Weekly Download #27

China trains models in Singapore to dodge chip bans + US AI Czar has 449 reasons to deregulate

Author

Dev Chandra
December 01, 2025 • Estimated Reading Time: 18 minutes

This week in Silicon Valley, Startups & Tech:

  • The Chip Ban Vacation: Alibaba and ByteDance training frontier models in Singapore and Malaysia; turns out export controls are just a 2,000-mile detour

  • DeepSeek's Gold Medal Sweep: Chinese lab drops V3.2 claiming GPT-5 parity while its math model literally outscores Ivy League students

  • OpenAI's Engagement Crisis: Users now spend more time chatting with Gemini than ChatGPT; 80% of US startups are building on free Chinese models

  • Sacks' 449-Company Conflict: Trump's AI czar is writing deregulation rules for industries where he holds nearly 500 investments

  • New York's Algorithm Shaming Law: First US state to force retailers to admit when AI is setting your personalized price

As we continue to build at Context Ventures, we've launched Startup Intros with a simple mission: to help early-stage founders find the right investors, faster and smarter.

🔥 Context VC & Startup Intros Events Coming Up

Check it out, stay plugged in, and don't miss these hot upcoming events:

🐉 The Chip Ban Backfire: China Does More With Less

🏝️ When U.S. Chips Are Banned, Book a Southeast Asian Vacation

  • Singapore Boom: Local data center operators report surging demand from Chinese AI firms since April

  • Models in Exile: Both Qwen and Doubao are being trained abroad, and they're topping global LLM benchmarks anyway

  • 100B+ Tokens: Alibaba's November 2025 Qwen-SEA-LION-v4 was enhanced with Southeast Asian language data, now ranking #1 on regional evaluations

Translation: This move turns US export restrictions into a minor geographic inconvenience as US chip restrictions assume geography equals control. Since Trump's April 2025 embargo on H20 chips to China, Singapore and Malaysia have become the hottest destinations for Chinese AI training runs.

Reality Check: Trying to contain Chinese AI development with export controls is like building a fence around one country's data centers, which is pointless when the entire region is wired with the same chips. The blockade didn't slow China's models down; it just spread the technology further and is accidentally turbocharging Southeast Asia's own AI ecosystem

🥇 DeepSeek Drops Gold-Medal Models While Giving Away the Code

DeepSeek just released two new reasoning models claiming GPT-5 and Gemini-3.0-Pro parity and brought receipts:

  • DeepSeek-V3.2: "GPT-5 level" for daily use, available on app, web, and API

  • V3.2-Speciale: Achieved gold-medal performance in IMO, CMO, ICPC World Finals, and IOI 2025

  • 85K+ Complex iInstructions: Trained on 1,800+ environments for agent-specific tasks

The Math Capabilities: DeepSeekMath-V2 solved 5/6 IMO 2025 problems as the first open-weight AI to hit the gold-medal tier. It scored 118/120 on Putnam 2024, demolishing the best human score of 90. The secret? A self-verifiable reasoning system that peer-reviews its own work before submitting.

Reality Check: While Western labs obsess over billion-dollar compute deals, DeepSeek ships models that beat Ivy League students at advanced mathematics and gives them away for free under Apache 2.0. The "AI is a US monopoly" narrative is looking increasingly quaint.

👁️ Alibaba Builds the Full Stack: From Vision Models to Smart Glasses

Alibaba's Qwen3-VL is making OpenAI's flagship look myopic, beating both GPT-5 and Gemini 2.5 Pro on specialized vision benchmarks:

  • 85.8% vs 81.3%: Qwen3-VL's lead over GPT-5 on MathVista visual mathematics

  • 100% Accuracy: Perfect needle-in-a-haystack on 30-minute videos

  • 39 Languages: OCR text recognition, 4x predecessor coverage

Beyond the Models: The Quark S1 smart glasses launched for $537, which undercut Meta's $800 glasses by a third. Dual Micro LED displays overlay real-time Qwen-powered contextual information onto your vision. International versions are coming in 2026.

Translation: Chinese tech giants aren't content building competitive models anymore; they're racing to own the entire stack, from silicon to hardware strapped to your skull. When an e-commerce company manufactures AI eyewear at mass-market prices, they're betting they can leapfrog Western hardware entirely.

💔 ChatGPT Won The Battle, Lost The War

⏱️ Users Are Spending More Time With Gemini

OpenAI is facing its most intense competitive pressure since ChatGPT's debut three years ago: According to Similarweb, users are now chatting longer with Gemini than with either ChatGPT or Claude. In the attention economy, time is the only metric that matters, and OpenAI is losing it.

Beyond Parity: When users choose to spend more time with a competitor, they're voting with the most finite resource they have. Google's deep integration with Search, Gmail, and Docs is doing exactly what it was designed to do: keep users in the Gemini ecosystem longer.

Reality Check: ChatGPT popularized conversational AI, but popularization and dominance are different things. The first mover taught everyone how valuable this tech could be, but despite OpenAI spending three years building the market, its competitors are spending 2025 capturing it.

🎣 Claude's Hook: 60% of Customers Can't Use Just One

Anthropic says 60%+ of its business customers now use more than one Claude product, a sticky pattern the company started noticing in summer 2025 after Claude Code exploded in popularity.

Moving Beyond Chatbots: When customers start with Claude for general tasks, then add Claude Code for development, then layer in API access for custom workflows, they're building dependencies across their entire tech stack. Each additional product makes switching to a competitor exponentially harder.

Translation: While OpenAI has been focused on flashy model releases and hardware partnerships, Anthropic has been quietly building the product suite that makes leaving painful. It's the enterprise software playbook: get them on one product, expand to three, and suddenly migration costs become a board-level discussion. This multi-product stickiness is exactly what made Microsoft Office untouchable for decades.

Bottom Line: OpenAI still has the brand recognition, but Anthropic is building something more dangerous: habit formation across multiple use cases. When your developers code in Claude, your analysts research in Claude, and your APIs call Claude, switching becomes a company-wide migration project, not a subscription cancellation.

🇨🇳 American Startups Are Building on Chinese AI

US startups are ditching OpenAI's premium APIs for free Chinese open-weight models, as roughly 80% of startup pitches now incorporate Chinese open-source model foundations, which is a stunning reversal from just 18 months ago.

Why Startups Are Switching:

  • $0 vs. $1000s: Chinese models (Qwen, DeepSeek, etc.) are freely downloadable, eliminating the subscription and API costs bleeding startups dry

  • Performance Parity: Alibaba's Qwen-2.5-Max now matches or exceeds GPT-4 on key benchmarks, destroying the "you get what you pay for" argument

  • Full Customization: Open weights mean startups can fine-tune for specific use cases instead of begging OpenAI to adjust their black box

The Irony: Sam Altman spent 2023 warning about Chinese AI dominance while pricing his own customers directly into Beijing's arms. OpenAI's moat was supposed to be capability leadership so overwhelming that price didn't matter. 

Reality Check: But when free Chinese models reach "good enough" for 80% of use cases, suddenly that $200/month ChatGPT Plus subscription funds a competitor ecosystem instead of customer loyalty.

🧩 The Great Regulatory Fracture: Tech Rules Splinter Everywhere

🤝 When the Fox Guards the AI Henhouse

Trump's AI czar David Sacks is writing rules for industries where he holds 449 AI investments out of the 708 total tech investments in his disclosed portfolio. The NYT investigation reveals regulatory capture in action: the guy dismantling AI restrictions profits from their removal.

The Brutal Math:

  • $1M Sponsorships: What his All-In podcast sought at a White House AI summit

  • 10x Mandate: His goal to "cut 10 times more regulations than new ones"

Sacks received two ethics waivers and claims divestment, but filings don't specify when he sold or what remains. Senator Warren called it "explicit conflict of interest." Legal ethics expert Kathleen Clark: "graft." Sacks' response? "Nothing burger."

Axios' Dan Primack argues these policies might genuinely help the US compete with China, even as they benefit Sacks' portfolio. Federal ethics rules permit this, which is the regulatory equivalent of grading your own homework.

Reality Check: Whether Sacks is a conflicted insider or visionary patriot depends on whether deregulation works. If it backfires, he's the poster child for why the Valley shouldn't write its own rules.

⚡ Other Regulatorgy Quick Hits

🤖 USPTO Says AI Can't Be an Inventor: New guidelines clarify that generative AI is legally just a tool; even if Claude wrote 90% of your breakthrough algorithm, only humans get inventor credit. Meanwhile, South Africa granted AI inventor status in 2021, and the EU's still debating.

🇪🇺 EU Accuses US of "Blackmail" on Tech Rules: Antitrust chief Teresa Ribera says Commerce Secretary Lutnick offered to ease steel tariffs if Brussels guts the Digital Markets Act and Digital Services Act. Her response: "Our rules are not a bargaining chip."

🇫🇷 Macron Blames Washington for Stalled Big Tech Probes: The French President says EU officials "are afraid" to enforce the DSA because of American pressure. The evidence: five open investigations into X, Meta, AliExpress, Temu, and TikTok; zero fines levied in two years.

🏷️ New York Forces Algorithmic Pricing Disclosure: Starting December 1st, retailers using AI to personalize prices must warn customers: "This Price Was Set By An Algorithm Using Your Personal Data." Ten-plus states are drafting copycat bills.

⚡ Startup Quick Hits

🎨 Black Forest Labs Triples to $3.25B on Open Source Bet: The German AI lab behind FLUX raised $300M Series B led by Salesforce Ventures. The founders invented Stable Diffusion, watched Stability AI implode, then rebuilt it better. Adobe, Canva, Meta, and Microsoft now build on their models.

🚁 Quantum Systems Hits €3B as War Drones Become Hot: Peter Thiel-backed German drone maker raised €180M led by Balderton, tripling valuation since May. Ukraine proved surveillance drones are battlefield necessities. They're projecting €500M revenue by 2026 and launching counter-drone systems next year.

🤖 Apptronik's $350M to Prove Humanoids Aren't Hype: Austin robotics startup raised a massive Series A from B Capital, Capital Factory, and Google after previously raising just $28M total. Apollo is already at Mercedes-Benz. Target price: under $50K. Currently: not there yet.

💰 Harmonic Hits $1.45B Promising "Mathematical Superintelligence": Robinhood CEO Vlad Tenev's side project raised $120M Series C from Ribbit, Sequoia, and Kleiner Perkins. Their Aristotle model won gold at the Math Olympiad using formal verification to eliminate hallucinations. Three rounds in 14 months.

🖥️ SF Compute Raises $40M as "Stock Exchange for GPUs": Raised Series A at $300M from DCVC and Wing to solve AI's dirty secret: startups locked into GPU contracts they might not survive. They let companies trade compute like commodities, taking 10% per transaction.

💸 Range Raises $60M to Replace Financial Advisors: Wealth management startup raised Series C from Scale Venture Partners for AI-powered flat-fee advice ($6K/year). Their chatbot already cut human advisor workload in half. Targeting the 99% who've never worked with an advisor.

💰 Investor Quick Hits

🇪🇺 6 Degrees Capital Closes €154M Fund III: The London/Antwerp VC closed its third fund since launching in 2023, targeting seed and Series A in enterprise software, AI, and fintech. Writing €1-5M checks with €15M follow-on reserves.

🇮🇱 Entrée Capital Raises $300M, Now at $1.5B AUM: The Israeli VC behind monday.com, Rapyd, and Coupang raised new funds for pre-seed through Series A. They've returned $3B+ to LPs with 43+ exits.

🇨🇳 Monolith Management's $289M Bet on China AI: The Chinese VC (run by ex-Sequoia China and Boyu partners) closed Fund II above its $265M target to back AI startups including DeepSeek rival MoonShot AI.

💳 Foxe Capital Raising $50M Fintech Debut Fund: Ruth Foxe Blader (ex-Anthemis, exits include Lemonade) is raising her first independent fund. DC-based, global focus on fintech and vertical SaaS. Notably, she's also chair of Unlock VC, the women-in-VC org.

🌱 Emerging Manager Watch: Castle Fund (Daniel Gould, ex-Amazon/Roku) targeting $20M debut; SpringTime Ventures of Austin raising $20M for Fund III.

💸 IPO & M&A Quick Hits

⚡ Nvidia's $2B Design Tool Investment: Nvidia bought $2B of Synopsys stock and formed a partnership to integrate CUDA libraries into EDA tools, aiming to slash chip verification times from weeks to days.

🔐 ServiceNow's $1B Security Blanket: In advanced talks to acquire identity and access management startup Veza for over $1B, potentially closing as early as next week; just months after Veza raised $108M in Series D funding.

🎯 Quest Global's Pre-IPO Warmup: Temasek is buying a 5% stake at a $4.6B valuation ahead of the engineering services firm's India IPO, up from $1.8B when Carlyle invested in 2023. Nothing says IPO-ready like a 2.5x markup and a move to Singapore.

🤖 China's Robot Bubble Warning: China's top economic planning agency warned that 150+ manufacturers flooding the humanoid robotics market are creating oversaturation with nearly identical products and declining quality.

🔄 OpenAI's Reverse Uno: Taking a stake in Thrive Capital's Thrive Holdings and embedding AI agents in its accounting and IT portfolio companies. Your investor's portfolio is now your product testbed.

💸 Other Bay Area Funding News

Here's a roundup of notable recent funding rounds across the Bay Area:

🤖 Artificial Intelligence

  • Harmonic: Raised $120M Series C led by Ribbit Capital for AI mathematical reasoning engine

  • vijil: Raised $17M Series A led by Brightmind Partners for trustworthy AI agent development tools

  • Momentic: Raised $15M Series A led by Standard Capital for AI-powered web application testing automation

  • Jeeva AI: Raised $9M Seed led by JLL Spark for AI-powered contact and lead finder

  • Onton: Raised $7.5M Seed led by Footwork for AI e-commerce search using knowledge graphs

  • Mira: Raised $6.6M Seed led by General Catalyst for audio-first smart glasses with memory

  • Palo AI: Raised $3.8M Seed for AI assistant platform for social media creators

  • Interface: Raised $3.5M Seed led by Defy.vc for AI industrial safety in energy sector

💻 Infrastructure & Hardware

🧬 Biotech & Healthcare

  • Akura Medical: Raised $53M Series C led by Qatar Investment Authority for next-gen stroke thrombectomy devices

  • Cordance Medical: Raised $8M Seed led by Sonder Capital for CNS pathology therapies

  • Intellihealth: Raised $5.2M Venture Round for obesity-related disease diagnostic software

⚡ Energy & Climate

🎮 Other

  • Fifth Door: Raised $20M Seed led by Daft Capital for instant online game deployment platform

  • Thermopylae Aerospace: Raised $1.6M Pre-Seed led by Naval Ravikant for rapid-response defense systems

🌟 Editor's Note

At Startup Intros, our mission is to bring the latest founder-investor news straight to your inbox, keeping you ahead in the fast-paced world of Silicon Valley.

💭 Parting Thoughts

So there you have it: a week where Michael Burry and Peter Thiel are shorting AI's golden children while the industry burns $371B annually and pitches lunar data centers.

Remember: when your bull case requires ignoring investors, breaking climate pledges, and launching GPUs into orbit, you're not disrupting the future, you're in the final act of a very expensive bubble.

Forward to a friend or hit reply to let me know what you're seeing in your world.

Till next time!

Dev Chandra
CEO @ Startup Intros
Associate @ Context VC
LinkedIn: /in/devchandra

P.S. Raised this week, and we missed you? Want to be featured? Have tips or funding questions? Reply or DM us as we’re here to help.

Startup Intros Weekly Download: Your trusted source for founder-investor insights, delivered with clarity and focus.