Category: Technology & AI

  • The Refinery Count: Fact-Checking the Global Infrastructure Pattern

    The Refinery Count: Fact-Checking the Global Infrastructure Pattern

    A post on X went viral last week: “10 oil refineries have blown up in 21 days.” It racked up over 23,000 views. The replies were split between people screaming sabotage and people dismissing it as conspiracy noise.

    Both camps are wrong. The real number is lower, the timeframe is longer, and the story is more nuanced — but the underlying pattern is real, and it matters more than the headline suggests.

    In my earlier piece on the convergence, I mapped the collision of energy disruption, food insecurity, and geopolitical fracture happening simultaneously across multiple systems. This is the data-driven follow-up on one thread of that pattern: the global refinery incidents.

    The Verified Count

    I went through every incident I could verify with at least one credible news source. The actual tally since the US-Iran conflict escalated in mid-March 2026 is 6-7 significant refinery incidents across 5 countries in roughly 45 days. Not 10 in 21 days. Still remarkable — but accuracy matters if you want anyone serious to listen.

    Here is every confirmed incident, sourced:

    1. Russia — Multiple Oil Infrastructure Strikes (Late March 2026)

    Ukraine launched one of its largest-ever drone campaigns against Russian energy infrastructure, with over 200 drones deployed in a single overnight wave. Multiple refineries and fuel depots were hit across southern and central Russia. This was not a single event — it was a sustained, multi-target campaign that continued into April.

    2. Russia — Tuapse Refinery

    The Tuapse refinery on the Black Sea coast was struck by Ukrainian drones in a separate, confirmed attack. Anadolu Agency and Reuters both reported the strike, which caused fires and forced partial shutdowns at one of Russia’s key southern refining facilities.

    3. Mexico — Dos Bocas Refinery (April 9, 2026)

    Mexico’s troubled Dos Bocas refinery — the flagship project of the previous administration — suffered a significant fire on April 9. The facility, located in Tabasco state, has been plagued by delays and cost overruns since construction began. The cause of the fire remains unclear, and Mexican authorities have released limited information. Reuters covered the incident alongside broader concerns about Mexican energy infrastructure reliability.

    4. Australia — Geelong, Viva Energy (April 15, 2026)

    A massive fire broke out at the Viva Energy refinery in Geelong, Victoria — one of only two remaining refineries in Australia. Emergency crews battled the blaze for hours. The significance here is structural: Australia’s refining capacity is already critically thin. Losing even partial output from Geelong puts genuine pressure on domestic fuel supply in the Asia-Pacific region.

    5. India — Rajasthan Refinery (April 20, 2026)

    Perhaps the most dramatic incident. A fire broke out at India’s .5 billion refinery in Rajasthan — literally the day before its planned inauguration. The timing could not have been worse. This was meant to be a showcase of Indian energy independence. Instead, it became front-page news for the wrong reasons.

    6. Texas — Port Arthur, Valero

    An explosion was reported at Valero’s Port Arthur refinery in Texas. Port Arthur is one of the largest refining complexes in the United States, and any disruption there ripples through US Gulf Coast refining capacity — the backbone of American fuel production.

    The Distinction That Actually Matters

    Here is where the viral narrative falls apart — and where the real story begins.

    The Russian incidents are deliberate acts of war. Ukraine has been systematically targeting Russian refining capacity as a strategic military objective for over two years. These are not mysterious explosions. They are drone strikes with clear attribution and military logic.

    The Australia and India incidents appear to be accidents. Industrial fires at refineries are not uncommon — these are complex, high-temperature chemical processing facilities. The Mexico incident remains unclear.

    The “coordinated attack” narrative that spread on social media conflates deliberate military strikes with what appear to be coincidental industrial accidents. That conflation is the problem. It transforms a real but nuanced pattern into conspiracy fuel, which makes it easy for serious people to dismiss entirely.

    But Here Is What IS Significant

    Strip out the Russian strikes entirely. You are still left with 3-4 significant refinery fires across different countries — Australia, India, Mexico, the United States — within a 45-day window, during a period of unprecedented global tension.

    Is that statistically anomalous? Possibly. Refineries have incidents regularly. But the clustering, combined with the geopolitical context, is notable. India Today ran a piece today examining the same pattern — the fact that mainstream outlets are now covering the “5 countries in flames” discussion tells you the signal is breaking through the noise.

    The Compounding Effect on Oil Markets

    None of this happens in isolation. The Strait of Hormuz disruption I wrote about in the convergence piece has already constrained global oil transit. Now add refinery outages — both deliberate and accidental — across five countries on four continents. Each incident alone is manageable. Together, they compound pressure on an oil market that was already stretched thin by the US-Iran conflict.

    Refining capacity is not like crude supply — you cannot just reroute it. When a refinery goes offline, the processed fuel it would have produced simply does not exist until it comes back online. Multiple simultaneous outages create bottlenecks that take weeks or months to clear.

    The Bottom Line

    The viral claim is wrong on the numbers. But the pattern it points to is real. Six to seven verified incidents across five countries in 45 days — a mix of deliberate military strikes and apparent accidents — during the most volatile geopolitical period in decades.

    That is not a conspiracy. It is a convergence. And if you want to understand why energy security, food systems, and geopolitical stability are fracturing at the same time, start with the original piece.

    The world is not experiencing a coordinated attack on refineries. It is experiencing what happens when multiple systems are under maximum stress simultaneously — and fragile infrastructure starts breaking in clusters.

    Pay attention to the pattern, not the panic.

  • ChatGPT Images 2.0: What It Actually Means for CFOs and PE Professionals

    ChatGPT Images 2.0: What It Actually Means for CFOs and PE Professionals

    OpenAI just shipped ChatGPT Images 2.0 — a major upgrade to image generation inside ChatGPT. Sharper outputs, clean text rendering, multilingual support, complex layout understanding, and a model that genuinely “thinks before it creates.”

    Most coverage has been product announcements and tech demos. I’m more interested in a different question: what does this actually unlock for finance professionals — specifically CFOs, FDs, and PE teams who spend real money on visual communications every quarter?

    The answer is: quite a lot.

    The Cost of Looking Professional

    If you’ve worked in PE-backed businesses or run finance functions in mid-market companies, you know the drill. Every board pack, investor update, results presentation, and annual report needs visual support. Infographics, branded charts, org charts, KPI dashboards rendered as images, portfolio company brand assets — the list is long.

    Historically, that means one of three things: an in-house design team (expensive), an external agency (expensive and slow), or a finance person wrestling with PowerPoint clip art at 11pm (free but painful). A decent design agency charges £2,000–£5,000 for a set of investor deck visuals. A full brand refresh for a portfolio company? £15,000–£50,000 depending on scope. Turnaround? Weeks, sometimes months.

    For PE firms running value creation plans across multiple portfolio companies, multiply those numbers accordingly. It adds up fast.

    Why Previous AI Image Tools Didn’t Cut It

    AI image generation has existed for a couple of years now. DALL·E, Midjourney, Stable Diffusion — they could all produce impressive artwork. But they had a fatal flaw for business use: they couldn’t reliably render text.

    Think about that for a moment. Almost every financial visual contains text — axis labels on charts, section headers on infographics, company names on org charts, metric callouts on KPI dashboards. If the AI mangles “Revenue Growth” into “Revnue Groth,” the output is useless for anything professional.

    That single limitation kept AI image generation firmly in the “interesting toy” category for finance teams. Images 2.0 fixes it. Clean, accurate text rendering is now standard. That changes the calculus entirely.

    What This Unlocks for CFOs

    Financial communications have always been bottlenecked by visual production. The numbers are ready, the narrative is written, but you’re waiting three days for the design team to turn it into something presentable. That bottleneck just disappeared.

    With Images 2.0, a CFO or their team can now:

    • Generate investor deck visuals from a text prompt in seconds
    • Create board pack infographics — waterfall charts, bridge diagrams, strategic roadmaps — with accurate labels and professional polish
    • Produce scenario visualisations for stakeholder presentations
    • Build org charts and operating model diagrams without touching Visio
    • Draft annual report graphics for review before briefing a designer on final production

    The time saving is obvious. The cost saving is significant. But the real win is speed-to-decision: when you can visualise a scenario in real time during a board discussion, the quality of the conversation changes.

    The PE Angle: Portfolio Company Branding at Speed

    Private equity value creation plans frequently include brand and communications workstreams. A newly acquired platform company needs refreshed branding. A carve-out needs its own identity built from scratch. A portfolio company going through transformation needs updated investor materials every quarter.

    Previously, each of those was an agency engagement with a timeline measured in weeks. Now, the first 80% of that work — concept exploration, visual direction, draft assets — can happen in an afternoon. The PE operating partner or portfolio CFO can generate brand concepts, test visual directions, and arrive at the agency brief with far more clarity about what they actually want.

    That doesn’t eliminate the agency. It compresses the engagement, reduces iteration cycles, and dramatically cuts cost. For a PE firm managing ten portfolio companies, the aggregate saving is material.

    Available Now — No Waiting List

    Unlike many AI announcements, this isn’t vapourware. ChatGPT Images 2.0 is live now. Free users get access; paid plans (Plus, Team, Enterprise) get the best version with higher limits and the ability to pull in live web information for context-aware generation.

    If you have a ChatGPT account, you can test this today. Open a conversation, describe the infographic or visual you need, and see what comes back. The results will surprise you.

    The Honest Caveat

    This is still a tool, not a replacement for strategic design thinking. A poorly described prompt produces a poor image — garbage in, garbage out. Brand consistency across a full suite of materials still needs a human eye. Complex data visualisations with precise numerical accuracy still need proper charting tools.

    But for the 80% of visual work that currently sits in a queue waiting for a designer — concept visuals, draft infographics, presentation graphics, quick-turn brand assets — this is a genuine step change. The AI toolkit for finance professionals just got significantly more capable.

    What To Do Next

    Open ChatGPT. Describe a board pack infographic you’ve been meaning to create. See what happens. Then think about what that means for your next investor update, your next portfolio company rebrand, or your next annual report cycle.

    The economics of financial communications just shifted. The CFOs and PE teams who notice first will move fastest.

    If you want to discuss how AI tools like this fit into your finance function or PE operating model, get in touch.

  • The Agentic Economy Has Already Started — 167 Million Transactions In

    The Agentic Economy Has Already Started — 167 Million Transactions In

    While most of the market is still debating whether AI agents will eventually transact autonomously, 167 million x402 transactions have already settled on-chain. That is not a projection. It is not a roadmap slide. It is done.

    Eighty-five percent of those transactions ran through Base, Coinbase’s Layer 2 network. The agentic economy is not arriving — it has arrived, and it is scaling fast.

    I run a live x402 service on Base. Most commentary on this topic comes from spectators. This is from a practitioner.

    What x402 Actually Is

    The x402 protocol is elegantly simple. An AI agent makes an HTTP request to a service. The server responds with HTTP 402 — Payment Required. The agent’s wallet pays the specified amount in USDC on Base, attaches proof of payment to the request header, and the service delivers the response.

    No account creation. No API key provisioning. No OAuth flows. No invoices. No 30-day payment terms. The agent pays per request, in a stablecoin, and gets what it paid for in milliseconds.

    This is what machine-to-machine commerce looks like when you strip out every piece of friction that was designed for humans.

    Agentic.market — The Discovery Layer

    Agentic.market is Coinbase’s discovery layer for x402 services. Think of it as a marketplace where AI agents can browse, evaluate, and purchase services from other agents and providers — programmatically, without human involvement.

    Today it lists services across data retrieval, computation, content generation, and verification. An agent needing market data, sentiment analysis, or specialised computation can find a provider, check pricing, pay, and consume — all in a single automated flow.

    This is not a concept demo. Services are live, priced in USDC, and settling on Base right now.

    What This Means for Finance and Business

    The implications for commercial operations are significant and immediate.

    Autonomous procurement. An AI agent managing a portfolio company’s data pipeline can independently source, evaluate, and pay for third-party intelligence. No procurement team. No vendor onboarding. No purchase orders. The agent identifies the need, finds the cheapest or highest-quality provider on Agentic.market, pays per request, and moves on.

    Real-time cost control. Every transaction is on-chain and auditable. Spend is granular — per request, per task, per agent. Finance teams get cost visibility that traditional SaaS contracts cannot offer. No more annual licenses for services used sporadically.

    Agent-to-agent supply chains. When agents buy from agents, you get supply chains that form and dissolve in seconds. A research agent pays a data agent, which pays a scraping agent. Each transaction is atomic, priced, and settled. The entire chain is transparent.

    For PE-backed businesses running lean, this is a step change. Operating costs shift from fixed contract commitments to variable, usage-based, and fully auditable micro-transactions.

    The Investment Angle — Rails vs Cargo

    This is where most analysts are getting it wrong. The instinct is to bet on the cargo — the individual services and agents. That is a mistake. The cargo will be commoditised. The rails will not.

    Coinbase (COIN) owns Base, the network processing 85% of x402 volume. Every transaction generates fee revenue. As agent transaction volume scales — and 167 million is just the beginning — COIN captures a percentage of every settlement. This is infrastructure-layer economics with network effects.

    Ethereum (ETH) benefits as the settlement and security layer underpinning Base. More Base activity means more value flowing to Ethereum. But the connection is indirect, and ETH carries broader market beta.

    USDC is the working capital of the agentic economy. Agents hold it, spend it, receive it. But stablecoins are pegged — they do not appreciate. USDC is essential infrastructure, but it is not an investment.

    COIN is the cleanest listed play on the agentic economy. It owns the dominant rail, captures fee revenue at scale, and is already publicly traded with institutional-grade liquidity.

    This Is Happening Now

    One hundred and sixty-seven million transactions. Eighty-five percent on a single L2. A functioning marketplace. Live commercial services settling in USDC.

    The agentic economy is not a whitepaper. It is not a keynote prediction. It is an operational, scaling, measurable economy — and the window for building conviction before the market fully prices this in is narrowing.

    If you are a CFO, a GP, or an operator and you are not tracking this, you are already behind.

  • The AI Just Got a Computer — And Your Competitors Still Think It’s a Chatbot

    The AI Just Got a Computer — And Your Competitors Still Think It’s a Chatbot

    Yesterday Changed Everything

    On 17 April 2026, xAI launched Grok 4.3 Beta. Most headlines focused on benchmark scores. They missed the point entirely.

    Grok now has a full Ubuntu shell built into the product. Not a sandboxed code snippet runner. Not a “try this Python” widget. A genuine Linux computing environment where the AI can execute commands, install packages, write and run code, and manage files — with a persistent file layer that survives between sessions.

    To demonstrate the capability, Grok encoded the xAI logo into audio frequencies, rendered a spectrogram video from the result, and saved the finished MP4 to persistent storage. No human touched the keyboard after the initial prompt.

    If you’re a CFO reading this and thinking “interesting, but not relevant to me yet” — you’re already behind.

    This Isn’t an Upgrade. It’s a Category Shift.

    For the past three years, AI has been a sophisticated text box. You type a question, it gives you an answer. Useful, yes. Transformational? Only if your definition of transformation is “slightly faster email drafting.”

    What happened yesterday is different in kind. AI stopped being a tool you query and became an agent that executes. It can now build a financial model, test it, debug it, save the output, and iterate — autonomously. That’s not a chatbot. That’s a computing environment with intelligence baked in.

    Claude Code from Anthropic and OpenAI’s Codex are racing in the same direction. But Grok 4.3 Beta delivers this natively, inside the product, with persistent state. No API configuration. No developer setup. It just works.

    What This Means for Portfolio Companies

    If you sit on a PE board or run a finance function inside a portfolio company, here’s the translation:

    Automation just got autonomous. Previously, automating a finance process meant scoping a project, hiring a consultant, building an integration. Now, an AI agent can take a description of what you need, write the code, test it, and deliver the output — in minutes. Month-end reconciliation workflows that took weeks to automate can now be prototyped in an afternoon.

    The talent gap just narrowed — and widened. The CFO who understands how to direct an AI computing environment will deliver more with a team of five than a competitor delivers with fifteen. The CFO who doesn’t will need those fifteen people just to keep up.

    Build vs. buy just flipped. When your AI can write, test, and deploy code, the business case for buying off-the-shelf SaaS tools weakens dramatically. Why pay six figures annually for a reporting platform when an AI agent can build a bespoke one tuned to your exact data structure?

    The Two-Year Illusion

    Every board deck I’ve seen in the past twelve months includes some version of “AI roadmap — 18-24 month horizon.” That timeline assumed AI would keep improving incrementally. It assumed you’d have time to hire a Head of AI, run a pilot, form a committee.

    Grok didn’t improve incrementally. It gained the ability to use a computer. That’s not a point on a curve. That’s a step function.

    The companies that will win from here are the ones whose leadership understands a simple truth: AI is no longer something your team uses. It’s something that works alongside your team — writing code, running analyses, building tools, and saving its work for next time.

    The Finance Function Specifically

    For CFOs and FDs, this is where it gets concrete. An AI with a persistent computing environment can:

    • Pull data from multiple sources, clean it, and produce a consolidated management pack — every month, automatically
    • Build and maintain bespoke Python-based forecasting models that improve with each iteration
    • Run scenario analyses across portfolio companies in parallel, saving outputs for board review
    • Automate the grunt work of audit preparation — file organisation, reconciliation testing, variance analysis

    None of this required a developer. None of it required a software vendor. The AI did the work.

    What To Do On Monday Morning

    Stop treating AI as a future initiative. It became a present capability yesterday.

    Three actions for this week:

    1. Try it yourself. Go to grok.com, open the shell environment, and ask it to build something specific to your business. A cash flow model. A data reconciliation script. See what happens.

    2. Identify one process. Pick a single finance process that’s manual, repetitive, and painful. Brief your AI on it. Let it prototype a solution.

    3. Rewrite your AI roadmap. If your current plan assumes AI is 18 months away from being useful, your plan is wrong. Rewrite it with the assumption that AI can execute work today — because it can.

    The AI just got a computer. The question is whether your competitors noticed before you did.


    If your portfolio companies need help understanding what AI computing environments mean for their finance functions and operations, get in touch.

  • The Convergence: Why Energy and Food Security Are Failing at the Same Time

    The Convergence: Why Energy and Food Security Are Failing at the Same Time

    April 2026 will be remembered as the month when energy and food security converged into a perfect storm. On the 15th, Australia’s Viva Energy refinery in Geelong exploded—one of two left standing Down Under. Days earlier, the UK exposed Russian GUGI submarines hovering over vital undersea cables. A year prior, Spain and Portugal endured a massive blackout exposing renewable grid frailties. Meanwhile, H5N1 has culled 200 million U.S. birds since 2022 and infiltrated dairy herds.

    These aren’t coincidences. They’re the intersection of deliberate state aggression, systemic brittleness, and biological disruption. Energy powers food production and distribution; food sustains energy workers and economies. When both falter simultaneously, cascading failures ensue.

    Geelong: Australia’s Fuel Heart Ripped Out

    Australia has shuttered refineries aggressively, dropping from 10 in 2000 to two by 2026. Geelong (120kbpd) was critical, supplying Victoria and Tasmania.

    The April 15 explosion—cause under investigation (rumors of cyber or insider)—sent a 100m fireball skyward. No fatalities, but operations halted indefinitely. Fuel imports, already 90% of supply, face tanker bottlenecks amid Red Sea tensions.

    Prices: diesel up 28%, jet fuel rationed. Knock-ons: mining halts, grocery delivery delays. The Age reports.

    Russia’s Sabotage Blitzkrieg

    2025 saw 321 confirmed sabotage acts in Germany: rail arson (Deutsche Bahn chaos), chemical plant fires, grid hacks. Cost: €2bn+.

    Escalation: April 2026, UK Defence Secretary revealed GUGI subs (Yantar-class spy ships) mapping cables carrying 99% of transatlantic data/power. Akula decoy trailed. RAF/ Navy shadowed; Russians fled.

    Baltic parallel: C-Lion1 and BCS East-West cables severed, Russian shadow fleet implicated. Breaking Defense | MOD | Baltic cables (prior)

    Iberian Collapse: Renewables’ Reckoning

    April 2025: 50M in dark. REE autopsy: 17 factors—inter-solar faults, wind drop-off (from 40% to 2%), hydro drought, interconnector trips. Renewables hit 70% pre-fail; fossils lagged.

    Lessons: Inertia shortfall, no spinning reserves. Cost: €10bn GDP hit. Renewables push ignored baseload needs. El Confidencial analysis.

    H5N1 Pandemic: Protein Supply Implodes

    100M+ layers culled 2022-2025; total birds 200M+. Egg prices tripled. Now cattle: 100+ herds, milk discard mandates. 5% U.S. dairy output at risk.

    Supply chain: Poultry 40% U.S. protein; disruptions ripple to feedlots (soy/corn via energy-intensive transport). USDA detections | FDA

    Convergence Dynamics: Modeling the Chaos

    Energy-food nexus:

    Threat Energy Impact Food Impact
    Sabotage Fuel/power loss Transport halt
    Blackouts Direct Refrigeration fail
    Flu Feed/transport strain Protein shortage
    Explosion Fuel scarcity Farm ops stop

    AI Monte Carlo: Base case 25% food inflation; stress 50%+. PE exposure: ag (15% drawdown), logistics (20%).

    PE Action Plan: Resilience Over Resilience

    1. Reprice assets: Add 5-10% hybrid risk premium to infra yields.

    2. Capital shift: 20% to small modular reactors (SMRs), 15% vertical ag, 10% subsea hardening.

    3. AI diligence: Simulate portfolio cascades.

    Mark Hendy | PE CFO | Geopolitics + AI

  • Visa Just Gave AI Agents a Credit Card. CFOs Should Be Paying Attention.

    Last week, Visa announced Intelligent Commerce Connect — a platform that lets AI agents initiate purchases, handle tokenisation, enforce spend controls, and authenticate payments on behalf of users. Not users clicking a button. Not users confirming a pop-up. Agents. Autonomously. On Visa’s network.

    I’ve been building with AI agents for the better part of two years now. I’ve got agents managing my inbox, drafting blog posts (hello), monitoring portfolios, and scheduling meetings. But the moment you give an agent a credit card, something fundamentally shifts. This isn’t automation anymore. This is delegation of financial authority. And if there’s one thing twenty years in finance has taught me, it’s that delegating financial authority without governance is how you get fired.

    What Visa Actually Built

    Intelligent Commerce Connect is designed as a network-agnostic, protocol-agnostic on-ramp for agentic commerce. Through a single integration via Visa’s Acceptance Platform, it enables AI agents to initiate payments using both Visa and non-Visa cards. It supports multiple agent protocols — Trusted Agent Protocol, Machine Payments Protocol, Agentic Commerce Protocol, and Universal Commerce Protocol — which tells you a lot about where the industry expects this to go. There isn’t one protocol yet. There are four, and Visa is hedging by supporting all of them.

    The platform is in pilot with partners including AWS, Highnote, Mesh, and Payabli, with general availability expected by June. That’s not a research paper. That’s a shipping product on the world’s largest payment network.

    The CFO’s Nightmare Scenario

    Here’s where my CFO brain starts twitching. OutSystems research published this month found that 94% of organisations deploying agentic AI are already concerned about sprawl — agents proliferating across the business faster than governance can keep up. Now imagine those ungoverned agents have purchasing authority.

    In any well-run finance function, there’s a concept called a delegation of authority matrix — a document that says who can approve what, up to what amount, under what conditions. It’s boring. It’s bureaucratic. And it’s the single most important control preventing your procurement team from accidentally (or deliberately) buying things they shouldn’t. Every auditor checks it. Every PE firm’s due diligence team asks for it.

    The question Visa’s announcement forces us to ask is: what does a delegation of authority matrix look like when the “who” is an AI agent?

    Spend Controls Are Necessary But Not Sufficient

    To be fair, Visa’s platform does include spend controls and authentication. And the enterprise AI governance frameworks I’m seeing — Deloitte’s CFO tech guide is a decent starting point — generally recommend that any AI action above a monetary threshold requires human sign-off. The ERP stays the system of record. The agent proposes, the human approves.

    That’s fine for the first generation. But it won’t hold. The whole point of agentic AI is removing humans from routine decision loops. If every purchase order still needs a human clicking “approve,” you haven’t automated procurement — you’ve just added an extra step. The economic pressure to raise those thresholds, to widen the autonomy boundaries, will be relentless. McKinsey estimates that autonomous procurement agents can capture 15–30% efficiency improvements by eliminating non-value-added activities. No CFO under PE ownership is leaving that on the table.

    So the thresholds will creep up. The approval requirements will relax. And one morning, a finance director will discover that an agent negotiated a twelve-month SaaS contract at 3am because it determined the vendor’s dynamic pricing was optimal at that hour. Good luck explaining that to the audit committee.

    What I’m Actually Worried About

    It’s not fraud — Visa’s tokenisation and authentication should handle that tolerably well. What concerns me is the combination of three things:

    Compounding commitments. An agent optimising for one metric (say, cost-per-unit) might make individually rational purchasing decisions that collectively create an overcommitted balance sheet. No single purchase triggers a threshold. But the aggregate position is one a human would have caught.

    Vendor lock-in at machine speed. If agents are negotiating contracts, they’ll optimise for the parameters they’re given. Unless you’ve explicitly programmed “maintain optionality” as a constraint, they’ll happily lock you into the cheapest long-term deal every time. Strategic flexibility is a human judgement call that agents don’t naturally make.

    Audit trail legibility. Right now, if I ask a PE firm’s portfolio company why they spent £2m with a particular vendor, someone can explain the reasoning. When an agent made that call based on a multi-variable optimisation that factored in 47 data points at 3am, the “reasoning” is a probability distribution. Try putting that in a board pack.

    The Parallel With Prediction Markets

    There’s an interesting echo here with what’s happening on Polymarket. This month, Bloomberg reported that $170 million flowed through Iran ceasefire bets, with at least 50 freshly created accounts placing suspiciously well-timed wagers before Trump’s announcement. Lawmakers are now asking whether prediction markets can even govern themselves.

    The pattern is the same: a powerful new mechanism for allocating capital, moving faster than the governance structures around it. Prediction markets and agentic commerce are both examples of what happens when you remove humans from the decision loop in systems that move money. Speed goes up. Friction goes down. And the attack surface for bad actors — or simply for well-intentioned software making collectively poor decisions — expands dramatically.

    What I Think Happens Next

    Visa’s move is the starting gun, not the finish line. By the end of 2026, I’d expect to see:

    Agentic audit frameworks becoming a real discipline — not just “we log what the agent did” but actual real-time monitoring of agent purchasing patterns against policy. The Big Four will sell this as a service by Q4.

    CFOs who’ve been ignoring AI agents will suddenly care a great deal, because an agent with a Visa card is no longer an IT project — it’s a financial control issue that sits squarely in their remit.

    And at least one spectacular failure. Some company will give agents too much rope, the agents will collectively do something that’s individually rational but organisationally stupid, and it’ll end up in the FT. That’s not pessimism. It’s pattern recognition. Every new financial instrument goes through this cycle.

    For now, I’m building my own agent governance framework — spend limits, approval chains, real-time anomaly detection — because I’d rather design the guardrails before Visa’s platform goes GA in June than after. If you’re a CFO reading this and you’re not thinking about what happens when AI agents can spend money on your company’s behalf, you’re already behind.

  • When the Machine Fires the Customer

    When the Machine Fires the Customer

    Mo Gawdat, former Google [X] Chief Business Officer, lands a haymaker via a viral thread by @r0ck3t23: AI dismantles capitalism’s core—labor arbitrage. The model? Pay humans $1 for their time, sell the output for $2. AI drives the input cost to zero. Game over.

    In my world as a PE interim CFO, I’ve stress-tested portfolios against AI disruption. Gawdat’s arithmetic is spot-on. But the true dynamite is the vicious cycle he uncovers: workers are customers. Fire them, and you torch demand. No hysteria—just cold finance logic. Let’s break it down.

    Labor Arbitrage: The Engine of Empire

    Strip capitalism bare: it’s arbitrage on human effort. Factories arbitrage $12/hour assemblers into $100 widgets. SaaS firms turn $150k devs into $10M ARR. The spread—profit—fuels empires. Private equity thrives here: buy labor-intensive businesses, optimize (read: squeeze), flip.

    Gawdat puts it bluntly: “The very base of capitalism, which is labor arbitrage… is going to disappear.” Global supply chains exemplify: iPhones assembled in Foxconn for pennies on the dollar, sold at premiums. AI? LLMs generate code at amortized $0.01/query. Humanoids like Figure 01 (~$20k amortized over lifetime, not $9k day-one, but close enough) run endless shifts.

    Result: Production cost floor vanishes. No wages, no unions, no sick days. Your LBO model? Redo the cost of goods from 40% margin to 95%.

    The Deadly Feedback Loop

    Gawdat’s killer insight: Employees = End consumers. Automate white-collar? Lawyers, analysts, marketers—gone. Goldman Sachs saved $1B with AI, laid off juniors. Scale to economy: IMF projects 40% jobs exposed. Unemployment spikes to 30-50%? Consumer spending craters.

    Your portfolio company automates DC ops, cuts 20% headcount. Short-term EBITDA pop. Long-term: Those ex-workers skip Black Friday. Infinite AI supply chases evaporating demand. Companies quietly assassinate their customer base.

    Gawdat’s Hits and Misses

    Hits Hard: Feedback loop unassailable. Echoes Keynes’ technological unemployment, but turbocharged. Rust Belt 2.0, global.

    Oversimplifications:

    • $9k robots: Hype. Real capex + opex higher short-term. But trajectory undeniable—costs plunging.
    • Scarcity evaporates: Bold leap. Compute, rare earths, energy constrain. Abundance requires ITER-scale fusion.
    • Transition blindspot: UBI from robot taxes? Congress gridlock. “Hell before heaven,” Gawdat says—12-15 years chaos per BI.

    CFO Pivot: From Ops to Oracle

    Not apocalypse—repositioning. Human value migrates upstack.

    Judgment Trumps Jargon: AI spits forecasts. You discern signal from noise: “This uptick? Competitor distress sell.” Relationships seal deals. Authority greenlights bets. AI can’t schmooze VCs or stare down boards.

    Operational CFOs Vulnerable: Rollups, variance reports—AI devours. McKinsey flags 45% finance exposure.

    Capital Ownership Essential: Wage slaves sink. Equity kings rise. PE CFOs: Deploy into AI infra, not legacy labor.

    AI Fluency = 2x Comp: Model agentic workflows, capex ramps. Prompt strategic insights. Demand surges—I’ve seen packages jump 50%.

    Survivors arbitrage human judgment + AI horsepower.

    Ready for the Reckoning? Audit your PE portfolio’s AI exposure. Book 30-min call. Let’s harden your numbers.

  • GLM-5.1: The Chinese Open-Source Model That Just Beat GPT and Claude at Their Own Game

    GLM-5.1: The Chinese Open-Source Model That Just Beat GPT and Claude at Their Own Game

    Something significant happened in the AI landscape this week, and I suspect it hasn’t got the attention it deserves outside of developer circles. Z.AI — the platform behind the GLM model family, developed by Zhipu AI in China — released GLM-5.1, a 754 billion parameter open-source model that has just topped the SWE-Bench Pro leaderboard with a score of 58.4, beating GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro.

    Let that land for a moment. An open-source, MIT-licensed model, trained entirely on Huawei Ascend 910B chips — no Nvidia, no American silicon — has beaten the flagship closed models from OpenAI, Anthropic, and Google on one of the most respected software engineering benchmarks in existence.

    What Makes GLM-5.1 Different

    The headline number is impressive, but what actually interests me is the architecture of how this model works. GLM-5.1 isn’t just better at answering questions — it’s designed for sustained autonomous execution. In testing, it completed an eight-hour uninterrupted coding session: plan, execute, test, optimise, repeat. 655 iterations. Built a Linux desktop environment from scratch. Increased vector database query throughput by 6.9 times.

    This is a different category of capability. We’re not talking about a better chatbot. We’re talking about an AI that can hold a task in mind, work through it independently, hit dead ends, correct course, and deliver a finished result — the way a competent junior engineer would, but without stopping for the night.

    The technical foundation is a Mixture-of-Experts architecture with 40 billion active parameters per token (not all 754B are active at once, which is what keeps inference costs manageable). It supports a 200,000 token context window with up to 128,000 output tokens. API access is priced at $1.00 per million input tokens and $3.20 per million output tokens — a fraction of what the US frontier models charge.

    Why This Matters Beyond the Benchmarks

    I’ve written before about AI moving from a tool you prompt to a system that acts. GLM-5.1 is a concrete illustration of that shift happening faster than most people expected, and from a direction many in the West weren’t watching closely.

    The geopolitical dimension is real. This model was trained on Huawei hardware using Huawei’s MindSpore framework — a deliberate demonstration that China’s AI development pipeline is no longer dependent on US export-controlled chips. The export restrictions that were supposed to slow Chinese AI development have instead accelerated domestic alternatives. That is a significant strategic development, regardless of where you sit on the AI competition question.

    The open-source dimension is equally significant. With weights published under an MIT licence, GLM-5.1 can be downloaded, fine-tuned, and deployed by anyone. The closed-model advantage that OpenAI and Anthropic have built commercial moats around is being systematically eroded — not just by each other, but by well-resourced open-source releases like this one.

    What I Take From This

    I use AI heavily in my work — for financial analysis, document preparation, research, and increasingly for autonomous background tasks. The pace at which these systems are improving is not slowing down. If anything, GLM-5.1 suggests the competitive field is widening: more players, more approaches, more open options.

    For anyone running a business or advising one, the practical implication is straightforward: the cost of access to frontier-level AI capability is falling rapidly, and the choice of provider is expanding. The question is no longer whether to use these tools — it’s which ones, for what, and how to build processes around them that compound over time.

    GLM-5.1 is worth watching. Not because it’s the final word, but because it’s a clear signal that the race is genuinely global, the open-source movement is closing the gap faster than expected, and the next twelve months are going to be interesting.


    GLM-5.1 is available via z.ai on the GLM Coding Plan, with weights on Hugging Face under MIT licence.

  • Polymarket’s Insider Trading Scandal Is the Best Case for Prediction Markets

    Fifty brand-new Polymarket accounts placed large bets on a U.S.–Iran ceasefire in the hours before Trump announced it on social media. Some of those accounts turned a few thousand dollars into six figures overnight. Congress is furious. Senator Blumenthal wants investigations. The hot take is that prediction markets are broken.

    I think the opposite is true. This is prediction markets working exactly as designed — and the reaction tells you more about the people reacting than it does about the platform.

    The Numbers Are Genuinely Wild

    Let’s be clear about what happened. Over $170 million flowed through Polymarket’s Iran ceasefire markets, making it one of the largest geopolitical wagers in prediction market history. Blockchain analytics firm Lookonchain flagged three freshly created accounts that collectively pocketed more than $480,000 by betting on a ceasefire before selling at the top. One account — the now-infamous “Magamyman” — placed its first-ever trade seventy-one minutes before news broke, when the market implied only a 17% probability. It walked away with roughly $553,000.

    A Harvard study published last month went further, screening over 93,000 Polymarket markets and nearly 50,000 wallet addresses from 2024 to early 2026. Their finding: across 210,000 flagged wallet-market pairs, suspicious traders achieved a 69.9% win rate — more than 60 standard deviations above chance. Total estimated profits from potentially informed trading: $143 million.

    Those numbers look damning. But here’s the thing that nobody seems to want to say out loud.

    We Only Know Because It’s on a Blockchain

    Every single one of these trades is visible. Timestamped. Publicly auditable. A Harvard research team could sit down and systematically screen two years of trading data because it’s all on-chain. The accounts are pseudonymous, sure, but the trades themselves are transparent in a way that nothing in traditional finance comes close to matching.

    Now think about how insider trading works in equities. Someone at a law firm hears about a merger. They tell their cousin. The cousin buys call options through a brokerage account that takes the SEC years to subpoena — if they ever notice at all. The SEC estimates it catches a fraction of actual insider trading. Academic studies suggest informed trading precedes something like 25% of major M&A announcements. We just don’t see it, because the infrastructure is designed for opacity.

    Polymarket’s “scandal” is that informed trading is visible in real time. That’s not a bug. That’s the entire point of building financial infrastructure on transparent ledgers. Congress is effectively complaining that the system is too transparent.

    The Real Question Isn’t Whether People Traded on Information

    Of course people traded on information. That’s what markets are for. The interesting question is: who had the information, and should they have been allowed to trade on it?

    If White House staffers or intelligence officials were betting on ceasefire outcomes they helped negotiate, that’s a serious problem — but it’s a problem of government ethics, not of prediction markets. The White House reportedly warned staff in late March against trading on prediction markets. Which rather suggests they knew it was happening.

    But if journalists, diplomats, or well-connected analysts were trading on information they’d gathered through legitimate means? That’s price discovery. That’s exactly how you want markets to function. The ceasefire contract moving from 17% to near-certainty before the official announcement is the market doing its job — aggregating dispersed information faster than any single news outlet could.

    When Polymarket’s election markets outperformed every major polling model in 2024, we celebrated prediction markets as an information revolution. Now that the same mechanism is surfacing uncomfortable truths about who knows what in Washington, we want to shut it down. You can’t have it both ways.

    What This Means for the Platform

    Polymarket clearly knows it’s at a crossroads. Last week they announced a full exchange upgrade — a rebuilt trading engine, updated smart contracts, and a new USDC-backed collateral token called Polymarket USD. They’re preparing for serious U.S. expansion, which means they’re preparing for serious U.S. regulation.

    The smart move would be some form of voluntary KYC tier for large positions on geopolitically sensitive markets. Not because prediction markets should be restricted, but because demonstrating that you can identify bad actors — government officials trading on classified information, for instance — is how you survive regulatory scrutiny. The blockchain already gives you the trade data. Add identity to the large positions and you’ve got something the SEC can only dream of for equity markets.

    The CFO Angle (Because There’s Always a CFO Angle)

    I spend my days working with PE-backed businesses where forecasting accuracy is everything. We obsess over budget variance, rolling forecasts, scenario models. And here’s a market that just priced a geopolitical event more accurately and more quickly than any intelligence briefing, any analyst note, any Bloomberg terminal alert.

    If you’re a CFO running scenario planning on geopolitical risk — and if you’re in any business exposed to energy prices or supply chains, you should be — prediction markets are increasingly a better input than traditional sources. Polymarket’s finance predictions already cover oil prices, rate decisions, and major policy outcomes. The Iran ceasefire market moved Bitcoin from $67,000 to $72,700 and back. WTI crude is sitting above $110. These aren’t abstract bets — they’re real-time consensus probability that feeds directly into the models I build every week.

    The Harvard researchers found statistical anomalies. Congress found a talking point. But what I found was a market that processed a ceasefire probability faster than Reuters could file a story. And I’ll take that signal every time.

    Where This Goes

    Prediction markets aren’t going away. Polymarket is a $20 billion platform now. The CLARITY Act markup later this month will shape the U.S. regulatory framework, and the smart money — no pun intended — is on a framework that legitimises these markets with guardrails rather than banning them.

    The insider trading debate will rage on. But every time someone points to suspicious Polymarket trades as evidence that prediction markets are dangerous, remember: the only reason we can see those trades is because the system is transparent. The alternative isn’t a world without insider trading. It’s a world where insider trading happens in the dark.

    I know which one I prefer.

  • The SaaSpocalypse Is Real — But the Market Is Panicking About the Wrong Thing

    I spent last week reviewing the tech stack costs across three portfolio companies. The exercise used to be straightforward: count seats, multiply by price, negotiate volume discounts. This time, two of the three CFOs asked me the same question: "Should we be cancelling licences and moving to agents?"

    That question — and the speed at which it has gone from fringe to board-level — tells you everything about where we are in April 2026.

    $285 Billion in 48 Hours

    If you have been anywhere near a Bloomberg terminal this year, you know the term SaaSpocalypse. It started in late January when Anthropic shipped Claude Cowork with industry-specific agent plugins — legal contract review, financial analysis, sales automation — and the market did what markets do: it extrapolated to infinity.

    Bloomberg reported roughly $285 billion wiped from SaaS valuations in a single 48-hour window. Thomson Reuters dropped 15%. LegalZoom cratered nearly 20%. By mid-March, the IGV software ETF was down over 21% year-to-date, and analysts were calling it the largest AI-triggered repricing in software history.

    The logic was brutally simple. If an AI agent can do the work of five humans, why pay for five seats? The per-seat pricing model — the entire economic foundation of B2B SaaS since Salesforce invented it — was suddenly an existential vulnerability.

    What the Market Got Right

    Let me be clear: the structural thesis is correct. Per-seat pricing is dying. I have seen it in our own portfolio.

    One of our companies ran a pilot replacing three junior paralegals' document review work with an AI agent pipeline. The agent does not need a LegalZoom subscription, a DocuSign seat, or a Westlaw login in the traditional sense. It calls APIs, processes documents, and routes exceptions to a human. The annual software cost for those three "seats" — roughly £45,000 — dropped to about £8,000 in LLM API costs.

    That is not a marginal improvement. That is a different business model.

    The survey data backs this up: 40% of IT budgets are reportedly being reallocated from traditional SaaS subscriptions to agentic platforms and token-based usage. CIOs are not asking "how many employees will use this?" anymore. They are asking "how many tasks can this complete?" That is a fundamental shift in procurement psychology, and SaaS companies built on headcount-correlated revenue should be worried.

    What the Market Got Wrong

    Here is where it gets interesting — and where I think the panic has overshot.

    The market treated the SaaSpocalypse as if every SaaS company is equally exposed. They are not. There is a massive difference between a company that sells seats for humans to click buttons and a company that sells the underlying data, workflow engine, or integration layer that agents also need.

    Thomson Reuters does not just sell a UI for lawyers. It sells access to legal databases, case law, and regulatory intelligence. An AI agent doing contract review still needs that data. The delivery mechanism changes, the underlying value does not. Same story with ServiceNow — the Motley Fool piece this week calling it a bargain has a point. Workflow orchestration becomes more valuable when you have agents that need orchestrating, not less.

    The companies that are genuinely toast are the ones that were essentially selling a graphical interface on top of commodity functionality. If your product is a pretty wrapper around CRUD operations and your moat was user habit, then yes, an agent that calls the same APIs without the wrapper is an existential threat. But that is not every SaaS company — it is maybe 30% of them.

    What This Means If You Are a CFO

    Here is my practical take, from someone currently navigating this across multiple PE-backed businesses:

    Audit your stack ruthlessly, but intelligently. Do not just cancel licences because agents are trendy. Map each SaaS tool to what it actually provides: is it data, workflow, integration, or just interface? The first three categories will likely survive the transition. The fourth will not.

    Start modelling token-based costs now. The shift from per-seat to per-task pricing is real, but token economics are volatile and opaque. I have seen API costs swing 30% month-on-month as providers adjust pricing. You need a cost model that accounts for this, and you need someone on your team who understands it — not just a vendor's sales estimate.

    Watch the middleware layer. The real winners of the agentic transition might not be the agent builders themselves. Microsoft's Agent Framework 1.0, released last week, unifies Semantic Kernel and AutoGen into a production-ready orchestration layer. That is the plumbing that enterprises will standardise on. If you are making build-vs-buy decisions on agent infrastructure, this is the framework to evaluate first.

    Do not mistake a repricing for a revolution — yet. Most enterprises are still in pilot mode. The 40% budget reallocation figure is aspirational, not actual. In our portfolio, the company furthest along has moved maybe 12% of its SaaS spend to agent-based alternatives. The rest are running proofs of concept. The gap between "we are exploring AI agents" and "we have decommissioned Salesforce" is about three years of integration work and change management.

    The PE Angle

    For anyone in private equity, the SaaSpocalypse is creating a genuinely interesting buying opportunity. High-quality SaaS businesses with real data moats and sticky integration layers are trading at 2022-era multiples. If you believe — as I do — that the best SaaS companies will successfully transition to hybrid pricing models (seats plus tokens plus outcomes), then the current discount is mispriced fear.

    The businesses to avoid are the ones with high seat counts, low switching costs, and functionality that an off-the-shelf agent can replicate. You know the type: they raised a Series B on "AI-powered" features that were really just a ChatGPT wrapper bolted onto a form builder.

    The SaaSpocalypse is real. But like most market panics, it is painting with too broad a brush. The death of per-seat pricing does not mean the death of software businesses. It means the death of lazy software businesses. And frankly, most of those were overdue a correction anyway.