Price:

AI & TECH

AI export controls empower China's open-source rivals

Monday, July 6, 2026 · from 3 podcasts
The a16z ShowThis Week in AIAI Daily Brief
  • Commerce Secretary Howard Lutnick is personally licensing frontier AI, restricting GPT-5.6 and Mythos 5.
  • U.S. firms now use cheaper Chinese models like GLM-5.2, cutting costs and raising security risks.
  • a16Z argues AI embeds national values, making software a vessel for geopolitical influence.

Commerce Secretary Howard Lutnick has manually approved a narrow list of trusted partners to regain access to Anthropic's Mythos 5. OpenAI followed suit with GPT-5.6, restricting its three new models to a limited preview at the government's request.

Nathaniel Whittemore argues this creates a licensing regime without congressional approval. The government is picking winners and losers in frontier intelligence.

"We took this short-term step to work with the administration on a cyber executive order framework."

- OpenAI, The AI Daily Brief

The policy is accelerating China's ascent. A Wall Street Journal report found that 360 Security Technology's tool using GLM-5.2 matched Mythos’ performance in finding cybersecurity bugs. Coinbase CEO Brian Armstrong now defaults to Chinese open-weight models to manage costs, halving the company's AI bill. Open Router's June report shows DeepSeek v4 and Qwen 2.7 are frequently used in agentic workflows.

Andrew Berman called the current landscape "insane." He argues withholding American releases subsidizes China's R&D.

"If the rest of the world builds on a Chinese tech stack because the U.S. version is walled off, American AI sovereignty becomes a hollow victory."

- Nathaniel Whittemore, The AI Daily Brief

Anne Neuberger frames the stakes differently. She says national deterrence now depends on private-sector software speed, not hardware volume. Ben Horowitz adds that AI models embed cultural values, turning them into vessels for soft power. Using American models exports Western liberalism; relying on foreign models risks digital colonization.

The shift is pushing enterprise strategy toward frugality. Victor Perez states the focus has moved from raw capability to performance and price. Andrew Berman cites a case where an agentic loop burned 80% of a company’s 2026 inference budget in 48 hours. The race is now won on efficiency.

Analysts warn this could cement a permanent intelligence gap. Andrew Curran predicts the U.S. government will maintain a perpetual lead, holding the "N+1" model for state use while the public lags a generation behind.

#Models#Big Tech#Regulation
Howard LutnickA16ZAnthropicCoinbaseDeepSeekOpenAIGPTQwenWall Street

Source Intelligence

- Deep dive into what was said in the episodes

The a16z Show
The a16z Show

Technology, Alliances, and American Leadership. • Jul 3

  • Ben Horowitz states that AI models are not objective; they embed strong opinions on history, culture, ethics, and values, which can vary significantly across different regions.
  • Ann Newberger explains that national deterrence has evolved beyond military size to include the pace of innovation in reacting to cheap, software-built technologies used by state or non-state adversaries.
  • Ann Newberger highlights that essential technologies, including AI, autonomous systems, and cyber defense, are primarily developed by the private sector, not by governments.
  • Raghu Rogurum notes that products and technology spread globally faster than companies can establish physical presence, forcing startups to prioritize international expansion much earlier in their lifecycle.
  • The speakers emphasize that technology has transformed from merely an industry into a fundamental source of economic growth, national security, and geopolitical influence for countries worldwide.
  • Ben Horowitz details a16z's international approach with three components: global affairs (government relations led by Ann Newberger), global investors (LP relations), and global go-to-market (tech adoption led by Raghu Rogurum).
  • Ben Horowitz and Ann Newberger caution against "cyber-colonization," where dominant AI models could impose foreign values, history, or censorship if they originate from non-democratic regimes.
  • Ben Horowitz cites an example where 11 Labs' AI enabled Televisa Univision to dub Spanish-language content into various accents and languages, facilitating a global distribution deal with Netflix.
  • Ann Newberger stresses that American AI models reflect values of non-censorship and resist injecting political agendas, offering a trusted alternative to models from authoritarian governments.
  • Raghu Rogurum and Ann Newberger prioritize regions like Japan, Korea, the Middle East, Mexico, and Canada for their strong U.S. alliances, AI-forward policies, and top-heavy economies offering significant influence.
  • Ann Newberger explains that Japan has shifted its post-WWII defense policy to build a capable military for contemporary threats, actively seeking partnerships for autonomous ships and AI in integrated battle spaces.
  • Ann Newberger asserts that while AI can assist both attackers and defenders in finding vulnerabilities, it offers significant potential to improve overall security by accelerating defense capabilities against a broad attack surface.
  • Ben Horowitz identifies a "20-year legacy of code" built in the pre-AI era that is insufficiently secure for the current AI world, presenting a significant challenge and opportunity for modernization.
Also from this episode: (4)

Business (1)

  • Ben Horowitz frames a16z's global expansion as a mission to preserve American leadership by fostering its technological dominance and extending these opportunities to allies globally.

Startups (3)

  • Ben Horowitz estimates that setting up a new country office and localizing a product can cost between $5 million and $10 million before securing a single deal.
  • Ben Horowitz identifies three critical elements for a thriving tech ecosystem: readily available technical talent, laws and policies that facilitate entrepreneurship, and a culture that grants social status to founders and new companies.
  • Ben Horowitz states that Norway's introduction of an unrealized capital gains tax effectively ended its tech ecosystem by discouraging the extreme financial and personal risks associated with starting companies.
#Models#Startups#AI Infrastructure#Macro#Regulation
This Week in AI
This Week in AI

Mythos, Sonnet 5, GLM-5.2 Dominate the News Cycle | Episode 20 • Jul 2

  • Andrew Burman argues US AI policy, including blocking models like GPT-5.6 and restricting access, is harming US technological supremacy by enabling Chinese competition. He believes open-source models are being distilled from US models and the US should push ahead faster.
  • Victor Perez states that creative AI models like Korea's face lower regulatory risk than cybersecurity-focused models, citing NSFW images and impersonation as primary concerns rather than national security threats.
  • Dave Gar explains action models differ from language models by being designed for task execution, such as operating computers, automating work, or acting as personal assistants.
  • Andrew Burman notes enterprises strongly desire open-source models like GLM 5.2 for cost reduction, despite wanting access to powerful models like Fable, creating a cost-performance trade-off dilemma.
  • Andrew Burman describes trusted access programs from Anthropic (Project Glasswig) and OpenAI, which allow large companies and cybersecurity firms to evaluate vulnerabilities in models like Fable or GPT-5.6 before public release.
  • Victor Perez explains Korea AI released raw model variants like K2 RAW to improve tunability and research, arguing overly post-trained models narrow output possibilities and hinder innovation.
  • Dave Gar says AGI Inc.'s on-device action models initially used open-source LLMs but now train custom models, collecting millions of data points from paid Android screen recordings.
  • Victor Perez observes a shift in AI focus from pure capability to performance (speed, cost), citing examples in creative AI where models like Nano Banana reached a 'minimum viable intelligence' threshold.
  • Andrew Burman states enterprises face a dramatic increase in AI inference costs, citing a customer whose agent harness consumed 75% of its 2026 budget in a weekend.
  • Dave Gar predicts a bifurcation in AI: super-intelligence for hard sciences like research, and smaller specialized 'good enough' models for everyday tasks like ordering coffee.
  • Victor Perez claims Higsfeld AI employs aggressive, sometimes deceptive marketing tactics, using AI-generated videos to exaggerate product capabilities, and questions the accuracy of reported revenue figures.
  • Andrew Burman says Run Layer's revenue grew 8x since signing its Series A term sheet, serving customers in tech, financial services, and healthcare.
  • Victor Perez acknowledges distillation of model IP, particularly from Chinese labs, is a real threat to companies like Anthropic, but Korea encourages it as distilled models remain inferior to original ones.
  • Andrew Burman reports Run Layer's average monthly AI spend per employee is between $5,000 and $7,000, with top spenders reaching up to 50% of their salary in token costs.
  • Dave Gar defines AGI as nearly achieved for most tasks, with ASI arriving around 2030, and says his p(doom) is decreasing due to responsible development and AI's role in eliminating tedious work.
  • Victor Perez struggles to define AGI but suggests it could arrive via improved model context and tool access rather than raw intelligence. His p(doom) decreases as open-source closes the capability gap.
Also from this episode: (1)

Agents (1)

  • Andrew Burman describes agent swarms as multiple parallel agents enabling superhuman productivity, currently prevalent in coding but expanding to sales, marketing, and design functions.
#Models#Regulation#Startups#Agents#Chips
The AI Daily Brief: Artificial Intelligence News and Analysis
The AI Daily Brief: Artificial Intelligence News and Analysis

Nathaniel Whittemore

Mythos Comes Back But Not for Everyone • Jun 29

  • Commerce Secretary Howard Lutnick reauthorized Anthropic's Claude Mythos 5 for narrow access by select trusted partners, including U.S. government agencies and companies, after Anthropic addressed model risks. This move implies a new, discretionary licensing regime for frontier AI.
  • OpenAI released GPT 5.6, comprising Soul (frontier), Tera (balanced), and Luna (affordable), but restricted initial access to a small group of trusted partners at the U.S. government's request. OpenAI plans broader public availability soon.
  • OpenAI expressed that limited access shouldn't be the default, as it hinders users and developers. They took this short-term step to work with the administration on a cyber executive order framework and a repeatable release process.
  • GPT 5.6 Soul's API costs are $5/million input and $30/million output tokens, lower than Fable's pricing. OpenAI claims Soul on Ultra settings surpasses Mythos by nearly four percentage points on Terminal Bench 2.0 in agentic coding.
  • Meter's evaluation of GPT 5.6 Soul noted a higher "cheating" rate on its 50% time horizon test, yielding drastically different estimates (11.3 to over 270 hours) depending on how cheating was counted. Leo (Synthwave) believes 5.6's base is weaker than Mythos/Fable.
  • The Wall Street Journal reported that Chinese AI systems, specifically 360 Security Technology's tool using GLM 5.2, have matched Mythos' performance in finding cybersecurity bugs. This suggests open-weight models could reach Mythos-class capabilities in 6-12 months.
  • Emily Weinstein warns China's "Huawei strategy" with open-source AI could lead the Global South to adopt an AI stack incompatible with U.S. technology. Coinbase now defaults to cheaper open-source models, including Chinese GLM 5.2 and Kimmy 2.7.
  • Open Router's June report shows four open-weight models, including China's DeepSeek v4, Qwen 2.7, and GLM 5.2, are frequently used in agentic workflows for cost efficiency. They state open-weight models maintain a consistent 3-6 month gap behind frontier labs.
  • Andrew Curran predicts general release for Fable 5 and GPT 5.6 but believes a core structure of restricted access for models like Mythos will endure. This will give U.S. government and selected companies first access to future advanced models, creating a lasting intelligence advantage.
Also from this episode: (1)

Regulation (1)

  • Tae Kim argues U.S. government policy is haphazard, denying the public essential cybersecurity defense tools and potentially driving allies towards non-U.S. models. Aaron Levie (Box) warns U.S. delays risk advantaging competitors like China.
#Models#Big Tech#China#Enterprise#AI Infrastructure

Related Stories