Ninety percent of companies surveyed say artificial intelligence has not affected their productivity, according to the National Bureau of Economic Research. Yet venture capital funds invested $242 billion in AI companies in a single quarter — 80% of all global venture investment. That represents a four-fold increase from the previous year.
Both figures are accurate and both describe April 2026. Together they define the moment we occupy: technology racing ahead faster than markets, government or society can follow. In the Bay Area, home to the headquarters of companies creating this technology, the gap between those two numbers is tangible. On Market Street, one building hires AI engineers for $400,000 while the neighboring office cuts entire departments those engineers have rendered obsolete.
Stanford University released its ninth annual AI Index on April 13 — a 423-page report tracking the industry’s coordinates each year. This year those coordinates shifted so dramatically the document reads less like an academic report than a battlefield dispatch.
On Humanity’s Last Exam, a benchmark composed of experts’ most challenging questions in their professional fields, the best model scored 8.8% a year ago while current leaders have crossed the 50% threshold. On SWE-bench Verified, which measures AI’s ability to write functioning code, results jumped from 60% to nearly 100% of human performance in one year. Generative AI reached 53% of the global population in three years — faster than personal computers, faster than the internet. Eighty-eight percent of tech sector organizations already use AI in at least one key function, while four of five university students work with generative models. These are measurements, not projections.
One of 2026’s key trends involves secrecy rather than power. Google, Anthropic and OpenAI stopped disclosing parameters of top models, training data sizes and training duration. The Foundation Model Transparency Index, measuring developer openness, collapsed from 58 to 40 points in one year. More than 90% of significant models come from private companies, while 80 of 95 most notable 2025 models launched without source code. The pattern is simple: the more powerful the model, the less known about it.
Anthropic restricted access to its Mythos Preview model to a narrow group of organizations. According to the company, the model proved capable of finding and exploiting tens of thousands of software vulnerabilities, reproducing them in more than 80% of cases. Cybersecurity experts warned that competitors will develop similar capabilities within months. Models their creators feared to release represent a new genre of AI news that simply did not exist a year ago.
The same Stanford report shows the quality gap between American and Chinese models has narrowed to 2.7 percentage points. In early 2023, OpenAI’s ChatGPT pulled ahead of all competitors. By 2024, Google and Anthropic closed the distance. In February 2025, China’s DeepSeek-R1 briefly matched the best American model. As of March 2026, Anthropic leads, followed by xAI, Google and OpenAI, with Chinese companies DeepSeek and Alibaba trailing minimally.
The competition runs on different tracks. The United States releases more models, attracts more funding and operates 5,400 data centers — ten times more than any other country. China leads in patents, scientific publications and industrial robotics. South Korea files more AI patents per capita than anyone, while 44 nations now operate government-funded supercomputing clusters.
But one detail sounds quietly in the Stanford report while carrying heavy weight: the number of AI researchers moving to the United States fell 89% since 2017, with an 80% drop in the past year alone. The country that built its AI industry largely on imported talent is losing those talents.
As capabilities grow, so does the bill.
Microsoft alone spent $34.9 billion on infrastructure in one quarter, while Goldman Sachs estimates the four largest hyperscalers — Amazon, Google, Microsoft and Meta — will spend around $527 billion in 2026. More than 80 major data centers are under construction simultaneously worldwide, with McKinsey estimating cumulative data infrastructure investment at $7 trillion by 2030.
Training one large model in 2025 generated carbon emissions equivalent to 16,000 flights between San Francisco and New York, according to Stanford AI Index data. Operating GPT-4o alone consumes annually as much water as needed for drinking supplies in Los Angeles and San Francisco combined, researchers cited in the report estimate. In Northern Virginia, farmers discovered their land surrounded by six-story computing buildings that hum around the clock, while several U.S. counties have suspended construction permits because power grids cannot handle the load.
Despite this, OpenAI earns $20 billion annually and plans to spend $1.4 trillion on data centers over eight years, though only 3% of users pay for AI tools. Whether the artificial intelligence market represents a bubble or the beginning of a supercycle remains unanswered even among those profiting from the debate. JPMorgan tested the market against five classic bubble criteria and found no matches. Nobel laureate Daron Acemoglu of MIT says models are overvalued, investor Ray Dalio sees parallels with dot-com, while Fed Chair Jerome Powell argues AI companies, unlike dot-com firms, generate real revenue. Microsoft has already paused some data center construction projects, signaling that even major players are starting to count costs.
In the first quarter of 2026, AI’s impact on the job market stopped being a discussion topic and became statistics: the tech sector cut 78,000 to 80,000 positions, three-quarters of them in the United States. According to Nikkei Asia, companies directly attributed 47.9% of cuts to AI and automation. Jack Dorsey’s Block reduced staff from 10,000 to fewer than 6,000, Atlassian eliminated 1,600 positions, Oracle announced cuts of 20,000 to 30,000 jobs.
But here begins territory where numbers say one thing and reality says another. Marc Andreessen called the layoff wave AI-washing, arguing companies that inflated staff during the pandemic are cutting for economic reasons while blaming progress. According to Forrester, 55% of employers already regret AI-driven cuts. Klarna replaced 700 workers with a model, but quality collapsed, customers left, and the company began rehiring.
Real AI impact on employment in 2025 was minimal, according to a CFO survey by Duke University and Federal Reserve Banks of Atlanta and Richmond, while 2026 projections suggest roughly 500,000 job losses — not through layoffs but by leaving positions unfilled. Openings appear and simply remain vacant.
The impact hit entirely different targets than expected. Five years ago, AI seemed likely to replace drivers and janitors first, but the opposite occurred. Stanford data shows employment among 22-25 year old developers fell nearly 20% since 2022, while entry-level programming positions dropped from 43% to 28% of job postings. Those AI was supposed to replace last were hit first. Meanwhile, CNN reported in early April, citing Citadel Securities data, that developer openings on Indeed increased 11% year-over-year, IBM tripled entry-level hiring, and Goldman Sachs documents 12-18% annual salary growth for senior developers with AI skills. Companies fire and hire simultaneously, but hire different people — not those who write code, but those who can direct AI agents that write code.
On GitHub in China, a tool called colleague.skill gained 10,000 stars in two weeks by creating digital copies of departed employees from their messages and documents, followed by anti-distill, which strips files before handover to management. The digital arms race between workers and employers unfolds in real time with open source code.
AI has stopped answering questions and started taking action. Visa launched a platform allowing AI agents to search products, compare prices and complete purchases on users’ behalf. Microsoft updated Copilot for multiple models collaborating on single tasks, Salesforce transformed Slackbot into an autonomous work assistant, while China’s Z.ai released an open model that developers claim can work on one task for up to eight hours straight. Analysts project agentic AI will comprise 10-15% of all corporate IT spending by year’s end.
Jensen Huang, Nvidia’s CEO, declared on Lex Fridman’s podcast March 23: “I think we’ve achieved AGI,” though he immediately clarified he meant AI’s ability to create a billion-dollar product, albeit briefly, rather than a Nvidia-scale company. Dario Amodei, Anthropic’s CEO, gave a specific timeline at a World Economic Forum panel: AI will replace programmers completely, end-to-end, within 6-12 months. Sam Altman told Axios that superintelligence is “this close” and represents not new technology but societal restructuring.
MIT, Stanford and Google DeepMind researchers immediately countered that models still hallucinate, struggle with novel reasoning types and lack human-level understanding — all true. But when CEOs of the world’s three largest AI companies speak aloud within three weeks words previously heard only in narrow circles, that signals more than marketing.
OpenAI published a 13-page document April 6 titled “Industrial Policy for the Intelligence Age,” arguing existing economic systems cannot survive the transition to superintelligence without restructuring. The document’s logic is straightforward: government relies on payroll taxes — if AI eliminates payrolls, government loses its base. Social benefits tie to employers — if employers replace people with models, benefits disappear with positions.
Proposals include shifting taxes from labor to capital, eliminating income tax for those earning under $100,000, creating sovereign wealth funds paying dividends to all citizens, four-day work weeks and recognizing AI access as a basic right alongside electricity.
Critics immediately noted contradictions: the same OpenAI lobbied to weaken the EU AI Act, fought California’s SB1047 and shifted from nonprofit to for-profit status — now proposing government wealth redistribution. Cambridge researcher Eryk Salvaggio called the document advertising disguised as policy, though the content merits attention if only because no one else has written similar proposals.
Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced The AI Data Center Moratorium Act of 2026 — halting new data center construction until federal AI regulation passes, while legislators in at least 11 states proposed similar pauses. Stanford data shows Americans trust their government on AI regulation at 31% — the worst score among surveyed countries.
Stanford summarizes with one phrase: “AI is running while everyone else tries to find their shoes.”
For the Bay Area, $242 billion in quarterly venture investment represents more than a report line item. It means Mountain View housing prices, Castro Street coffee shop lines filled with people who received Anthropic offers yesterday, and empty floors in buildings where support departments sat in January.
The opening presented two numbers: 90% of companies say AI changed nothing; $242 billion quarterly investment says everything changed. Both are true. The gap between them defines where we stand now.