Apple dominated the day's news with a long-delayed overhaul of Siri unveiled at Tim Cook's final WWDC as CEO, built in unusually close collaboration with Google. The financial story of the moment is OpenAI's confidential IPO filing, which lands a week behind Anthropic's and just ahead of SpaceX's expected debut, signaling that the largest AI companies are all moving toward public markets at once. Elsewhere, a Chinese lab posted a striking inference-speed milestone, the compute build-out deepened across NAVER, Google, SpaceX, and Samsung, Microsoft reshaped the economics of AI agents at its Build conference, and a fresh wave of security incidents hit developer tools and password managers.
Apple's AI Reset
Apple introduced a substantially rebuilt Siri, now branded "Siri AI," alongside broader Apple Intelligence upgrades that will arrive with this fall's operating system releases. The assistant can draw on past interactions, on-screen visual context, and broader knowledge to handle multi-step tasks such as researching concert tickets or brainstorming recipes, and Apple is launching a dedicated AI assistant app that moves its product closer to standalone offerings from OpenAI and Anthropic. The most notable structural change is under the hood: Apple's new architecture runs Foundation Models co-developed with Google, both on-device and on its Private Cloud Compute infrastructure, marking a significant reliance on an outside model partner. Apple also released Core AI, a developer framework for building, running, and deploying models across the CPU, GPU, and Neural Engine on Apple silicon, with a Swift API for controlling specialization, caching, and inference performance. The event doubled as a leadership milestone, with Tim Cook making his final WWDC appearance as CEO before handing the role to hardware chief John Ternus in September.
The IPO Wave
OpenAI confirmed it has confidentially submitted a draft S-1 to the SEC, while stressing that no timing for a public offering has been set. Executives framed the filing as preserving strategic flexibility rather than signaling an imminent listing, noting there are still initiatives easier to pursue as a private company; OpenAI also plans to facilitate a tender offer for existing shareholders. The move comes about a week after rival Anthropic filed its own IPO paperwork and just before SpaceX is expected to go public, a clustering that analysts say will intensify competition for capital as frontier-model development grows more expensive. Underscoring the trend, Perplexity signaled it intends to pursue an IPO in 2028 regardless of how the Anthropic and OpenAI offerings play out.
Models and Compute
Xiaomi, working with inference partner TileRT, unveiled MiMo-V2.5-Pro-UltraSpeed, a one-trillion-parameter model clocking roughly 1,000 tokens per second on a standard eight-GPU commodity node — a pace it describes as about 15 times faster than ChatGPT and Claude. The gains come from FP4 quantization on the model's expert layers and a technique called DFlash speculative decoding, which proposes a full block of tokens in a single pass; the model is available through a limited API trial from June 9 to June 23 at roughly three times the standard rate for about ten times the output.
The infrastructure race continued on several fronts. NAVER announced plans to expand sovereign AI capacity from 55 megawatts toward gigawatt scale using NVIDIA's DSX platform, starting at its data center in South Korea. Google agreed to pay SpaceX roughly $920 million per month from late 2026 through mid-2029 to rent data-center capacity, contingent on SpaceX supplying 110,000 NVIDIA chips. And Samsung is reportedly in early-stage production of AI "space computing" chips for SpaceX's plan to process and beam satellite data from orbit, with TSMC also named as a partner.
Enterprise AI and Agents
At Build 2026, Microsoft shipped Scout, an always-on work agent, on top of the open-source OpenClaw runtime — and made the runtime itself free, choosing instead to monetize the surrounding "control plane" of identity, governance, auditability, and enterprise management. Microsoft also previewed MXC, a platform-agnostic sandbox that runs untrusted and agentic code at the OS level rather than the application level. Amazon added interactive, terminal-style shells to its Bedrock AgentCore Runtime, giving developers persistent access into running agent sessions. The day's analysis pieces echoed a recurring theme — that the model is no longer the bottleneck, with value increasingly determined by the workflow, data, and tooling built around it.
Security
Microsoft removed roughly 70 GitHub-hosted open-source projects after discovering malware in tools tied to Azure, Claude Code, Gemini's CLI, and VS Code that captured developer credentials, an incident apparently linked to an earlier compromise of Microsoft's Durable Task project. Researchers also disclosed a now-patched remote-code-execution flaw in the Cursor editor that could trigger simply by opening a malicious repository, and detailed an attack on Dashlane's device-enrollment APIs that let intruders download encrypted vaults from a small number of accounts. Microsoft has begun delaying automatic VS Code extension updates by two hours for non-trusted publishers to blunt fast-moving supply-chain attacks.
Quick Takes
Google reported that Search queries hit an all-time high last quarter, with AI Mode surpassing 1 billion monthly active users and AI Overviews topping 2.5 billion, helping drive 19% year-over-year revenue growth. OpenAI launched an Economic Research Exchange, a program for outside researchers studying AI's effects on workers, firms, and the broader economy, and separately published a plan from Sam Altman and Jakub Pachocki framing its goals as building an automated AI researcher, broadly sharing economic gains, and giving everyone access to a personal AGI. New research suggests AI adoption has lifted pull-request throughput by roughly 10 to 15 percent at many organizations, with a median gain closer to 8 percent, as non-coding work continues to cap overall velocity. A new benchmark called FrontierCode aims to measure how well models produce mergeable, production-quality database code. Airbnb CEO Brian Chesky is reportedly starting an independent AI lab focused on user interaction and product design. OpenAI added a Lockdown Mode to ChatGPT that restricts outbound network requests to guard sensitive data against prompt-injection exfiltration. Commonwealth Fusion Systems said five peer-reviewed papers validate the physics of its planned commercial plant, with grid power targeted for the early 2030s and nearly $3 billion raised to date. And Situational Awareness, the AI-focused fund started by 24-year-old Leopold Aschenbrenner with no prior professional investing experience, has reportedly grown to more than $20 billion in under two years, helped by an early stake in Anthropic.
What This Means for Your Business
Apple's decision to lean on Google's models is a reminder that even the most vertically integrated companies are now assembling AI from outside parts. For small and midsize businesses, the practical takeaway is that a dedicated, system-level AI assistant is about to be standard on hundreds of millions of devices your customers and employees already own. Owners should start thinking about how their apps, content, and support flows will be surfaced — or summarized — by these assistants, much as they once optimized for search engines. The arrival of Apple's Core AI framework also lowers the bar for adding on-device AI features to your own apps without sending data to third-party clouds, which matters for any business handling sensitive customer information.
The simultaneous march toward IPOs by OpenAI, Anthropic, and SpaceX signals that the cost of building frontier AI is rising faster than private capital can comfortably fund. For buyers, that has two implications: pricing pressure as vendors chase revenue and profitability ahead of public scrutiny, and a continued premium on the "control plane" — governance, identity, and auditability — that Microsoft is betting will be where the money is. If you are evaluating AI agents, weigh not just model capability but the management, logging, and permission controls around them, because those are increasingly what you are actually paying for.
The security incidents this week deserve direct attention from any business whose teams use AI coding tools. Malware planted in packages tied to Claude Code, Gemini's CLI, and VS Code, plus a zero-interaction exploit in Cursor, show that the developer toolchain has become a prime target. A reasonable response is to slow down on auto-updating extensions and packages, pin trusted publishers, and treat any repository opened in an AI-enabled editor as potentially hostile until reviewed. The Dashlane vault-theft technique is a fresh argument for strong, unique master passwords and for confirming that your password manager hardens stored data against brute-force attacks.
Finally, the data on AI's measured productivity impact is a useful corrective to inflated expectations. A roughly 8 to 15 percent gain in software throughput is meaningful but far from the wholesale transformation often promised, and it is concentrated in coding rather than the planning, review, and coordination work that consumes most of a team's time. The lesson for businesses adopting AI is to target the surrounding workflow — clearer requirements, faster reviews, better handoffs — rather than expecting the model alone to deliver outsized returns. Xiaomi's speed milestone points the same direction: as inference gets cheaper and faster, competitive advantage will come less from access to a model and more from how well it is wired into the way your business actually runs.