Oil Shock and World Models Reset Growth-Equity Pipelines

◆ DEEP DIVES

1. 01

   Stagflation Trap Closes — Every Deal Model in Your Pipeline Is Stale

   <h3>The Macro Setup</h3><p>Yesterday's Fed decision was technically a non-event — rates held at <strong>3.5-3.75%</strong>. The context around it is a regime shift. Three forces converged simultaneously that haven't been present since early 2022:</p><ul><li><strong>Energy shock</strong>: Iran declared Gulf energy infrastructure "legitimate and prime targets" and struck a Qatar fuel hub. Oil spiked above <strong>$111/barrel</strong>. The Strait of Hormuz carries 20% of global oil supply.</li><li><strong>Inflation reacceleration</strong>: Wholesale prices rose more than <strong>2x faster than expected</strong>. The Fed projects 2.7% year-end inflation — and energy pass-through hasn't hit CPI yet.</li><li><strong>Rate persistence</strong>: The dot plot held at <strong>one cut for 2026</strong>. Markets had priced two. The gap between hope and reality just widened.</li></ul><p>The administration's response signals severity: a <strong>Jones Act waiver</strong> (a rarely-used emergency lever) and VP Vance personally convening oil executives. These aren't confidence-building measures — <em>they're crisis management</em>.</p><hr><h3>What Broke Yesterday</h3><p>Every asset class sold off in correlation — the signature of genuine risk repricing:</p><table><thead><tr><th>Asset</th><th>Level</th><th>Move</th></tr></thead><tbody><tr><td>S&P 500</td><td>6,624.70</td><td><strong>-1.36%</strong></td></tr><tr><td>Nasdaq</td><td>22,152.42</td><td><strong>-1.46%</strong></td></tr><tr><td>10-Year Treasury</td><td>4.259%</td><td><strong>+6.0 bps</strong></td></tr><tr><td>Oil (Brent)</td><td>>$111/bbl</td><td>Spiking</td></tr><tr><td>Bitcoin</td><td>$71,328</td><td><strong>-4.61%</strong></td></tr></tbody></table><p>Bitcoin's 4.6% drop is notable — it's supposed to be a geopolitical hedge. It didn't hedge anything. Meanwhile, <strong>Micron nearly tripled revenue</strong> on a memory-supply crunch, confirming AI infrastructure bottlenecks are creating winner-take-most dynamics even in a risk-off environment.</p><hr><h3>The Fed Leadership Vacuum</h3><p>Powell's term expires in May. His nominated replacement, <strong>Kevin Warsh</strong>, is blocked by a GOP senator until the DOJ drops an investigation into Powell over — remarkably — the Fed's headquarters renovation. Powell stays indefinitely as a lame duck during a war-driven economic crisis. Whether or not the DOJ probe is political, <em>the perception that monetary policy independence is compromised is itself a risk factor</em>.</p><hr><h3>Portfolio Implications</h3><blockquote>Any deal model assuming 2+ rate cuts in 2026 is stale as of yesterday. The 10-year above 4.25% and rising means discount rates need to go up 50-100bps across your pipeline.</blockquote><p>Three immediate actions:</p><ol><li><strong>Late-stage growth equity</strong> valued on 2024-25 public comps needs <strong>15-25% haircuts</strong> in realistic exit scenarios under a no-cut, $100+ oil environment.</li><li><strong>Energy cost exposure</strong> across portfolio companies needs auditing. Any company where energy/logistics exceeds 15% of COGS faces margin compression at sustained $111+ oil.</li><li><strong>Defense tech and energy security</strong> companies that seemed expensive 6 months ago may now be fairly valued — the geopolitical premium is structural, not cyclical.</li></ol>

   Action items

   • Stress-test every active deal model against a 'no cuts in 2026' scenario with oil sustained above $100/barrel by end of this week
   • Audit portfolio company energy cost exposure and flag any company where energy/logistics > 15% of COGS within 10 business days
   • Build scenario model for Fed leadership outcomes — Warsh confirmed, Warsh blocked indefinitely, or third candidate — by quarter-end

   Sources:Oil at $111, inflation reaccelerating, one cut in '26 — your portfolio needs a stagflation hedge now · MSFT down 25% but still not cheap — the AI capex bet is the real call you're making

2. 02

   World Models — $4B+ Deployed Into AI's Next Foundation Category at Zero Revenue

   <h3>A New Platform Category Is Being Capitalized</h3><p><strong>World Models</strong> — AI systems that learn the causal structure of environments by predicting next states conditioned on actions, not next tokens conditioned on text — just attracted <strong>$4B+ in funding</strong> across 8+ startups in roughly 12 months. This is not a feature of LLMs. This is a parallel compute paradigm for embodied intelligence.</p><p>The funding velocity is staggering:</p><table><thead><tr><th>Company</th><th>Round</th><th>Valuation</th><th>Architecture</th><th>Key Data Moat</th></tr></thead><tbody><tr><td>Wayve</td><td>$1.2B</td><td>$8.6B</td><td>Generative (GAIA-2)</td><td>Proprietary driving fleet data</td></tr><tr><td>AMI Labs</td><td>$1.03B</td><td>$3.5B</td><td>Latent (JEPA)</td><td>LeCun's Meta research pipeline</td></tr><tr><td>World Labs</td><td>$1.0B</td><td>$5.4B</td><td>3D / generative</td><td>Fei-Fei Li's ImageNet pedigree</td></tr><tr><td>Physical Intel.</td><td>$600M</td><td>$5.6B</td><td>VLA (π₀ series)</td><td>Multi-robot task data (68 tasks)</td></tr><tr><td>General Intuition</td><td>$133.7M (Seed)</td><td>Undisclosed</td><td>Generative (DIAMOND)</td><td><strong>1B+ gaming clips/yr, ground-truth action labels</strong></td></tr></tbody></table><hr><h3>The Architecture War</h3><p>The field is split three ways, each with fundamentally different data requirements, compute profiles, and investment characteristics:</p><ul><li><strong>Generative World Models</strong> (General Intuition, Wayve, Runway): Predict future frames directly. Interpretable but compute-intensive. DIAMOND built a playable Counter-Strike engine from 87 hours of footage on a single GPU.</li><li><strong>Latent World Models</strong> (AMI Labs/JEPA, Embo/Dreamer): Predict in abstract representation space. Efficient but opaque — <em>you can't see what the model is thinking</em>. No agent demonstrations from AMI Labs despite their $1.03B raise.</li><li><strong>Vision-Language-Action models</strong> (Physical Intelligence, Skild): Piggyback on LLM infrastructure. Shipping real demos today. But architecturally suboptimal for spatial-temporal reasoning.</li></ul><p>The <strong>Hardware Lottery thesis</strong> is the critical lens: markets often converge on the technology that fits existing infrastructure, not the one that's technically superior. VLAs inherit the entire LLM stack. World Models need new infrastructure — and that's both their risk and their opportunity.</p><hr><h3>Where the Alpha Is</h3><p>The talent migration is the strongest signal. When <strong>Danijar Hafner</strong> (DreamerV2, DeepMind), <strong>Ian Goodfellow</strong> (GANs inventor), <strong>Bob McGrew</strong> (OpenAI CRO), and <strong>Anthony Hu</strong> (GAIA-1 lead) leave prestigious positions to found World Model companies, they're voting with their careers. This pattern preceded the founding of Anthropic, Cohere, and Mistral in the LLM wave.</p><blockquote>AMI Labs CEO Alexandre LeBrun explicitly warned: 'World Models will be the next buzzword and in six months every company will call itself a World Model to raise funding.' When the CEO of a $3.5B company tells TechCrunch the category is about to be flooded with pretenders, believe him.</blockquote><p><strong>Google DeepMind is the existential threat</strong>. They're pursuing every approach simultaneously — Genie 3, SIMA 2, Waymo fleet data — with near-infinite compute and talent. The startups that survive are those with <strong>data assets Google can't easily replicate</strong>: General Intuition's 15M-user Medal gaming platform (ground-truth action labels), Wayve's driving fleet, Physical Intelligence's multi-robot task data.</p><p>The hardware layer is the de-risked proxy play. Decart's migration from Nvidia GPUs to Etched's custom Sohu ASICs signals real-time World Model inference needs specialized silicon. <strong>You profit from the sector's growth regardless of which architecture wins.</strong></p>

   Action items

   • Build a World Models sector map with three architecture columns (latent, generative, VLA) and score each company on data moat, team pedigree, and sim-to-real evidence by end of Q2
   • Take a meeting with General Intuition within 30 days — their $133.7M seed is done but Series A will be massively oversubscribed given the 1B+ gaming clips/year data flywheel
   • Commission technical diligence on ground-truth vs. inferred action labels — the single highest-leverage question for the entire category
   • Evaluate Etched and custom ASIC plays as hardware-layer proxies that de-risk the architectural uncertainty

   Sources:$4B+ deployed into World Models in 12 months — here's where the moat actually forms and which bets are mispriced · Nvidia's robotics lock-in play just accelerated — three investable sectors and the platform risk reshaping your deal flow

3. 03

   AI Agents Target the SaaS Add-On Layer — NRR Compression Is the Unmodeled Risk

   <h3>The Proof Point</h3><p>Brian Spanswick, CIO of <strong>Cohesity</strong> ($2B+ revenue, 400-person IT department), just gave the market a preview of SaaS cannibalization — and it's not the 'AI replaces Salesforce' narrative everyone's been debating. It's <strong>surgically worse</strong>: AI agents eat the highest-margin automation add-ons while enterprises keep paying for core platforms.</p><p>The specifics are damning:</p><ul><li>A Cohesity cybersecurity executive built a ServiceNow ITAM equivalent — priced at <strong>hundreds of dollars per user per month</strong> — using Claude Code in <strong>under two days</strong></li><li>Cohesity replaced Splunk's security monitoring with a consultant-built AI agent at <strong>lower operating cost</strong></li><li>Spanswick plans to keep Salesforce, Workday, and ServiceNow for 1-2 years but <strong>will not spend on their automation add-ons</strong></li><li>He estimates <strong>50% cuts</strong> to automation add-on budgets based on early tests</li></ul><hr><h3>The Macro Confirmation</h3><p>This isn't isolated. Enterprise AI spend is surging <strong>81% YoY</strong> while total IT budgets grow just <strong>3.4%</strong>. Anthropic and OpenAI are extracting <strong>$40-50B annually</strong> from enterprise budgets — the equivalent of 2-3 Salesforces' annual revenue. The SaaS incumbents' responses confirm they see it too:</p><table><thead><tr><th>Company</th><th>Strategy</th><th>Signal</th></tr></thead><tbody><tr><td>Salesforce</td><td>$50B buyback funded by $25B in bonds</td><td>Financial engineering over innovation — <em>capitulation</em></td></tr><tr><td>SAP</td><td>Dual alliances with Nvidia + Foxconn</td><td>Partnership-forward — positioning as enterprise AI context layer</td></tr><tr><td>Workday</td><td>$1.1B Sana AI acquisition</td><td>Buy-vs-build premium signals urgency</td></tr><tr><td>Okta</td><td>"Okta for AI Agents" launching April 30</td><td>Category creation — AI agent identity governance</td></tr></tbody></table><p>The divergence between Salesforce (buyback-defensive) and SAP (partnership-forward) will compound. SAP's ERP data <em>is</em> the business context that enterprise AI needs — its moat actually strengthens in the AI era.</p><hr><h3>The Unpriced Risk: NRR Compression</h3><blockquote>Wall Street models ServiceNow and Salesforce on 120%+ net revenue retention. If CIOs retain core platforms but zero out add-on budgets, NRR drops to 105-115% — a 10-15 point compression that triggers 30-40% multiple contraction even without headline revenue declines.</blockquote><p>This is a <strong>2026-2027 earnings risk</strong> that current multiples ignore. ServiceNow's defensive response — citing compliance, integrations, and auditability as moats — reveals the exact fault line. Those moats are real today but <em>codifiable</em>. The moment a startup ships SOC 2-compliant AI agent orchestration, ServiceNow's most defensible argument evaporates.</p><p>The credit market is already reacting. JPMorgan's suspension of Qualtrics' <strong>$5.3B debt deal</strong> because investors balk at AI displacement risk to survey software confirms this has jumped from equity to debt markets. Three of the five largest enterprise SaaS companies simultaneously executing debt-funded buybacks is a sector-wide signal we haven't seen since 2008.</p><hr><h3>The Three Investable Categories Emerging</h3><ol><li><strong>AI Agent Governance & Identity</strong>: Okta's April 30 launch, 88% of organizations reporting agent security incidents, Workday's Sana performing write actions. The compliance/governance gap is the next $1B+ platform opportunity.</li><li><strong>AI-Native Enterprise Automation</strong>: Workday's $1.1B Sana acquisition sets the valuation benchmark. The key differentiator is safe write actions within HR, finance, and supply chain workflows.</li><li><strong>Agent Orchestration & Compliance Tooling</strong>: ServiceNow's defensive moat (compliance, auditability, integration) is simultaneously the best argument for incumbents AND the clearest market map for what startups should build. The company that ships enterprise-grade compliance for AI-built internal tools becomes the <strong>Terraform of the agent era</strong>.</li></ol>

   Action items

   • Stress-test NRR assumptions in every portfolio company with SaaS add-on revenue models — run scenario analysis with add-on revenue declining 20-50% over 24 months this quarter
   • Build deal sourcing pipeline around AI agent governance and compliance tooling — identity, security, observability, kill-switch infrastructure — by end of Q2
   • Evaluate a SAP long / Salesforce underweight thesis and present at next IC meeting
   • Use Workday's $1.1B Sana AI acquisition as the valuation comp for any AI-native enterprise automation company in your pipeline

   Sources:$40-50B SaaS budget destruction is here — your portfolio's winners, losers, and the 3 new categories forming in the wreckage · AI agents are eating SaaS add-on margins — the $100B+ expansion revenue thesis for ServiceNow, Splunk, and Workday just cracked · AI disruption fears freeze $5.3B debt deal — software LBO financing just got repriced while OpenAI breaks from Microsoft · Simplifying AI