Amazon's AI code broke AWS for 13 hours. Nvidia paid $20B to admit GPUs aren't enough.

Amazon confirmed this week that AI-generated code from its Kiro tool caused a six-hour retail outage and a thirteen-hour AWS disruption. The response was the first formal governance retreat from a hyperscaler on AI-assisted development: mandatory senior sign-off on all AI-generated changes from junior and mid-level engineers. Internally, they called the incidents "high blast radius." That is the operative phrase for the rest of this year.

In the same news cycle, Lovable crossed $400M ARR with 146 employees — $2.74M per head, roughly 10x a well-run traditional SaaS company at comparable scale. Replit tripled to $9B on Fortune 500 "vibe coding" adoption. Anthropic hit $19B annualized revenue and started pricing Claude Code Review at $15–$25 per PR, explicitly targeting the review bottleneck their own industry created.

Those two facts are not in tension. They are the same fact.

The benchmark-to-production gap is now measured

Claude Opus 4.5 scores 92% on Stripe's eleven-task integration benchmark. METR ran 296 real pull requests past actual maintainers of scikit-learn, Sphinx, and pytest, and roughly half of the SWE-bench-passing patches got rejected. So you have a coding agent that passes automated checks at 92% and gets accepted by senior humans at closer to 50%. Amazon lived that delta for thirteen hours. If you are procuring coding agents against public leaderboards, you are shopping on a metric that predicts about as much about production safety as "does it compile."

The New York Times ran the same technology through different guardrails and got the inverse outcome: test coverage from 28% to 83% across six web projects, with an estimated 70% effort reduction. Three constraints made the difference — AI generates tests only, source code is read-only, every artifact gets human review before merge. The worst case is a failed CI run. Not thirteen hours of AWS.

Yes, but — the counter-read is that Amazon's incident is a one-off from a specific tool at a specific maturity level, and hyperscalers ship risky changes all the time. Fair. Except METR's rejection data, McKinsey's Lilli platform getting rooted through a textbook SQL injection this same week, and Amazon's own decision to formalize sign-off gates all point at the same underlying pattern: AI code passes syntax and unit tests, then fails on the implicit invariants senior engineers carry as mental models. The tool improves; the gap doesn't close on its own.

The hardware layer just split

Nvidia licensed Groq's LPU for roughly $20B and is shipping racks of 256 inference-specialized chips, with Intel processors bridging communication because NVLink doesn't integrate with the LPU yet. This is the first time Nvidia has put another company's silicon inside its own server architecture. The first named buyer is OpenAI, specifically for coding agents. AWS separately partnered with Cerebras on cloud inference. Samsung, not TSMC, will manufacture first-generation LPUs.

When the dominant AI hardware vendor pays $20B to acknowledge that GPUs alone can't serve the inference load, the training-era assumptions in your infrastructure contracts are already stale. Racks ship H2 2026. The GPU-plus-LPU fused die (Feynman generation) is 2027-2028. That window — roughly eighteen months — is when your inference cost per token starts moving in ways your current serving stack was not designed to exploit.

Most inference frameworks — vLLM, TensorRT-LLM, Triton — assume GPU cluster topology. The Intel bridge chip in the Groq rack is a red flag that heterogeneous compute is going to fragment the software layer before it consolidates. If your model serving code has hard CUDA dependencies, you are accumulating lock-in against a market that is actively pulling apart.

What to do this week

One thing, specifically. Classify every AI coding pathway in your org by blast radius, and put the classification in writing before your next sprint planning.

High blast radius — anything touching auth, payments, data pipelines, feature stores, model serving, infra config, or migration scripts — requires mandatory senior review and staged rollout. That is Amazon's lesson, borrowed for free. Low blast radius — test generation, documentation, boilerplate, internal tooling — ships aggressively with the NYT constraints: read-only source, output confined to non-production files, human review on merge. Everything in between gets file-permission scoping and blast-radius CI gates that count files changed and services affected.

While you are in the pipeline anyway, grep every GitHub Actions workflow in your org for pull_request_target. Forty-eight repos including Trivy — a container security scanner running with elevated privileges inside probably your CI right now — were compromised through that trigger this week. If any of your workflows check out github.event.pull_request.head.sha in a pull_request_target context, any external contributor can execute code with your secrets and write permissions. Fix or replace with pull_request plus workflow_run. Verify Trivy binary signatures against Aqua's published values before you ship anything else.

The teams that win the next eighteen months are the ones that segment AI coding by blast radius and abstract their inference layer for hardware heterogeneity. The teams that don't will either learn Amazon's lesson at their own scale, or pay a premium to run the wrong workload on the wrong silicon. Both mistakes are avoidable this week.