TikTok's $80B divorce finalized + 3 ways to revolutionize drug discovery
Keep reading to join our events, and get the latest news!
Dev Chandra
Trending Thursday #16
TikTok's $80B divorce finalized + 3 ways to revolutionize drug discovery
Dev Chandra
September 18, 2025 โข Estimated Reading Time: 12 minutes
This week, OpenAI released GPT-5-Codex and watched it dominate their platform in a single day. Meanwhile, Google quietly built the payment infrastructure for the entire agent economy while everyone else was distracted by biological models. The TikTok deal proved that geopolitics is just venture capital with a twist.
As we continue to build at Context Ventures**, we 've launched **Startup Introswith a simple mission: to help early-stage founders find the right investors, faster and smarter.
๐ฏ BLUF: Bottom Line Upfront
-
TikTok 's Final Form: ByteDance keeps <20%, Oracle/a16z/Silver Lake gets control, algorithm gets licensed, Trump claims victory
-
The Virtual Cell Wars: Arc Institute says we're at "GPT-1 stage" while Demis claims AlphaFold solved biology
-
OpenAI 's Coding Domination: GPT-5-Codex becomes the majority-used model in Codex after ONE day, proving developers will instantly adopt better tools
-
Google 's Infrastructure Genius: Launches Agent Payment Protocol with Mastercard, Amex, PayPal - becoming the toll booth for the $1T agent economy
-
YouTube Goes Full AI: Veo 3 turns everyone into Spielberg while building likeness detection to prevent deepfake disasters
๐ฌ The TikTok Resolution: Everyone Wins, Nobody's Happy
The White House: Further Extending The TikTok Enforcement Delay
After years of political theater, the TikTok deal is landing exactly where everyone expected: a messy compromise that lets both sides claim victory.
The Final Ownership Structure:
-
US Investors: 50% (Oracle, Silver Lake, a16z)
-
Existing Investors: 30%+ (Sequoia, KKR, General Atlantic, Coatue)
-
ByteDance: <20% (just under the legal threshold)
-
US Govt: Possibly getting a board seat (still unclear)
The Awkward Details:
-
Algorithm licensing: ByteDance keeps the secret sauce, licenses it to a US entity
-
Data sovereignty: Oracle hosts everything in Texas (Larry wins again)
-
The CEO question: Shou Chew out, Trump-aligned exec probably in
-
User migration: Everyone moves from TikTok to "TikTok US" (good luck with that)
The Existing Investor Chaos: Current ByteDance shareholders "still in the dark" about their stakes. Imagine owning millions in TikTok shares and finding out about your dilution from the WSJ.
Strategic Takeaway: This isn't about TikTok. It's the template for every Chinese tech company operating in the US: <20% ownership, algorithm licensing, domestic data storage. The new cost of doing business is letting the Valley own most of your US ops.
๐งฌ The Virtual Cell Wars: Biology's $100B Philosophical Battle
These past weeks crystallized the fundamental debate in AI-biology: Demis Hassabis celebrates his Nobel for AlphaFold, Patrick Hsu's Arc Institute pushes virtual cells, and Geoff von Maltzahn's Lila Sciences says everyone's doing it wrong.
๐ The DeepMind Way: "Solve One Problem at a Time"
Google DeepMind CEO Demis Hassabis on AI, Creativity, and a Golden Age of Science
Fresh from signing the Nobel book next to Einstein, Hassabis admits we're nowhere near AGI for science.
-
AlphaFold 's Success: Solved the 50-year protein folding problem with ~90% accuracy, created an open database used globally, and won the Nobel Prize in record time.
-
Isomorphic 's Next Play: Building "adjacent AlphaFolds" for each drug step, hybrid models mixing AI with physics, partnerships with Eli Lilly and Novartis.
-
Reality Check: AI still can't generate new hypotheses, make reasoning leaps, or answer its benchmark: Could AI rediscover Einstein's relativity given 1901 knowledge?
๐ญ The Lila Sciences Way: "Generate New Data or Die"
Von Maltzahn's $235M bet: stop mining old research, start generating new experiments.
-
Core Thesis: "Training on past data limits you to the past." Solution: autonomous labs running thousands of concurrent experiments.
-
The Irony: Backer Flagship Pioneering also invests in traditional AI drug discovery by betting against their own portfolio.
๐ฌ The Arc Institute Way: "Simulate Everything"
A16Z Podcast: Faster Science, Better Drugs
Patrick Hsu's vision: don't solve problems, simulate entire cells.
-
The Promise: Test millions of perturbations in silico, compress years to days, create biology's "ChatGPT moment."
-
Current Reality: At "GPT-1/2 stage" producing "blurry pictures of life." Missing protein/metabolite data, as we only have RNA, the "low-resolution mirror."
-
The Benchmark: Can it rediscover Yamanaka factors (adult cells โ stem cells)? Not even close. Running Virtual Cell Challenge with $100K prizes to accelerate progress.
๐ Why Biology Refuses to Be Disrupted
-
The Data Crisis: We measure RNA (low resolution) but lack protein/metabolite data. Can't model what we can't measure.
-
The Wet-Lab Wall: Biology needs physical validation. Cells grow on their schedule, not Moore's Law. 90% of drugs fail in human trials for unpredictable reasons.
-
The GLP-1 Reality Check: Biggest pharma win (Ozempic/Wegovy, $1T+ value) came from traditional R&D;, not AI. Decades of incremental research still beat algorithms.
๐ฎ What Happens Next
Hsu predicts new architectures beyond transformers are "overdue." Three possible outcomes:
-
One approach breaks through, triggers a discovery cascade
-
The hybrid methodology emerges by combining all three
-
Biology remains resistant; we redefine success downward
The Bottom Line: While Silicon Valley races to build AGI for science, biology doesn't care about our quarterly cycles. The virtual cell wars have begun, but biology remains stubbornly analog, infinitely complex, and patient.
๐๏ธ Quick Hits: The AI Agent & Infrastructure Gold Rush
๐ค OpenAI's Lightning Strike: GPT-5-Codex became the majority-used model in Codex after ONE day. Sam Altman's flex: "Already on pace to become the majority-used model inside of Codex after a single day." Code review catches critical bugs before shipping, works everywhere (terminal, IDE, GitHub, iOS), and is included with ChatGPT Plus ($20/month). Microsoft spent years building GitHub Copilot, and OpenAI just ate their lunch in 24 hours.
๐งโ๐ผ AI Gets Office Jobs: OpenAI and Anthropic are building "RL gyms" or fake versions of enterprise tools to train AI coworkers. They are paying domain experts $150-250/hour to demonstrate complex tasks (NASA scientists, PE VPs teaching Excel). OpenAI expects to spend $8B annually on this by 2030. The goal is that AI will use Salesforce, Gmail, and Excel alongside you, not just chatbots.
๐ณ Google's $1 Trillion Toll Booth: While everyone's building agents, Google built the payment rails they'll run on. The Agent Payment Protocol with Mastercard, Amex, and PayPal creates "mandates" for pre-authorized spending. The use cases include auto-purchase restocks (bye PS6 scalpers), price monitoring, and B2B procurement. This is right out of the classic Google playbook: give away the protocol, tax every transaction.
๐บ YouTube's Veo 3 Creator Revolution: Text-to-Shorts generation now live in the US, UK, Canada, Australia. And on the way is raw footage editing, photo animation, and style transfers. They are also launching likeness detection to prevent deepfakes (lawsuit prevention disguised as safety). Production value is no longer a moat, with ideas and personality becoming everything.
๐ Meta's Face Computer Ships: Zuck launched the Ray-Ban Display glasses ($799) with built-in screen, gesture wristband, six-hour battery. Unfortunately, the live demo failed (authentic > polished), but the consensus is that Meta beat Apple to market. The EMG wristband translates muscle movements to clicks to be the biggest interface breakthrough since touchscreens.
๐ Salesforce's Agent Reality Check: Agentforce adoption is below 5% of customers despite Benioff's hype. The technical team warned that it required heavy planning, but it was disbanded for contradicting the CEO, only to be quietly rehired nine months later when the warnings proved true. So they slashed prices and bundled it with other services to juice numbers.
๐ธ What Other Startups are Popping Off
Here's a roundup of this week's trending startup activity:
๐ค AI Agents & Automation
-
Replit Agent 3: Our most autonomous agent yet
-
Claude Memory: Bringing memory to teams at work
-
Plander.ai: Your personal AI project manager that turns goals into steps
-
Blocks: Turn plain language into smart work apps and agents
-
CodeWords: Turn ideas into automations by chatting with AI
๐ ๏ธ Developer Tools & Infrastructure
-
Aikido Security: Secure everything you build, host, and run
-
Cosine CLI: AI pair programming directly in your terminal
-
Stash MCP Server: Make AI IDEs even smarter with your team's knowledge
-
GraphBit: Developer-first, enterprise-grade LLM framework
-
tambo: An open-source AI orchestration framework for React
-
tweakcn: Design your perfect shadcn/ui theme with real-time preview
-
Google AI Edge Gallery: Gallery of on-device ML/GenAI demos to try locally
-
Endform: Playwright end-to-end tests in seconds
-
Elasticsearch DevTools for VS Code: IDE DevTools for Elasticsearch and OpenSearch
-
Your Next Store: Launch your store in minutes using AI with Stripe built in
-
Clix : Mobile push for modern builders
๐จ Creative & Content Tools
-
Mozart AI v0.5: Cursor for making music
-
Bazaar V3: Vibe-code your software demo video
-
Simply Piano for Apple Vision Pro: A whole new way to learn piano
-
Retrotype: Feels like a real typewriter
๐ข Sales & Marketing Tools
-
The New Userlist: Email automation for sophisticated product marketers
-
viral.app: tracking & analytics for TikTok, Instagram & Youtube
-
StoryChief: Paste content and optimize for AI Search automatically
-
AI SEO Score: Check the AI SEO score of your company
-
SnowSEO 2.0: Rank #1 in search & AI, while you sleep
-
Munch Studio: The AI Marketing Partner for SMBs
๐ผ Business & Productivity
-
useDaily : Turn scattered thoughts into clear actions via WhatsApp and calls
-
Magic Mic by Guidde : Think out loud, and we'll do the rest
-
Cal.ai Phone Agent: Supercharged Cal.com scheduling with AI-powered calls
-
RizzCalc: AI Financial Modeling Agent in Google Sheets
-
Capalyze: ChatGPT for datavores: scrape โ ask โ visualize
-
DeepTagger: From Documents to structured data with interactive labelling
๐ฐ Fintech & Trading
-
Fundlify: Crowdfund & pre-sell products on Shopify
-
My Financรฉ: The dashboard of your finances
๐ง Email & Communication
-
Greenmor Mail: Email, simplified
-
HeyHelp: Gmail with a brain
๐ฑ Lifestyle & Entertainment
-
Luvu: Your unhinged fitness accountability buddy
-
Rocket Journal: Your AI-powered mental health companion
-
Oboe: Learning, for the curiously minded
๐ 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. Our free weekly newsletter delivers curated insights, deals, and trends to help you navigate the startup ecosystem.**
Stay tuned for our upcoming paid daily newsletter and fundraising app to help you close your next round!
๐ญ Parting Thoughts
Infrastructure tends to beat innovation every time: Google's building payment rails while others build agents, OpenAI dominated GitHub Copilot in 24 hours, and Oracle's making $300B just hosting everyone else's AI. Meanwhile, despite Nobel prizes and billions invested, we still have zero AI-discovered drugs, as biology doesn't care about Silicon Valley speed.
The real winners aren't building the future; they're charging rent on it.
Forward to a friend and 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 weโre here to help.
**Startup Intros**Trending Thursday : Your trusted source for founder-investor insights, delivered with clarity and focus.