A trillion dollars of AI cloud spend is walking in a circle

Anthropic committed $200B to Google Cloud this week — more than 40% of GCP's disclosed backlog — while Google put up to $40B of equity back into Anthropic. Add OpenAI's $688B across Microsoft, Oracle, and Amazon, and the two labs have now booked $1.018T of cloud spend with the same four companies that have written $88B+ of equity into them. That's roughly half of the $2T+ cloud backlog the hyperscalers are reporting. The sellers handed the buyers the money, and the buyers are now booking it as revenue growth.

Dario Amodei's defense of the math: "The other player is pretty confident they'll have the revenue at the right time." That sentence is load-bearing for a trillion dollars.

Yes, but — vendor financing is an old complaint and it has been wrong before. Cisco 2000 is the reflex comparison and it's imperfect: unlike telecom equipment, inference capacity gets consumed even when the buyer is losing money, and Anthropic's finance vertical and OpenAI's ad business are producing actual revenue on a curve most SaaS companies would kill for. The loop can keep spinning for a long time if end demand shows up. The reason I still think it matters this quarter is that the demand assumption has a specific dependency, and the dependency isn't holding.

The reliability curve that isn't moving

Kapoor, Rabanser, and Narayanan benchmarked fourteen frontier models over eighteen months and found the same thing everyone shipping to production already knew: capability rose sharply while reliability barely moved. GPT-5.5 Instant jumps +15.8 points on AIME 2025. Opus-4.5 with web search still produces roughly 30% ungrounded claims in multi-turn conversations. Only about 15% of enterprises have the data foundation to run agentic AI in production. The other 85% are funding pilots that will die on contact with their own data estates.

The capacity commitments assume an adoption curve, the adoption curve assumes a reliability curve, and the reliability curve is the flattest line in the room. Oracle is the cleanest instrument to watch — its backlog is growing 438% year over year, almost entirely on OpenAI consumption. When its realization rate against reported backlog starts to slip, the read is that consumption is lagging commitments, and that dynamic will be playing out at Azure, GCP, and AWS with more accounting cover.

The two labs picked opposite futures

OpenAI is becoming an ad and hardware company. $100M in ads ARR six weeks into a beta, a $100B run-rate target by 2030 against 2.75B weekly users, and a 30M-unit AI phone with dual-NPU silicon in 2027-2028. Anthropic is becoming Bloomberg with a compliance department — a $1.5B services JV with Goldman, Blackstone, and H&F, an explicit ad-free posture, and finance already disclosed as its number-two revenue segment. Both stood up services arms this week that will bid directly against their customers' internal AI teams and against the systems integrators those customers already pay. Brad Lightcap moved from OpenAI COO to run the $4B Deployment Company, which is how you signal a side bet is actually the central one.

If your product is a model wrapper, the iOS 27 selector screen this fall — one billion devices, one choice, never revisited — is where your distribution ends. If your product is a workflow, the vendor you signed with last year is deciding whether to compete with you, and the answer depends on which vendor.

And meanwhile, the operational fires

CVE-2026-0300 is a PAN-OS buffer overflow, actively exploited, no patch until mid-to-late May. The management plane is the entry point. DAEMON Tools installers have been shipping a legitimately-signed China-nexus backdoor since April 8, with a QUIC RAT dropped selectively to a dozen high-value targets — the same CCleaner-to-Notepad++-to-DAEMON-Tools playbook, third iteration in nine years, same publisher-allowlisting control failing every time. Anthropic's MCP STDIO transport carries an architectural RCE that inherits into every downstream framework — 150M+ downloads, 10+ CVEs, one root cause. North Korean APTs are registering the package names your coding agent hallucinates. CISA is floating a compression of the critical patch SLA from fourteen days to three.

Multi-token prediction shipped production-ready across vLLM, SGLang, llama.cpp, and Ollama with day-zero support this week. A 78M-parameter draft head against a 27B+ target reports ~75% acceptance and headline 2-3× throughput. The honest number under real batch sizes and loaded servers is 1.3-1.5×. Still the cheapest inference win available this sprint, provided you measure acceptance per traffic slice rather than in aggregate.

What to do this week

Pull PAN-OS management interfaces off the public internet today and hunt DAEMON Tools installer hashes dated after April 8 across your endpoint estate. Then do the harder thing: build a counterparty stress test that assumes a 30% token-price increase from your primary lab and a 40% cut in hyperscaler AI capex. Both labs need to service their cloud commitments, and the cleanest lever is API pricing, which lands on your gross margin. Whichever numbers your infrastructure roadmap is quoting for Q3, run them again against Oracle's next realization print. If the model doesn't survive it, the model is what needs to change, not the print.