Weekly Download #30

Hi !

This week in Silicon Valley, Startups & Tech:

• Nvidia's Strategic Retreat: Folds DGX Cloud into engineering, pivoting from AWS competitor to internal AI R&D platform while shipping H200s to China

• China Cracks the Code: Reverse-engineered ASML's EUV machines with ex-employees, while Tencent routes $1.2B through Tokyo to access 15K Blackwells

• TikTok's $45B Handoff: Deal closes January 22; Oracle, Silver Lake, and MGX take 45% while ByteDance keeps 20% and its algorithm

• Databricks Hits $134B: Raised $4B Series L as AI and data warehousing each cross $1B ARR; revenue growing 55%+ YoY

• Alphabet's Compute Panic Buy: Pays $4.75B for Intersect after internal council forms to ration chips among Kurian, Hassabis, and other execs

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

• Jan 20: AI Startup Pitch Night: Apps to pitch are open! Join us in SF for an exciting pitch night featuring the hottest AI startups. Sign to pitch here!

• May 27: Early Stage Summit: Due to various logistical factors, we had to reschedule the event but plan to make it better than ever. Register on Luma!

🎁 Megadrop: $50K in Founder AI Tools

I’m excited to partner with Notion to give you $50K in AI Founder tools to help startup founders go from idea to MVP to raising funds with confidence!

What you get:

• $50k in PostHog credit to track real user behavior

• $60 in v0 by Vercel credit to vibe code your front-end

• $50 in Supabase credit to spin up your backend

• Up to 6 months of Notion Biz + AI to run your startup

But wait, there's more:

• A Vetted List of 2,000+ early-stage VCs & accelerators

• 50 Pitch Decks from YC, a16z & Sequoia startups

• Guides on pitching and raising from the best founders!

• Comment "Founder" on my LinkedIn post, connect with me & I'll DM you the tools!

Why Notion? They just crossed $600M in ARR (half from AI products) and are preparing for an IPO at an $11B valuation. They know what founders need to scale.

Repost & Share it & I'll jump on a 15-min call with you (for the first 40 reposts)!

🎯 Nvidia's Strategic Pivot: From Cloud Competitor to Internal Powerhouse

💨 DGX Cloud Folds Into Engineering as Nvidia Abandons AWS Fight

Nvidia is merging its DGX Cloud group into its core engineering organization, effectively abandoning its push to compete with AWS and other hyperscalers as a stand-alone enterprise cloud provider. The platform will now focus on powering Nvidia's own AI model and chip R&D.

The Numbers:

• The Pivot: From selling cloud to enterprises → internal AI development platform

• The Signal: Even Nvidia can't fight the hyperscaler wars while running an AI chip monopoly

• The Strategy: Focus the compute on making better chips rather than reselling them

Translation: When the company that makes 90% of AI training chips decides it can't compete in cloud, it tells you everything about hyperscaler moat depth. Nvidia's betting its future on silicon, not services.

🇨🇳 H200 Shipments Resume to China While Workarounds Multiply

Nvidia plans to resume H200 exports to China before mid-February 2026, with initial shipments expected to range from 40K to 80K units. The H200 is Nvidia's second-most powerful AI chip, and the move comes after a pause under earlier export controls.

The Loophole Economy: Tokyo-based Datasection signed a $1.2B+ contract giving Tencent access to 15,000 Blackwell chips through Japanese cloud data centers. Chinese companies can't buy the chips directly, so they're renting them through offshore "compute proxies."

The Policy Contradiction: A Council on Foreign Relations analysis shows the H200 is more than six times more powerful than any previously approved chip for China, and better than anything Huawei plans to make for at least two years. The rationale that Huawei is a "viable competitor," justifying looser controls, doesn't hold up to the data.

Reality Check: US export policy is simultaneously blocking direct sales while approving workarounds that achieve the same result at a higher cost. China's getting the chips either way; the only question is whether American companies or Japanese intermediaries capture the margin.

🔬 China's Semiconductor Manhattan Project: Reverse-Engineering the Future

⚛️ China Builds Working EUV Prototype with Ex-ASML Engineers

Chinese scientists in Shenzhen have developed the nation's first Extreme Ultraviolet Lithography machine, a highly complex piece of equipment required to produce state-of-the-art AI chips. The massive machine, occupying an entire factory floor, was reportedly reverse-engineered using ASML's technology with assistance from former ASML employees.

The Status:

• Operational: The prototype works but isn't yet producing chips at scale

• The Gap: Technical challenges remain before commercial production

• The Timeline: Years behind ASML, but accelerating

The DUV Retrofit: China is also retrofitting older ASML DUV machines to produce advanced smartphone and AI chips, exposing cracks in US-led export controls. When you can't buy the latest equipment, you squeeze more out of what you have.

Translation: Export controls were designed to keep China a decade behind in semiconductor manufacturing. China's response has been to throw unlimited resources at the problem, recruit talent from Western companies, and treat the gap as an engineering challenge rather than an insurmountable barrier. The controls are working, just not as well as advertised.

🔍 Google's Compute Crunch: When You Can't Get Enough Chips

🏛️ Executive Council Forms to Ration AI Compute

Google has formed a small internal executive council to ration and allocate AI computing capacity because demand for chips and data center resources is outpacing supply, even after massive capex increases. Thomas Kurian (Cloud CEO) and Demis Hassabis (DeepMind CEO) are among the key members.

The Tension: Cloud revenue demands GPUs for paying customers. Core products need compute for Search and YouTube AI features. DeepMind needs capacity for frontier research. Everyone's fighting over the same silicon.

🏗️ Alphabet Buys Intersect for $4.75B to Build Faster

Alphabet agreed to acquire data center company Intersect for $4.75B in cash, plus existing debt. Intersect pairs data centers with on-site power generation, exactly what Google needs to accelerate its AI infrastructure buildout.

The Context: Google had previously partnered with Intersect in a multi-party deal announced in late 2024. When a partnership isn't fast enough, you buy the whole company.

🧮 The Innovator's Dilemma Hits Mountain View

Analysis from SiliconANGLE argues that Google faces the ultimate innovator's dilemma because its search business is tightly linked to advertising revenue. Moving to a chatbot-like experience increases cost to serve by 100x. The alternative, shifting toward integrated shopping, requires a trust relationship with users that Google doesn't currently possess.

The Bright Spot: Josh Woodward, head of Google Labs, is credited with driving Gemini app growth from 350M to 650M MAUs in 2025. Consumer AI isn't hopeless; it's just expensive.

Reality Check: The company that invented the Transformer architecture is now rationing compute while buying data centers at premium prices. Google's not losing the AI race, but it's definitely not winning it either.

📱 TikTok's $45B American Makeover Finally Closes

TikTok's long-awaited US sale will formally close on January 22, 2026. An investor group including Oracle, Silver Lake, and MGX will take a 45% stake in the new entity.

The Ownership Structure:

• 45%: Oracle, Silver Lake, MGX (new investor consortium)

• 30%: ByteDance's existing investors

• 20%: ByteDance itself

• 5%: Still-unnamed investors

The Fine Print: According to an internal memo, TikTok's global business will still handle ecommerce, advertising, and marketing. The algorithm gets retrained on US user data. Oracle oversees data protection.

Translation: ByteDance keeps 20% direct ownership plus influence through its existing investors' 30% stake. The "divestiture" looks more like a financial restructuring than a clean break. The algorithm that Washington claimed was a national security threat will be retrained, not replaced.

🤖 AI Models & Research: The Year in Review

📊 Karpathy's 2025 Verdict: RLVR Is the New Scaling

Andrej Karpathy's 2025 "LLM Year in Review" frames this year as a transition from scaling-only progress to optimization-driven progress, with RLVR (Reinforcement Learning from Verifiable Rewards) as the central mechanism.

The Shift: Instead of learning from subjective human-preference labels, models are trained in environments with automatically verifiable rewards, math puzzles, code, and logic tasks, where outputs are either correct or incorrect. This enables much longer RL runs than typical RLHF, as the reward signal is cheap, automated, and consistent.

The Jagged Intelligence Problem: Karpathy describes current systems as "summoned ghosts" rather than growing animals. Intelligence under RLVR becomes extremely spiked: brilliant on tasks resembling training environments, unreliable everywhere else. Benchmarks become optimization targets rather than intelligence measures.

The GUI Future: Karpathy singles out Google's "Nano Banana" as a paradigm-shifting example, not for image quality, but for hinting at an LLM-native GUI layer where text, image generation, and world knowledge share the same weights. Today's chat interfaces are the command lines of AI; tomorrow's will be interactive canvases.

🔍 OpenAI's Monitorability Framework

OpenAI released a framework for evaluating chain-of-thought monitorability, assessing whether a model's verbalized reasoning reliably reflects its internal decision-making, particularly for safety-relevant behaviors such as deception or reward hacking. The 13-evaluation suite across 24 environments aims to make CoT a scalable safety control mechanism.

⚡ Startup Quick Hits: When $4B Is Just Another Tuesday

📊 Databricks' $4B at $134B Makes Data the New Oil: Closed Series L from Insight Partners, Fidelity, and J.P. Morgan Asset Management. Revenue run rate hit $4.8B, growing 55%+ YoY. Both data warehousing and AI product lines have each surpassed $1B ARR.

🏦 Palmer Luckey's Erebor Raises $350M at $4.35B for Crypto-Friendly Banking: The Anduril founder's neo-bank secured FDIC approval last week. Lux Capital, 8VC, and Founders Fund are backing the bet that debanking during the Biden years created demand for tech-friendly alternatives.

💻 Lovable's $330M at $6.6B Proves "Vibe Coding" Is a Category: The Swedish AI coding startup raised Series B led by CapitalG and Menlo Ventures, tripling its July valuation of $1.8B. One of the largest AI funding rounds in Europe this year.

⚛️ Radiant's $300M+ to Mass-Produce Nuclear Microreactors: Series D led by Draper Associates and Boost VC, just six months after closing Series C. Breaking ground on R-50 factory in Oak Ridge, Tennessee early next year.

🧠 Neurable's $35M for Brain-Computer Interfaces You Can Buy Today: While Neuralink grabs headlines, Neurable has BCI hardware consumers can actually purchase. New capital earmarked for "new devices and new industries."

🤖 Manus Hits $100M+ ARR in Eight Months: The AI agent startup crossed the milestone faster than arguably any software company in history. Total revenue run rate now over $125M, growing 20%+ MoM since the release of Manus 1.5.

💰 Investor Quick Hits: Follow the Money to Public Safety

🚔 Sequoia, a16z Poured $1.2B+ Into Police Tech: The VC giants backed police and public safety tech startups heavily over the past year, following Palantir's success. When the government is your customer, regulatory risk is your business model.

🐯 Tiger Global's $12.7B 2021 Fund Now Up 16%: After fueling the COVID-era unicorn bubble and getting wrecked in 2022, the crossover fund that wrote down billions has clawed back to positive territory. Survivors get to raise again.

🚀 Lightspeed's $9B War Chest Sets Firm Record: The 25-year-old VC raised its largest haul ever across six funds, including a $3.3B opportunity fund for doubling down on winners. The firm has backed 165 AI companies, including Anthropic, xAI, Databricks, and Mistral. 

🤠 S3 Ventures' $250M Fund VIII Marks 20 Years of Patient Texas Capital: Austin's longest-serving VC crossed $1B AUM with its eighth fund, backed by its only LP, a multibillion-dollar philanthropic family foundation.

💸 M&A Quick Hits: Platform Plays & Consolidation

🔮 Coinbase Acquires The Clearing Company: After integrating Kalshi-powered prediction markets, Coinbase snaps up the infrastructure startup that raised just $15M earlier this year from USV. Building its own prediction market requires owning the plumbing.

📚 Coursera + Udemy Merger Valued at Just $2.5B: Two edtech companies that lost most of their value since ChatGPT's release are combining. The new entity's pitch: "skills for the AI era." Translation: the market that AI disrupted is now selling AI training.

🤖 Salesforce Acquires Qualified for Agentic AI Marketing: Qualified's AI worker transforms websites into conversational lead-gen machines. Bringing it into Salesforce lets customers "quickly deploy fully-featured marketing agents that autonomously generate pipeline."

🌟 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.

🎁 $50K Founder Tools Giveaway

Partnering with Notion (just crossed $600M ARR, prepping for IPO) to give you: PostHog, Vercel & Supabase credits, 6 months of Notion Biz + AI, 2,000+ VC contacts, and 50 pitch decks from YC/a16z startups.

Comment "Founder" on my LinkedIn post to grab it.

💭 Parting Thoughts

So there you have it: a week where Nvidia retreated from cloud wars to focus on what it does best, China reverse-engineered the most complex manufacturing equipment on Earth with help from the company that built it, and Google formed an executive council to decide which of its own divisions gets to use its own chips.

The theme of 2025 might be this: every company is either building AI infrastructure, buying AI infrastructure, or routing around export controls to access AI infrastructure. The rest is just commentary.

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

Till next time!

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.