1. Printing one AI chip needs a $400 million machine. The revenue to pay it back is what nobody can find.

Jos Benschop climbs a ladder to reach the top of his newest machine, and it is a schlep. The thing stands as tall as a double-decker bus: more than 150 tons of precision-milled aluminum, wrapped in thousands of snaking tubes, colored cables, and pressurized tanks. His employer, ASML, builds the lithography systems that pattern the silicon inside advanced AI chips. One machine runs about $400 million.

That number sits at the bottom of the AI economy. Above it stack the fabs, the data centers, the GPUs, and the models that enterprises now pay to run. The hardware works, and the precision is real. The open question waits at the other end of the chain: whether the output ever earns back the machine.

David Cahn of Sequoia Capital started asking in September 2023, in a post titled "AI's $200B Question." Nine months later he reran the math under the title "AI's $600B Question." His estimate of the revenue gap had tripled inside a year. Independent writers including Ed Zitron had flagged the same hole earlier.

The gap shows up in unit costs. Estimates of how much platforms spend to produce $1 of revenue cluster between $8 and $14, according to figures Zitron compiled. To probe it, SemiAnalysis, a semiconductor analyst Zitron describes as strongly pro-AI, ran random long-horizon coding tasks until they hit the usage ceilings on OpenAI's and Anthropic's paid tiers.

The results, as Zitron reported them, point one way. Assuming the platforms aren't selling tokens below cost, Anthropic subsidizes its enterprise customers by up to 40 times. OpenAI, by his reading, subsidizes by up to 70 times. The heaviest users cost the most to keep.

The business press has started to notice. Companies are complaining about the cost of the tokens their employees burn, and the complaints have gone from a trickle to a flood. The machine on the factory floor holds its tolerances to the nanometer. The price of running what it prints holds only as long as the funding does.

Why it matters: Enterprise AI priced up to 40–70x below provider cost, per Zitron; heaviest-usage customers are the most expensive to serve; current token pricing survives only while platforms keep funding the gap

2. The reasoning Claude Code saves to your disk is encrypted, summarized, and not what drove the model

A developer spent a weekend inspecting Claude Code's session logs, which record the model's reasoning as it works. Inside the "thinking blocks" he found a signature roughly 600 characters long and no readable text. The docs explain why: Claude encrypts its reasoning into that signature, Anthropic holds the key, and the local machine never receives it. The API hands back a summary of the reasoning, not the reasoning. Getting the full thinking output requires an enterprise agreement.

The practical upshot, according to the writeup, is that anyone promising an audit trail from those local files cannot produce one. The text surfaced through ctrl+o is a compressed account of whatever logic actually ran, with the lossy conversion that compression implies.

A separate post, summarizing a paper on prompt injection, lands on a parallel point about a different layer. It reframes prompt injection as "role confusion" and argues the attacks stem from a flaw in how LLMs perceive roles. The chat interface shows distinct turns. The model receives one continuous string holding system prompts, user messages, tool outputs, and its own prior responses. Edit that string and you edit the model's reality; delete a turn and the exchange never happened.

The two findings come from unrelated authors and point at one gap. What a model presents to the user is a rendering, not the computation beneath it. The displayed reasoning is a summary of the real reasoning. The role boundary the model appears to honor is a perception the model can get wrong, not a barrier enforced somewhere in the stack.

That gap reaches two common developer habits. Debugging an agent by reading its self-reported reasoning means trusting a summary the vendor generated, not the trace that produced the actions. Security designs that assume a firm wall between trusted system instructions and untrusted tool output rest on the same role perception the injection research shows is unreliable. Both treat the visible layer as the mechanism.

Why it matters: Claude Code logs cannot serve as a reasoning audit trail; full thinking output gated behind an enterprise agreement; agent debugging via self-reported reasoning trusts a vendor summary; injection defenses assuming firm role boundaries rest on flawed model perception

3. Midjourney wants to dunk you in a vat of water and call it an MRI, with no evidence it works

Midjourney built its reputation on an image generator. Last week the startup announced a body scanner, and the gap between those two products is the story.

The pitch, according to The Verge, is a futuristic ultrasound machine that would submerge users in a tank of water. Midjourney says the device could produce "something as powerful as MRI" while staying "as casual as a trip to the spa." What the company has not produced is evidence. The Verge reports nothing to support the comparison to an actual MRI scan. The product reads as a vision dressed up as a medical instrument.

Google's Fitbit Air arrives the same season selling the opposite. Its Health Coach does not promise spa-grade diagnostics. It tells the wearer their sleep is poor and their readiness score is low. It flags heart rate variability below baseline and notes too much time in a hot, humid environment. The Verge calls this the smarter response to a market crowded with AI health noise. The claims stay small because the data is ordinary, drawn from metrics wearables have tracked for years.

Two companies place two bets on AI entering health. One leads with a spectacle it cannot yet back. The other narrows its promises to what a wrist sensor can measure. Midjourney sells the future; Fitbit sells the recovery score you can check this morning.

For the buyer, that difference is hard to see from a launch page. A water tank that scans like an MRI and a band that flags low HRV both carry the same AI framing. The Verge's reviews separate them on one axis: whether the company shows its work. Midjourney has shown a concept. Fitbit has shipped a tracker that, the reviewer says, mostly tells him he is exhausted.

Why it matters: No clinical data behind Midjourney's MRI-grade ultrasound claim; Fitbit confines promises to wrist-sensor metrics it already collects; buyers can't separate spectacle from proven diagnostics at the launch page

News Roundup

Oracle cut 21,000 jobs to fund debt-financed AI data centers Oracle eliminated roughly 21,000 positions while borrowing billions to build AI data center capacity. The layoffs offset infrastructure spending the company is financing through debt rather than profit. Source

Meta sold smart glasses under its own brand, dropping Ray-Ban exclusivity Meta showed glasses in three styles and seven colors without Ray-Ban branding, including one line tied to Kylie Jenner. The shift ends three years of pairing every Meta face-wearable with Ray-Ban. Source

GPT-5 Pro helped an immunologist crack a three-year T cell question Researcher Derya Unutmaz used GPT-5 Pro to generate a hypothesis explaining T cell behavior that had stalled his lab for three years. He said the insight could feed cancer and autoimmune research. Source

NVIDIA pitched always-on AI agents for telecom network operations NVIDIA released tooling for AI agents that run network management, customer care, and back-office tasks without step-by-step human direction. The company aims to move telecom operators from task automation to agents that correlate data and decide next steps. Source

Google Home will recognize tagged people facing away from cameras Starting June 23rd, Google expanded Familiar Faces so smart home cameras identify tagged people by clothing and other cues, not just frontal faces. The update targets misidentification when subjects turn from the lens. Source

Sony's AI Camera Assistant produced poor photos on the Xperia 1 VIII After a week of testing, The Verge found Sony's new AI Camera Assistant degraded image quality on the Xperia 1 VIII. Sony had promoted the phone using shots widely criticized as among its worst. Source

Fika Jobs raised $4M for AI agents that interview job candidates Stockholm-based Fika Jobs combines AI interview agents with short-form video profiles in a hiring platform. The $4M round funds a product the founders compare to a mix of LinkedIn and TikTok. Source

OpenAI backed the Appia Foundation to set AI evaluation standards OpenAI is funding the Appia Foundation to build shared evaluation frameworks and safety practices for advanced AI. The effort positions OpenAI inside the bodies that will write industry test standards. Source

OpenAI published a workflow for running Codex on multi-session projects OpenAI detailed how engineer Jason Liu uses Codex to preserve context across sessions and continue complex work beyond a single prompt. The guide targets developers running agents on long projects. Source

EnterpriseClawBench tests AI agents on 852 real workplace tasks Researchers built EnterpriseClawBench from proprietary archives of real agent sessions, producing 852 reproducible tasks with fixtures, role classes, and grading rubrics. The benchmark measures agents that read files, call tools, and deliver business artifacts. Source

Cory Doctorow argued for attacking the AI bubble at its economic roots Doctorow's new book, The Reverse Centaur's Guide to Life After AI, lays out where he thinks AI deployment fails workers and how to dismantle its business model. He frames current AI as augmenting surveillance over labor. Source