Mara Reyes
Below Cost
On the same day Anthropic filed to go public, it extended its Mythos security model to about 200 of the most important organizations on earth — and confirmed it is selling that model below cost. The loss-making product is the IPO asset. The subsidy is the moat.
The Supplier
In February the Pentagon called Anthropic a supply chain risk. By May the Pentagon was deploying its model across the US government, the Vatican was launching Pope Leo's first encyclical with the co-founder, and the G20's Financial Stability Board was waiting on a briefing — every institution that tried to push Anthropic away, and every institution that didn't, reaching for the same company in the same month.
The Token Meter
Amazon set weekly AI usage targets so its workforce would actually use the tools the company had bought. Eighteen weeks later, employees have built a tool whose only purpose is to fake hitting them — the fastest Goodhart cycle on record.
The Reform
In June 2025, the Trump administration dissolved the US AI Safety Institute into the Center for AI Standards and Innovation, scrubbed "safety" from the mission, and oriented the body toward "US dominance." Eleven months later, every frontier US lab is granting CAISI pre-release evaluation access and the White House is drafting an executive order requiring it.
One Rule
In December 2025 the Trump White House signed the "One Rule" executive order, designed to keep the federal government out of AI model review and to sue states that imposed their own. On May 4 the same administration was sourced as drafting an executive order to vet AI models before release.
The Floor
On the same Sunday, Maryland became the first US state to ban algorithmic price discrimination in grocery stores and a Chinese court ruled that companies cannot fire workers to replace them with AI. Neither government was waiting for its national legislature. The federal preemption architecture designed to silence the states made the states the architecture.
The Rails
On a quiet Saturday, five stories landed that looked unrelated — an AI super-PAC paying TikTok influencers, retail traders unleashing AI agents on Polymarket, an Iranian crypto exchange exposed as a sanctions conduit, the Trump family token sold privately in "white glove" transactions, and a Ripple exec putting $3.5M behind an AI-regulation congressional race. The shared infrastructure is the story.
The Toll
Apple was supposed to be the AI loser — two years behind, no model, capex moving the opposite direction from everyone else. On the day it posted the best quarter in its history, the company couldn't ship Mac minis fast enough because the model-builders were buying them all.
The Treehouse
A late-2025 amendment redefined founder-controlled stock as taxable wealth — converting California's billionaire tax from a liquidity nuisance into an existential threat to Silicon Valley's founder model. Four months later, Larry Page had moved his assets out of state, Peter Thiel had donated $3M to anti-tax groups, and Sergey Brin — the most apolitical billionaire of the founding tech generation — had spent $58 million-plus mobilizing a network of fellow tech leaders to choose California's next governor. The tax was designed to extract from billionaires. It produced the most coordinated billionaire political coalition in California history.
The Bifurcation
For three years, every frontier AI release was cheaper than the one before it, and the global market for model inference behaved like a single commodity. On April 24, 2026, OpenAI launched GPT-5.5 at $5 per million input tokens — exactly double GPT-5.4's price six weeks earlier. The same day, DeepSeek released V4 Pro at $1.74 per million input tokens. The race to zero ended where every commodity race ends: at the moment buyers in different regions stop being treated as the same buyer.
The Reseller
On April 22, Google split its eighth-generation TPU into separate training and inference chips — a hyperscaler move proving it intends to compete with Nvidia on the silicon. The same day, Mira Murati's Thinking Machines Lab signed a multi-billion-dollar agreement with Google Cloud to rent Nvidia's newest chips. The chip built in 2015 to reduce Google's dependency on Nvidia now sits alongside Nvidia in the same cloud, rented by the same customers.
The Silicon CEO
On April 21, Apple named John Ternus — the SVP of Hardware Engineering who led the Mac's transition to Apple Silicon — as Tim Cook's successor, and promoted chip chief Johny Srouji to Chief Hardware Officer the same day. The same quarter, Craig Federighi was reported to be routing the revamped Siri to other companies' models. Every peer building the AI era picked a software CEO. Apple picked the chip guys. The two moves are the same bet — that models commoditize, and silicon doesn't.
The Starting Line
On April 18, OpenAI lost three executives in a single day — Kevin Weil, Bill Peebles, Srinivas Narayanan. The same day, Recursive Superintelligence — a four-month-old lab founded by ex-DeepMind and OpenAI engineers — was reported to have raised $500 million at a $4 billion valuation from GV and Nvidia. OpenAI's 2015 charter was written to prevent a concentrated AGI race. Eleven years later, the charter underwrites the payroll of every frontier lab that competes with it. The org became the starting line it was designed to refuse.
The Verdict
The Department of Justice settled its antitrust case against Live Nation in March 2026 — divesting 13 amphitheaters, which Senator Klobuchar called "weak." One month later, a jury of twelve citizens found the same company guilty of illegally maintaining monopoly power. The system designed to enforce antitrust gave up. The backup system didn't.
The Tax
Samsung is one of three companies that manufacture the world's memory chips. This week, it raised the price of its Galaxy tablets by $280 — because memory chips are too expensive. Microsoft's Surface laptops are up $500 since launch. Both cite the same cause: AI data centers consuming 70% of high-end memory production. The boom that was supposed to make software smarter made hardware more expensive.
The Standard
In March 2026, the Linux Foundation raised $12.5M to help open-source maintainers defend against low-quality AI-generated code. Three weeks later, the Linux kernel — the most conservative, most scrutinized codebase on earth — accepted AI-generated contributions. The quality bar didn't move. The definition of who clears it did.
By Invitation
Anthropic built a framework to govern its most dangerous models. On April 8, it launched Claude Mythos Preview — too capable for public release, available only to 40+ critical infrastructure partners. Hours later, OpenAI sent investors a memo about compute advantage, then announced it was building a competing restricted product. "Similar to Anthropic's."
Cognitive Debt
AI coding tools were built to reduce the time developers spent writing code. They did. The time went to something else — reviewing code nobody wrote, debugging code nobody fully understands.
The Anti-Distillation Skill
Skill documentation was built on one assumption: that knowledge worth capturing can be captured. The adversarial cycle it spawned is a live test of that assumption.
The Pressure Test
In 2018, US export controls were designed to prevent China from mass-producing advanced chips. In 2026, Huawei's domestic AI chip is in mass production with prices rising 20% on demand, Chinese semiconductor companies just reported record revenue, and DeepSeek V4 runs on domestic hardware. The containment is creating the thing it was supposed to contain.