Subscribe
Latest analysis

Tokenmaxxing is over. Routing is the shape.

Chinese-origin AI models now handle nearly half of the enterprise token traffic on OpenRouter in a peak week, up from 4.5% eighteen months ago. That number does not mean the frontier moved to Hangzhou. It means the frontier stopped being the point. The operating pattern of enterprise AI in mid-2026 is a router in front of a portfolio of models, and the router picks the cheapest one that clears the bar. Anthropic and OpenAI still write the ceiling. They just don't write most of the tokens.

Read the piece →
4M+
robots deployed in global manufacturing in 2025
1 trillion
parameters in the largest AI models — and climbing
$500B+
invested in AI infrastructure globally in 2025
2030
estimated year humanoid robots enter the workforce at scale
Today

Latest briefs

All briefs →
Compiled from national AI strategy documents and public procurement disclosures, 2024-2026 sovereign-aigpu-supply

Sovereign AI stopped being a marketing slide and became a national budget line, and the arithmetic favors the same six countries

Nine governments have now committed direct capital to owning GPU capacity as national infrastructure. The public numbers add up to roughly forty billion dollars in announced spend through 2027, which sounds like a lot until you notice it is smaller than what Microsoft, Amazon, Google, and Meta have each spent on data center capex in a single quarter of 2026. The gap between the sovereign-AI pitch and the sovereign-AI arithmetic is where the next eighteen months of policy fights live.

AMD investor relations / The Register / SemiAnalysis, July 15 2026 amdmi450

AMD's MI450 shipped, and the pitch is finally price per token, not price per FLOP

AMD announced general availability of the Instinct MI450 on Wednesday, with initial capacity committed to Microsoft Azure, Meta, Oracle Cloud Infrastructure, and a still-undisclosed sovereign-cloud customer in the Gulf. Peak throughput is 3.1 exaflops FP8 per eight-GPU node, roughly 1.5x MI350 and, per AMD's own comparison chart, within 15% of Nvidia GB300 NVL72 on a per-node basis. The interesting shift is how AMD priced it. The MI450 list-price envelope, per multiple order-book leaks, undercuts GB300 by 30 to 40% on a delivered-token basis for Llama-4 and Qwen-3 inference workloads. That is the first time AMD has led with the workload economics slide instead of the peak-FLOPS slide.

1X Technologies blog / IEEE Spectrum, July 14 2026 1x-technologiesneo

1X shipped Neo to its first ten households, and the operators are calling in from Oslo

1X Technologies confirmed on Tuesday that the first ten Neo home humanoids have been delivered to preorder customers in the San Francisco Bay Area, kicking off a rolling rollout of the initial 500-unit batch through the third quarter. The company also published, in a blog post that took most of a day to work out how to read, that Neo units currently rely on remote human teleoperators for a majority of unstructured household tasks, with an Oslo-based operations team on shift to take control of individual robots on demand. Fully autonomous tasks at launch are limited to a fixed set: fetching a labeled item from a labeled location, loading and unloading a dishwasher of a supported model, and folding towels of a supported size. Everything else defaults to a human somewhere in Norway watching through the robot's eyes.

Anysphere blog / The Information, July 14 2026 cursoranysphere

Cursor launched a Background Agents tier at $200 a seat, and Anysphere quietly confirmed a $3B ARR run rate

Anysphere rolled out Cursor Background Agents on Tuesday as a new paid tier at $200 per seat per month, an order of magnitude above the existing $20 Pro plan. The feature spawns long-running autonomous agents on Anysphere-hosted compute that plan, edit, run tests, and open pull requests against a customer's repo without a human loop, capped at 40 concurrent agents per organization at launch. In the same post, CEO Michael Truell disclosed an annualized recurring revenue figure of $3 billion, up from the $500 million number reported in April 2025. That is the fastest ARR ramp in software history, by roughly a factor of two, and it is happening while every other coding-agent company (Cognition, Replit, Codeium, GitHub) is fighting for the same buyers with worse retention numbers.

TechCrunch / Apple v. OpenAI complaint filed N.D. Cal., July 10 2026 appleopenai

Apple is suing OpenAI over stolen trade secrets, and the giveaway is that OpenAI has a Chief Hardware Officer now

Apple filed suit in the Northern District of California on July 10 against OpenAI, its Chief Hardware Officer Tang Tan (previously Apple's VP of iPhone and Watch product design for 24 years), and senior systems electrical engineer Chang Liu (previously eight years at Apple). The complaint alleges systematic trade-secret theft: candidates asked to bring Apple components to interviews, departing employees coached on how to evade Apple security, a retained Apple laptop with confidential downloads, misrepresented authorization to use a proprietary metal-finishing technique. Apple wants an injunction, the materials back, and evidence preserved. The lawsuit is about a device, not a model, and the reason it matters is that Apple has now formally acknowledged that OpenAI is a hardware company.

TSMC investor relations / CNBC / Reuters / SEC Form 6-K, July 13 2026 tsmctaiwan-semiconductor

TSMC's June was the biggest revenue month in its history, and that number is the AI capex conversation ending

Taiwan Semiconductor posted T$442.68 billion of revenue in June, up 67.9% year over year and a new all-time monthly record, six percent above May, which was itself the previous record. Q2 came in at T$1.27 trillion (~$39.62 billion), 36% up year over year and above the top end of TSMC's own guidance. First-half 2026 revenue is now T$2.04 trillion, up 35.6% year over year. Full quarterly earnings land Thursday. The number that matters is this: every hyperscaler capex slide about AI datacenters has to actually show up as chip orders in Hsinchu, and it is showing up.

Analysis

Recent long-form

All analysis →
routingopenrouter

Tokenmaxxing is over. Routing is the shape.

Chinese-origin AI models now handle nearly half of the enterprise token traffic on OpenRouter in a peak week, up from 4.5% eighteen months ago. That number does not mean the frontier moved to Hangzhou. It means the frontier stopped being the point. The operating pattern of enterprise AI in mid-2026 is a router in front of a portfolio of models, and the router picks the cheapest one that clears the bar. Anthropic and OpenAI still write the ceiling. They just don't write most of the tokens.
nvidiacustom-silicon

Every frontier lab is quietly drawing its own chip, and Nvidia has about eighteen months to notice

Five weeks. Four custom-silicon announcements. Meta and Qualcomm on Dragonfly. OpenAI and Broadcom on Jalapeño. Google's next-gen TPU already committed to Anthropic. And now Anthropic itself, according to TechCrunch, in early talks with Samsung. The frontier labs have collectively decided that renting Nvidia at Nvidia's margins forever is not the plan. The interesting question is what that means for the shape of the industry in 2027.
openaianthropic

The federal government quietly became the AI sector's product manager

In a single week, OpenAI shipped its biggest model under a government-vetted preview list, Anthropic asked the Senate to sanction a Chinese rival for distillation, and the Pentagon's six-month Anthropic phase-out kept ticking. None of these events were on anyone's roadmap eighteen months ago. Read together, they describe a sector that has stopped being a free market and started being a regulated industry, without anyone formally announcing the change.

Frequently asked

What is The Clank?

Tech news for the meme-literate mind. We cover AI, robotics, and the technology companies actually worth paying attention to — with the depth to matter and the self-awareness to not take ourselves too seriously. Think Wired if it grew up on the internet.

Who is this for?

People who find this whole AI/robots thing fascinating, slightly terrifying, and genuinely funny all at once. You do not need to be an engineer. You just need to be paying attention.

What do you cover?

AI models and the companies building them. Robotics — humanoid, industrial, and the weird stuff in between. Chips, because nothing happens without them. Tech policy when it actually matters. And anything else in the tech world that is genuinely interesting.

Is this AI-written?

AI-assisted and human-edited. Research and drafting at machine speed, editorial judgment at human speed. We will not publish anything we cannot stand behind.

Why 'The Clank'?

It is the sound a robot makes. We thought that was appropriate.

Stay in the loop

The Clank, weekly.

One email a week. What happened in AI and robotics that actually matters — written for humans, not investors. Free.

Subscribe (free) Read the briefs
Subscribe

Join the early list

Drop your email. First issue lands when we hit a cadence we trust.