GPT-5.4 Pro cost $180 per million output tokens, while DeepSeek V4-Flash cost 66 cents per million output tokens during off-peak hours and $1.32 during peak hours. That creates a more than 135-fold difference in output-token pricing at DeepSeek's peak rate. The models are optimized for different purposes, however, so the figures should not be interpreted as an apples-to-apples comparison of equivalent models.

The broader point is that Chinese open-source and open-weight models can deliver impressive intelligence and performance at substantially lower prices than proprietary U.S. models.

This enormous difference in cost is threatening the billions of dollars of capital investment of VCs and tech firms in US proprietary frontier models. And those VCs and tech firms want to receive an ROI on their investments.

But water flows downhill. And any time you have a significant difference in price for roughly equivalent products and services the market will move. This is leading organizations to optimize workflows for what I call token arbitrage.

Token Arbitrage

Companies are increasingly using AI workload placement or model routing to optimize each task based on 15 factors, such as:

Table outlining Jim Harris's 15 factors for AI workload placement and model routing. The factors and their determinants are: 1. Accuracy/quality — How good does the answer need to be? 2. Cost — Acceptable cost per task, transaction or million tokens. 3. Speed/latency — Must the answer be rendered in 200 milliseconds, 2 seconds or 2 minutes? 4. Task importance — Summarizing meeting notes or making a $10 million decision? 5. Consequence of error — What happens if the AI gets it wrong? 6. Data sensitivity — Does the prompt contain customer, employee, financial, intellectual property or confidential data? 7. Privacy — Can the provider retain, inspect or use prompts or data? 8. Security — What's the risk of leakage, attack or unauthorized access? 9. Data sovereignty — Must the data or model remain in the U.S.? 10. Regulation — HIPAA, financial regulation or government requirements? 11. Reliability — How critical is continuous availability, and what happens when the model or provider goes down? 12. Model capability — Requirements such as reasoning, coding, vision, voice or huge context. 13. Integration — How easily does it connect to existing data, cloud services, applications and workflows? 14. Scalability — Can it predictably handle 10 requests or 10 million? 15. Geopolitical risk — Can the provider, jurisdiction and supply chain be trusted?

15 factors companies can consider when routing AI workloads among models. Framework by Jim Harris.

The Impact

In late 2024 Chinese open-source models had a negligible US market share: as low as 1.2% according to OpenRouter and Andreessen Horowitz's 100-trillion-token study. On July 13, 2026 64% of tokens used by US users on OpenRouter were open-source Chinese models. It’s important to note that OpenRouter doesn’t track the entire US AI market, but it’s a good indicator of which way the market is moving.

Summary

If the Washington lobbying efforts to restrict access to Chinese open-source models were successful it would radically increase the cost of US companies’ AI workloads. But the financial backers for large frontier models would be winners. Ultimately businesses, academics, and researchers want freedom of choice and ability to control costs. Restricting the market will only slow innovation and drive up costs.

Sources:

About Jim Harris:

Jim Harris is a #1 International Bestselling author published in 80 countries. His new book The Age of Fakes!: How AI Abuse, Fake News, and Deepfakes Threaten Business and Society is a #1 International bestseller in multiple categories and countries. You can see it at https://www.amazon.com/Age-Fakes-Deepfakes-Threaten-Business/dp/3982801001/ref=sr_1_1

Follow Jim on Linkedin.com

Interested in having Jim as an AI keynote speaker at an upcoming event? Email him at [email protected]

Keep Reading

View more
caret-right

Home | About Us | Our Approach | FAQ | ARCHIVE | DRIVING TOMORROW PODCAST | CONTACT US | ADVERTISE WITH US

© 2026 AiNews.com.
beehiivPowered by beehiiv