Synthesized by Clarity (Claude) from 33 sources · May contain errors — spot one? [email protected] · Methodology →
MIT: 95% of AI Budgets Burn on Unchanged Workflows
- Sources
- 33
- Words
- 1,315
- Read
- 7min
Topics Agentic AI AI Capital LLM Inference
◆ The signal
The contradiction IS the insight: the model layer is commoditizing and budgets are growing, but 95% of companies are burning both on unchanged workflows.
◆ INTELLIGENCE MAP
Intelligence map
01 Enterprise AI's 95% Failure Paradox: Budgets Growing, Returns Not
act nowMIT NANDA quantified enterprise GenAI failure at 95% (n=300 deployments). Yet RBC's CIO survey confirms AI spend is net-additive with 50%+ in production. The blocker isn't models — it's organizational 'learning gap.' Vendor-buy succeeds 67% vs. 22% internal-build. Real ROI hides in back-office, not sales/marketing AI.
- Vendor success rate
- Internal build success
- In production now
- Novice uplift
- Expert uplift
02 Regulated Rails Cross Platform Scale: Kalshi, Chainlink, x402
act nowKalshi nearly doubled to $40B in 7 weeks on CFTC approvals. Chainlink's Project Pangea onboarded 50+ banks across 16 countries for T+0 atomic settlement on the $9.6T/day FX market. x402 AI-agent payments hit 500k daily tx (5x June growth). Meta named prediction markets a top priority. Capital is rotating from crypto-beta into regulated infrastructure with real-world rails.
- Kalshi mark-up
- Pangea banks
- FX TAM (daily)
- x402 daily tx
- Strategy mNAV
03 Model Commoditization: First Public-Company Proof Points Arrive
monitorCoinbase cut AI spend ~50% while increasing token usage by defaulting to GLM 5.2 and Kimi 2.7. GLM-5.2 benchmarks near Opus 4.8 on SWE Bench and runs a 45-min agentic session for $3.36. The 'frontier access = moat' thesis is now empirically falsified at the application layer. Value migrates to inference routing, workflow lock-in, and proprietary data.
- Agentic session cost
- Zhipu vs Opus cost
- Token cache rate
- Model-swap time
- Frontier API (Claude/GPT)100%
- Open-weight (GLM 5.2)20%-80%
04 PQC Federal Mandate: 2030/2031 Deadlines Create Non-Discretionary Category
monitorTrump's June 22 executive order sets hard deadlines: PQC key establishment by Dec 2030, digital signatures by Dec 2031, with phase-out teeth for non-compliant systems. Pentagon calls quantum computers an 'existential threat.' First-dollar spend lands in cryptographic discovery and crypto-agility automation — not algorithm IP. Keyfactor positioning for category leadership.
- Key establishment
- Full PQC use
- Scope
- First spend
- EO signedJun 2026
- Crypto inventory2027-28
- Key establishmentDec 2030
- Full PQC useDec 2031
05 Agentic Commerce & Agent Identity: Two Pre-Consensus Categories Forming
backgroundAI shopping agent traffic grew 7,851% YoY with 4.4x human conversion rates, yet most companies misclassify or block it. Okta launched agent-identity governance (FedRAMP/HIPAA-grade). SaaStr's inbound agent booked 614 qualified meetings across 2.25M sessions with zero headcount. The attribution/governance infra for non-human actors is a category forming in real time.
- Agent conversion vs human
- Direct product page
- Agent-booked meetings
- Added headcount
◆ DEEP DIVES
Deep dives
01 Enterprise AI's Paradox: 95% Fail, Budgets Still Growing — Finding the 5%
act nowThe Contradiction That Reprices Your Book
Four independent studies converged this week on an uncomfortable truth, and one corporate survey contradicts the bear case they imply — together creating the most actionable intelligence pattern in today's briefing.
MIT NANDA (300 deployments, 150 exec interviews, 350 employee surveys): 95% of enterprise GenAI pilots delivered no measurable P&L impact. Only 5% produced rapid revenue acceleration. The blocker isn't model quality — it's the 'learning gap': organizations bolting AI onto unchanged workflows.
RBC CIO Survey: enterprises are creating net-new AI budgets, not cannibalizing existing software spend. 50%+ already run AI in production; another 35% within six months. Token costs aren't slowing adoption.
The market is simultaneously proving AI budgets are expanding AND that 95% of those dollars generate zero return. The 5% that works is the only thing worth funding.
What Separates the 5% From the 95%
Three structural findings rewrite sourcing priorities:
Finding Data Investment Filter Buy beats build 3:1 ~67% vendor success vs ~22% internal Tailwind for vertical AI SaaS; headwind for 'enterprises self-assemble' infra ROI hides in back-office Sales/marketing AI: most budget, least return Contrarian alpha in underfunded finance ops, compliance, document processing Self-reported gains are fiction METR RCT: devs 19% slower, believed 20% faster Discount any pitch with sentiment-based KPIs The productivity data adds nuance: AI lifts novices +34% but veterans ~0% (Brynjolfsson, n=5,179). Stanford's Canaries dashboard — built on ADP data covering 1-in-6 US workers — already shows employment falling for 22-25-year-olds in AI-exposed roles.
The Coinbase Counter-Example
Against the 95% failure backdrop, Coinbase just proved the opposite extreme: it cut AI spend roughly 50% while increasing token usage by defaulting to Chinese open-weight models (GLM 5.2, Kimi 2.7). This isn't a pilot — it's a public company optimizing at the P&L level. The key: Coinbase didn't bolt AI onto an unchanged process. It restructured its inference economics as a deliberate operational move.
The pattern becomes clear: the 5% that works either forces workflow redesign (the MIT finding) or owns its inference cost structure (the Coinbase finding). Most of your deal flow does neither.
What This Means for Your Portfolio
The net-additive budget finding dismantles the 'AI just eats SaaS' bear thesis — TAM is expanding, not zero-sum. But 95% failure means most of those dollars will churn. The alpha is in identifying which companies force the workflow change that makes value stick.
Action items
- Rebuild AI app-layer diligence around three gates: (a) does the product force workflow redesign, (b) vendor-deployed not customer-built, (c) inference contribution margin at scale
- Audit existing portfolio for self-reported productivity KPIs and request controlled output metrics plus token-cost-per-unit trends from each AI company
- Re-weight sourcing toward back-office automation AI (finance ops, compliance, document processing) and away from sales/marketing AI
- Open a thesis on AI-agent validation/testing as a category — the 'testing the 95%' wedge
Sources:95% of enterprise AI pilots return zero P&L — your app-layer multiples are mispriced · Enterprise AI budgets going net-additive + agent governance is a fresh category · Chinese open-weight models just halved Coinbase's AI bill · AI hit revenue>quarterly-depreciation · It's a conference promo — but the 'AI transformation gap' is a thesis worth tracking
02 Regulated Rails Cross Platform Scale — The Crypto Bifurcation in Real Time
act nowTwo Diverging Trajectories in One Market
The crypto/fintech market just cleaved into two completely different risk profiles, and the velocity of the divergence demands repositioning this week.
Going up: Kalshi nearly doubled from $22B to ~$40B in seven weeks — the fastest mark-up in a regulated fintech this cycle, driven by CFTC approvals, not crypto-beta. Chainlink's Project Pangea brought 50+ banks across 16 countries onto cross-chain infra integrated with Swift, targeting T+0 atomic settlement on the $9.6T/day FX market. Meta named its prediction-market app Arena a top internal priority and is actively courting Polymarket and Kalshi for partnerships.
Going down: Strategy's enterprise mNAV fell below 1.0 for the first time, common stock is -85% from its November 2024 peak, and STRC preferred trades 25% below par despite an 11.5% coupon. The leveraged-Bitcoin-treasury model is structurally broken.
Capital is rotating out of crypto-beta wrappers and into regulated infrastructure with real-world rails — that rotation is your alpha map.
The x402 Signal Nobody's Underwriting
The most asymmetric finding is the quietest: x402, the HTTP 402 micropayment protocol for AI agents, hit ~500k daily transactions in June — a 5x jump in a single month. This establishes machine-to-machine payments as a real on-chain category with live, compounding traction. Almost no one is underwriting this at seed/Series A.
The convergence is structural: as AI agents proliferate (7,851% traffic growth in commerce alone), they need payment rails. x402 is the early wedge where agentic commerce meets crypto infrastructure. Protocol-level network effects accrue to early movers.
Meta's Strategic-Buyer Signal
Meta's Arena uses points, not real money, targeting 18-34 — an engagement funnel, not a monetized betting book. That means Meta likely needs the regulated liquidity and resolution infrastructure that Polymarket and Kalshi already own. The incumbents' moat (real-money liquidity + regulatory positioning) is complemented, not commoditized. This establishes a strategic-buyer floor for the prediction-market category.
Vehicle Direction Signal Action Kalshi $22B → $40B (7 wk) Platform-scale premium Comp pipeline deals before round reprices Chainlink Pangea 50+ banks, Swift integration TradFi adopts infra, not assets Source settlement-layer deals x402 500k tx/day, 5x growth Pre-consensus agentic wedge Open thesis at seed/A immediately Strategy/STRC mNAV < 1.0, -85% Leveraged model broken Exit/mark down; flag copycats Airwallex Confirms the AI-Fintech Premium
Separately, Airwallex closed a $320M Series H at $11B — up 38% in six months. The re-rate is driven by product narrative: T:0 (AI-native platform automating bookkeeping, tax, compliance) and Airi (agentic consumer wallet). Private markets are paying a premium for autonomous finance as a category. This is your comp for any AI-fintech in pipeline.
Action items
- Open a thesis exploration on AI-agent payment infrastructure (x402 and competitors) at seed/Series A — source 3-5 candidates this month
- Pull comps and ownership maps on Polymarket, Kalshi, and 2-3 adjacent event-contract startups; gauge whether Meta partnership talks imply near-term round
- Mark down or exit exposure to leveraged Bitcoin-treasury vehicles and crypto-leveraged structured products
- Re-comp every prediction-market deal in pipeline against Kalshi's $40B; pressure-test whether 82% mark-up is demand-driven or froth
Sources:Kalshi 2x'd to $40B in 7 weeks while Strategy's mNAV broke below 1.0 · Airwallex marks AI-fintech up 38% in 6mo · Meta validates prediction markets · The Download from MIT Technology Review · Compute just became a weapon
03 Post-Quantum Mandate: A Federal Deadline Just Created a Multi-Year Category
monitorThe Rarest Demand Signal in Security Software
On June 22, 2026, the White House signed a post-quantum executive order setting firm federal deadlines: key establishment by December 31, 2030 and digital signatures by December 31, 2031. In parallel, the Pentagon released a PQC Strategy calling cryptographically relevant quantum computers an 'existential threat' and mandating that DOD systems support PQC by 2030 — or be phased out — and fully use it by 2031.
For investors, this is the rarest type of demand signal: a mandated, deadline-enforced, government-wide procurement cycle with phase-out teeth. Compliance isn't discretionary. Non-compliant systems get retired. That converts a decade-long research narrative into contractually forced spend across every federal and defense system.
A federal mandate with phase-out teeth just made post-quantum crypto-agility a non-discretionary, deadline-driven buy — the alpha is in the discovery layer, and the window closes as the market prices in 2030.
Where First Dollars Actually Land
The mistake is assuming dollars flow to PQC algorithms. They don't — at least not first. The binding constraint is cryptographic inventory: the DOD strategy explicitly directs a full inventory of quantum-vulnerable cryptography. You cannot migrate what you cannot see.
Segment Demand Driver Switching Cost Revenue Model Crypto discovery/inventory 2030 mandate, full asset inventory High First-dollar, recurring Cert lifecycle/crypto-agility Architectural mandate (DOD) High Platform, compounding PQC algorithm IP NIST standards adoption Low Downstream, commoditized Keyfactor is already running the thought-leadership land-grab — prescribing CISO governance, living cryptographic inventories, and multi-year funding. That's the textbook precursor to category commercialization and consolidation. Recall the Venafi/CyberArk precedent in machine-identity.
The Comp Set to Build Now
Map private targets before the mandate fully prices in: DigiCert, Entrust, SandboxAQ, PQShield, plus any crypto-agility automation startups at seed/A. The structural analog is GDPR-tooling in 2016-2018 — a compliance deadline that created Vanta, Drata, and OneTrust at scale. PQC's mandated refresh cycle is broader (all federal systems) and deeper (physical infrastructure, not just data handling).
The market knows 'quantum is coming someday' but hasn't fully absorbed that a hard 2030/2031 enforcement clock now exists with explicit phase-out consequences. That gap between awareness and pricing is your entry window.
Caveat: deadline hype will inflate near-term valuations on anything quantum-adjacent. Discriminate ruthlessly between durable, switching-cost-rich platforms (discovery, lifecycle) and point solutions riding the mandate's coattails.
Action items
- Build a private-market comp set for crypto-agility (Keyfactor, DigiCert, Entrust, SandboxAQ, PQShield) by end of July — identify which are raising or acquirable
- Update cybersecurity thesis to add 'mandated compliance cycle' as a distinct category — prioritize upstream discovery/inventory plays over algorithm IP
- Stress-test any portfolio exposure to government ad-tech / commercial-location data brokers against data-broker loophole legislation
Sources:PQC just became a federal procurement mandate · Cyber's recurring revenue thesis: supply-chain & PQC are where the alpha is · Top Enterprise Technology Stories
◆ QUICK HITS
Quick hits
Update: Model commoditization — Coinbase cut AI spend ~50% while increasing usage by routing to GLM 5.2 and Kimi 2.7; first public-company proof point for the open-weight substitution thesis
Chinese open-weight models just halved Coinbase's AI bill
Airwallex closed $320M Series H at $11B (up 38% in 6 months) on autonomous finance narrative — hard comp for AI-native fintech pipeline
Airwallex marks AI-fintech up 38% in 6mo
AI shopping agent traffic grew 7,851% YoY, converts at 4.4x human rates, and 77% lands directly on product pages — yet most companies misclassify or block it
Agentic traffic +7,851% YoY: the AI-commerce infra wedge your portfolio isn't pricing yet
Okta launched agent-identity governance (FedRAMP/HIPAA-grade) — validates 'non-human IAM' as a category; source independent challengers before GA crowds the space
Enterprise AI budgets going net-additive + agent governance is a fresh category
Two of China's top hedge fund managers publicly called AI rally a 'super bubble ready to burst' as ON Semi cratered -24% in a single session to $90.65
AI 'super bubble' warnings + $270B speculation unwind: time to stress-test your AI book
Malware authors now embed prompts causing LLM security tools to refuse analysis — the core moat of AI-SOC startups is being actively tested in the wild
AI-SOC startups just took a hit: malware now jailbreaks LLM analysts
Gusto shipped a tier-one product with 5 people in 10 weeks and zero PM/Figma/Jira — AI-native teams are retiring the process-tooling SaaS stack
Open-weight just hit frontier parity at 1/10th cost — your AI infra multiples are mispriced
Adobe acquiring Topaz Labs (Emmy-grade AI enhancement) to bundle into Firefly/Photoshop/Premiere — standalone AI-enhancement startups face closing exit window
Adobe's Topaz grab + Figma's canvas dilemma: the design-AI consolidation just started
NVIDIA's ENPIRE framework hit 99% on dexterous manipulation tasks using commodity arms + RTX 5090 — teleoperation-data moats in robotics are depreciating
NVIDIA quietly moved up the robotics stack — and China hit GPU parity
Italy's AGCM opened first EU antitrust probe on AI bundling — investigating Microsoft for forcibly integrating Copilot into M365 and auto-upgrading pricing
US just nationalized frontier-AI distribution — your AI GTM models need a regulatory haircut
Snowflake publishing defensive 'benchmarks don't matter' content against Databricks' Reyden — classic incumbent posture signaling share-loss pressure
Snowflake's defensive benchmark blog signals Databricks is winning the data-platform war
◆ Bottom line
The take.
The AI market's most dangerous assumption just got empirically destroyed twice in one week: 95% of enterprise pilots deliver zero P&L (MIT, n=300), yet Coinbase proved the 5% path by cutting AI spend 50% through open-weight models while increasing usage. The gap between those two numbers is where mispricing lives — fund companies that force workflow redesign and own their inference economics, short everything that bolts AI onto unchanged processes, and redirect attention to the regulated-rails category (Kalshi $40B, Chainlink 50+ banks, x402 500k daily tx) where real revenue infrastructure is being built beneath the hype.
Frequently asked
- What separates the 5% of enterprise AI deployments that work from the 95% that don't?
- Two patterns: they either force workflow redesign rather than bolting AI onto existing processes, or they restructure their inference economics directly. MIT NANDA found vendor-deployed solutions succeed ~67% of the time versus ~22% for internally-built, and back-office use cases (finance ops, compliance, document processing) show better ROI than the crowded sales/marketing segment.
- Why should investors discount self-reported AI productivity metrics in pitch decks?
- A METR randomized controlled trial found developers were actually 19% slower using AI while believing they were 20% faster — a 39-point gap between perception and reality. Brynjolfsson's study (n=5,179) further showed AI lifts novices +34% but veterans ~0%. Any diligence relying on sentiment-based KPIs is systematically inflated and should be replaced with controlled output metrics and token-cost-per-unit trends.
- What's the x402 protocol and why does it matter for early-stage sourcing?
- x402 is an HTTP-based micropayment protocol enabling AI agents to transact on-chain. It hit ~500k daily transactions in June 2026 — a 5x monthly jump — establishing machine-to-machine payments as a live category with compounding traction. It sits at the intersection of agentic AI and crypto rails, and is largely unpriced at seed/Series A, making it one of the more asymmetric wedges currently open.
- Is the Strategy (MSTR) mNAV break below 1.0 a temporary dislocation or a structural break?
- Structural. When mNAV falls below 1.0, the equity-issuance flywheel that funded Bitcoin accumulation reverses — new issuance becomes dilutive rather than accretive. Common stock is -85% from its November 2024 peak and STRC preferred trades 25% below par despite an 11.5% coupon. Copycat leveraged-treasury vehicles face the same mechanical unwind and warrant markdowns or exits.
- How large is the post-quantum cryptography procurement opportunity and where do first dollars land?
- The June 22, 2026 executive order mandates federal key establishment migration by December 31, 2030 and signatures by December 31, 2031, with DOD systems facing phase-out if non-compliant. First dollars flow to cryptographic discovery and inventory tooling — you can't migrate what you can't see — followed by certificate lifecycle and crypto-agility platforms. Algorithm IP is downstream and commoditizing.
◆ Same day, different angle
Read this day as…
◆ Recent in investor
Keep reading.
Spot an error? [email protected]