A bi-weekly speculative fiction suggesting the shape of things to come.
(sourced from trustworthy trade pubs, think tanks + frontier science news)
This fortnight brought 3,033 signals across corporate, startup, thinktank, and research sources into focus, and the throughline is disintermediation — the layers that used to sit between a company and its outcome quietly dissolving. The storefront got disintermediated: Walmart opened a gateway for outside AI agents to shop on customers' behalf, Google launched a Universal Commerce Protocol to standardize how machines read catalogs, and Ulta and Fanatics moved the buy button inside ChatGPT — leaving brands to wonder whether they still own the customer relationship or merely a SKU in someone else's answer. The factory worker got disintermediated: Tesla began humanoid production in California, Siemens and NVIDIA switched on an autonomous electronics plant, and robots started learning tasks by watching rather than being programmed. The doctor's gatekeeping got disintermediated: an FDA official floated judging generative-AI tools 'like a physician,' CMS proposed paying for software that reads your scans, and hospitals pushed to hold developers — not doctors — liable. The grid got disintermediated from its own planning assumptions, as $228 billion of new gas plants were conscripted to feed data centers whose demand forecasts had been off by 600%. And most fundamentally, the open web got disintermediated: ChatGPT now sends less than 1% of Google's referral traffic even as shopping journeys increasingly start there, airlines and retailers found themselves invisible inside the answer box, and publishers from Reddit to Time began walling content or serving ads directly to the agents. Against all this, a counter-current: a mounting ROI reckoning, with Apollo finding zero AI payoff outside Big Tech and Gartner reporting companies aren't actually saving money. Twenty fortnights in, the pattern is clear — AI isn't just automating tasks, it's removing the middle. The question every enterprise now faces is whether it's the layer being removed.
AI agents started shopping on your customers' behalf, and the companies that spent decades owning the storefront are now fighting for a mention inside someone else's answer.
The most consequential retail shift of the fortnight wasn't a new store or a new app — it was the quiet arrival of the customer's proxy. Walmart launched the Walmart Agent Gateway, an early beta that connects outside AI agents directly to its commerce platform, paired with Agent Profiles that let shoppers spin up dedicated AI shoppers with spending limits and purchase-approval controls — and updated its terms to bar agents from touching payment-card data. Google went structural, announcing the Universal Commerce Protocol (UCP) with Shopify, Etsy, Wayfair, Target, and Walmart — a global standard for how AI agents read a store's catalog, pricing, and availability — and positioned itself not as a retailer but as a 'commerce matchmaker,' forcing brands to adapt to a protocol that prioritizes machine discoverability over the brand's own front door. Consumer brands rushed to get inside the conversation: Ulta Beauty shipped a ChatGPT plugin that builds editable, personalized product bundles (with a 10% discount to close the sale on Ulta.com), and Fanatics rebuilt its 'Shop the Look' outfit tool on OpenAI models. But the banks drew a line: JPMorgan's Jamie Dimon said the firm won't let outside agents like Meta's into customer accounts, and the FDIC flagged the deposit-flight risk of agentic banking — a reminder that when the agent becomes the shopper, the question of who it's loyal to, and who's liable, is suddenly very real.
Humanoids crossed from demo reel to production line this fortnight — and the leap wasn't better hardware, it was robots that learn a task by observing instead of being programmed.
Physical AI hit an inflection point. Tesla began Optimus humanoid production in California, with next-generation AI5 chips from its Texas Terafab slated to supercharge the robots' on-board compute. Siemens and NVIDIA switched on an autonomous future at Siemens' Erlangen electronics factory, where AI agents built on NVIDIA's blueprints coordinate production and quality inspection — cutting engineering effort on production ramp-up by up to 85% through fully virtual commissioning. The deployments got real and repeatable: GXO and Agility Robotics expanded Digit humanoids across US warehouses, BMW's Landshut plant began developing open-platform software and training data for humanoids, and Audi and FAW-Volkswagen started testing them in logistics. New factories exist only to build the builders: Italy's Generative Bionics is erecting a 'Physical AI Data Factory' to generate the training data its humanoids need. And the form factor reached the living room, as New York's Tuesday Labs debuted Voxel, a semi-autonomous 'droid butler' for household chores. The common thread across all of it: these machines increasingly learn from demonstration, simulation, and world-model data rather than hand-coded instructions.
The fight over AI's raw material went from abstract to brutal this fortnight — 1,050 songs in a new lawsuit, a studio buying a chatbot while dodging its liabilities, and an AI actress doing press.
The creative industries spent the fortnight litigating, acquiring, and bracing. Reservoir Media sued Suno and Anthropic on October 21, alleging they copied 1,050 of its compositions — songs by Joni Mitchell, Johnny Cash, and Madonna — to train AI models. In the parallel Universal Music Group and Sony case, Suno admitted it has not deleted all the sound recordings used to train its models, while the labels moved to compel Suno investor Lightspeed and distributor DistroKid to hand over data on the 'legal risk of training data.' The strategy to acquire the capability while quarantining the risk went on full display: Disney bought CharacterAI's software assets but structured the deal to explicitly exclude the startup's legal liabilities. New synthetic talent kept arriving — Particle6's AI actress Tilly Norwood gave press interviews — and the terms of future training quietly widened as Epic Games updated its EULA to permit AI training on Unreal Engine and MetaHuman outputs. Abroad, the ICMP accused AI firms of trying to 'strong-arm' Australia over copyright, revealing Kylie Minogue's songs had already been used in training. The counter-move was licensing: Spotify inked a deal with Believe (and UMG, Merlin, Kobalt) for an AI covers-and-remixes tool that links every remix back to the original.
Regulators stopped treating clinical AI as software and started treating it as a practitioner — proposing to judge it like a physician, pay it like a service, and sue its maker like a doctor.
A quiet but profound reclassification of medical AI unfolded across the fortnight. At The MedTech Conference, Rick Abramson, director of the FDA's Digital Health Center of Excellence, said generative-AI-enabled tools may need to be evaluated 'like a physician' rather than as traditional software — a conceptual leap from device to practitioner. The money followed the logic: CMS proposed a new payment category, 'Software as a Medical Service,' to reimburse AI tools that read scans, EKGs, or charts and return diagnoses or risk scores, with 36 new codes in the first wave. The liability question sharpened as the American Hospital Association, AMA, and American Telemedicine Association told the FDA that developers, not doctors, should be held accountable for generative-AI medical devices, while the AMA separately pressed for risk-based oversight with ongoing performance monitoring. The capability is already in the exam room: the FDA cleared SimBioSys' TumorSight Risk for breast-cancer recurrence, Doximity launched DoseCheck after finding popular general AI models get medication doses dangerously wrong, and Google published a clinical study of AMIE, an AI agent that conducts patient conversations before appointments.
The electricity math finally caught up with the compute math — $228B of gas plants got conscripted to feed data centers whose demand had been underestimated by 600%.
The physical constraint on AI became impossible to ignore. US grid planners revised their estimate of additional electricity demand within a decade from 24 GW in 2022 to 166 GW by 2025 — an underestimate of roughly 600% — and the industry scrambled for generation. Natural gas reemerged as the fastest path to power: $228 billion in new gas-plant projects are now in play, with hyperscalers increasingly building, owning, or co-developing generation and turbine makers like GE Vernova and Mitsubishi Power sold out for years. Infosys argued data centers can 'delay, not dodge' the crunch through efficiency, AI-driven energy management, and peak-load optimization — buying time, not a reprieve. The compute vendors moved into the power layer directly: NVIDIA launched its DSX platform to integrate grid, facility, and compute and 'break the power bottleneck,' with GE Vernova supplying grid-to-rack expertise. Capital flowed to the constraint, as Apollo and Blackstone structured multibillion-dollar private-credit deals around data-center power and GPU financing. And the network itself began routing around scarcity: Lumen's Emerald AI shifts AI workloads to alternate locations based on power availability and cost.
Spending hit $2.7 trillion while the payoff stayed stubbornly invisible — Apollo found zero AI benefit outside Big Tech and Gartner found companies aren't actually saving money.
A counter-narrative to the AI boom hardened into data this fortnight. Apollo Global Management released an analysis finding AI investments have delivered zero financial payoff to corporate America outside of Big Tech — revenue per worker and margins are rising for the tech giants but flat for everyone else. Gartner reported that companies are not saving money with AI, even as it projected global AI spending to reach $2.7 trillion in 2026, a 49.5% year-over-year jump concentrated in infrastructure. The confidence gap showed up at the top: an EY report found only 16% of CEOs have detailed, real-time visibility into AI costs and high confidence in assessing ROI. And the capital markets started pricing the risk — Crunchbase framed the AI IPO wave as 'the dot-com bubble's evil twin,' noting even Anthropic's IPO delay and a New York City comptroller publicly urging it to pause. The through-line is a widening gap between the money going in and the returns coming out, everywhere except the handful of companies selling the picks and shovels.
Answers replaced links this fortnight — ChatGPT sends under 1% of Google's referral traffic even as journeys start there, and brands discovered they'd gone invisible inside the answer box.
The open web's core economic engine — a search that sends you to a page — started seizing up. A MarTech-reported study of 100 sites over 12 months found ChatGPT sends only 0.62% as many sessions as Google organic search, with all LLMs combined at 0.73% — yet shopping and research increasingly begin inside those chatbots, meaning the demand is real but the click-through is gone. Entire categories found themselves invisible: a Bain and Profound analysis of 12,000 AI-assistant prompts and 64,000 cited URLs found only ~6% of travel results lead to an airline's own site, with online travel agencies dominating the citations. The platforms moved to control the funnel on their terms: OpenAI now feeds partner content into ChatGPT as pre-generated summaries ranked ahead of everyone else's raw scraps, Reddit and other first-party data owners began walling logged-out content while licensing it directly to LLMs, and Time magazine started serving ads to AI agents — treating the bot, not the human, as the audience. The web is bifurcating into licensed, machine-readable content inside the answer and an unmonetized long tail outside it.
Underneath the enterprise headlines, the research slice this fortnight showed AI moving from tool to instrument of discovery — proving centuries-old conjectures, designing antivenoms from scratch, hearing the patterns that connect birdsong to human speech, and reading the planet's forests from orbit. These are the signals that don't fit a collision but are too remarkable to skip.