AI

Anthropic stacked three releases worth a trillion dollars in market signal. A $65B Series H. Opus 4.8. Dynamic Workflows in Claude Code that just rewrote Bun in 11 days. The SpaceX compute deal we covered last week turns out to have a 90-day exit. Musk and the S-1 disagree on which framing is real. And a quiet finding that strengthens every Western frontier-lab IPO narrative this fall. 15 stories. Read More:

Cognition closes $1B+ at $26B, with Devin AI Engineer at $492M ARR and Mercedes compressing an 8-month modernization to 8 days. NVIDIA commits $150B a year to Taiwan, not the US, the political ambition of onshoring is running ahead of the manufacturing reality. And Biohub opens protein biology with 6.8B sequences, putting an AlphaFold 3 challenger in every researcher's hands. Plus Simon Willison's read on why coding is the only AI vertical where the pricing math closes, OpenAI's Codex building a tax agent that patches its own failures, Ramp pointing 10,000 coding agents at its own backend, and 9 more. 16 stories. All sourced. All on the record. Read More:

Anthropic argues the way to keep agents safe isn't training them better. It's locking down the environment around them. Users approved 93% of permission prompts. One controlled phishing test exfiltrated AWS credentials 24 times out of 25. Harvey shows frontier models pass just 7.1% of legal tasks under an all-pass standard. Legal work is far from saturated. OpenRouter doubles to $1.3B betting the routing layer matters more than any single model. Chinese models went from 1% to 60%+ of its volume in 18 months. Plus: SpaceX's S-1 tells two AI compute stories, only one generates revenue. China extends travel curbs to private-firm AI talent. NVIDIA squeezes 15% more from already-optimized GPU kernels. And more:

Two voices the AI industry doesn't usually hear from showed up today. The Vatican published Pope Leo XIV's first full encyclical on AI: environmental cost, algorithmic exclusion, and the concentration of power. Doctrinal weight for 1.4 billion people, and a procurement document for Catholic-majority markets within a year. DeepSeek's 75% price cut last week turned out to be a wedge into a $10 trillion strategy: commoditize the model layer to capture the Chinese AI hardware platform underneath it. Plus: AlphaProof solves nine decades-old Erdős problems for a few hundred dollars each, xAI ships its first credible coding agent, and a new meta-benchmark finds only GPT 5.2 can design a test other frontier models actually find hard. And more:

We just hit a massive milestone, and it’s all because of YOU. The AI Insiders newsletter has officially crossed the 10,000 subscriber mark!

Two things landed today that say more about the next year of AI than any benchmark. DeepSeek made its 75% discount permanent. What was a promotion is now the price. The reasoning-tier gap between DeepSeek and Western flagships is now roughly 7-to-1 for the same evaluation. Procurement teams locking multi-year contracts at frontier rates are signing a price floor at exactly the moment the floor is moving. Anthropic shipped its own answer to that pressure in the same week. Compute cost per revenue dollar fell from 71 cents in Q1 to 56 cents in Q2. Claude Code alone is at $2.5 billion in annualized revenue. The "AI labs burn money forever" story finally has a number that breaks it. Underneath that, Anthropic is staging the Mythos 1 release. The closed-consortium model is now appearing in Google Cloud and AWS vulnerability scanners while the marketing rolls out as capability disclosure: a new paper showing Mythos can chain V8 bugs into working exploits, outperforming every other tested model. And in Washington, David Sacks made a phone call. Trump shelved the binding AI executive order he was hours from signing. The U.S. now has no federal AI framework on the 12-to-24-month horizon. The labs' own safety publications become the de facto record. Read More:

The "AI is taking developer jobs" narrative has a major flaw. Despite AI writing more code than ever, the demand for human software engineers is actually surging. Here is what the data is showing right now: ▪️Up 11% YoY: Software engineer job postings (per Citadel Securities). ▪️67,000+ open roles: Tracked by TrueUp, marking a 3-year high. ▪️42% of companies: Report needing more engineers because of AI, not fewer. How is this possible? It comes down to a classic economic principle known as the Jevons Paradox. When technological progress makes a resource cheaper and more efficient to produce, the demand for that resource expands rather than shrinks. Think of the Industrial Revolution: highly efficient steam engines didn't reduce coal consumption—they exploded it. AI is currently having the exact same effect on software development. ▪️AI Made Code Cheap Because AI has drastically lowered the cost and time required to generate code, the barrier to entry has plummeted. The result? Every business now wants bespoke, tailored software. Organizations that previously couldn't afford custom tech solutions are now actively building them, driving an unprecedented demand for engineering talent to oversee, architect, and deploy these systems. ▪️The Bottom Line The doom-and-gloom takes about the future of tech work confused two very different things: the cost of typing code and the value of engineering judgment. AI is an incredible tool for the former. But human expertise, architectural foresight, and complex problem-solving remain entirely indispensable for the latter.

Cursor hit a $3 billion revenue run-rate. OpenAI's Q1 came in at $5.7 billion. Anthropic is lining up a fourth compute supplier. The money is not the story anymore. Compute supply is. Every one of these companies is selling faster than it can find chips to serve. Then there's the deal nobody expected to see undone. Beijing has ordered Meta's acquisition of the AI agent startup Manus reversed, months after it closed. The founders are now raising a billion dollars to buy their own company back. Completed acquisitions are not supposed to come apart like this. And underneath it all, the bill. A survey of 7,000+ developers found AI now writes 56% of their code, up from 28% a year ago. In the same news cycle, the unemployment gap between entry-level and experienced workers keeps widening in exactly the jobs AI touches. The graduates booing commencement speakers this spring were not confused about technology. They had read the numbers. And more:

Three things landed today that don't usually land together. Anthropic committed $45B over three years to SpaceX for compute. The same week, it disclosed Q2 revenue tracking toward $10.9B and told investors it still doesn't expect full-year profit. The compute bill is rewriting the math. OpenAI lined up a September IPO after the Musk lawsuit was dismissed. Goldman and Morgan Stanley are already retained. And an OpenAI reasoning model disproved a 1946 Erdős conjecture in combinatorial geometry. Nine mathematicians, including Noga Alon and W.T. Gowers, signed off on the proof. The pricing question got two contradictory answers on the same day. CNBC says cheap models are eating the frontier. Tomasz Tunguz says flagship prices are climbing. Both are true, depending on which tier you're buying. And More:

OpenAI just turned compute into a contract. The new Guaranteed Capacity offering lets customers lock in inference compute for one, two, or three years at a discount. AWS played this game with Reserved Instances in 2009. Sixteen years later, the model layer is borrowing the playbook. Andrej Karpathy joined Anthropic for frontier R&D, choosing one lab among three he could have written his own ticket at. The signal value sits independent of compensation. Content provenance moved from policy paper to infrastructure. OpenAI adopted C2PA credentials and Google DeepMind's SynthID watermarking for AI-generated images, joining Adobe and Meta. The standards finally have the surface area to matter. And more: