OpenAI filed a confidential draft S-1 with the SEC on June 8 — eight days after Anthropic made the same move. The two filings together produce the first paired data on AI lab unit economics at scale: OpenAI at $25 billion in annualized revenue running a negative 122% non-GAAP operating margin; Anthropic at $47 billion ARR on a projected faster path to profitability. Public markets are being asked to price entities that generate more revenue than most Fortune 500 companies while running deficits without precedent in major tech IPO history.
OpenAI has committed to roughly $600 billion in compute costs over the next five years — against $25 billion in current annualized revenue — making the capital structure the central investor question independent of revenue growth. Goldman Sachs and Morgan Stanley are set as underwriters; a September 2026 debut has been cited but not confirmed by the company. Rather, in a tweet yesterday, OpenAI said - "We recently submitted a confidential S-1. We expect it to leak so we’re just announcing it. We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company. But it’s a complicated set of tradeoffs and this gives us the option to go public sooner if that ends up being best".
Google ordered more than 3 million custom TPUs from Intel's foundry for delivery by 2028, and reports emerged that Nvidia is evaluating Intel's 18A process node for a future multi-chip design. Intel closed at $99.17 on Friday — down 11% in the Broadcom-triggered chip selloff — then reversed the full move in a single session on Monday. A hyperscaler outsourcing 3 million units of a production workload to Intel is the strongest public signal yet that the 18A process has reached commercial trust. TSMC's capacity constraints are the forcing function; watch for additional hyperscaler commitments as CoWoS lead times remain extended.
Apple's WWDC keynote confirmed the rebuilt Siri routes its heaviest inference queries to Google Cloud running Nvidia Blackwell B200 GPUs — a direct named commitment from a platform with approximately 2.2 billion active devices. The performance gap on Apple's own Private Cloud Compute hardware drove the outsourcing decision. Modest day-over-day percentage gain in a session where Intel dominated; full structural analysis in Emerging below.
Apple's three-tier inference stack. The new Siri architecture announced at WWDC routes queries through three layers: on-device models, Apple's Private Cloud Compute, then Google Cloud running Nvidia Blackwell B200 GPUs for the heaviest reasoning tasks. Apple attempted to run the 1.2-trillion-parameter Gemini model on its own PCC hardware and found performance insufficient at Siri query volumes. Nvidia's confidential compute — hardware encryption of model weights, user input, and inference results inside GPU memory — met Apple's privacy bar; Apple's contract with Google also prevents Siri query data from being used to train future Gemini models.
Apple's privacy constraint was solved by hardware, not contractual architecture alone. Nvidia's confidential compute is now on record as meeting the privacy standard of the most privacy-focused consumer hardware platform in the market — a precedent directly applicable to enterprise, healthcare, and government customers facing equivalent constraints. The rebuilt Siri targets general availability with iOS 27 in September. Watch whether Samsung, Google Pixel, and other OEM platforms move toward equivalent third-party inference stacks before or shortly after the iPhone 18 launch.
Midjourney moves into hardware. In a tweet yesterday, Midjourney CEO notes they have a new hardware launch coming this week. Watch for the announcement, noting a shift for Midjourney beyond a model provider.