Tech Intelligence Brief

Apple filed suit against OpenAI in federal court in Northern California, alleging that two former Apple engineers — Tang Tan and Chang Liu, now at OpenAI — systematically stole confidential data on unreleased hardware products, technical specifications, and supply-chain vendor details, with Apple's filing alleging the scheme operated "at every level" of the company, from technical staff up to OpenAI's Chief Hardware Officer, according to CNBC. Apple also alleges OpenAI directed job candidates still employed at Apple to bring "actual parts" to interviews for show-and-tell sessions. The suit lands two months after OpenAI won its own trial against Elon Musk over his nonprofit-governance claims — and Musk waded back in Monday, needling OpenAI's defense before Sam Altman shot back that Musk is "the one selling investors on space datacenters." Relations between the two companies chilled after OpenAI bought former Apple design chief Jony Ive's startup, io Products, for $6.4 billion last year to build consumer AI hardware. This is now the defining legal fight of the AI-hardware era — whoever wins effectively sets the boundary on how freely talent and know-how can move between Big Tech and frontier AI labs. Bearish

Anthropic is lobbying policymakers directly — blog posts and letters to lawmakers — arguing Chinese labs including DeepSeek, Moonshot and MiniMax ran industrial-scale operations to extract Claude's capabilities via distillation, citing more than 16 million exchanges generated through roughly 24,000 fraudulent accounts, according to reporting cited by Techmeme. The pushback has been sharp: AI researcher Nathan Lambert argues distillation is an industry-standard practice used by American labs too, and that banning Chinese open-weight models would "demolish the open model economy emerging in the US" — framing Anthropic's campaign as regulatory capture that happens to financially benefit Anthropic specifically. Open-weight models are facing a genuine policy inflection point in Washington right now, and this fight is shaping the terms of it. Mixed

Applications chief Fidji Simo is stepping back from her full-time role to recover from a chronic illness, telling colleagues "I failed to make this decision many times before," according to Yahoo Finance; separately, Business Insider reports that OpenAI's safety and alignment leadership keeps departing, with several recent exits landing at Anthropic. Turnover at this level, on this team, is worth tracking regardless of the individual reasons cited. Neutral

Anthropic is keeping Claude Fable 5 available on all paid plans, and Claude Code's weekly rate limits 50% higher, through July 19, according to Techmeme — a modest but notable signal of how much compute Anthropic is willing to burn to keep adoption momentum through the summer. Bullish

TSMC posted June revenue of NT$442.68 billion, up 68% year-on-year and 6.2% from May — a sequential gain that runs directly against 4 straight years of historical June declines — pushing first-half 2026 revenue to NT$2.4 trillion ($74.99 billion), up 35.6% from a year earlier, according to CNBC. The foundry says it remains sold out on the N3 node feeding leading AI GPU and CPU makers, and a Taiwanese minister confirmed TSMC plans two additional advanced chip-packaging plants in Chiayi Science Park, bringing the total to four facilities expected to generate more than $9.35 billion in combined annual output, according to Techmeme, citing Reuters. Thursday's Q2 earnings call is now the highest-conviction print of the week for reading the true state of AI capex. Bullish [SEMI CYCLE]

SK Hynix fell more than 15% in Seoul Monday — its largest single-session decline on record — unwinding a 13% Friday Nasdaq-debut pop that had priced American depositary receipts at $170, roughly 14% above the $149 reference price, in an offering that raised $26.5 billion, according to Bloomberg. The proximate trigger was profit-taking plus disappointment that second-quarter HBM4 shipment growth didn't materialize at the scale investors had priced in — the first genuine evidence in months that AI-memory positioning, not just AI-memory demand, had become stretched. The move dragged the KOSPI down hard enough to trip a trading halt, and rippled into Tokyo, where Kioxia plunged alongside it while Advantest and Tokyo Electron slipped by smaller amounts — full Japan-listed detail is in today's Japan Intelligence Brief. Watch whether this stays contained to Korea-specific listing mechanics or becomes the catalyst for a broader HBM-cycle repricing. Bearish [SEMI CYCLE]

China temporarily halted helium exports Friday, compounding a global supply crunch that the Iran war had already worsened by forcing a major Qatari helium facility offline; helium is essential for wafer cooling, plasma etching, chemical vapor deposition and lithography support, and China imports roughly 85% of its own supply, according to Nikkei Asia. The US has quietly become the top helium supplier to Japan, South Korea and Taiwan as a result — a small but genuine structural advantage for US-based fabs and equipment suppliers buried inside a headline that reads, at first glance, as purely negative for the sector. Neutral

Researchers at Peking University published an all-optical interconnect system linking standard electronic chips that reportedly boosts distributed AI inference speed more than 100-fold while using roughly one-ninth the typical compute resources, according to the South China Morning Post. It's an early-stage academic result, not a shipping product, but it's a clear signal of how much research capital China is directing at photonic workarounds for US chip-export restrictions — worth tracking as a multi-year thesis, not a trade. Neutral

Institutional investors are paring back exposure to Asian semiconductor names following a run-up that outpaced fundamentals in several cases, according to the Financial Times — a theme that lines up directly with Monday's SK Hynix unwind and is worth watching for whether it broadens into Taiwan- and Japan-listed names next. Bearish

Meta's custom "Iris" accelerator — the first of a four-generation, in-house Meta Training and Inference Accelerator roadmap co-designed with Broadcom and manufactured by TSMC — cleared testing in six weeks and enters production in September, part of a plan to double Meta's data-center computing capacity to 14 gigawatts by 2027, according to a Reuters-reviewed internal memo carried by CNBC. Meta plans to deploy 7 gigawatts of infrastructure this year alone — 1 gigawatt already added in the first half, another 5.5 gigawatts forecast by year-end — and expects to spend as much as $145 billion on AI infrastructure in 2026, a meaningful slice of Big Tech's projected $700 billion-plus AI capex for the year. Custom silicon at this scale is now a genuine third leg for Meta alongside Nvidia GPU purchases, not a hedge. Bullish

Ireland's data-center buildout has reached a scale where it now accounts for nearly a quarter of national electricity consumption, according to the Register — a concrete, national-scale data point for the power-constraint thesis that keeps showing up across every AI infrastructure conversation this year. Neutral

Cursor is internally testing a general-purpose AI agent, codenamed Sand, that handles email, texts and spreadsheet organization for non-developer users — its first product aimed beyond its coding-tool base — positioning it to compete directly with Anthropic's Claude Cowork and OpenAI's ChatGPT Work, according to The Information. Whether Sand ever ships is genuinely uncertain: Cursor's pending $60 billion acquisition by SpaceX, expected to close in the third quarter, could rewrite its product roadmap before Sand gets anywhere near a public launch. Neutral

OpenAI, Meta and SpaceXAI may be able to pressure Anthropic competitively by emphasizing cost efficiency, as enterprise customers increasingly scrutinize AI spending line by line, according to Techmeme, citing Bloomberg. It's the clearest sign yet that frontier-model competition is shifting from a pure capability race toward a genuine price war for enterprise wallet share. Mixed

Apple is reportedly skipping the M6 Pro and M6 Max entirely, jumping straight from the M5 generation to M7 Pro and M7 Max models in the second half of 2027, with an M7 Ultra chip targeting up to 1.5 terabytes of unified memory — roughly double the M5 Ultra's planned ceiling — and an M8 generation also in planning for 2028, according to Bloomberg's Mark Gurman. New iPad Pro models and two new Apple Pencils, redesigned partly to satisfy EU battery-replaceability rules, are also in the pipeline for next year. Whether Apple actually ships the full 1.5TB configuration likely depends on how the ongoing memory shortage evolves — a direct link back to this week's HBM and DRAM pricing story. Bullish

The musician's blunt critique, covered by The Verge, is a small but real data point on the design and cultural-acceptance hurdle AI wearables still face outside the tech-enthusiast bubble, regardless of how good the underlying hardware gets. Neutral

TSMC's four-plant, $9.35 billion Chiayi packaging build-out and the US's emerging helium-supplier advantage — both covered in full above — are two separate stories pointing at the same trend: the chip supply chain is actively re-routing around geopolitical risk points, not just responding to demand. Bullish