TechLook
The frontier is moving in two directions at once. Up — SpaceX is selling orbital AI data centers that solve the gigawatt problem the terrestrial grid cannot. And out — OpenAI just joined Anthropic in the IPO queue, a money-losing nonprofit-turned-megacap telling retail investors to fund a burn no private wallet is big enough to carry. The same day, Apple finally shipped the Siri it promised two years ago, and the market decided, by a 2% drop, that even Cupertino's catch-up is not enough. Today's news is the AI era hitting its bandwidth limits — power, capital, distribution — and a handful of companies trying to break each.
SpaceX Wants to Put a Data Center in Orbit
Days before its IPO is expected to value the company near $1.75 trillion, Elon Musk told investors SpaceX plans a second Starlink — but this one is for AI compute, not broadband. The first orbital AI satellite, named AI1, will deliver about 150 kilowatts of peak power (roughly one Nvidia GB300 rack) using tech already proven in Starlink V3. A Bastrop, Texas factory should reach meaningful output by end of 2027.
Here's everything you need to know:
- SpaceX unveiled a prototype orbital AI data center satellite named AI1
- Each satellite delivers ~150 kW of peak compute power — about one Nvidia GB300 rack
- The Bastrop, Texas factory targets meaningful output by end of 2027
- A 1 GW orbital data center needs roughly 7,000 satellites, or about 280 launches
- SpaceX already flies a Starlink mission every three days and has lofted over 12,000 satellites
- The orbital thesis: free solar, free vacuum cooling, no grid interconnect wait
- The hard constraint is launch cadence, not physics — and SpaceX is the only operator that owns one
The deeper play here is that the gigawatt wall on Earth is not breaking. Permitting, transformers, water rights, and grid interconnect queues are getting longer, not shorter. A space-based data center sidesteps every one of them — the launch line is the only real bottleneck, and SpaceX is the only company in the world that controls one. As we noted yesterday, SpaceX has already locked in more than $70B of terrestrial compute contracts; orbital compute is the next leg of the same bet, with a much longer monopoly runway. For founders, the implication is not that you should build a space startup — it is that the AI infrastructure map is now three-dimensional, and the cheapest compute in 2030 may literally be above you.
OpenAI Files for IPO at $852B — One Week After Anthropic
OpenAI filed confidential IPO paperwork with the SEC on Monday at an $852 billion valuation — the third AI giant to line up for Wall Street in 2026, one week after Anthropic filed its own S-1. The same day, CEO Sam Altman published a manifesto promising "everyone on Earth a personal AGI" and gains that are "widely shared" — a message that lands days after he met Senator Bernie Sanders (who wants the public to own half of every major AI firm) and while the White House floats giving citizens a slice of AI's upside.
Here's everything you need to know:
- Confidential S-1 filed Monday at $852B post-money — one week after Anthropic
- OpenAI has not committed to a listing date; executives are weighing "things easier as a private company"
- Altman's manifesto lands the same week as his meeting with Senator Sanders
- The Trump administration is separately weighing a 1%–5% federal stake in OpenAI via a "Public Wealth Fund"
- OpenAI still loses more than it earns; private capital is no longer large enough to fund its burn
- The raise is being packaged as a "gift to humanity" before a single share changes hands
- Sriram Krishnan, the White House AI advisor, is also leaving the role at the end of June — removing a key Valley-to-Washington bridge mid-negotiation
The honest read: OpenAI is going public because the cheapest private money has run out. It loses more than it earns, its model lead is being chased by Google and Anthropic, and the only wallets left are mutual funds, sovereign wealth, and retail. The S-1 will eventually disclose governance, dual-class structure, and cap table — the things every founder will reference for the next five years. Today the question is not whether this IPO happens, but at what valuation, with what lockups, and under what government pressure. Build your API strategy assuming OpenAI's pricing becomes a public-market variable, not a private-market one.
Apple Finally Shipped an AI Siri — And the Market Shrugged
Two years after the smarter Siri it promised in 2024 never arrived, Apple used WWDC 2026 to relaunch Siri AI — a more conversational assistant that reads your screen, edits photos, and gets its own ChatGPT-style app. Tim Cook's keynote — billed as his last — also debuted iOS 27, macOS Golden Gate, and a homeOS preview for Apple's smart home hub. Wall Street reaction was muted: Apple stock fell about 2% on the day.
Here's everything you need to know:
- Siri AI ships English-only, skips China, and excludes EU iPhones at launch
- The underlying Apple Foundation Models were built with Google under this year's Gemini supply deal
- iOS 27 supports every iPhone from the iPhone 11 and iPhone SE 2nd gen forward — an unusually long support window
- Performance gains: app launches up to 30% faster, photos loading up to 70% faster, AirDrop up to 80% faster
- A new transparency slider lets users dial back the divisive Liquid Glass design language
- Apple is adding automatic blurring of inappropriate images and FaceTime content for child accounts
- iCloud Shared Albums now opens to Android and Windows
The catch is in the fine print. As we flagged before WWDC on June 5, the new Siri routes cloud inference through Google's Gemini — the first time Apple's AI stack has used a third-party model. Apple keeps the wrapper, the privacy story, the distribution; Google keeps the model. What shipped is assistant, not agent: Siri reads apps and suggests actions, but does not go do things across them. The 2% drop on a flagship AI reveal is the market telling Apple the bar has moved. John Ternus inherits a company that just joined the era it was supposed to define.
Microsoft Is No Longer Just OpenAI's Banker
For years, Microsoft was OpenAI's exclusive compute partner, biggest backer, and main distribution channel. As we noted after Build 2026 on June 3, that changed when the two companies formally split the arrangement in April. Now Microsoft is shipping its own frontier models — MAI-Thinking 1 plus six siblings, all trained on commercially licensed data and hosted on Azure, with customer fine-tuning on top.
Here's everything you need to know:
- MAI-Thinking 1 is in private preview and has not hit public leaderboards yet
- MAI-Image-2.5 currently sits at #4 for image generation
- The full lineup runs on commercially licensed data only — a meaningful IP position
- Models are hosted on Azure, giving Microsoft a built-in distribution edge
- The "commercially licensed + fine-tunable on customer data" combination targets enterprise procurement
- The OpenAI–Microsoft breakup is now a full-blown rivalry, not a renegotiation
For founders, this is the most important read in the story. The frontier model table is about to get more crowded, and Azure is being weaponized as a distribution edge for the in-house models specifically. If you are an Azure-native shop, expect preferential pricing, embedded endpoints, and Microsoft sales overlay. If you are multi-cloud, expect Microsoft to push hard on the "commercially clean training data" angle in every enterprise deal. Either way, the era of "we just resell OpenAI" at Microsoft is over.
Biology Just Proved the Boring Plumbing Is the Bottleneck
Anthropic published a research blog with an unfashionable claim: frontier AI agents race ahead in coding but crawl in biology — and the reason has nothing to do with how smart the models are. The problem is the data. Biology's databases were built for humans clicking through web pages, not for agents. When frontier models tried to retrieve virus sequences on their own, they returned 266 records on one Ebola task, then 106, then 5 — from the very same prompt.
Here's everything you need to know:
- Adding gget virus — a tool that makes retrieval deterministic — pushed every model above 90% accuracy
- The same fix erased the gap between cheap and expensive models
- A cheap model with the right tool matched the flagship
- Anthropic framed this as evidence that agent capability is gated by plumbing, not intelligence
- Karpathy has made the same complaint about software agents
The industry's instinct is to wait for the next, smarter model. This points the other way. The biggest applied-AI gains of the next year may not come from a bigger brain but from the auditable, deterministic retrieval and tooling layer no one wants to build. For founders, that is an underrated wedge: any vertical where the data is messy, the tools are ad-hoc, and the agents are failing is a vertical where a "boring infrastructure" startup can ship something that looks like magic without training a single new model. Biology is the demonstration case. Materials, chemistry, legal, and procurement are next.
WorkClaw and Kimi Work: Agents Stop Pretending to Be Tools
Two agent products landed in the same news cycle with the same pitch from different angles. WorkClaw deploys "AI coworkers" inside Slack and Microsoft Teams, each with a job description, a manager, and its own isolated computer, compatible with 3,000+ apps. Kimi Work is a desktop agent that can run up to 300 AI agents in parallel on a single computer.
Here's everything you need to know:
- WorkClaw is built by ex-Meta, Google, and Microsoft engineers and is targeting enterprise compliance buyers
- WorkClaw positions agents as teammates that onboard like hires, not as tools
- Kimi Work's 300-agent swarm is closer to a personal compute cluster than an assistant
- Both compete with the Claude Code self-prompting pattern we covered June 5 and the agent frameworks from Build 2026
- The category is fragmenting by workplace surface: Slack/Teams bots for chat-native teams, desktop swarms for power users
- The "one agent, one task" mental model is already aging
The implication for founders building in the agent space: the agent layer is splitting into at least three pricing surfaces — chat-native coworker seats, power-user compute clusters, and vertical agent frameworks. If you are picking where to build, the 300-parallel-desktop-agent story is the one to watch for power users, the WorkClaw story is the one to watch for enterprise, and the Claude Code story is the one to watch for developer-led adoption. The next six months will separate the "agent in a chat channel" demos from the ones that actually ship a measurable workflow outcome.
Pentagon Brands Alibaba, Baidu, Tencent, and BYD as "Chinese Military Companies"
The Pentagon reclassified the "Chinese military company" label to include Alibaba, Baidu, Tencent, and BYD — dragging three of Beijing's most familiar consumer-tech brands and its leading EV maker into the U.S.–China security standoff. The expanded roster now runs to nearly 200 Chinese firms.
Here's everything you need to know:
- The label stops short of sanctions but blocks access to U.S. defense contracts and research dollars
- Memory-chip makers YMTC and CXMT reappeared on the list after flickering on and off a February draft
- Beijing calls it discriminatory; the companies deny any military role
- Michigan lawmakers have separately proposed federal legislation to bar Chinese-branded vehicles from U.S. roads
- U.S. critics warn the label now brands nearly every major Chinese tech firm a de facto defense company
- The logic mirrors the TikTok-era hardware-security argument, but applied across the auto and consumer-tech supply chains
For founders with any cross-border supply chain, vendor list, or cloud dependency, this is the second "expand the perimeter" signal in two months. The first one was the TikTok ban; this one is broader and quieter. Any startup with Chinese-manufactured hardware, Chinese cloud dependencies, or Chinese OEM partners should be running a "what changes if my provider gets listed" scenario before the next reclassification, not after.
OpenAI's Side Bet, Tools for Humanity, Is Cutting Staff
OpenAI may be heading toward a blockbuster IPO, but CEO Sam Altman's other bet — iris-scanning startup Tools for Humanity, the company behind the Worldcoin Orb — is reportedly cutting staff after failing to turn biometrics into real revenue.
Here's everything you need to know:
- Tools for Humanity is downsizing despite a $2.5B valuation backed by a16z, Bain Capital, and Khosla Ventures
- Worldcoin's offer of roughly $50 in crypto for biometric data has triggered bans, fines, and privacy probes in South Korea and elsewhere
- The pitch — give us your eyeball, receive a token — has not crossed the regulatory threshold in most major markets
- It is a useful counterweight to the OpenAI IPO narrative: not everything in Altman's orbit scales
For founders in crypto, identity, and biometrics, this is a useful cautionary tale. A $2.5B valuation does not protect a product that asks users to hand over a biometric in exchange for a token, once regulators start asking questions. The bigger lesson is the funding-round-to-revenue gap: Worldcoin has spent years raising at a unicorn-multiple valuation without ever proving the unit economics. In a market that is repricing everything on revenue, that is the first thing investors now refuse to underwrite.
⚡ Quick Hits
- Perplexity: CEO Aravind Srinivas told CNBC the company is on track for a 2028 IPO, regardless of how the Anthropic and OpenAI listings fare — positioning Perplexity as the late-but-stable listing in a 2026–2028 AI IPO cluster.
- Instagram grid reorder: Long-press any post, tap "Reorder grid," drag posts into any order. Mobile-only (iOS and Android), with pinned posts staying locked at the top — turning the profile into a curated storefront rather than a chronological diary.
- Meta NameTag pulled: Meta removed hidden facial-recognition code from its Meta AI smart-glasses app after Wired exposed the unreleased system had already shipped — confirming facial recognition on face-worn cameras remains the third rail for smart-glasses adoption.
- ChatGPT in-chat graphics: ChatGPT can now generate graphics directly in conversation, turning the chat into a more interactive learning surface rather than a pure text response.
- PhysicsX: Raised $300 million to speed up hardware design with AI, joining a growing cohort of physics-informed model startups.
- UK AI supercomputer: The UK is backing a billion-dollar AI supercomputer to reduce dependence on U.S. tech — a sovereign-AI bet that mirrors France, Germany, and the Gulf.
- Karpathy AI usage video: Andrej Karpathy's "how I use AI day-to-day" walkthrough hit 1.5M views and 29K bookmarks in 48 hours, becoming the de facto reference for personal AI workflows.
- Instacart smart carts: AI-powered shopping carts are rolling into Weis Markets stores, expanding the computer-vision checkout footprint beyond the pilot.
- AI data centers in drought zones: New construction is increasingly landing in water-stressed U.S. regions — the resource constraint is no longer power alone.
- NASA Artemis IV / Prada: Astronauts will wear a Prada-designed garment under their spacesuits, with circulating chilled water and oxygen tubing for cooling — the first high-fashion partnership inside a NASA flight stack.
- Nintendo fined ~$40M: French authorities hit Nintendo with a ~€40M fine for misleading consumers about long-running stick-drift issues in the Switch's Joy-Con controllers — a hardware-quality enforcement precedent in consumer electronics.
- NASA X-59: NASA's quiet supersonic research aircraft has flown faster than the speed of sound for the first time — the first sustained data point in the low-boom revival.
- Laser-etched solar desalination: Researchers built a solar surface that turns seawater into drinking water while recovering salts as reusable solids instead of brine — flipping desalination's unit economics if the chemistry holds at scale.
Techlook — AI & tech signal for founders and builders.