Week 18, 2026•6 stories

The Week Google Wrote a $40 Billion Check

Google invests $40B in Anthropic. DeepSeek V4 drops on Huawei chips at 1/20 the cost. Google Cloud Next goes all-in on agents. Stanford's AI Index says junior dev jobs are down 20%. And Snap fired 1,000 people because AI writes 65% of their code.

April 27, 2026

01.Google Bets $40 Billion on Anthropic

FundingStrategy

Google is investing up to $40 billion in Anthropic. $10 billion now, in cash, at a $350 billion valuation. Another $30 billion if Anthropic hits performance targets. This is the largest single AI investment ever made.

$40B

Total Investment

$350B

Anthropic Valuation

$30B

Anthropic Revenue Run Rate

5 GW

Compute Committed

The numbers are insane

Anthropic's run-rate revenue has hit $30 billion, up from roughly $9 billion at the end of 2025. That's 3x growth in four months. The investment includes access to 5 gigawatts of Google Cloud compute starting in 2027, which is roughly the power consumption of a small country.

Why Google is paying its competitor

Google already uses Claude inside its own developer platform (announced at Cloud Next this week). It competes with Anthropic on models but depends on it for enterprise AI credibility. The $40 billion isn't just an investment. It's an insurance policy against OpenAI running away with the enterprise market, and a compute deal that locks Anthropic into Google Cloud infrastructure.

Amazon's shadow

Amazon invested $5 billion in Anthropic just weeks ago, with up to $20 billion more tied to milestones. Anthropic is now simultaneously funded by both Google and Amazon, who are direct competitors. Anthropic is playing them off each other, and it's working. The Motley Fool called Google's deal "a screaming bargain."

thinkidiot take: Anthropic is doing what Switzerland did during the Cold War, except instead of neutrality, it's selling compute contracts to both sides. At a $350B valuation with $30B revenue, they're valued at ~12x revenue, which is actually reasonable for a company growing 3x in a quarter. The real story: the AI infrastructure arms race has moved past models. It's about who can physically build enough data centers. 5 gigawatts is nuclear-power-plant territory.

02.DeepSeek V4 Drops on Huawei Chips

ChinaOpen Source

One year after upending Silicon Valley with V3, DeepSeek is back. On April 24, the Chinese lab published DeepSeek V4 on Hugging Face under an MIT license. 1.6 trillion parameters. One million token context window. And the part that should worry every Western AI lab: it runs entirely on Huawei Ascend 950PR chips.

1.6T

Total Parameters

49B

Active Parameters

Context Window

$1.74

Per M Input Tokens

The technical details

V4 uses a Mixture-of-Experts architecture with 1.6 trillion total parameters but only 49 billion active per inference. A lighter V4-Flash variant runs 284 billion total with 13 billion active. The hybrid attention mechanism (Compressed Sparse + Heavily Compressed) cuts KV cache to 10% of V3 and inference FLOPs to 27% of V3 at the million-token context length.

The pricing shock

V4-Pro API: $1.74 per million input tokens, $3.48 per million output. Claude Opus 4.7 charges $15 per million output. That's roughly 1/5th the price for competitive performance. V4-Flash drops to $0.14/$0.28. Effectively free.

The chip story

This is the first frontier-class AI model built entirely on Chinese domestic semiconductor infrastructure. Despite US export controls on NVIDIA chips, DeepSeek trained V4 on Huawei silicon. MIT Technology Review called it one of three reasons "why DeepSeek's new model matters." The export control strategy is not working as intended.

thinkidiot take: DeepSeek V4 is the sequel nobody wanted. The first time DeepSeek shocked the industry, it was dismissed as a fluke. This time they did it again, on domestic hardware, under sanctions, at a fraction of the cost. The open-weights release means every company on Earth can run a frontier model for free. That's the real disruption. Not the benchmarks. The fact that the API to frontier intelligence is trending toward zero cost.

03.Google Cloud Next Goes Full Agent

AgentsInfrastructure

Google Cloud Next 2026 in Las Vegas was entirely about agents. Not models, not training, not benchmarks. Agents. The company unveiled a full stack for building, deploying, and orchestrating AI agents across enterprises.

The big announcements

Agent2Agent Protocol v1.2. The open protocol for cross-platform agent communication, now governed by the Linux Foundation, has reached 150 organizations in production. Version 1.2 adds cryptographic signatures for domain verification. If MCP connects agents to tools, A2A governs how agents talk to each other across company boundaries.

Project Mariner. Google's web-browsing agent, powered by Gemini 2.0, scores 83.5% on the WebVoyager benchmark and can handle ten concurrent tasks on cloud VMs. It automates shopping, research, and form-filling. Available to Google AI Ultra subscribers in the US. Roadmap: Mariner Studio (Q2), cross-device sync (Q3), agent marketplace (Q4).

No-code agent builder for Workspace. Build agents without code, directly inside Google Docs, Sheets, and Gmail. Each agent gets its own dedicated inbox to post progress reports. This is the "AI teammate" pitch made tangible.

200+ model platform. The redesigned developer platform now hosts models from Google, Anthropic (Claude), Meta, and others. Managed MCP servers across Google Cloud. Enterprise customers can pick the best model per task without switching platforms.

thinkidiot take: Google is not trying to win the model race. It's trying to be the platform where all models run, where all agents operate, and where all agent-to-agent communication passes through. That's a classic Google move: don't own the content, own the infrastructure. A2A at 150 orgs in production, not pilot, is the real signal. The agent interop standard is solidifying and Google is at the center of it.

04.Stanford AI Index: Junior Dev Jobs Down 20%

ResearchWorkforce

The Stanford HAI AI Index Report 2026 dropped this month with 500+ pages of data. The headline that hit hardest: employment for software developers aged 22 to 25 has fallen nearly 20% since 2022.

-20%

Junior Dev Jobs

53%

Population Using GenAI

$172B

US Consumer Surplus

-89%

AI Researchers Moving to US

The capability story

Several models now meet or exceed human performance on PhD-level science questions, multimodal reasoning, and competition math. On SWE-bench Verified (a real-world coding benchmark), performance rose from 60% to near 100% of the human baseline in a single year. The best models score above 50% on Humanity's Last Exam, a test designed to be unsolvable by AI.

The adoption story

Generative AI reached 53% population adoption within three years, faster than the personal computer or the internet. US consumer surplus hit $172 billion annually, up from $112 billion a year ago. The median value per user tripled.

The uncomfortable story

The junior developer employment drop is the first large-scale, data-confirmed instance of AI replacing a specific professional demographic. Not automating tasks. Replacing people. Meanwhile, studies show AI boosts output by 14-15% in customer support, 26% in software development, and 50% in marketing. The productivity gains are real. So are the job losses.

The talent flight

The number of AI researchers moving to the US dropped 89% since 2017, with an 80% decline in the last year alone. At the same time, US-China model performance gap narrowed to just 2.7%.

thinkidiot take: The junior dev number is the one to watch. It's the canary in the coal mine for white-collar AI displacement. But read it carefully: the drop is concentrated in 22-25 year olds, not senior engineers. What's happening is that companies are hiring fewer juniors because AI handles the work that juniors used to do (boilerplate, tests, simple features). The path to seniority just got harder. If you're early-career, the move is to go deep on AI-native workflows, not compete with AI on the tasks it already owns.

05.Snap Fires 1,000. AI Gets Their Jobs. Stock Jumps.

LayoffsBusiness

Snap laid off 1,000 employees (16% of its workforce) on April 15, citing AI efficiencies. The stock jumped 8% the same day. CEO Evan Spiegel said "rapid advancements in artificial intelligence" allow smaller teams to achieve the same output.

1,000

Jobs Cut

65%

New Code Written by AI

$500M

Annual Savings

+8%

Stock Jump

The math Spiegel presented: AI generates 65% of new code at Snap. The company assigns more critical work to focused teams and AI agents. The cuts reduce annualized costs by $500 million and establish "a clearer path to net-income profitability."

Snap joins Meta (~8,000 cuts), Oracle, and Amazon in what is now an unambiguous pattern: companies are scaling headcount down and AI infrastructure up simultaneously. The restructuring costs ($95-130 million) are a rounding error against the $500 million annual savings.

US employees get four months of severance, healthcare, and equity vesting. That's better than most. But the message is clear.

thinkidiot take: The Snap story crystallizes what the Stanford Index showed with data. AI-generated code at 65% is past the tipping point where you need fewer humans to ship the same product. The stock jumping on layoff news tells you exactly how the market values this equation: fewer people + more AI = more profit. That's the world we're in now. Whether you think it's good or terrible, the incentives are locked in.

06.Quick Hits

Anthropic + Amazon: 5 GW of compute. Anthropic expanded its Amazon partnership to secure up to 5 gigawatts of compute for training and deploying Claude. Nearly 1 gigawatt expected online by end of 2026. That's on top of the Google deal. Anthropic is stockpiling compute like a nation-state.

Merck + Google Cloud: $1B AI partnership. Merck is making a multiyear investment worth up to $1 billion to become an "AI-enabled enterprise" with Google Cloud. Pharma is going all-in on AI for drug discovery, clinical trials, and operations.

Alibaba drops Qwen3.6-Max-Preview. Early preview of Alibaba's next flagship model with improved agentic coding, world knowledge, and instruction following. The Chinese frontier is crowding fast: DeepSeek, Qwen, Moonshot, MiniMax all shipping within weeks of each other.

EU AI Act: 98 days to compliance. The EU AI Act's general application date is August 2, 2026. Companies are scrambling to build Explainable AI modules and governance frameworks. Compliance spend is spiking. The regulatory clock is real.

40% of enterprise apps will have AI agents by year-end. Up from 5% in 2025. Microsoft and MYOB announced a five-year partnership deploying customer-facing agents for cash flow forecasting and compliance. The agent deployment curve is hockey-sticking.

Human scientists still beat AI agents on complex tasks. A Nature study found that while AI agents handle routine research tasks well, human scientists significantly outperform them on novel, complex problems requiring creative hypothesis formation. The machines are fast. The humans are still creative.