The week the cloud got hit and the moat got cheap

On April 1, the IRGC put me-south-1 on a target list with a time attached. Then they hit it. AWS is still recovering capacity in Bahrain, with unconfirmed reports of a second strike on an Oracle facility in the UAE. Eighteen US tech companies were named as military targets.

This is the first documented kinetic strike by a state military actor on commercial cloud infrastructure. Your DR plan was built for AZ failures and fiber cuts. It was not built for missiles. The 99.99% SLA has a force majeure clause you have never read, and today is the day to read it.

If you run anything in me-south-1, Azure UAE North, or GCP me-central1, the action this week is concrete: test cross-region failover end-to-end, get legal on the phone about data residency waivers, and ask your account team for their actual physical recovery timeline in writing. Not "review your DR posture." Fail over a workload. Watch what breaks.

Yes, but — cloud regions have been physically threatened before (fires, floods, one memorable backhoe) and the industry absorbed it. The counter doesn't hold here because military targeting is correlated across sites in a theater in a way natural disasters are not, and because the compute market is already tight. AWS just lost a $10M Fortnite contract it couldn't fulfill on capacity. H100 rentals are at an 18-month high. The failover queue when everyone in the region evacuates at once will not be short.

The moat got cheap on the same day

While the physical layer was on fire, the model layer collapsed. H Company's Holo3 — Apache 2.0, downloadable now — scored 78.85% on OSWorld-Verified, beating GPT-5.4 and Opus 4.6 on autonomous computer use at roughly one-tenth the inference cost by activating 10B of 122B parameters. Google's TurboQuant cuts KV cache to 3 bits with zero retraining, 6x memory and 8x attention speedup on H100s. Gemma 4 runs multimodal function-calling on a Raspberry Pi under Apache 2.0. Alibaba's Qwen 3.6-Plus ships drop-in OpenAI and Anthropic API compatibility — a one-line base URL change reroutes your traffic to Chinese infrastructure, and your API gateway inspects format, not destination.

The consequence downstream: 235,800 new apps hit the App Store in Q1, up 84% YoY, reversing an eight-year decline of 48%. Codex went from 100K to 2M developers in three months. Apple pulled the vibe-coding app Anything under Guideline 2.5.2 with no technical workaround available, because dynamic code generation and static pre-review are architecturally incompatible. Salesforce, ServiceNow, and Snowflake each dropped about 30% in Q1 as the market repriced them for agent substitution.

If your product moat is code complexity or a proprietary vendor API you pay a premium for, that moat is gone. Not eroding — gone. A motivated non-technical founder can now ship a functional v1 of most SaaS products in a weekend. The defensible layer moved to proprietary data, distribution control, integration depth, and — the underrated one — output trust.

The trust gap is where the money is

Anthropic's 81,000-person study is the clearest positioning signal I've seen this quarter. Users' top desire from AI isn't time savings (that's fourth). It's professional excellence — 19% ranked it first, meaning "help me produce work I couldn't produce alone." Their top fear is hallucination. The promise and the failure mode are the same axis.

That's your product. The AI feature that closes the gap between "AI can do this" and "AI can be trusted to do this without me babysitting the output" is worth more than another capability demo. Perplexity's Model Council — three frontier models plus a synthesizer that flags divergence — is already a paid tier. At today's inference costs, running three models in parallel is cheaper than running one was six months ago. The divergence signal is where you route to human review. That's a real product, not a feature.

What to do this week

Three things, in order.

Fail over one production workload out of a Middle East region and document what breaks. If you have nothing there, run the drill anyway with a US-East to US-West pair — the muscle memory is the point.

Audit your Claude Code and Copilot permissions in the SDLC. The 512K-line Claude Code source leak revealed deny rules silently stop enforcing after 50 subcommands to save Anthropic's inference tokens, and Copilot demonstrated a working write-path into code reviews by injecting ads before rolling back. If an AI tool can modify your repo without a human gate, that's the fix.

Run a clone-risk audit on your roadmap. Tag every item as defensible (proprietary data, network effects, integration depth, trust) or replicable (a solo dev with Claude Code ships this in a weekend). If more than 40% is replicable, you don't have a prioritization problem. You have a strategy problem, and the market is going to find it before you do.