SpaceX locks in the missing layer. The boldest promise in the SpaceX filing just got something to stand on.

The deal. Signal 005 flagged the Cursor option; this week it became binding — US$60 billion, all-stock, targeted to close in the third quarter. Paying in stock rather than cash keeps SpaceX's balance sheet free for the build-out. But it also means public shareholders absorb the dilution, and the integration story becomes part of what they're being asked to believe when they buy in.

Why now. The IPO roadshow still has to sell a climb in segment revenue that the filing can't yet support on organic growth alone. Starlink is the profitable engine; the AI side — xAI — is the loss-making slice: roughly US$3.2 billion in 2025 revenue against about US$6.4 billion in operating loss, and most of that revenue is disclosed compute leasing rather than a finished application product. The gap isn't only dollars. It's product shape.

What Cursor buys the stack. Cursor brings what SpaceX could not build on its own run-rate. An enterprise coding product — the cleanest return-on-investment field in the entire stack. A growing record of how large organisations actually write, review, and ship code. And customer relationships deep enough to expand inside. Under one owner, compute, model, and application become a single full-stack offer, with an operating history no rival can match on a shorter runway. The attached compute is the hardware half; Cursor is the enterprise half. And as data regulation moves into the buyer's checklist, a single-vendor path for enterprise coding — one audit trail, one residency answer, one contract — starts to look like a feature rather than overhead. None of that is in the filing yet.

What to watch. After close, the binding question is feed rate: does attached compute actually compound the model's capability inside the stack? Segment reporting and Cursor's usage numbers become the visible proof. Watch for deeper enterprise adoption, the economics of Composer, and a go-to-market built to land and expand. The integration test is now on a public clock.

What's growing. Two earlier threads just converged. Signal 002 measured how big the spending had become; Signal 005 caught the credit market starting to charge more for it. This week Nvidia — the company selling the chips, not buying them — issued its first-ever bond. The borrowing has reached the one layer that didn't need to borrow. The chart shows where the pressure is rising, year over year.

Why. Spending is climbing for two separate reasons, and it's worth keeping them apart. One: they're building more — more data centres, more capacity. Two: each unit costs more than it did — chips, power, and time in the interconnect queue have all gotten more expensive. So the bill rises even where the plan hasn't changed.

Management has an answer for this: the spending is demand-backed. Customers have signed forward orders, and those orders justify the build. The credit market is asking a sharper question — how much of this is just inflation on a build that was already coming, and what happens if the signed contracts turn out too small to cover it?

Oracle showed the split in real time last week. Backlog held. Revenue held. Capex stepped up. And free cash flow went sharply negative — strong enough to post good numbers, weak enough on cash to get repriced by lenders, at the same time. Both things are true. That's the whole tension in one company.

What happens next. Every kind of financing — debt, equity, vendor guarantees, money moving between related companies — is pouring into the same build-out, and the chart shows every layer speeding up at once. The cash leaves now. The cost of it shows up later: depreciation and debt service hit the income statement around 2028.

So the whole thing rests on a single bet. Real customers, paying for real applications, have to spend fast enough to cover that bill before it arrives — actual end demand, not hyperscalers selling compute to each other in a loop. If that revenue shows up, the build pays for itself. If it's the loop, the math doesn't close.

What to watch. Three clocks, on three horizons. Near term: the next CoreWeave or Oracle print — and specifically whether it's cash flow, not operations, that forces the reprice. Through 2026–27: capex guidance, the pace of new bond issuance against the rough ~$570B market estimate, and whether Nvidia's deal closes. Around 2028: depreciation finally landing on the income statement. The single tell across all three: does customer revenue actually show up in segment reporting — or does money keep circling between the same companies, funding the loop?

The Federal Energy Regulatory Commission (FERC) voted unanimously to make grid operators fast-track data-centre connections. The terms: data centres pay for the grid upgrades they require, and accept being throttled when the grid is under stress.

Interconnect queues — the wait to plug into the grid — have been the real bind on the build-out. A unanimous federal order on these terms is what that bind looks like when it stops being a forecast and becomes policy. It's the strongest read short of the grid actually failing. Power generation is still the industry's binding constraint — but now with a federal can of Red Bull behind it.

Other watch items. Water now travels with power in the public conversation. A UN-backed projection has data-centre water use in 2030 matching the basic needs of 1.3 billion people; community acceptance and local load limits are starting to bind as hard as queue time. Nuclear supply hardened in the same week: Oklo and Centrus on uranium, Oklo building for Meta in Ohio, Amazon on Talen. And where the grid is simply too slow, capital is routing around it — Verse on off-grid battery and solar, gas plants fast-tracked off-grid, UK regulators weighing mandatory throttling at peak demand.

Two different things are happening at once, and most coverage blurs them together.

The first is a story about a relationship breaking down. Anthropic refused to cross a line on autonomous weapons; much of the world applauded; Washington read it as defiance. The second is harder — a real question about how capable these models have actually become. That capability is perceived, but it was arguably validated from the outside when Project Glasswing put Fable and Mythos to work fixing live security vulnerabilities, right up until Commerce switched the models off. Merge the two threads and you misread both. Neither is overstated. Each needs its own read.

In the week after Signal 005's Commerce order, both threads accelerated. Cybersecurity executives pushed for relief. The breakdown in trust with the White House became the spine of the narrative. Business press started calling it an AI Strait of Hormuz — a single US chokepoint on frontier models that could cut off global access the way Hormuz can cut off oil. AFR's Chanticleer column argued Anthropic's own safety warnings had boomeranged on it. Late in the week the fight went global: Amodei appealed to the G7, allies pushed back on unilateral US action, Congress demanded the legal basis, and even insiders struggled to parse the export rules.

Washington acted on national-security grounds — the models reportedly beat human experts at cybersecurity tasks, which is the strongest read yet on how far autonomous frontier capability has come. The restriction cuts both ways on the China gap. Capability concentrates in US-controlled hands instead of diffusing worldwide — even as allies push for coordination rather than a unilateral wall.

The capital-markets version landed the same weekend. Anthropic's path to an IPO ran straight into the political reality the export action created: disclosure and geopolitical risk are now part of the governance story, not footnotes to it.

China's push for self-reliance moved on two fronts this week: silicon and capital markets.

On silicon: ByteDance is in talks for 50,000 Iluvatar CoreX chips — a third domestic supplier coming online. Zhipu released GLM-5.2 as open source, and its stock jumped 48%.

On capital: Shanghai relaxed its IPO rules to let unprofitable AI labs list on the mainland, giving DeepSeek, Zhipu, and MiniMax a domestic route to public funding. Meanwhile the price war among ByteDance, Tencent, DeepSeek, and Xiaomi keeps deepening — capability and commoditisation are moving together.

The mirror to the export-control story above is clean. As the US restricts frontier models, China builds its own stack and its own funding rails. The gap narrows on cost and benchmarks, even as export controls concentrate the very top of US capability behind a wall.

This week, the reckoning. Metered AI pricing isn't new. The invoice is. This week it landed, and enterprise finance is scrambling to catch up — budgets blown, caps imposed, named companies suddenly forced to show a return under the new math. And finance is still a stage behind the vendors. Bundled seats are gone. Seat-plus-metered-API is what blew the budgets this round. Next come model tiers, tool calls, and connector traffic stacked on top of the seat. The number finance will insist on is cost per completed task — not the list price of a token.

Two ROI patterns. RBC, Cisco, and Uber went all-in on metered stacks — agentic coding, API-heavy workflows — with no caps. Usage surged, mid-year invoices blew past annual budgets, and finance pushed back. The productivity may well be real, but on a metered stack the return is usage × price, and both move fast.

UnitedHealth is the counter-pattern. Bloomberg reports US$3 billion over two years to automate scheduling and claims — bots and workflow, not a token-invoice reset. The company claims a 2:1 return (Bloomberg, company figure, not independently verified here): the payback shows up as cost cut and faster cycle time.

Where smart budgets go next. Caps and routing are now table stakes — the baseline, not the strategy. The companies that come out ahead are investing in their own internal structure: clean context, prepared workflows, semantic layers — so that fewer tokens buy a finished task. Defending margin is a matter of context discipline and auditable ROI tracking, not waiting for the next pricing letter to arrive. The borrowing is uniform across the build-out; the payback is not — and at this layer, it has to prove out in pockets.