The return of tech earnings — which kick off with ASML (ASML: NASDAQ) and Netflix (NFLX: NASDAQ) this week — would normally offer a peek at demand, strategy or at least narrative. This quarter, expect something quieter. Guidance will be vague. Forecasts will lean cautious. Executives will talk about “volatility” and “macro headwinds.” hoarding capital. That’s not just earnings etiquette—it’s survival instinct. Meanwhile, tariffs did what no Fed hike could: they broke the market. The most interesting question now isn’t what’s priced in. It’s who flinches next.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

• Companies Discussed

• Meta Panda-ring = Beijing's AI Windfall

• Tweet O’ The Week

• Epistrophy In The News

• 📆 of Epistrophy Events

• Availability This Week

Meta Panda-ring = Beijing's AI Windfall

Understanding how China has accelerated its development of artificial intelligence requires more than tracking chip exports or monitoring Beijing’s military disclosures. It requires following the flow of algorithms—especially those built in the United States. This research note focuses on Meta Platforms (META: NASDAQ) and the whistleblower testimony of Sarah Wynn-Williams, who alleges the company’s open-source models and quiet engagement with Chinese officials helped seed critical elements of China’s AI infrastructure. Meta never launched in China. But according to Wynn-Williams, it didn’t need to. The company’s algorithms—and its ambitions—crossed the Pacific anyway.

Wynn-Williams, Meta’s former director of global public policy, is set to testify before Congress that company executives—including Mark Zuckerberg and Sheryl Sandberg—met regularly with Chinese government officials, beginning as early as 2015. She claims these meetings included technical briefings on emerging technologies, including artificial intelligence, with an explicit goal: help China outcompete U.S. companies. In her prepared remarks, she characterizes Meta’s behavior as a deliberate campaign to gain favor with Beijing, wrapped in secrecy and denial. She also alleges the company pursued a covert program, “Project Aldrin,” to build infrastructure—specifically, a transpacific cable between Los Angeles and Hong Kong that cost as much as $400M —that could have allowed the Chinese Communist Party to intercept U.S. user data. Only Congressional intervention, she says, stopped it.

Meta has denied these allegations. Its spokesperson called the testimony “divorced from reality.” The company emphasizes that it does not operate services in China and insists that any military use of its AI models would violate its license terms. That may be true, but it’s also functionally irrelevant. Meta’s Llama model—released in multiple open-source versions beginning in early 2023—has already been adopted by Chinese institutions affiliated with the People’s Liberation Army to build military-oriented AI systems. The initial Llama 1 model included 7B, 13B, and 65B parameter variants, trained on over 1.4 trillion tokens across a range of curated datasets including Common Crawl, C4, and GitHub repositories. Llama 2, released in July 2023, improved performance significantly through pretraining on 2 trillion tokens and instruction tuning with over 1 million human-annotated examples. Chinese academic papers—including at least three from institutions tied to the Academy of Military Science and the National Innovation Institute of Defense Technology—confirm that researchers fine-tuned the Llama 13B checkpoint using military-domain dialogue data to create models optimized for operational planning, surveillance analysis, and real-time battlefield simulation. These adapted models, including one codenamed ChatBIT, emphasize task-specific reasoning, low-latency inference, and alignment with PLA strategic objectives, and in internal evaluations were claimed to match or exceed the capabilities of GPT-3.5 in domain-specific tasks.

ChatBIT, used only 100,000 military dialogue records—small by modern LLM standards—but still achieved performance levels reportedly rivaling commercial models. Meta’s models, like other large language models, are trained on trillions of tokens and use hundreds of billions of parameters, costing tens of millions of dollars to develop. But once released, they can be copied, retrained, and recompiled at a fraction of that cost. China’s DeepSeek model, which we now know was built using Meta’s Llama architecture and launched for just $6 million, is now seen by analysts as a rival to GPT-3.5. Its architecture closely mirrors Meta’s open weights, and Chinese developers openly credit Meta’s contributions in the model documentation.

Meta’s defense rests on license restrictions—Llama may not be used for military or surveillance purposes—but enforcement is practically impossible. Unlike semiconductors or servers, which are tracked through customs and export licenses, open source foundational models like Llama are digital, duplicable, and nearly untraceable. There is no international framework capable of stopping a foreign adversary from downloading, fine-tuning, and deploying a U.S.-developed LLM.

This raises a strategic problem. Since 2021, the U.S. has placed strict export controls on advanced GPUs and AI chips to limit China’s access to training infrastructure. But those controls are moot if the models themselves are handed over for free. Meta’s AI leadership, particularly Joelle Pineau and Yann LeCun, have defended open sourcing as an innovation driver. That may be true domestically, but the temptation is global. You can’t keep China out of the cookie jar.

Wynn-Williams links Meta’s AI collaboration with Beijing to broader national security risks. In her testimony, she argues that there’s a “straight line” from the company’s 2015 briefings to China’s current use of Llama-based tools in military planning. Meta has dismissed that logic, pointing out that the PLA’s use of its models violates terms of service. But such defenses feel naive. Llama’s architecture has already seeded at least two military-aligned Chinese models, including DeepSeek and ChatBIT, both of which claim superior cost-performance ratios over Western equivalents. The technical design of these systems—Transformer-based encoders, autoregressive generation, parameter-efficient fine-tuning—mirrors U.S. innovation nearly line for line.

Meanwhile, Meta remains financially entangled with Chinese firms. In 2023, Chinese advertisers contributed roughly 10 percent of Meta’s global advertising revenue—an estimated $13.7 billion. PDD Holdings, parent of the Temu app, reportedly spent over $2 billion on Facebook and Instagram ads alone. Shein, a fast-fashion firm based in Singapore but operating largely out of China, accounted for another multibillion-dollar tranche. According to Meta CFO Susan Li, two-thirds of Chinese ad revenue came from outside the top 10 advertisers, suggesting broad exposure to any disruptions in cross-border e-commerce. 

The deeper concern, though, isn’t economic. It’s systemic. Wynn-Williams alleges that Meta worked with the CCP to build censorship tools tailored to China’s information control apparatus. These tools, she says, included facial recognition capabilities, kill switches for content takedown, and custom filters built into the core of Meta’s platform. While the company never officially launched in China, it prepared as if it might—and in doing so, reportedly handed Beijing technical infrastructure that could now be used to suppress dissent at home and abroad.

Project Aldrin, the transpacific cable plan, is a case in point. The proposed route, Los Angeles to Hong Kong, was eventually abandoned after U.S. security officials raised concerns about potential data interception. But for several years, the project proceeded under a veil of internal confidentiality. Wynn-Williams claims the initiative was limited to “need-to-know” staff and driven by a belief that technical cooperation would pave the way for political access.

The AI race has already entered its next phase. Models are smaller, cheaper, and more specialized. The cost of training a competitive language model has fallen by an order of magnitude since 2022. Tools like DeepSeek are now trained with $5–10 million in compute, fine-tuned with domain-specific datasets, and deployed across Chinese institutions without any formal oversight. The result is a proliferation of capable, semi-clandestine AI systems that are strategically aligned with China’s military and surveillance objectives. Meta helped open that door.

It’s unclear whether Wynn-Williams’ allegations will lead to sanctions, legislation, or hearings beyond the Judiciary Subcommittee. Meta will likely continue to deny wrongdoing and emphasize its commitment to transparency. But the stakes are now clear. The Llama model, developed in Menlo Park and released under the banner of open science, has become part of the Chinese military’s AI backbone. It’s not the first such case, and it won’t be the last.

The U.S. can’t put the genie back in the bottle. But it can redraw the boundary between open-source innovation and national security risk. Meta, for its part, still seems to think it can serve both goals at once. Congress may disagree.

Tweet O’ The Week

Epistrophy In The News

NewsNation tapped us three times in a week for commentary on the tariff-driven selloff, its implications for tech and the shocking Facebook testimony before the Congress. I also joined Fox News KTVU live from my New York hotel room to break down how the new tariffs may force a full rethink of U.S. tech supply chains. The segment was fast, sharp and I was—as always—broadcast-ready. That’s thanks to the kit I now carry everywhere: a collapsible key light, a lav mic, and a pocket-sized makeup kit that makes me look at least half awake at any hour. All in my carry on!

📆 of Epistrophy Events

Availability This Week

I’m in San Francisco this week, catching up with a few companies, writing about ASML and Netflix earnings and reacting to the insanity. Thanks to Good Friday (where I’ll be singing Mozarts Requiem with the church choir) the week’s shorter, but the stakes are not. 

Written reports are available to clients, with video summaries on YouTube and, of course our popular summaries of the summaries on Instagram and Tiktok.

I hope these notes are helpful to you. I’d love to discuss them further and, as always, comments, questions and ideas are appreciated.