Google has a department whose only job is to steal startups: inside the copying machine (2026)

"Google has a whole department whose only job is to steal startups."

That's not speculation—it's a direct quote from Alex Socoloff, a former principal designer at Google, posted publicly on May 22, 2026. "I was a principal designer at that startup - what we did is literally copy successful startups and launch them under the Google name." He explains how they find targets: "All your free apps on your phone collect data legally - VPN, ChatGPT, Notes... you name it." Google buys this data, identifies winners, and launches clones with Google's distribution.

The same week this leaked, developer YouTuber Theo released a 32-minute breakdown titled "You do not hate Google enough"—documenting how Gemini 3.5 Flash's "fast and cheap" marketing hides $9/M output tokens (3× more than Gemini 3 Flash), how the Antigravity CLI is so buggy it can't scroll, how Google killed the beloved open-source Gemini CLI to replace it with closed-source slop, and how Google Cloud banned Railway's $2M/month account mid-operation, taking all their services offline.

This isn't one bad week. It's systemic dysfunction at a trillion-dollar company that has infrastructure, talent, TPUs, and research but can't ship good products because internal politics kill projects before they finish and good people leave before shipping.

This article connects the dots: how Google identifies startups to copy, recent examples of the copying and dysfunction, why good teams get replaced, and what founders should do.

TL;DR

Primary sources: Alex Socoloff on X · Theo's video · Artificial Analysis benchmarks

The Copying Department: How It Works

Alex Socoloff's revelation

From the viral post:

"Google has a whole department whose only job is to steal startups. I was a principal designer at that startup - what we did is literally copy successful startups and launch them under the Google name."

"How do we find a successful startup you might ask? All your free apps on your phone collect data legally - VPN, ChatGPT, Notes... you name it."

"All the cool products you build with AI are gonna be stolen one day. You're literally paying to build and test stuff that all those huge companies are gonna steal from you."

Translation:

1. Free apps collect usage data (which users agree to in ToS)
2. Data brokers aggregate and sell this data to Google
3. Google identifies products with strong engagement (retention, growth, monetization signals)
4. Internal team replicates the product with Google branding and distribution
5. Launch kills or acquires the original startup

Recent Example: Gemini 3.5 Flash—Misleading Benchmarks

Google announced Gemini 3.5 Flash at I/O 2026 with charts showing it's "frontier intelligence with action" and 300+ tokens/second speed. What they didn't show: the price and token waste.

What Google claimed

• Speed: 300 tokens/second (fastest in class)
• Performance: Beats Gemini 3.1 Pro on Terminal-Bench, SWE-Bench Pro, MCP Atlas
• Best-in-class on agentic benchmarks

What Google hid

Why this matters:

• Gemini 2.0 Flash: $0.10/M in, $0.40/M out (Theo's favorite model, now deprecated)
• Gemini 3.5 Flash: $1.50/M in, $9/M out
• That's a 20× price increase from 2.0 to 3.5

Token efficiency comparison (Artificial Analysis benchmarks):

• GPT-5.5 Medium: 22M output tokens
• Gemini 3.5 Flash: 72M output tokens (3.3× more waste)
• Result: Even if 3.5 Flash is "2× faster per token," it generates 3× more tokens, so it's actually slower and more expensive in practice

Theo's coding test:

• Task: Rewrite his game "Fish Slop" with AI
• Result: Completely broken. Fish too big, feeding mechanic doesn't work, aging doesn't work, images have no transparency
• GPT-5.5 result on same task: Worked perfectly, then added 3D mode when asked

Quote from Theo:

"It is the only model I have run this test on that didn't manage to make it work. This is embarrassing... legitimately embarrassing for an allegedly state-of-the-art model."

The Antigravity CLI Disaster

What happened to Gemini CLI

Gemini CLI was an open-source project Google launched in early 2025 (after Claude Code / Codex CLI went viral). It had:

• 100K+ GitHub stars
• 6,000 merged PRs
• Hundreds of contributors
• Active community building skills and workflows
• Weekly releases with responsive team (Dimitri, Jack, Gal)

Theo's take:

"I actually had a lot of faith in this team... They have never been shy to DM me, to correct things I'm wrong about, but also thank me for the things I was right about... They are why I didn't talk shit."

The replacement: Antigravity CLI

Google announced on May 20, 2026:

"Transitioning Gemini CLI to the Antigravity CLI... Gemini CLI proved the terminal could be an incredible interface for agentic tasks, but your needs shifted."

What actually happened:

1. Google killed Gemini CLI (forced deprecation June 18, 2026 for Pro/Ultra users)
2. Replaced it with closed-source Antigravity CLI
3. Written from scratch in Go (not TypeScript like Gemini CLI)
4. Built by Windsurf acquisition team (Google bought Windsurf founders to build this)

How buggy is Antigravity CLI?

From Theo's live testing:

Bugs encountered:

• Scroll doesn't work (puts previous inputs in the input box instead of scrolling)
• Can't exit with Ctrl+C (must use /exit command)
• Email leaks in UI (no way to hide it)
• UI elements get stuck at the bottom, can't clear them
• Re-sign in every time you open it
• Input box moves around randomly
• Asks for feedback mid-task and breaks the input flow

Quote:

"It's the buggiest CLI I've used in a while. It was really, really bad. I ran into so many broken buggy states... I actually couldn't fathom it."

The Cursor copying

In Google's official Antigravity announcement video, they show a demo of adding folders. One of the folders visible in the demo is named "Cursor." They're not even hiding that they're copying Cursor.

Context: Antigravity UI looks identical to Cursor and Codeium. The Windsurf team (now Antigravity team) has a history of copying these UIs.

Railway Gets Banned: $2M/Month Doesn't Matter

Railway is a platform-as-a-service (like Heroku) that makes deploying apps easy. They chose Google Cloud for their infrastructure layer (CDN, web layer) while moving compute to their own metal.

What happened on May 20, 2026

Google Cloud banned Railway's account entirely.

Impact:

• All web-facing Railway services went offline
• Railway spends $2M+/month with Google Cloud
• No warning, no explanation, just banned

Jake (Railway CEO) context:

• Theo interviewed at Railway in 2021 (would have been 4th employee)
• Jake encouraged Theo to start his own thing instead (leading to this channel)
• Railway has been fighting Google Cloud issues for years (rate limits Google doesn't know exist, outages, etc.)
• Railway built their own metal specifically to get off Google Cloud but still depend on it for web layer

This isn't the first time

UniSuper incident (2024):

• $135 billion Australian pension fund
• Google Cloud accidentally deleted their entire account
• Quote from Google Cloud CEO: "unprecedented sequence of events whereby an inadvertent misconfiguration... resulted in the deletion of UniSuper's private cloud subscription"
• Only survived because they had backups on a different cloud provider
• People thought it was a cyberattack due to severity

Quote from Theo:

"I have heard stories like this far too many times... Azure might be shit but they know how to make money. AWS still has their stuff together. Google Cloud is a joke."

Why Google Can't Fix This: Internal Politics

Theo names three people who built trust and were doing good work:

• Dimitri
• Jack
• Gal

What they did:

• Built Gemini CLI as open-source
• Engaged with community (DMs open, responsive to feedback)
• Shipped weekly releases
• Built good faith with creators like Theo

What happened to them:

• Roles replaced by the Antigravity team (Windsurf acquisition)
• Their work (Gemini CLI) killed in favor of closed-source rewrite
• Years of community-building thrown away

Quote from Theo:

"All three of these people have been incredible to interact with... they built a ton of good faith from me... All of their roles have been replaced by the Antigravity team."

The pattern

Google's cycle:

1. Buy a startup or hire talented people
2. Give them a project
3. Internal politics shift priorities
4. Kill the project before it's finished
5. Talented people leave
6. Launch broken replacement with worse team

Result: Google has "the infrastructure, the talent, the ecosystem, the TPUs, the research" but "no one's ever allowed to work together long enough to make anything good."

My Take: Google Is Structurally Broken

This isn't about one bad product or one bad week. It's systemic.

What I've learned from friends at Google

I know multiple people who've worked at or left Google. The pattern is consistent:

1. Projects get killed arbitrarily

• You build something for 2 years
• Leadership changes priorities
• Your project gets cancelled or handed to a different team
• You start over or leave

2. Internal politics > product quality

• Promotions require launching something big
• Maintenance doesn't count
• Result: People launch half-baked products, get promoted, then the product dies

3. Teams don't talk to each other

• Google Cloud, Google AI, and Google Workspace might as well be different companies
• Integration is painful
• Nobody takes responsibility when things break

4. Good people burn out and leave

• Talented engineers join, build something promising, then watch it get killed
• After 2-3 cycles, they leave for startups where they can actually ship

Why this matters for founders

If you're building on Google:

1. Google Cloud: Expect outages, bans, and "unprecedented" errors. Have backups on AWS/Azure.
2. Google APIs: They will deprecate with <6 months notice (see Gemini 2.0 Flash deprecation pushing users to 30× more expensive models)
3. Google partnerships: They will copy you if you succeed. Have an exit strategy.

If you're competing with Google:

1. They will copy you eventually
2. They will mislead with benchmarks (hide pricing, cherry-pick metrics)
3. They will have distribution advantages (Chrome, Android, Search)
4. But they can't maintain products long-term—internal politics will kill even good Google products

What Should Founders Do?

1. Don't build critical infrastructure on Google Cloud

Alternatives:

• AWS: Industry standard, reliable, expensive but works
• Azure: Microsoft's enterprise focus, slower but stable
• Self-hosted: Railway's approach (own metal + cloud for edge)

If you must use Google Cloud:

• Have backups on different providers (like UniSuper did)
• Expect outages
• Don't spend $2M/month—apparently that's not enough to prevent bans

2. Assume Google will copy you

Defense strategies from our previous post:

• Network effects (Slack's moat vs Google Chat)
• Proprietary data (Harvey's legal memos vs general models)
• Regulatory capture (HIPAA compliance Google doesn't have yet)
• Brand/community (Linear's passionate users won't switch to Google's clone)
• Speed (ship features faster than Google's internal politics can copy)

3. Don't trust Google's benchmarks

Red flags in Gemini 3.5 Flash announcement:

• Chart showing speed but no dollar signs
• "Fast" claim without mentioning token waste
• Comparison to old models (3.1 Pro) not competitors (GPT-5.5, Claude)
• Total cost hidden (2× more than 3.1 Pro despite being "flash")

How to evaluate honestly:

• Test on your actual use case (like Theo's Fish Slop test)
• Measure total cost (tokens used × price per token)
• Check Artificial Analysis for independent benchmarks
• Watch for "reasoning tax" (extra tokens used for thinking)

4. Support the good people, not the company

Theo's message resonated: Dimitri, Jack, and Gal built trust and did good work. When companies screw over good people, support the people.

Where this happens:

• Google (killing Gemini CLI team's work)
• Anthropic (sometimes—Theo has been critical when they don't respond to feedback)
• OpenAI (Theo says they usually fix issues when reported privately)

Who's doing it right:

• Cursor: Keeps shipping, listens to users
• Railway: Builds in public, admits mistakes
• OpenAI (when Greg Brockman/Sam respond directly): Actually fixes reported issues

The Uncomfortable Truth

Google isn't evil—they're incompetent.

They have:

• ✅ World-class AI researchers
• ✅ Billion-dollar budgets
• ✅ TPUs and infrastructure
• ✅ Chrome, Android, Search distribution
• ✅ Talented designers and engineers

They don't have:

• ❌ Ability to ship consistently
• ❌ Internal coordination
• ❌ Long-term product support
• ❌ Respect for open-source communities
• ❌ Reliable cloud infrastructure

Result: They copy startups not because they're strategic, but because they can't innovate internally. The copying department exists because internal teams can't ship fast enough.

Related on ExplainX

• You're beta testing ideas for billion-dollar companies
• Andrej Karpathy joins Anthropic: when good people leave
• Cohere Command A+: when open source beats closed
• oh-my-pi: when indie devs build better tools than big tech

Sources

• Alex Socoloff (former Google designer) on X: Public post about Google's copying department
• Theo's video: "You do not hate Google enough" (32 minutes)
• Artificial Analysis benchmarks: Token usage and cost data
• Railway incident: Public outage reports (May 20, 2026)
• UniSuper incident: Google Cloud blog post (2024)
• Gemini CLI deprecation notice: Google's official announcement (May 20, 2026)

Google's internal politics, product decisions, and cloud reliability continue to evolve (usually downward). Treat this as May 22, 2026 context. Verify current pricing, product status, and cloud SLAs before committing to Google infrastructure. The pattern of killing good projects and copying competitors has held for decades, but individual outcomes vary.

Final note: Theo is risking his YouTube career (built on Google's platform) to publish this. That's how bad it is. If you're building on Google, have a backup plan.