If your mental picture of AI is still "ChatGPT that sometimes hallucinates" — you are out of date by about two product generations. Here's the current reality, in plain English, with the receipts.
The most important thing that happened in early 2026 was not a new chatbot. It was that AI systems became good enough at coding, fine-tuning, and optimization to materially speed up the production of the next generation of AI systems.
This is the threshold that researchers have warned about for a decade. And we've crossed it without most people noticing.
Most striking: AI systems can now optimize their own training code by 52x. A human researcher takes 4-8 hours to achieve 4x. Frontier AI is now ~13x past the human baseline at making AI faster — and accelerating.
The three biggest AI labs in the world have all stated this is their explicit goal:
Plus at least two new neolabs — Recursive Superintelligence ($500M raised) and Mirendil — exist solely to automate AI research itself.
That's the working estimate of a senior researcher at a leading AI safety lab, based entirely on public benchmark data — not insider claims. 30% by end of 2027.
If you've been waiting for "the singularity moment" before taking this seriously, you may be waiting from the wrong side of it.
Between April 27 and May 7, 2026, the AI industry produced more strategic restructuring than the previous five years combined. The shared driver: compute pressure, capex pressure, and a race to define the new industry standard.
On May 6, Anthropic locked up all 220,000+ GPUs at SpaceX's Colossus 1 data center in Memphis (300 megawatts). Just months earlier, Musk publicly accused Anthropic of "massive-scale training data theft."
The next day (May 7), Musk dissolved xAI as a separate company and folded everything — Grok, Colossus, the team — into SpaceX. Anthropic now buys compute from its biggest critic.
On April 27, Microsoft and OpenAI restructured their famously exclusive partnership. Microsoft lost AI exclusivity, the AGI clause, and Azure-only revenue. Microsoft kept a 20% revenue cut through 2030 and a $250 BILLION Azure spending commitment from OpenAI through 2032.
One day later (April 28), OpenAI's models went live on Amazon's AWS Bedrock. The most exclusive partnership in AI is now a multi-cloud handshake.
On May 4, OpenAI closed a $10 billion joint venture called "The Deployment Company" with TPG, Bain, Brookfield, Advent, SoftBank, Dragoneer, and 13 other PE firms. The catch: OpenAI promised the PE backers a 17.5% guaranteed annual return over five years.
The play: embed OpenAI engineers directly inside PE-owned portfolio companies. Top-down AI mandates that bypass the corporate immune system.
Also May 4: Anthropic launched a $1.5 billion AI services firm with Blackstone, Goldman Sachs, Hellman & Friedman, Apollo, GIC, Sequoia, General Atlantic, and Leonard Green. Target: directly compete with McKinsey, Bain, BCG for corporate AI transformation revenue.
An AI lab is now a consulting firm, financed by Wall Street, eating the consulting industry.
On May 1, the Pentagon awarded frontier-AI contracts to eight companies: Amazon, Google, Microsoft, OpenAI, SpaceX, NVIDIA, Reflection, Oracle. Anthropic was excluded — Trump's administration severed ties because Anthropic refused autonomous-weapons and mass-surveillance terms.
The Pentagon labeled Anthropic a "supply chain risk" (a label normally reserved for foreign adversaries). Anthropic sued. A federal judge in California blocked the exclusion. Same week Anthropic locked up the SpaceX compute deal and grew revenue past OpenAI.
The job-impact story is usually told in vague aggregates: "some jobs will disappear, new ones will appear." A May 2026 study from the National Partnership for Women & Families breaks the aggregate apart — and the picture is sharply asymmetric.
The mechanism is mundane and brutal: women are over-represented in clerical, administrative, and customer-support work — the exact tasks frontier AI handles well today. These roles also tend to have low pay, low manager visibility, and low employer investment in transition support.
Workers in vulnerable roles also report distrust of the AI systems being rolled out because they were "largely left out of creating" them. The systems feel "foreign, awkward, or even hostile."
Anthropic's own Economic Index (March 2026) measured real-world Claude usage by occupation. Hiring rates for under-25 workers in AI-exposed occupations dropped 14% compared to pre-ChatGPT baseline. The entry-level white-collar squeeze is no longer a forecast. It's a measurement.
The implication is uncomfortable: "AI will create new jobs" is probably true on aggregate over 10-20 years. "AI will create new jobs for the specific people whose old jobs disappear" is much harder to defend on the data we have.
If frontier AI is "all working," you would expect the lab in the lead to charge straight at an IPO. Instead, something different is happening.
OpenAI's CFO, Sarah Friar, has privately argued for delaying the IPO from 2026 to 2027. Her reasoning, per multiple reports:
CEO Sam Altman is still publicly pushing 2026. Wall Street banks are reportedly telling both OpenAI and Anthropic that "the first to market defines the industry" — creating pressure to rush regardless.
When the CFO — the person whose actual job is making the IPO go — is the one applying the brakes, that's a tell. Either the financial picture doesn't survive contact with public-market scrutiny, or somebody is racing to exit at a valuation the cash flows don't yet support.
Neither of those readings is bullish for retail investors buying lab equity at IPO.
While OpenAI was renegotiating its plumbing and Anthropic was buying Musk's GPUs, Google was just printing money.
11th consecutive quarter of double-digit growth. AI-targeting turning every model improvement into ad revenue. DeepMind admitted publicly they're compute-constrained inside Google itself — Search, Cloud, and DeepMind are fighting each other for new GPU capacity.
The under-priced AI winner thesis is no longer hypothesis. It's reported earnings.
The honest answer: most people are doing nothing, because the rate of change has exceeded the cognitive bandwidth available to process it. The brain treats "this much changing this fast" as background noise and reaches for the comfortable framing — "AI is overhyped," "it's just a chatbot," "the bubble will pop."
Some of that may turn out to be true. But staking your career or your savings on the comfortable framing is a bet, not a strategy.
Here is the asymmetric version: if AI plateaus, your life doesn't change much. If AI doesn't plateau — and the benchmark data above is hard to reconcile with a plateau — your life changes profoundly within 18-36 months, and the people who started adapting in 2025 are several rungs ahead of the people who start in 2027.
Six concrete things you can do this month:
The AI Revolution Intelligence Database tracks 531 structured entries across breakthrough technology, industry disruption, labor data, expert credibility, and the financial mechanics of all of the above. Searchable, AI-chat enabled, downloadable.
Open The AI Revolution Database The AI Transition TimelineI'm Scott Covert. I build AI-powered tools and write analysis. I don't take ad money. I don't sell certainty. I track this stuff because I'm trying to figure out the same things you are — only earlier, and with a paper trail.
This page borrows from 531 entries in my public AI Revolution database. Every claim above has a primary source behind it. If something here is wrong, tell me.
If you want a tighter framing of the financial side — which industries get repriced, by how much, on what horizon — I track that separately at aistockmarketimpacts.com.