The Investor

The Problem No One's Modeling

Three independent signals converged this week confirming the same structural break: agentic AI workflows consume 3-5x more tokens than conversational AI and run autonomously for hours — yet most AI SaaS products still charge flat monthly seats. This isn't a pricing choice; it's an unmodeled margin collapse happening in real time across every portfolio company shipping agent features.

The data points are unambiguous. Anthropic's Sonnet 5 shipped with materially higher per-query costs than its predecessor. Tesla began capping employee AI spend — a Fortune 10 company implementing cost governance because the burn is real. And the analysis is explicit: agentic workflows burn multiples more tokens than chat while operating without human session boundaries.

Every AI app charging flat monthly seats while shipping autonomous agents is converting revenue into negative-margin compute right now. The gross-margin surprise hits in Q3/Q4.

Where Value Migrates

Microsoft's commitment of 6,000 engineers to 'Frontier Company' for enterprise AI integration declares the thesis: services and workflow integration — not model access — is the enterprise revenue battleground. Meanwhile, DeepSeek's DSpark cut inference latency ~85% and Mistral's Leanstral 1.5 grinds API margins further down. The model layer is commoditizing from both ends — cheaper inputs and higher consumption — while the integration layer captures the spread.

The emerging category with best risk/reward is Spec-Driven Development (SDD): AWS launched Kiro, GitHub shipped Spec Kit, and startup Tessl is positioning independently. When both hyperscaler incumbents enter a category in the same cycle, TAM is validated — but the independent window narrows fast.

The Circular-Financing Tell

Adding urgency: J.P. Morgan flagged red flags across the AI market the same week Meta began renting 'excess' compute and Nvidia continued financially backing young cloud providers. When the largest hardware beneficiary funds its own buyers, the market is pricing supply ahead of durable demand. This means the agentic compute surge could collide with overcapacity, creating a squeeze on anyone caught between rising COGS and flat pricing while their cloud provider dumps excess inventory into the spot market.

What To Do Now

The defensive move is immediate and mechanical: run a COGS-per-account stress test across every portfolio AI app shipping agentic features. Model 3-5x token consumption against current flat pricing. Identify which companies flip negative-margin at scale before they surface as bridge-round requests. The offensive move is repositioning toward the integration and verification layers where Microsoft just signaled the next $10B+ market will be built.

The Framework

Martin Ford — who called automation-driven job loss in 2009 when economists dismissed him — has produced the most portable deal-screening tool of this cycle, which is a low bar and he clears it anyway: score every AI target on (a) volume of clean historical training data and (b) cost to check output against ground truth. It is not a blue-collar/white-collar split. A financial analyst doing routine quant work is more exposed than a plumber. A coding agent is worth funding; a burger-flipping robot dies on edge-case ROI.

Screen AI deals on how cheaply the output can be checked — fund the ones creating markets below the old price floor, and discount any ARR that assumes adoption before the org chart is redrawn.

The Electrification-Lag Model

The contrarian bit is the one worth sitting with: economists studying the LLM rollout find no measurable disemployment effect yet. Ford maps this to factory electrification, where plants got nothing from electric motors until they physically redesigned around distributed power. The bottleneck is organizational, not technical. Adoption is bottom-up — employees use the tools and quietly pocket the slack, Ford's '2011 of smartphones' equivalent — until management formally folds three roles into one or two.

That has direct underwriting consequences. Any portfolio company assuming instant enterprise adoption for labor-replacement ARR is mismodeling the timeline. The restructuring wave arrives, but it is org-chart-gated, not capability-gated, and those are very different clocks. The Turing Post analysis says the same thing from another direction: 'no AI-native enterprise exists yet' because the obstacle is hidden workflows, politics, and institutional habit, not the model.

Where the Alpha Sits: Market Creation vs. Share Shift

The cleanest use of the framework is separating share-shift plays that attack incumbents from market-creation plays that serve demand below the old price floor. Self-service contract review is the proof case. Nobody paid $500 for a lawyer to read a routine agreement, so near-zero marginal cost opens TAM that never existed. That is growth, not the kind of cannibalization that invites incumbents to fight back.

One gate does the real work here: liability architecture — not technical capability — is the ceiling in regulated professions. There is no framework for a model that makes the same systematic error across thousands of cases at once. Radiology and law are technically feasible and structurally blocked, which is the whole point. This is probably where most people will lose money — underwriting those TAMs before the legal infrastructure catches up.

Scoring Matrix for Deal Screening

The Demand Signal

This isn't a breach report — it's a capital allocation signal. North Korea's PolinRider campaign shipped 108 malicious packages across four ecosystems (npm, Packagist, Go, Chrome) in a single coordinated operation, plus a separate fake-Rollup-polyfill campaign surfaced by JFrog. This is industrialized supply-chain warfare at a scale that converts developer dependency management from best-practice hygiene into board-mandated spend.

Simultaneously, Avalon's modular framework with CrownX ransomware demonstrated multi-stage phishing that bypasses traditional EDR and email gateways — the same commoditization pattern that killed signature-based antivirus a decade ago is now repricing legacy endpoint detection. And Armored Likho's BusySnake Stealer hit power grids in Russia, Brazil, and Kazakhstan, broadening OT/ICS targeting beyond the Western-target axis investors usually monitor.

Attacker sophistication is the security sector's demand engine — and this week it pointed capital at supply-chain, OT, and behavioral detection while quietly compressing legacy EDR.

The Sub-Sector Map

Where This Gets Investable

The key insight is timing relative to regulatory catalysts. Supply-chain security (SCA/SBOM) is already monetizing — JFrog is the public comp proving enterprise willingness to pay. OT/ICS is pre-mandate: Armored Likho's power-sector targeting is the kind of incident that precedes energy-security regulations by 12-24 months. The entry window on OT/ICS is open now and closes when regulation arrives.

FatFs — bundled in millions of embedded devices — disclosed seven vulnerabilities with no clean remediation path. Combined with Bad Epoll (CVE-2026-46242) giving unprivileged-to-root on all Linux/Android, the IoT/embedded device attestation category faces persistent structural demand because the problem literally cannot be patched away at scale.

De-risk watch: audit any portfolio exposure to legacy signature-based EDR or email-gateway incumbents. Every 'bypasses traditional controls' disclosure is a margin-compression event for those vendors. The repricing pattern from AV → next-gen EDR is now playing out at the next layer.