$2B AI Infra Week Meets 10x Cheaper Open-Source Inference

◆ DEEP DIVES

1. 01

   The AI Infrastructure Paradox: $2B+ Flooding In While Costs Collapse 10x

   <h3>The Week Capital Showed Its Hand</h3><p>Seven AI infrastructure companies raised over <strong>$2 billion in a single week</strong> — and the deployment pattern reveals where institutional capital is placing its structural bet. Mistral raised $830M in debt to build a data center. Rebellions raised $400M pre-IPO at $2.34B for non-NVIDIA inference chips. ScaleOps closed a $130M Series C at >$800M for Kubernetes compute orchestration. Qodo raised $70M for AI code governance. Starcloud hit $1.1B at Series A for orbital data centers. Sarvam AI raised $300-350M at ~$1.5B for India's 22-language AI stack.</p><p>The message is unambiguous: capital has migrated from <strong>"who has the best model" to "who controls the infrastructure every model needs."</strong> These aren't speculative bets — ScaleOps claims 80% cost reduction for enterprise Kubernetes, Qodo counts NVIDIA, Walmart, and Red Hat as customers, and Rebellions is expanding to the US, Japan, and Saudi Arabia.</p><hr><h3>The Paradox: Why Build More Infra If Costs Are Collapsing?</h3><p>Here's where the synthesis gets interesting. At the <strong>exact same moment</strong> capital floods into compute infrastructure, a separate cluster of signals shows inference costs in freefall. H Company's Holo3 outperforms GPT-5.4 on OSWorld by activating only <strong>10B of its 122B parameters</strong> — at 1/10th the cost. Google DeepMind's TurboQuant achieves <strong>6x KV memory reduction and 8x faster attention computation</strong> on H100s with zero retraining. Sparse MoE architectures have fundamentally broken the "bigger compute = better model" equation.</p><p>The resolution of this paradox is where the alpha lives. The infrastructure buildout isn't wrong — it's just targeting a different bottleneck than raw compute power. The winning investments are in the <strong>orchestration, governance, and reliability layers</strong> that sit between raw hardware and production AI workloads:</p><ul><li><strong>Compute orchestration</strong> (ScaleOps) — managing GPU allocation as costs fluctuate</li><li><strong>Code governance</strong> (Qodo) — ensuring AI-generated code meets compliance standards as agentic coding proliferates</li><li><strong>Model routing</strong> (Perplexity's Model Council, the emerging category) — running queries across multiple models and synthesizing results</li><li><strong>Agent reliability</strong> — the guardrails, monitoring, and testing layer every enterprise needs</li></ul><blockquote>The capex-to-performance curve just kinked. Any portfolio company or deal premised on compute scarcity as a permanent moat needs re-examination — the moat is moving to orchestration.</blockquote><h3>Sovereign AI Goes Unicorn-Scale</h3><p>Geographic diversification of AI capital deserves its own investment category. <strong>India</strong> (Sarvam AI, $300-350M at $1.5B), <strong>South Korea</strong> (Rebellions, $400M at $2.34B), and <strong>France</strong> (Mistral, $830M debt) all raised unicorn-scale rounds in the same week. Sarvam's thesis — covering 22 Indian languages that Western frontier models struggle with — creates a <strong>durable linguistic moat</strong>. Similar plays in Arabic, Bahasa, and other complex language markets remain structurally underserved.</p><h3>The Acqui-Hire Signal</h3><p>Anthropic's ~$400M stock acquisition of Coefficient Bio — an <strong>8-month-old, fewer-than-10-person biotech AI startup</strong> with ex-Genentech talent — establishes a new per-head valuation of ~$40M+ for domain-specific AI researchers. This reprices every bio-AI team in deal flow and validates healthcare as the next frontier AI vertical. Meanwhile, Yupp's shutdown (<strong>$33M raised from a16z crypto, dead in under 12 months</strong>) confirms that crowdsourced evaluation tooling is a dead category in the agentic era.</p>

   Action items

   • Reprice all AI infrastructure pipeline deals using new comps: ScaleOps >$800M, Rebellions $2.34B, Starcloud $1.1B — these are the floor valuations for AI infra with traction
   • Build a thesis memo on sovereign/regional AI as a standalone investment category, anchored by Sarvam ($1.5B) and Mistral ($830M) data points, by end of Q2
   • Audit portfolio for 'Yupp-pattern risk' — any company whose thesis depends on crowdsourced data labeling or human evaluation loops in pre-agentic workflows
   • Map the model orchestration/routing category for Series A-B deal flow — Perplexity's Model Council validates this as an emerging investable category

   Sources:$2B+ deployed in one week across AI infra — here's the capital map reshaping your deal flow · AI compute rationing is real — labs killing products, turning down revenue · AI inference costs just collapsed 10x · Inference costs just got 6-8x cheaper and Chinese frontier models went API-compatible

2. 02

   Short-Side Intelligence: SEC Investigations, Quantum SPAC Fraud, and a CFO Churn Wave

   <h3>SEC FOIA Logs: 11 Companies Under Investigation</h3><p>The SEC's March 2026 FOIA logs reveal likely ongoing investigations into <strong>AppLovin, CrowdStrike, Danaher, Honeywell, Starbucks, Bristol-Myers Squibb, Boston Scientific, ASP Isotopes, Organon & Co, Sable Offshore, and Perimeter Solutions</strong>. FOIA logs are a lagging confirmation of investigation initiation but a <strong>leading indicator of formal enforcement action by 12-18 months</strong>. The market rarely prices this in until charges are announced.</p><p>For portfolio purposes: if AppLovin or CrowdStrike appear as public comps in active deal models, apply a <strong>15-25% regulatory risk discount</strong> until investigation outcomes are known. Both names carry premium multiples that assume clean governance.</p><hr><h3>Xanadu (XNDU): The Theranos of Quantum Computing</h3><p>Toronto-based Xanadu went public via SPAC at a <strong>~$3B valuation on approximately $3M in revenue and $60M in annual losses</strong> — roughly a 1000x revenue multiple. Former quantum physicist Pierre-Luc Dallaire-Demers publicly calls it <strong>"the Theranos of quantum computing,"</strong> claiming it's "15 years away" from error-corrected qubits that competitors already have.</p><p>The SPAC sponsor paid Outside the Box Capital <strong>$200,000 for a six-month stock promotion campaign</strong> targeting Reddit, YouTube, and Discord starting March 9. When the sponsor needs paid social media campaigns to sustain the stock price, you're watching the endgame unfold.</p><blockquote>At ~1000x revenue with paid Reddit promotion, credible insider fraud allegations, and unproven photonic quantum technology, Xanadu is the highest-conviction short candidate surfaced this quarter.</blockquote><h3>CFO Churn: The Canary Wave</h3><p>Six companies disclosed CFO or C-suite departures in a single week. The most alarming patterns:</p><table><thead><tr><th>Company</th><th>Market Cap</th><th>Signal</th><th>Historical Pattern</th></tr></thead><tbody><tr><td><strong>Clover Health (CLOV)</strong></td><td>$902M</td><td>CFO stepped down immediately</td><td>6 CFOs in 5 years</td></tr><tr><td><strong>Climb Bio (CLYM)</strong></td><td>$313M</td><td>Principal Accounting Officer separated</td><td>6 CFOs in 5 years</td></tr><tr><td><strong>T1 Energy (TE)</strong></td><td>$1.17B</td><td>Two board members resigned</td><td>Culper short report Jan, CAO quit Feb</td></tr><tr><td><strong>Aviat Networks (AVNW)</strong></td><td>$261M</td><td>Auditor downgrade Deloitte→Grant Thornton</td><td>Revenue recognition allegations</td></tr></tbody></table><p>T1 Energy's cascading pattern deserves particular attention: <strong>short report → CAO resignation → two board exits</strong> in three months. This sequence frequently precedes disclosure of the issues the short seller initially flagged.</p><h3>ProCap Financial (BRR): The Pompliano SPAC Implosion</h3><p>Anthony Pompliano's bitcoin-focused SPAC is <strong>down ~80% since its December 2025 merger</strong>. It has lost its CIO (8 months in), a board member (7 weeks in), and switched auditors from MaloneBailey to BDO — all within four months of going public. The compressed timeline of governance failures makes this a textbook SPAC mortality case.</p><h3>Activist Short Reports Worth Tracking</h3><p><strong>MakeMyTrip (MMYT, $3.79B)</strong>: Morpheus Research published a report based on <strong>103 industry expert interviews</strong> alleging illegal price parity practices and accounting manipulation. The methodology is unusually rigorous. <strong>Aviat Networks (AVNW, $261M)</strong>: GlassHouse Research alleges premature revenue recognition, supported by the Deloitte-to-Grant Thornton auditor downgrade — a standalone red flag that warrants deep diligence.</p>

   Action items

   • Screen portfolio and watchlist against the 11 SEC FOIA investigation targets immediately — if any appear as public comps in deal models, apply 15-25% regulatory risk discount
   • Add XNDU to short watchlist with entry trigger on lockup expiration or first insider sale — the $200K promotion campaign has a six-month runway from March 9
   • Run a systematic CFO turnover screen: flag any public company with 3+ CFOs in 5 years and cross-reference with auditor changes using VerityData or similar
   • Reassess quantum computing sector private valuations if XNDU is being used as a comparable — photonic approach may be a dead end per credible expert testimony

   Sources:SPAC graveyard expanding, SEC hunts AppLovin & CrowdStrike — your short book needs this map

3. 03

   Vibe Coding's Dual Shock: 84% App Surge Collides with Apple's Platform Kill

   <h3>The Supply Explosion</h3><p>After a <strong>48% decline from 2016 to 2024</strong>, new App Store submissions surged 30% in 2025 to nearly 600,000, then <strong>accelerated to 84% YoY growth in Q1 2026</strong> — 235,800 new apps in a single quarter. If Q1's pace annualizes, 2026 will see ~943,000 new apps, potentially the highest in App Store history. The inflection maps precisely to broad release of AI coding tools: Anthropic's Claude Code (May 2025) and OpenAI's Codex (October 2025), which grew from 100K to <strong>2M developers in 3 months</strong>.</p><p>This isn't incremental growth — it's a <strong>structural reversal of a decade-long decline</strong>. The cost of creating an app is approaching zero, and the data is now definitive.</p><hr><h3>The Platform Kill</h3><p>Apple simultaneously escalated enforcement against vibe coding apps, <strong>removing "Anything" entirely from the App Store</strong> under Guideline 2.5.2, which prohibits apps from introducing new functionality without Apple's review. The enforcement is deliberate and escalating: first blocking updates, then full removal. The architectural conflict is <strong>irreconcilable</strong> — vibe coding tools generate code dynamically at runtime; Apple's review model requires static, pre-approved code. There is no technical workaround.</p><p>This splits the AI app builder market along distribution lines:</p><table><thead><tr><th>Strategy</th><th>TAM Impact</th><th>Investment Posture</th></tr></thead><tbody><tr><td>iOS-first</td><td>-50 to -60% US TAM</td><td>Existential risk — pause or reprice</td></tr><tr><td>Web-first / PWA</td><td>Neutral to positive</td><td>Structurally advantaged — premium justified</td></tr><tr><td>Android-first</td><td>Captures iOS refugees</td><td>Screen for opportunities now</td></tr></tbody></table><blockquote>The biggest risk in AI app builders isn't that the models don't work — it's that the platforms controlling distribution decided they work too well.</blockquote><h3>Where Value Migrates</h3><p>When app creation costs approach zero and ~1M apps per year flood the store, value shifts decisively:</p><ol><li><strong>App compliance and quality infrastructure</strong> — Apple's crackdown creates urgent demand for tools ensuring AI-generated code meets review standards. This is a greenfield category with toll-booth economics. Every one of ~943K annual new apps is a potential customer.</li><li><strong>Distribution and discovery</strong> — Organic App Store discovery breaks down under volume. Mobile marketing, ASO platforms, and AI-native distribution tools become essential infrastructure.</li><li><strong>AI coding tool consolidation</strong> — An unnamed startup hitting a $1.3B valuation for AI model selection infrastructure confirms VCs are pricing the picks-and-shovels thesis. But the compliance and distribution layers remain underinvested relative to the opportunity.</li></ol><h3>The Competitive Self-Interest Question</h3><p>Apple may be acting from <strong>competitive self-interest</strong>, not governance principle. Vibe coding tools commoditize app creation, threatening the developer ecosystem and the <strong>30% commission model</strong> built around it. If Apple launches its own AI dev tools at WWDC (June 8) while blocking third-party alternatives, the bundling narrative becomes a regulatory catalyst under the EU's Digital Markets Act. <em>This is a 12-18 month timeline, not near-term — but it changes the long-term category risk calculus.</em></p>

   Action items

   • Audit every AI app builder and vibe coding startup in pipeline for iOS-dependent TAM assumptions — flag for immediate re-diligence on distribution strategy
   • Source deals in app compliance/quality infrastructure — companies building automated review, testing, and policy-compliance tooling for AI-generated apps
   • Update CAC assumptions for any mobile-first consumer deal in active diligence — pre-2025 organic discovery models are structurally broken
   • Monitor Apple WWDC June 8 for AI dev tool announcements — if Apple launches competing tools while blocking third-party alternatives, the antitrust narrative becomes investable

   Sources:Vibe coding just 2x'd App Store supply in one quarter · Apple just killed iOS distribution for vibe coding startups · Three mega-IPOs are draining Big Tech liquidity

4. 04

   Kalshi's 20x Growth vs. Criminal Charges: The Prediction Market Binary

   <h3>Growth Metrics That Mask Existential Risk</h3><p>Kalshi's numbers are extraordinary by any standard: <strong>$10.4B in monthly trading volume</strong> (up 20x from $521M in March 2025), <strong>5.1M+ monthly active users</strong> (up 8.5x from 600K). These are comparable to Robinhood's 2020 breakout or Coinbase's 2021 crypto supercycle. But there's a critical detail: <strong>sports betting accounts for ~90% of activity</strong>. This isn't a prediction market with some sports betting — it's a nationwide sportsbook wearing a CFTC-regulated commodity trading costume.</p><hr><h3>The Legal Assault Is Escalating — From Civil to Criminal</h3><p>Arizona filed <strong>criminal charges</strong> against Kalshi — the first-ever criminal prosecution of a prediction market platform. Nevada issued a temporary restraining order on March 20. In total, <strong>30+ active lawsuits from 8 states and 2 tribal governments</strong> are now open. The coordinated nature of these actions signals that the traditional sportsbook lobby — DraftKings, FanDuel, and the American Gaming Association — is mobilizing its state-level political relationships.</p><p>The entire category rests on a single legal theory: event contracts are CFTC-regulated financial instruments, not state-regulated gambling. Kalshi <strong>self-certified</strong> to expand into sports in January 2025 — using a mechanism originally designed for commodity hedging. Former CFTC special counsel Peter Sanchez Guarda warned: <em>"The mechanisms to police their activity don't exist, which creates more opportunities for people with asymmetric information to do unscrupulous things with it."</em></p><h3>Three Regulatory Scenarios for Portfolio Models</h3><ol><li><strong>Federal preemption wins (~30% probability)</strong>: Prediction markets become nationally licensed fintech. Kalshi's trajectory implies $100B+ annual volume. At 2-5% take rates, that's $2-5B in sector revenue. Kalshi and Polymarket become $10B+ companies.</li><li><strong>State-by-state fragmentation (~45% probability)</strong>: TAM shrinks 30-50%. Compliance costs surge. Sports betting may need to be carved out. Viable but smaller — mid-single-digit-billion valuations.</li><li><strong>Reclassification as gambling (~25% probability)</strong>: State gaming licenses required, 21+ age restrictions, election/political contracts banned. Category effectively destroyed at current form.</li></ol><h3>The Insider Trading Wildcard</h3><p>The Trump family's financial exposure spans the entire ecosystem: <strong>1789 Capital's eight-figure Polymarket investment</strong> (Trump Jr.'s advisory board seat), <strong>Truth Predict launch</strong> (Trump family majority-owned), and policy control over outcomes users bet on. Contracts exist on whether the President will say specific words or visit specific countries. If a manipulation scandal breaks — and structural incentives make one almost inevitable — it triggers the kind of <strong>sudden regulatory overcorrection</strong> that destroys category valuations overnight.</p><blockquote>Kalshi's 1,900% volume growth proves massive consumer demand for prediction markets — but with criminal charges filed, you're not investing in a fintech platform, you're betting on a regulatory outcome.</blockquote><h3>The Historical Parallel</h3><p>The most precise analogy is the <strong>2015-2016 daily fantasy sports crisis</strong>: DraftKings and FanDuel faced near-identical regulatory challenges (state AGs arguing illegal gambling, federal regulators providing cover). The resolution — state-by-state licensing — ultimately <em>helped</em> incumbents by creating regulatory barriers to entry. If prediction markets follow the DFS path, the survivors will need deep pockets and political patience. <strong>Traditional sportsbook incumbents may be the asymmetric beneficiary</strong> of either outcome.</p>

   Action items

   • Stress-test any prediction market portfolio exposure against a scenario where self-certification is narrowed — model the impact of losing sports betting (90% of activity)
   • Evaluate DraftKings and Flutter/FanDuel as relative value plays that benefit regardless of prediction market regulatory outcome
   • Place prediction markets (Kalshi, Polymarket) on watchlist but do NOT initiate new positions until Arizona criminal case and at least 2-3 major state lawsuits resolve

   Sources:Kalshi's 20x volume growth masks an existential regulatory cliff · Kalshi's 1,900% volume surge masks existential regulatory risk