Claude Code Burns 33,000 Tokens Before Your Prompt; Apple's iPhone Chip Started Inside a Car

01On the same model and task, Claude Code spends 33,000 tokens before it reads your prompt

Someone ran Claude Code and OpenCode on the same model, the same machine, and the same tasks, then logged every byte sent and received. Asked for a one-line reply, Claude Code spent roughly 33,000 tokens of system prompt, tool schemas, and injected scaffolding before the user's request arrived. OpenCode spent about 7,000, according to the writeup.

The first test used Sonnet 4.5. Re-running on Claude Fable 5 narrowed the gap to about 3.3x, because Claude Code sends newer models a smaller system prompt. The multiple depends on the model. The direction does not.

Caching widened the bill. OpenCode's request prefix was byte-identical across every captured run, so it cached once per session and read it back cheaply. Claude Code rewrote tens of thousands of prompt-cache tokens mid-session, and on one task wrote up to 54 times more cache tokens than OpenCode, the test found. Cache writes carry a premium, which the author tied to a climbing usage dashboard.

Configuration compounds it. A production repository's 72KB instruction file adds an average 20,000 tokens to every request. Five modest MCP servers add 5,000 to 7,000 more. A real working setup runs 75,000 to 85,000 tokens deep before anyone types a word.

That is the builder's view of the machine. Set it against the industry's louder story.

George Hotz, writing on his blog, described himself as a former believer. He read Yudkowsky and expected recursive self-improvement and hard takeoff. Then he built things. At comma, he ships a hardware product with roughly a phone's complexity, and calls it genuinely hard. "Reality has lots of finicky details," he wrote, doubting a superintelligent ChatGPT could change a bike tire.

His argument is narrow and physical. No quantity of high-quality tokens turns lead into gold, he wrote. Intelligence is the current bottleneck for a few tasks, not the end of all things. "You cannot take over the world with tokens," he added. Software did not eat the world; it removed one layer of friction, then reintroduced it for a few companies.

One side promises machines that rewrite themselves. The other measured a coding tool burning five figures of tokens before the first keystroke. For developers billed per token, that overhead is a cost multiplier worth measuring before standardizing a team on any harness.

02SK Hynix's $26.5 Billion IPO Set a US Record. The Next Demand Is American Fabs.

SK Hynix raised $26.5 billion listing shares in the United States, the largest foreign IPO ever completed on an American exchange, according to TechCrunch. No foreign company had raised more in a US debut. The memory maker supplies the high-bandwidth chips that sit beside Nvidia's accelerators, and its Wall Street haul did not settle the demands on it.

TechCrunch reports SK Hynix and Samsung are both being urged to build new fabrication plants on US soil. The pressure points the capital back at physical output: the plants that etch silicon, not the balance sheets that fund them. A fab costs tens of billions and takes years to bring online. That lead time sits under every forecast of AI compute growth.

The friction downstream is already organized. The Verge's newsletter column, titled "The fight against AI data centers is just beginning," tracks community opposition forming against the buildout of the facilities that house those chips. The piece frames the resistance as early, not spent. It follows the same construction wave from the far end of the supply chain.

Line up the two signals and they point one direction. The binding constraint on AI is moving off the model and onto the plant: chip capacity at the top, data center siting at the bottom. Capital cleared the first bottleneck in a single record listing. It cannot clear a zoning board or pour concrete on a fab floor.

Both ends now run on timelines the software cycle never faced. Fab construction is measured in years and permits. Data center approvals route through local governments that can slow or stop them. The money has arrived ahead of the ground it needs to stand on.

03The AI chip inside every new iPhone started as a brain for a car Apple never shipped

Apple's self-driving car program never reached a road. What it built along the way now sits in the phone in your pocket.

Early in designing the driving platform, Apple's engineers hit a hard constraint. The vehicle would need heavy AI processing done on the device, not sent to a distant data center. According to reporting by Mark Gurman, that requirement pushed Apple's silicon group to chase a workload it had never targeted. The car processor itself was never finished. The engineering habits it forced survived the project.

Those habits now carry Apple's AI ambitions. The Verge traces a line from that abandoned automotive chip to the M-series parts Apple relies on for running models locally. The car was the reason to learn how to move that much computation onto a single piece of silicon. The car went away. The capability stayed and kept compounding through each chip generation.

That history frames Apple's posture this week. Apple has sued OpenAI, alleging that former Apple employees took trade secrets tied to unreleased hardware "for the benefit of OpenAI," according to the complaint. Apple says significant evidence points to individuals now at OpenAI wrongfully taking confidential information about its processes and products. OpenAI has responded separately; the specifics remain contested.

Strip away the courtroom and one detail stands out. The secrets Apple describes are hardware and process knowledge, the same category of work that turned a dead car project into a live silicon advantage. Apple built that advantage almost by accident. It is now going to court to keep rivals from copying the parts it never meant to develop in the first place.

The company that could not ship a car ended up shipping the chips that let it run AI without the cloud. The lawsuit is Apple treating that outcome as something worth defending in front of a judge.