The $1 Trillion Gap

These three companies are attacking the same problem from different architectural levels. NVIDIA is building dedicated inference hardware (a concession that training GPUs aren't optimal for inference). OpenAI is building purpose-built small models that make cheap, fast API calls good enough for the millions of subagent calls agentic systems generate. Meta is co-designing custom silicon with its own workloads, cutting out the GPU vendor entirely for classification and detection tasks. The combinatorial effect is multiplicative, not additive. Our March 13 prediction of 5-10x inference cost reduction in 12 months now looks conservative. A dedicated hardware improvement stacked on top of algorithmic optimizations and pricing model changes puts 10-20x reduction in range. The inference economy didn't arrive gradually. It arrived in a week.

The agents aren't failing because the technology is bad. They're failing because enterprises are bolting agents onto workflows designed for humans in 2019. Nobody is redesigning the actual work. I've seen this pattern across dozens of deployments: teams take an existing SOP, add an AI step, and wonder why it doesn't scale. The companies that will make agents work are the ones treating them as a new category of worker that needs purpose-built processes, not the ones automating existing procedures. This is the same pattern we saw in early cloud adoption. Lift-and-shift failed; cloud-native won. The 77% failure rate isn't a technology verdict. It's an organizational design verdict. And it sits in direct tension with the $1 trillion hardware bet happening simultaneously.

This isn't regulatory caution. It's a structural problem. AI capabilities are evolving faster than regulators can define categories to regulate. The US bill's carve-outs are the tell: by exempting child safety and state procurement, it creates a two-tier regulatory system that reintroduces the patchwork it claims to eliminate. Builders shipping products that touch both tiers, think education or healthcare, face the same fragmented compliance problem the bill was supposed to solve. The bipartisan opposition from governors (DeSantis and Newsom agreeing on something is a signal) suggests the real fight is federal vs. state power, not left vs. right. For builders, the paradox: regulatory paralysis benefits large companies with legal teams who can navigate ambiguity and punishes smaller teams who need clear rules to plan product roadmaps. Three data points make a pattern, and we're naming this one.

53% sounds modest until you register the qualifier: 100% of these bugs passed through human review. This isn't replacement. It's a genuinely new capability. Two weeks ago, Anthropic found 22 vulnerabilities in Firefox that human reviewers had missed. Now Sashiko catches half the bugs that slip through the kernel's review process. Two different AI systems, two of the most carefully reviewed codebases on earth, same result. AI-as-security-auditor is now two-for-two on major codebases, which makes it a pattern, not an anomaly. The implication for production teams is direct: if AI catches bugs that the Linux kernel's reviewers miss, it is absolutely catching bugs in your codebase. Treat AI review as a parallel track alongside human review. You run both unit tests and integration tests. Same logic applies here.