The GitHub Infrastructure Crisis Is the Best Evidence We Have — And It's Still Not Enough to Resolve This Forecast

Let's start with the resolution problem, because it matters before we get to the evidence. The forecast as currently defined — 'autonomous agents widely deployed in enterprise workflows' — cannot be independently resolved by a reader. We're fixing that. The updated resolution criteria for this forecast: at least 100 Global 2000 companies deploying autonomous agents (defined as systems capable of multi-step task completion without human approval at each step) across at least three distinct workflow categories beyond software development, with deployment confirmed through earnings disclosures, vendor case studies, or independent analyst audits. That's the bar. The GitHub data doesn't clear it on its own. It matters as a signal, not a verdict.

So what does the GitHub data actually prove? AI-opened pull requests went from 4 million in September 2025 to 17 million by March 2026 — a roughly 4x increase in six months. Microsoft's infrastructure buckled under the load. This is direct evidence of one thing: autonomous coding agents are operating at production scale inside the GitHub ecosystem, which means inside the software development workflows of thousands of companies. That's meaningful. But it's developer tooling, the earliest and most permissive adoption category in enterprise AI. Software engineers are the population most likely to adopt AI agents first, most likely to have IT clearance to do so, and least constrained by the regulatory and liability structures that govern finance, healthcare, HR, and legal workflows. The GitHub number tells us the leading edge is real. It doesn't tell us whether the middle of the adoption curve has moved.

The capex signals — Goldman's $7.6 trillion projection, Oracle's $638 billion committed backlog, Anthropic's $1.25 billion monthly Colossus spend — are proximate evidence, not direct proof. We said so in our last analysis and then partially contradicted ourselves by suggesting the spend confirms production workloads are arriving. To be direct: infrastructure commitments at hyperscaler scale are driven by anticipated demand and competitive positioning, not confirmed arrival. Google, Microsoft, and Amazon have historically over-committed infrastructure. The more honest signal would be inference revenue broken out by workload type — Azure AI Services ARR, AWS Bedrock consumption growth, Google Cloud AI API call volumes. We don't have those numbers at the granularity we need. What we have is supply-side commitment, which is consistent with production deployment but does not prove it.

The counterargument we haven't fully reckoned with: Gartner and similar analyst data consistently show 50–80% of enterprise AI pilots failing to reach production. The GitHub evidence is atypical — Microsoft-internal, developer-native, single-workflow. The baseline failure rate for enterprise AI deployments outside dev tooling remains the most important structural counter to this thesis. We weight against it because the GitHub signal shows that at least one workflow category has genuinely crossed the pilot-to-production threshold at massive scale, and because the capex environment suggests companies are betting real money on production value materializing. But we're not dismissing it. The 77% reflects: strong direct evidence in dev tooling, proximate evidence of broad enterprise commitment, offset by an unresolved base-rate problem in non-developer verticals. What would drop us below 65%: if Q3 enterprise software earnings calls show AI agent project cancellations or pauses at rates consistent with Gartner's failure projections. What would push us above 85%: direct disclosure from three or more major banks, insurers, or healthcare systems confirming autonomous agent deployment in operations workflows with measurable headcount impact.