The AI frontier split into two directions this week: the money went physical while the software went cheap. A record-shattering memory IPO priced the hardware boom in public markets, even as Chinese models quietly captured nearly half of US enterprise usage on price alone — and Apple dragged OpenAI into court over the device meant to define AI's next form factor. Here's what mattered.
The memory boom goes public — and breaks a record
The biggest AI story of the week wasn't a model; it was a memory chipmaker. SK Hynix raised $26.5 billion in its Nasdaq debut, pricing near $149 a share and popping about 13% on the first day, in what became the largest IPO by a foreign company in US history — surpassing Alibaba's 2014 listing. The significance runs deeper than the number. High-bandwidth memory is the component stacked alongside every AI accelerator, and SK Hynix is its dominant supplier, which makes this listing a public-market milestone for the AI-memory supercycle rather than just a big finance headline. With the company also under pressure to build fabs on US soil, the message is clear: the scarce input in AI is shifting from raw compute toward the memory that feeds it — and capital is repricing the whole hardware stack accordingly.
Apple takes OpenAI to court over its secret AI device
The race to build AI's first breakout hardware device just moved from the lab to the courtroom. Apple filed suit in the Northern District of California alleging that OpenAI stole trade secrets "at every level," poaching more than ten Apple engineers and lifting unreleased hardware secrets to build the Jony Ive-designed device at the heart of its roughly $6.4 billion "io" acquisition. The suit points at OpenAI's hardware chief Tang Tan — himself a former Apple iPhone and Watch VP — and claims one named engineer downloaded confidential hardware files before leaving. Whatever the merits, the practical effect is friction: discovery, potential injunctions, and disputes over engineering talent are the kind of drag that can slow a hardware roadmap as surely as any technical hurdle. Apple is signalling it sees OpenAI's device as a genuine threat — and it's willing to fight it on legal ground.
Chinese models now handle nearly half of America's AI usage
While US labs chase the frontier, American enterprises have been quietly voting with their wallets. A CNBC investigation found that Chinese-origin models reached a weekly peak of about 46.4% of routed tokens on OpenRouter, edging past the roughly 35.7% share held by US-origin models. DeepSeek alone is the single largest vendor at around 17.6% — some 5.13 trillion tokens a week — and Chinese models undercut their US rivals by 60 to 90 percent. That price gap is the whole story: when buyers route real production traffic on cost, a frontier lab's brand counts for little, and the pressure on US margins becomes structural rather than temporary. The defensible advantage now has to be a durable capability lead or genuine lock-in — not simply being the best-known name.
DeepSeek's V4 turns AI inference into a metered utility
The model behind that adoption pressure just went generally available. DeepSeek shipped V4 in mid-July, with the flagship V4-Pro weighing in at 1.6 trillion total parameters (49 billion active), scoring around 80.6% on SWE-bench Verified and offering a one-million-token context window as standard. The most novel part isn't the benchmark — it's the price tag. V4 introduces peak and off-peak API pricing that roughly doubles rates during busy windows (mid-morning and mid-afternoon), effectively pricing intelligence like electricity, by time of day. That reframes the economics of always-on AI agents: workloads that can be shifted to quieter hours get materially cheaper, rewarding whoever schedules and batches their compute well. The open question is whether Western labs, staring down the same price pressure, will follow with metered pricing of their own.
A venture capitalist gets a seat at the Fed's table
The last shift this week was institutional. New Federal Reserve Chair Kevin Warsh has stood up external task forces, and a16z co-founder Marc Andreessen will co-lead the "Productivity and Jobs" panel, working alongside Stanford economist Charles I. Jones and Microsoft/Xbox executive Asha Sharma, with recommendations due by the end of 2026. It marks the first time a prominent venture capitalist has held a Federal Reserve policy seat — a polarizing appointment given Andreessen's outspoken, acceleration-first views. How the Fed chooses to frame AI's impact on productivity and jobs shapes the macro backdrop for every AI company in the country, which makes the identity of the person holding the pen genuinely consequential.
The through-line: money is flowing into AI's physical foundations even as the software layer commoditizes from below. SK Hynix's record listing prices the memory boom in public markets; Chinese models and DeepSeek's utility-style pricing squeeze US labs on cost; and the Apple-OpenAI clash shows the hardware race is now being fought in court as well as in the lab. Worth watching next: whether SK Hynix's raise translates into US fabs, whether any Western lab answers DeepSeek's peak/off-peak pricing, and the first real procedural moves in Apple v. OpenAI. Signal, not advice — and no live prices.