The Product Desk

The Largest-Scale Goodhart's Law Failure in Tech History

New reporting from The Pragmatic Engineer reveals what may be the most expensive measurement failure in modern tech: Meta's 85,000 employees burned 60.2 trillion tokens in 30 days at an estimated cost exceeding $100M — and insiders describe much of the output as 'throwaway, wasteful.' This isn't an isolated cultural quirk. Microsoft has operated token leaderboards since January 2026, with VP-level executives appearing in the top 20 despite rarely writing code. Salesforce sets minimum weekly spend targets ($100/week on Claude Code, $70/week on Cursor) and flags engineers who fall short.

Token usage metrics are the new 'lines of code' — a vanity metric being gamed at scale, inflating the demand signals that AI vendors use to justify pricing and capacity allocation.

The damage isn't just financial. Meta engineers report that careless AI code generation has already caused production SEVs — actual customer-facing incidents from developers who prioritized token volume over code correctness. Microsoft engineers explicitly admit to asking AI questions already answered in documentation and prototyping features they'll never ship. Newer, more junior Microsoft engineers tokenmax not to climb leaderboards but to avoid being seen as using too few tokens — AI usage has become a proxy for job security.

Why This Matters for Your Product Decisions

If your organization tracks AI adoption rates, AI-assisted code volume, or developer tool utilization, those numbers are almost certainly inflated. Your '40% of code is AI-generated' metric in exec decks is measuring compliance theater, not productivity. Worse, when GitHub Copilot and Anthropic ration individual users because business demand has '10x'd in recent months,' how much of that 10x is tokenmaxxing waste? This contaminates your procurement negotiations and capacity planning.

One long-tenured Meta engineer suspects the token leaderboard was intentionally designed to generate real-world coding traces for training Meta's next-generation coding model — making employees unwitting data labelers. If true, Meta's $100M+/month in token 'waste' is actually a data generation investment. Check your vendor agreements.

Shopify's Governance Model: The Playbook to Steal

Shopify under Farhan Thawar stands out as the only mature governance approach in the industry. Three specific moves worth copying:

1. Renamed 'token leaderboard' to 'usage dashboard' — a subtle but important reframing that discourages competitive waste
2. Implemented circuit breakers that auto-pause access when individual spend spikes anomalously, catching both runaway agents and infrastructure bugs
3. Discovered that per-token cost (not total volume) is the real quality signal — developers generating expensive tokens were doing deep, complex work; developers generating cheap tokens at volume were generating noise

The inversion of the obvious metric — cost per token rather than total tokens — is exactly the kind of insight that separates useful AI governance from bureaucratic overhead.

The Distribution Channel That Makes SEO Look Quaint

A cluster of data points this week confirms that AI agents are no longer a theoretical future user persona — they're your fastest-growing customer segment right now, and you probably can't see them in your analytics. Vercel's CTO revealed that 60% of their admin app traffic is bots, not humans. Claude recommends Resend as the default email provider ~70% of the time. Supabase is the default Postgres recommendation. When coding agents 'break containment' into non-coding workflows, this agent-driven selection will expand into every software category.

The compounding dynamic is brutal: products already in training data get recommended → generate more usage data → get reinforced in future training. If you're not in this flywheel, you're facing a moat that gets deeper every month.

The AirOps study of 353,799 pages quantifies how this works in practice: headlines that directly answer a query get cited 41% of the time vs. 29% for loosely related ones. Domain authority doesn't predict AI citation — content quality does. Pages covering 26-50% of sub-queries outperform pages covering 100%. Your content strategy needs restructuring for focused authority over exhaustive coverage.

Non-Coders Are Building Your Competitors

Mario Gabriele, a writer with zero coding background, used Claude Code + Opus 4.5 to build a 45,000-chunk hybrid search system with Voyage-3 vector embeddings, SQLite FTS5 keyword search, and a locally fine-tuned cross-encoder reranker — trained on ~40K query-passage pairs. He went from opening a terminal 'by accident' to spending 70%+ of working hours there within three months. The system replaced Obsidian, Ulysses, Google Drive, Dropbox, and Readwise — which became data sources, not interfaces.

The distillation displacement pattern is the most alarming vector: Gabriele used Cohere's reranker as a teacher model, distilled it into a local equivalent, and replaced the paid service. If your AI/ML API is good enough to be a teacher but not cheap/fast enough to stay in production, you're a bootstrapping tool, not a platform.

MCP Is the Composability Layer That Enables All of This

A CTO case study details replacing ~1 hour of daily multi-tool ops reporting with a 30-second single command composing data from Linear, BetterStack, and an incidents database via MCP. The tools aren't being replaced — they're being recomposed. Products whose value is 'we aggregate data into a dashboard' face existential risk when a terminal command delivers a richer, more personalized briefing. The tools that become composable infrastructure win. The tools that sit on top as read-only dashboards lose.

Meta's mandatory keystroke-tracking program adds urgency: they've identified that AI models 'still lack some of the basic ways that humans use computers like choosing from dropdowns and keyboard shortcuts' — and are collecting training data to close this gap. CTO Bosworth's endgame: 'our agents primarily do the work and our role is to direct, review and help them improve.'

The Margin Compression That Breaks Your Financial Model

A convergence of signals this week makes the AI cost reckoning unavoidable. SaaS gross margins are compressing from the traditional 70-80% range toward ~52% as AI integration costs spike. At 52%, your Rule of 40 math breaks, your valuation multiples compress, and your venture model needs rework. For PMs, every AI feature you spec without a cost model is a bet against your company's margin profile.

The infrastructure supply picture makes this worse, not better: only 13.3% of the 114GW of promised AI data center capacity is actually under construction (15.2GW breaking ground). The supply-demand imbalance is structural, not temporary. Microsoft and Anthropic are already responding with token-based billing and tighter rate limits.

Microsoft's Token Billing Is the New Pricing Template

Microsoft's June 2026 shift to token-based Copilot billing deserves close study. The structure: $19/user/month buys $30 of pooled credits at Business tier, $39/user/month buys $70 at Enterprise. That ~58% credit-to-price ratio is designed so median users stay within allocation (good perceived value) while heavy users pay overages (real margin). This is the AWS playbook applied to AI features.

If you're a PM pricing AI capabilities, study this structure: subscription + credit pool + overages solves the 'power users cost me money' problem while maintaining accessibility. Expect this to become the standard AI pricing template by year-end.

Meanwhile, Anthropic's Claude uptime of 98.79%–99.25% over 90 days falls well below enterprise-grade 99.9%+ expectations — that's 53–87 hours of annual downtime. If your product has AI in the critical path, you're inheriting this reliability profile. Design features that degrade gracefully: intelligent when AI is available, merely functional when it's not.

The Counter-Signal: Costs Are Also Falling

There's a real tension in the data. Google's new inference TPUs deliver 80% better performance per dollar. Cohere's W4A8 quantization delivers 58% faster time-to-first-token on Hopper GPUs. And the competitive pressure from Google, AMD, and custom silicon (SpaceX manufacturing its own GPUs) means Nvidia's pricing power is eroding. Features you killed for unit economics reasons in Q1 deserve a second look — but model costs at 2-3x current pricing before committing, because the subsidy removal may outpace the hardware cost decline in the near term.

Commodity Models Match Frontier on Vulnerability Discovery

The most consequential security finding this week isn't another breach — it's the complete debunking of AI security exceptionalism. Multiple independent teams replicated Anthropic's flagship Mythos vulnerability findings at 100-800x lower cost using commodity models. AISLE's nano-analyzer — using gpt-5.4-nano at $0.20/M tokens with models as small as 3.6B parameters — detected Mythos's flagship FreeBSD RCE 2 out of 3 times. They scanned the entire FreeBSD kernel (35K files, 7.5M lines) in 10 hours for under $100.

Semgrep's Kurt Boberg ran the most rigorous benchmark across five models and concluded that the 'hotspot interrogator' architecture — deterministic pre-filtering paired with targeted LLM analysis — consistently outperforms naive whole-file prompting regardless of model. Meanwhile, Mythos's claimed 72.4% Firefox exploit success rate drops to 4.4% when two previously-known bugs are removed, and the test used already-patched Firefox with sandboxing stripped out.

The moat in AI-powered code analysis is not model access — it's the three-stage pipeline: context generation → vulnerability scanning → skeptical triage. The scaffold is the product.

MCP Is Shipping With Systemic Vulnerabilities

Three independent MCP implementations disclosed critical RCE vulnerabilities simultaneously: Flowise (CVSS 9.9), Upsonic (CVSS 9.8), and OpenAI Codex CLI (CVSS 9.8). OX Security's research explicitly calls out 'systemic MCP supply chain vulnerabilities.' This isn't one vendor's mistake — it's a protocol ecosystem shipping without adequate security patterns.

Add the Axios CVSS 10.0 (unrestricted cloud metadata exfiltration via the most ubiquitous JS HTTP client) and Apache Kafka's JWT bypass (CVSS 9.1, accepts ANY JWT token without validation), and the infrastructure layer is cracking at multiple points simultaneously. CVE volume grew 38% YoY (29K to 40K+), while only 5-7% are actually exploited — making EPSS-based triage the highest-leverage process change available.

RSAC 2026: Agent Governance Is a Confirmed Greenfield Market

Eleven main-stage RSAC 2026 keynotes agreed on exactly what AI agents need for security: asset management, user-patterned data permissions, observability, output validation, and integrity checks. And then: no speaker claimed a working solution. An AI governance startup founder confirmed all customers remain in monitor-only mode with zero enforcement. Cisco released 5 open-source agent-defense tools (AI BOM, MCP Scanner, A2A Scanner, CodeGuard, DefenseClaw) — clearly designed to become the default standard.

Claude Opus 4.7 was used to create a universal jailbreak against itself in under 20 minutes. If your product relies on model-level safety guardrails as your primary safety layer, you now have evidence that the model itself can systematically defeat them. You need application-layer output validation.