SpaceX IPO Escapes the Gravity of Valuation Models

SpaceX IPO: too big to fail

SpaceX has filed a 200,000-word prospectus, paving the way for the largest IPO of all time. With a valuation of $1.75 trillion, SpaceX would be worth more than almost any publicly traded company in the world. The S-1 document reveals two completely different businesses under one roof: Starlink generates $7.2 billion in EBITDA at a 63 percent margin, while the AI division xAI burns $30 billion annually, dragging the company into a net loss of $4.9 billion. Elon Musk controls 85.1 percent of the voting rights with no time limit, no independent board, and no possibility of removal. Three of the largest U.S. pension funds, with a combined trillion dollars in assets, have already demanded structural reforms – SpaceX rejected them all. → Linas's Newsletter

Synthszr Take: A $1.75 trillion valuation at 94x revenue and a governance structure that relegates public shareholders to mere extras – that's the price for Musk's Mars vision. Starlink is undoubtedly the most valuable connectivity asset of a generation, but the AI division is burning through those profits faster than a Falcon 9 rocket consumes fuel. The 37 pages of risk factors in the prospectus read like a warning to every institutional investor: you're buying into a one-man show where the Total Addressable Market of $28.5 trillion includes asteroid mining and Mars colonization. CalPERS and the New York pension funds get it; they are demanding reforms that SpaceX categorically rejects. At this price and with this structure, there is no margin of safety – only a bet on a man who buys $131 million worth of Cybertrucks from his own company.

Germany's Newest AI Billionaires: SAP Acquires Prior Labs

The number is still a secret, but it catapults three Freiburg natives directly into the top 500 wealthiest Germans: SAP is acquiring the 17-month-old AI startup Prior Labs. Founders Frank Hutter (48), Noah Hollmann (29), and Sauraj Gambhir (31) have developed a language model specifically for tabular data – exactly what SAP has been missing for years. CEO Christian Klein urgently needs an AI story after the stock price suffered for months under investor suspicion. Prior Labs is more than just a prestige acquisition: its technology could transform SAP's entire product portfolio, from accounting to supply chain optimization. The rapid integration shows how much pressure established software giants are under to buy relevant AI expertise instead of developing it themselves. → manager magazin – Der Tag

Synthszr Take: €435 million in annual profit, but no AI strategy of its own – that's SAP's dilemma in three numbers. Prior Labs solves a specific problem: while OpenAI and Anthropic focus on general language models, the Freiburg team understood that 80% of all corporate data is in tables. SAP is not just paying for technology here, but for lost time. The real punchline: a 17-month-old startup is saving a 52-year-old global market leader from AI irrelevance. We'll see this pattern more often – established companies buying innovation speed because their own structures have become too slow.

Microsoft and Uber Say Goodbye to Claude Code

Microsoft and Uber have weaned their engineering teams off Anthropic Claude Code. Microsoft is pulling the plug on the tool by June 30th, which developers internally preferred over its own GitHub Copilot CLI. Uber has already blown its AI budget for 2026. No wonder: Anthropic's revenue is projected to rise to $10.9 billion in the second quarter. The token-based costs became unsustainable for both corporations. Thousands of developers at Windows, Office, Outlook, Teams, and Surface must return to Microsoft's in-house tools, although Claude models will remain available via the Copilot CLI. In parallel, Alibaba is building a counter-model with Qwen3.7-Max: one million token contexts, four times larger than its predecessor, enough for medium-sized code repositories. → The Code

Synthszr Take: This is the first real reality check for AI coding tools at scale. The bill always comes due: Token-based pricing models scale brutally upwards when thousands of developers consume millions of tokens daily. Microsoft is doing the classic build-vs-buy calculation and deciding to go back to its own stack. The $10.9 billion quarterly revenue figure for Anthropic shows the other side of the coin. What's happening here is what we saw with cloud migration: First, everything runs on a credit card, then comes the CFO dashboard. Alibaba's million-token move is the right answer at the right time. Anyone who wants to compete in the coding AI market in 2026 will need either flat-rate models or in-house inference capacity.

Spotify Copies NotebookLM

Spotify is now also making AI podcasts. The new desktop app “Studio by Spotify Labs” generates personal audio briefings from emails, calendars, and web searches. The concept: An agent scours your digital footprint and crafts a podcast from it for your drive to Italy – including restaurant recommendations and suitable entertainment. The podcasts land in your personal Spotify library but are not publicly accessible. Spotify itself warns about unreliable AI content. The app is launching as a Research Preview in over 20 markets for selected users aged 18 and over. → Techpresso

Synthszr Take: Spotify is copying Google NotebookLM and everyone else peddling personal AI podcasts as the next big thing in 2026. Adobe is doing it, ElevenLabs is doing it, and now Spotify too. The pattern is always the same: an agent collects personal data, generates audio content, and voilà, the “magical experience” is ready. Except nobody listens to these podcasts (probably not even the developers themselves). The intent formulation “Create a podcast for my trip to Italy” sounds like a solution in search of a problem. Spotify would have been better off investing its compute resources in features people actually use. But in the current AI gold rush, every platform has to have its own podcast generator – even if it only fills the Spotify library with digital junk.

Is HTML the New Markdown?

Thariq Shihipar, Engineering Lead at Anthropic Claude Code, sparked a storm of concern with his blog post “The Unreasonable Effectiveness of HTML.” His thesis: Markdown comes from the era of token scarcity, while HTML offers interactive navigation, collapsible sections, embedded visualizations, and shareable links. 4.4 million views in 16 hours show that he's struck a nerve. The reactions are split into two camps: Team HTML sees Markdown's linear format as a relic (“A Markdown file you scroll past ceases to exist”), while Team Markdown warns of security risks from AI-generated JavaScript, unreadable code reviews, and token waste. Both sides have valid points, but the discussion highlights one thing above all: the developer community is ready for a debate about the future of documentation formats in the AI era. → Medium Weekly Digest

Synthszr Take: The Markdown debate is a proxy war for the real question: How much control are we ceding to AI systems? Shihipar's argument overlooks that Markdown is successful not in spite of, but precisely because of its limitations. It forces a clear structure, prevents JavaScript injection, and remains readable for decades. HTML may be practical for AI-generated reports, but who wants to do code reviews with hidden scripts? The real progress would be a format that combines the security of Markdown with the interactivity of HTML – without developers cluttering their Git history with DOM noise. Until then, anyone who declares Markdown dead doesn't understand why it was invented in the first place.

Figma Agent: Direct Manipulation Beats Prompt Engineering

Figma is bringing its own design agent directly onto the canvas. The special thing is: the agent knows the team's design system components, tokens, and patterns, and works in the same file as all other team members. Instead of prompting through external tools, you start directly from any design layer. You can explore different directions in parallel while the agent iterates. The integration is so deep that the agent automatically favors the most recently and frequently used components – but designers can also specifically control libraries and tokens via @-mentions. The workflow: Develop layouts and flows in Figma Design, then send it to Figma Make for code generation, iterate there, and then back again. → AI Valley

Synthszr Take: This is the right answer to the AI design tools popping up everywhere: integration directly into the workspace instead of separate playgrounds. The crucial point is direct manipulation – you stay in the Figma flow and use AI as a sparring partner, not a replacement. The parallelism (agent iterates while the designer iterates) solves the old problem of sequential handoffs. Figma is demonstrating what other tool makers still need to learn: AI agents only work if they understand and extend the existing work model instead of replacing it. The @-mention control for design system components? Brilliantly simple. This turns a generic AI tool into a specific tool for this exact design organization.

OpenAI is Visually Rewriting the Future of Advertising

OpenAI is systematically developing ChatGPT into an advertising platform. After months of infrastructure work, new ad formats are now following: larger images, customizable call-to-action buttons, and special e-commerce formats with prices and ratings. Particularly clever: The portrait version allows for carousel placements with three to four ads side-by-side. The self-service Ads Manager opens the platform to medium-sized advertisers, while features like retargeting and CPA optimization are still in the testing phase. Demand is already exceeding capacity – many accounts are still waiting to be activated. → STACKED MARKETER

Synthszr Take: OpenAI is building the next big advertising machine disguised as an AI assistant. The new formats show: ChatGPT is becoming a shopping mall with a conversational interface. The crucial point lies in Benji Shomair's remark about “creative variation” – in a high-intent environment where users are actively searching for solutions, ad formats work fundamentally differently than in a classic newsfeed. OpenAI is now collecting massive amounts of data on which creative formats work in which conversational contexts. That's the real competitive advantage: not the AI technology, but the understanding of contextual ad effectiveness in dialogue situations. The waitlists for the Ads Manager already show: the market is hungry for this new form of intent-based advertising.

Datasette Gets a Playful Agent Visualization

Simon Willison's Datasette ecosystem is expanding with an unexpected component: datasette-agent-sprites. The plugin visualizes AI agents as animated sprites in the database interface. When an agent navigates via SQL queries or transforms data, a small avatar shows its actions in real-time. Version 0.1a0 supports different agent personas with different animation sets – from the “Explorer” for read queries to the “Architect” for schema changes. The whole thing is based on Willison's MCP (Model Context Protocol) integration and makes the previously invisible reasoning process of the agents visually understandable. For example: Claude analyzes an SQLite database, and you see live how the agent sprite jumps from table to table, visualizes joins, and performs aggregations. → Simon Willison from Simon Willison's Newsletter

Synthszr Take: Willison is doing something clever here: He's giving the abstract work of agents a visual representation. This solves a core problem with using AI coding tools in companies – the lack of transparency. When an engineering team sees how the agent navigates its database, it builds trust. The sprite metaphor is deliberately chosen to be playful. This lowers the barrier to entry for non-technical users to engage with AI-powered data analysis. For the 200-engineer organization from my framework, this means the €546,000 annual investment in AI tools becomes more tangible when you can watch the agents at work. Datasette is thus positioning itself as an open-source alternative to the major cloud players – with the key advantage of complete control over one's own data.

Starbucks AI: Too Much Spilled Milk

Starbucks has discontinued its AI-powered inventory solution in all 11,000 North American stores after just 9 months. The provider, NomadGo, a 50-person startup from Bellevue, had pivoted in 2024 from people-counting to inventory-counting, using smartphone cameras, LiDAR, and on-device computer vision. The system was advertised with 99 percent accuracy and was supposed to work eight times faster than humans. The catch: The app constantly mixed up the different types of milk. → AI Secret

Synthszr Take: This is the perfect real-world test for the limits of computer vision. 99 percent accuracy sounds impressive until you realize: with thousands of products per day, a one percent error rate means hundreds of incorrect orders. Milk is milk? Maybe to a camera, but try explaining to a barista why whole milk was delivered again instead of oat milk. NomadGo made the classic mistake: They optimized the technology (eight times faster!) but underestimated the actual problem. In the food service industry, it's not speed that counts, but precision with critical ingredients. Starbucks learned a million-dollar lesson here: Sometimes, people with a clipboard are still the better solution.