Building an American TSMC
Research Rundown #147, plus new memos on Quora, Mixpanel, and more.
Research Rundown
The history of the semiconductor is the history of the United States: from inventor, to manufacturer, to consumer, to hostage. Today, America’s technological dominance rests on a single point of failure: Taiwan Semiconductor Manufacturing Company (TSMC). Responsible for 92% of the world’s most advanced chips, TSMC sits just 110 miles from China’s coast. And China is convinced Taiwan belongs to them. We’ve allowed ourselves to funnel our most important resource through a massive geopolitical flashpoint. And we can’t continue to pretend like its not going to be an issue.
Over the course of decades, as the complexity and cost of manufacturing leading-edge chips exploded, the US ceded much of its ability to contribute. Domestic chip fabrication fell from 100% in the 1960s to just 8% by 2024. The loss wasn’t just in capacity, but in the “chain of experience” essential to technological evolution. However, that gap was exposed by pandemic chip shortages, AI’s compute demands, and rising tensions over Taiwan. Now, things need to change. We need to build an American Semiconductor Manufacturing Company. But how?
There’s three options to rebuild domestic chip fabrication capabilities:
Build it from scratch
Borrow the capability from TSMC
Buy it from an existing player.
The first is too slow and talent-starved; the second keeps America dependent on Taiwan’s political will and supply constraints. That leaves Intel: the only US company still attempting to produce leading-edge chips. But Intel Foundry is in crisis, hampered by cultural drift from engineering to management and a business model that alienates potential customers. If Intel Foundry can’t secure major external customers then we’ll be facing the “death of America’s Foundry.”
Intel is our only viable shot at sovereign semiconductor capacity, but not Intel as it stands today. Becoming America’s Foundry will require visionary leadership, a capital structure built for growth over cost-cutting, and a return to technical leadership through bets like high-NA EUV and 14A chips. And even then, Intel can’t do it alone. It will take an entire ecosystem from government backing and regulatory reform to a loyal domestic customer base and new-age startups. That’s the only way to end America’s dependence on a flashpoint in the Taiwan Strait. Whether we rise to that challenge is the conclusion that remains to be seen
Read our full deep dive here for the full story...
Tines is quietly powering the future of security automation—without writing a single line of code. To find out how, read our memo here.
What if the secret to building breakout products was hidden in your user data—and only Mixpanel knew how to unlock it? Read our memo here for everything you need to know.
Quora is disrupting the AI and search landscape—behind the scenes, Poe is reshaping how we find, share, and interact with knowledge forever. We break it down in our memo here.
Speaking of Intel, the company’s CEO Lip-Bu Tan defended his record after President Trump called for his resignation over alleged conflicts tied to past investments in Chinese-linked semiconductor firms, saying he has always met legal and ethical standards. Intel said it is engaging with the administration and remains aligned with US national and economic security priorities.
OpenAI has released its first open-weight large language models since 2019, aiming to counter China’s DeepSeek and undercut rivals like Meta, Anthropic, and Google by offering high-performing, smaller, and easier-to-run alternatives that enterprises can deploy on-premise while betting GPT-5 will still command premium demand.
Anthropic has released Claude Opus 4.1, an upgraded version of its flagship model that offers improved coding, research, and data‑analysis capabilities including better handling of complex, multi‑step problems to position itself more competitively ahead of OpenAI’s GPT‑5 release
OpenAI has also launched GPT-5 for all ChatGPT users, including free-tier accounts, with improved reasoning, lower hallucination rates, and new “safe completion” safeguards. It’s rolling out to Microsoft 365 Copilot, Azure AI Foundry, and APIs, with early testers like Box citing major gains in complex logic.
Anduril Industries opened a Taiwan office, expanded partnerships with NCSIST and local firms, and delivered Altius loitering munitions within six months of contract signing, representing its first to Taiwan and underscoring a long-term commitment to accelerate autonomous defense capabilities in support of Taiwan’s security.
Cognition, the $4 billion AI coding startup behind Devin, offered nine months’ salary buyouts to roughly 200 Windsurf employees it acquired three weeks ago, following 30 layoffs, as CEO Scott Wu told staff the company demands 80-hour, six-day workweeks; the move highlights the turbulence and uneven outcomes in partial startup acquisitions when founders and key talent depart.
Google said its AI-powered vulnerability researcher Big Sleep, developed by DeepMind and Project Zero, identified 20 flaws in popular open source software like FFmpeg and ImageMagick demonstrating the promise and challenges of LLM-based bug hunting as human experts still verify results before disclosure.
Reflection AI, a New York startup founded by ex-Google DeepMind researchers, is in talks to raise over $1 billion to build open-source large language models that can rival Meta, Mistral, and China’s DeepSeek, aiming to position itself as the leading US provider amid growing corporate demand for low-cost, customizable alternatives to proprietary AI.
Cloudflare claims that Perplexity has been using stealth, undeclared crawlers that evade robots.txt directives and network blocks by rotating IPs, altering user agents, and accessing content from millions of domains; behavior it contrasts with OpenAI’s declared, compliant crawlers leading Cloudflare to block Perplexity as a verified bot.
Perplexity defended its AI assistant against Cloudflare’s claims of “stealth crawling,” arguing that its user-driven agents fetch content only in real time to answer specific user requests without storing or training on the data, and that Cloudflare misattributed unrelated BrowserBase traffic revealing, in Perplexity’s view, a fundamental failure to distinguish legitimate assistants from malicious bots.
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!