SpaceX bought its own tenant. Price the dependency now.

SpaceX is now running $28B a year in annualized compute revenue. Anthropic at roughly $1.25B a month. Reflection AI on a $6.3B, $150M-a-month deal against Colossus 2. Google in the mix. A $20B bond issuance queued to fund the next wave. And this week, the acquisition of Cursor — one of the compute business's own tenants and the most widely deployed AI coding tool in the industry.

That is the story. Not the model release, not the eval paper, not the Hollywood deals. Every other item in today's feed is downstream of a single structural fact: one company now sells GPUs to the frontier labs, owns a fourth-hyperscaler-scale substrate, and has started buying the application layer that runs on top of it.

Call the shape of the trade what it is. AWS Lambda, except the runtime provider also sells you the compute and, increasingly, the IDE that generates the calls. Anthropic's inference now shares a landlord with a product that competes with Claude-native coding. Cursor's telemetry — prompts, code context, repo structure — flows into a subsidiary of the company that hosts your competitors' training runs. None of this is illegal, or even surprising once you look at the customer list. It is just newly true, and most enterprise contracts were written before it was.

The contracts are the second tell. Every large SpaceX compute deal reportedly carries a 90-day out clause. Anthropic has one. Reflection has one. Frontier labs are paying premium — Blackwell above $10/hour in bulk — and still refusing to lock in past a quarter. When the buyers with the most information in the market won't sign multi-year paper, the message is that they expect the price to move against SpaceX, not for it. The 90-day clause is not a migration plan. It is a negotiating instrument. Copy it into your next renewal.

Yes, but — the counter-reading is that $28B on 90-day paper is the most structurally fragile revenue base in AI infrastructure, and one bad quarter from Reflection reprices the entire narrative. That's fair. It also doesn't change what you do this week, because the vertical integration through Cursor is already shipping regardless of whether the compute revenue holds at scale.

What sits on top of this

Two other items landed in the same window, and both look smaller than they are once you place them next to the SpaceX story.

GLM-5.2 shipped at $1.40/$4.40 per million tokens, ranked #3 on GDPval-AA at 1524 Elo, and beat Opus 4.8 on Cline's real bug-fix loop at $0.41 per task versus $0.81. Twenty-plus providers serving it. Baseten at 280+ tok/s and sub-second TTFT, fresh off a $13B round. This is the first open-weight model where senior practitioners are quietly swapping defaults rather than writing thinkpieces. It is not frontier on every axis — it uses more tool calls, runs slower, wins on verification-heavy work and loses on latency-critical surfaces. The routing story writes itself. Model choice is now a config value, not an architectural commitment, and any codebase where it is still a code change is one sprint away from being expensive.

And a 541K-judgment audit across 21 LLM judges and 9 providers found that Cohen's kappa runs 33–41 points below exact-match agreement on MT-Bench. That is not a rounding error. It is the difference between shipping a model and shelving it. Most eval harnesses in production today report exact-match. Which means most quality gaps in your recent A/B decks are overstated, most judge-based ship calls were made on inflated signal, and the GLM-vs-Opus comparison you're about to run will give you the wrong answer if you don't correct the metric before you press go.

Stack the three items and the operator move is obvious. The model layer is commoditizing. The compute layer is consolidating under a vertically-integrated new entrant with 90-day paper. The eval layer that was supposed to help you route between them has been quietly lying by 37 points.

Two things for security and one thing for procurement

While the compute story dominates the headline, two adjacent items deserve a line each. OpenAI's Daybreak has landed AI-authored patches into cURL, Go, CPython, Sigstore, and pyca/cryptography — 70K reviewer-confirmed fixes, 500K auto-detected. No SCA tool flags AI-authored upstream commits today. No SLSA field exists for it. This will be an audit question within two quarters. And the Polymarket homoglyph campaign — lowercase l swapped for uppercase I, paired with paid US-geofenced amplification — is rented infrastructure. Block the brand string and the kit reskins under another name in a week. Detect on kit signatures and drainer wallet clusters instead.

Both are downstream of the same trend: capability is diffusing faster than governance.

What to do this week

One move, not five. Open your top three AI vendor contracts — inference, IDE, agent runtime — and check three things: whether sub-processor changes require notification, whether there is a 90-day flexibility clause you could invoke, and whether Cursor is anywhere in the paper as either a tool or a data path. If any of those three answers is unsatisfying, that is your renegotiation surface for the next 30 days. Take the $150M/month Reflection number into the room as your anchor. The frontier labs are paying premium on 90-day paper. You should not be paying more than that on longer terms.