The Board Room

Deep dives

1. 01

   The AI Feature Sprawl Death Certificate — And the Margin Math That Killed It

   Microsoft Proved the Negative

   Microsoft's Copilot rationalization is the most instructive strategic signal in enterprise AI this quarter. A company with unlimited frontier-model access, 400 million Office users, and effectively unlimited capital concluded that broad AI distribution destroys value. Customer feedback produced the phrase 'functionally useless.' The earnings call confirmed that inference costs drag margins. Eighty-one distinct products were in flight. Nadella's sequence was consolidation under one executive (Andreou), then killing everything that failed both a customer-value test and a unit-economics test.

   If breadth-first AI feature sprawl does not work inside the largest software distribution on earth, the question of whether it works inside a smaller one answers itself.

   The Structural Margin Gap

   A reasonable skeptic would call this a maturity problem that scale resolves. The skeptic is wrong, and the numbers say so. BVP puts AI company gross margins at 50-60% against 80-90% for traditional SaaS. Reasoning models consume 10-100x the tokens of the prior generation for the same user-visible answer. OpenAI's 1,000x cost reduction over 14 months was eaten by its own model advances. Per-token cost falls. Tokens consumed per task rise faster. A company shipping AI on every surface is running thirty margin-negative line items to fund the one that pays for itself.

   The Opposite Bet Is Working

   Anthropic's shift to per-result pricing, charging for outcomes rather than tokens, is the structural alternative. It works because the agents complete the work, with a 90% autonomy target for Claude Code, which makes the vendor's eat-the-cost risk acceptable. Focused 365 Copilot, the part Microsoft is keeping, grew paying users 33%. Narrow, high-value surfaces where customers pay on purpose outperform broad feature spray by every measure that matters in year two.

   ---

   Portfolio Implications

   Every AI feature shipped on inference without corresponding willingness-to-pay is a standing cost against non-existent revenue. The audit is straightforward. Map every AI-powered feature to customer-perceived value and to inference cost. Anything that fails both tests is a margin leak that compounds with scale. Microsoft absorbed it for 18 months. Most organizations cannot absorb it for one quarter.

   The era of competing everywhere with undifferentiated AI is over. Microsoft proved it does not work with infinite resources, which is useful to know before spending finite ones.

   What to do

   1. →

      Audit every AI feature against customer willingness-to-pay AND unit economics — kill or pause anything failing both tests by end of Q2

      NowMicrosoft absorbed margin drag for 18 months before cutting. Most companies have one quarter before it shows on the P&L.
   2. →

      Model per-result vs. per-token pricing for your top 5 AI use cases within 30 days

      This sprintAnthropic's outcome-based pricing is attracting enterprise customers who reject per-seat or per-token models. Be ready to offer the alternative.
   3. →

      Centralize AI product ownership under a single executive this quarter

      This sprintMicrosoft's consolidation under Andreou preceded the rationalization. When six VPs each own AI variants, none has incentive to kill their own feature.
   4. →

      Establish inference cost budgets per feature — treated with the same seriousness as latency budgets

      This quarterThe 30-point margin gap is structural. Without explicit cost discipline, your fastest-growing AI features may be your least profitable.
2. 02

   Your Codebase Is $30 From Full Exposure — The Offense-Defense Gap Is Now Unfixable at Human Speed

   The Numbers That Changed

   Three data points landed this week, and taken together they rewrite the economics of defense rather than nudging them:

   • $30-$150 per codebase for AI-powered zero-day discovery (IronCurtain framework, open-weight models)
   • 95% success rate for autonomous red team agents achieving domain dominance in under 6 minutes (Dreadnode Ares benchmark)
   • 271 vulnerabilities found in Firefox, a mature and well-maintained project, in a single Mythos pass

   The defender's side of the ledger reads differently. 55-day average remediation against 135 new CVEs per day leaves a running deficit of roughly 7,400 unpatched vulnerabilities at any given moment. Human-speed processes do not close that arithmetic.

   When the attacker's economics and the attacker's clock are both on the same side of the ledger, and the defender's procurement cycle is still measured in quarters, the gap is structural, not operational.

   The Model-Agnostic Problem

   A reasonable skeptic would argue that restricting frontier models still buys time. The reasonable skeptic is already behind the evidence. Niels Provos has shown that older, widely available models with expert orchestration replicate frontier findings, and the UK AI Security Institute found GPT-5.5 (broadly available) may outperform Mythos (restricted) on cybersecurity tasks. Any security strategy that relies on attacker capability being gated by model access is pricing against a world that ended this quarter.

   What Changed in Cloud Detection

   The attack surface has migrated from exploits to abuse of legitimate APIs. The Mini Shai-Hulud worm crossed npm, PyPI, and Packagist in 48 hours, compromising official SAP and PyTorch Lightning packages. AWS Bedrock AgentCore's S3 access creates bidirectional C2 channels that AWS calls intended behavior. Traditional monitoring architectures are structurally blind to this class of attack, and language diversity in the stack no longer delivers security heterogeneity.

   ---

   The Strategic Response

   Google's VRP restructuring is the leading indicator worth watching. Rewards are dropping for bugs AI can find (commoditized) and rising for hardware exploitation and complex chains (up to $1.5M for Titan M2). The value is migrating from discovery to exploitation prevention and runtime protection. Security spend has to move off incremental improvement of legacy detection and toward AI-native defense that operates at machine speed, because the alternative is a budget defending last quarter's perimeter against this quarter's economics.

   Layer	Old Assumption	New Reality
   Discovery	Scarce, expensive	$30, commodity
   Exploitation	Requires expertise	95% automated
   Remediation	Manageable queue	55-day structural gap
   Detection	Network monitoring	Blind to API abuse

   What to do

   1. →

      Deploy AI-powered offensive testing against your own infrastructure at the $30-150/codebase price point using IronCurtain or equivalent — this week, not next quarter

      NowIf a consultancy's tool finds these bugs at $30, the adversary already has. Find them first.
   2. →

      Conduct a 'patch wave readiness' assessment: model whether your engineering org can handle 5-10x current CVE volume with 3-day response windows

      This sprintCISA is considering 3-day patch deadlines. The 55-day average means most organizations fail this test today.
   3. →

      Commission an AI agent infrastructure audit — inventory all exposed AI endpoints (MCP servers, Ollama instances, agent sandboxes) and implement VPC isolation

      This sprintBishop Fox's AIMap can fingerprint exposed AI infrastructure via Shodan. If the tool exists for consultants, it exists for attackers.
   4. →

      Present autonomous security investment thesis to the board — frame as operating model change, not tooling purchase

      This quarterBuying autonomous tools gated behind the same approval queue that produces the 55-day gap changes nothing. The operating model must change.
3. 03

   GitHub's 85% Uptime Is a Preview — Your Products Face the Same 30x Agent Load

   GitHub's 85% Uptime and the Agent Load Miss

   GitHub's measured uptime fell to 85%, which works out to two or three hours of downtime a day. CTO Vlad Fedorov attributed the shortfall to AI agent load being 'much bigger than expected' and disclosed that GitHub revised its scaling target from 10x to 30x in four months. A single developer running Claude Code or Codex now generates the load previously associated with 10-50 developers. Agents run flat around the clock, parallelize aggressively, ignore rate-limit etiquette, and convert failed requests into more requests rather than backing off.

   This Is a GitHub Failure, Not an Industry Condition

   GitLab, Bitbucket, Vercel, Linear, Railway, and Sentry are absorbing equivalent AI-driven growth without catastrophic reliability failures. Google's SRE teams were preparing for 10x code increases in July 2025, a full year before GitHub's crisis became public. The gap between Google's anticipation and GitHub's surprise is the cleanest available read on which organizations internalized agent-era scaling and which are reacting to it.

   Mitchell Hashimoto, HashiCorp founder and 18-year GitHub user, publicly declared it 'unfit for professional work.' That is less a single defection than a permission structure for enterprise procurement conversations that were previously unthinkable.

   The Mirror Problem: Portfolio Products Under Agent Load

   The harder question is whether products shipped from the same portfolio are architected for the equivalent 30x load explosion. The customers running agents against GitHub are the customers running agents against every other API they touch. A platform that passes the GitHub dependency audit and fails the 'can our own product handle agent traffic' audit has simply moved the outage from a vendor's status page to its own.

   ---

   Agent Traffic Is a Different Traffic Class

   • No diurnal pattern — load is flat 24/7
   • Aggressive retry behavior — failed requests multiply rather than queue
   • High parallelization — one user generates 10-50x historical load
   • No rate-limit politeness — agents hit limits then spawn parallel paths

   Every platform sitting in the AI-agent consumption chain was built against human usage patterns. Most will meet the new profile the same way GitHub did: by degrading first and re-architecting second. The strategic question is which products get re-architected before the load arrives and which get re-architected after. The two paths do not cost the same, and they do not look the same to customers twelve months from now.

   The security overlay compounds the problem. A critical vulnerability allowing full repository access via git push overlapped with the reliability failures, creating compound risk that should trigger immediate enterprise security review for any on-premises GitHub Enterprise deployment.

   What to do

   1. →

      Audit GitHub dependency across your engineering organization — map which workflows block when GitHub degrades and quantify productivity loss at 85% uptime

      NowTwo-three hours of daily downtime is the current ceiling whether your operating plan acknowledges it or not.
   2. →

      Stress-test your own product architecture against 30x current load with AI agent traffic patterns (high-frequency, programmatic, burst-heavy) by end of Q2

      This sprintYour customers are already running AI agents against your APIs. The load curve that broke GitHub is coming to every developer-facing platform.
   3. →

      Evaluate multi-vendor source control strategy or self-hosted alternatives (GitLab, Forgejo) for critical repositories and CI/CD pipelines

      This quarterA genuine switching window has opened due to incumbent failure. These close faster than they open.
   4. →

      Watch GitLab (GTLB) and developer infrastructure startups for acquisition or partnership as GitHub's crisis creates rare platform-switching opportunity

      WatchHashimoto's departure is a leading indicator of enterprise sentiment shift. The competitive window is 12-18 months.