Çelik here.

OpenAI filed its S-1 Friday at $852B–$1T, while still losing $1.22 for every $1 of revenue. The same week the four-star running NATO's transformation arm told Brussels Europe has nothing close to Palantir's Maven, five NATO capitals blocked a Ukraine aid floor, and a leaked Zuckerberg audio admitted Meta trained its frontier models on its own employees' keystrokes before firing 8,000 of them.

Four pieces, one thesis: the bodies meant to approve, fund, or compete walked off the field, and operators have a different Monday calculus because of it. The OpenAI filing is the public-markets refill for AI burn. The Vandier admission is the end of the EU's sovereign-tech pitch since CADA. The NATO block leaves Estonia and Poland carrying the 5% load alone. The Meta leak tells every knowledge worker their clicks are model fuel and their contract didn't say so.

OpenAI confidentially filed its S-1 Friday at a target between $852B and $1T. Q1 numbers show the company lost $1.22 for every $1 of revenue and projects a $14B operating loss for 2026. The ask is for retail and institutional money to absorb a multi-year burn window in exchange for the upside if the loss curve crosses zero.

The positional shift is who's writing checks. Strategic rounds at this scale required a syndicate of chip vendors and a Saudi or two. The IPO route hands the same risk to mutual funds, pensions, and 401(k)s, right after Hark raised $700M at $6B pre-product with NVIDIA, AMD, Intel, Qualcomm, Salesforce, and Brookfield all on one cap table. The next time AI prices drop, the seller is a fund manager facing quarterly redemptions, not a hyperscaler on a ten-year compute swap.

If you sell into enterprises or raise from US strategics, the next six months are the window. The prospectus will disclose customer concentration, revenue cohorts, and gross margins by product, giving competitors a Bloomberg view of OpenAI's price floors.

Build your 2026 pricing model assuming OpenAI defends revenue per user once the filing is public; the S-1 amendment cycle will print the first real AI gross margins by late summer.

Admiral Pierre Vandier, NATO's Supreme Allied Commander Transformation, told POLITICO on May 25 that "Europe has no real competitor" to Palantir's Maven Smart System. The admission landed the same week Maven completed its rollout to NATO's Joint Force Command Norfolk, the alliance command that integrates US Atlantic forces and operates over a portion of NATO's eastern flank C2. The man tasked with allied tech independence used the lectern to admit it's years away on the alliance's most important integration layer.

This is the end of the sovereign-tech pitch Brussels has been selling since CADA. The CEE counter-move was Germany's Quantum Systems acquiring Estonia's SensusQ earlier in May, bolting a Tallinn data-fusion team onto a German drone maker. That is the right ambition and the wrong order of magnitude: SensusQ to Palantir is like Stripe Atlas to JPMorgan. The continent's defense AI moat is a few mid-cap acquisitions and a handful of MoDs piloting Mistral instances, against a US incumbent already inside NATO's C2 plumbing.

If you build defense or dual-use tooling in Europe, the procurement door just got wider for the US incumbent and narrower for the local startup. Plan for a sovereign-AI fork in your roadmap: a SaaS instance for commercial and a sovereign instance for any government deal that survives Vandier's correction. Watch the EU's CADA implementation acts for what counts as "European" when the body that's supposed to enforce it has already conceded the technical race.

On May 24 the Telegraph revealed the UK, France, Italy, Spain, and Canada had blocked Mark Rutte's plan to peg NATO member contributions to Ukraine at 0.25% of GDP, the first concrete burden-share proposal since the alliance committed to a 5% defense floor. Rutte had foreshadowed the rejection two days earlier; the five-capital coalition formalized the kill quietly enough that the press only confirmed it on the weekend. The fixed-share concept is dead at the alliance level.

The capitals carrying the load now are the ones already over the 5% line. Estonia, Latvia, Lithuania, and Poland are all above it; Romania and Bulgaria are climbing toward it; and Ukraine itself spends nearly every dollar it receives the month it lands. The "ReArm Europe" architecture was always going to need a way to push the cost off the front line and onto the rear; that was the entire point of pegging support to GDP. The block tells the front-line members they pay in cash and conscripts while the western five negotiate bilaterally.

If you sell defense or dual-use into CEE, the procurement signal flipped. Eastern MoDs that expected the rear capitals to mutualize cost now have to plan for sovereign funding cycles, which means more domestic-preference clauses, longer payback windows, and harder asks on price. The first revised 2027 procurement framework out of Warsaw, Tallinn, or Vilnius will say the quiet part out loud, and Ukrainian export negotiations will start pricing in the new burden math.

A leaked audio from a Zuckerberg all-hands surfaced May 25 and confirmed Meta's "Model Capability Initiative" trained its frontier models on employees' Gmail, VSCode, and Metamate keystrokes before the company eliminated 8,000 roles in late April.

The recording is the first explicit admission that a frontier lab has used its own knowledge workers as training labor and then automated them. The April layoff round, which read as a routine reorg at the time, now reads as the harvest after the harvest.

The positional shift is the contract every operator signed without knowing. If you run a knowledge-work product, your employees' clicks may be model fuel for whoever provides your AI stack, and your employment agreements almost certainly do not say so out loud. The Meta leak gives the unions and works councils in Germany, France, and the Nordics a primary-source admission to point at, and it gives every prospective hire a question your offer letter has to answer.

Two moves before the end of the month. Audit which AI tools your team uses with productivity-telemetry enabled by default (Microsoft Copilot, Glean, and any LLM with "memory" turned on), and decide explicitly which sessions get logged for training. Tell your team what you decided, in writing, before someone else's leaked audio decides for you.

Five tools worth installing this week. Marketing sits out the rotation; Dev gets a double.

Productivity: Microsoft Agent 365 lands the governance layer. Microsoft's May update ships policy-based controls, runtime blocking, and Intune integration for every Copilot, custom, and third-party agent inside a tenant. If you sell into Microsoft-shop enterprises, this is the procurement gate your AI vendor walks through. Stand up a sandbox tenant this week and audit who can spin up an agent before IT does.

Design: Figma's AI Design Agent goes native on the canvas. Figma launched a canvas-level AI agent on May 20 that generates and edits inside your design system, runs multiple in parallel, and routes through OpenAI and Anthropic. Free in beta for Pro, Org, and Enterprise. If your team still treats design as the bottleneck, this is the week to test agent-driven variants on a live brand brief.

Sales: Salesforce Agentforce starts running your pipeline updates. Agentforce now surfaces high-intent accounts in Salesforce or Slack and updates pipeline stages and next-step fields in either "suggestive" review mode or fully automated. Already at 18,500 customers and 3 billion workflows a month. Flip one rep's pipeline to suggestive mode this week and measure the delta in stage hygiene against your manual baseline.

Dev: Cursor 3.3 ships Bugbot for autonomous triage. Cursor 3.3 adds durable canvases for multi-step plans and Bugbot, an in-editor agent that triages and fixes bugs without a human in the prompt loop. Treat it the way you treated dependabot: turn it on for low-risk repos first, watch the false-positive rate for a week, then graduate it. The point is whether your team trusts the diff, not whether the model can write one.

Dev: Claude Code 2.1.149 turns /usage into per-skill cost attribution. Anthropic's May 22 build breaks down spend by skill, subagent, plugin, and MCP server inside a single /usage call. This is the first mainstream coding agent that makes agent ROI measurable line-by-line; budget conversations stop being theological. Run /usage Friday afternoon on your busiest repo and you will know which skill to deprecate by Monday.Upcoming Events