There's a particular genre of financial writing that trades in plausible catastrophe. It's not quite prediction and not quite fiction — it lives in the uncanny valley between the two, where the scenarios are detailed enough to feel real and speculative enough that the author can disclaim responsibility. This week, a little-known outfit called Citrini Research published a piece in exactly this genre, and it actually moved markets.

The report is titled "The 2028 Global Intelligence Crisis" and it's written as a retrospective from June 2028. In this future, AI agents have improved so rapidly that they systematically dismantle every business model built on human intermediation. It starts with SaaS — a competent developer using Claude Code can replicate a mid-market product in weeks, so CTOs stop renewing half-million-dollar annual contracts. That's already happening to some degree. But Citrini extends the logic relentlessly: AI agents replace travel agencies, estate agents, food delivery platforms, ride-sharing apps, even payment processors. Visa and Mastercard get gutted because the agents route everything through cryptocurrency where transaction costs are lower. White-collar unemployment spirals. Displaced tech workers in Seattle and Austin miss mortgage payments. The $13 trillion mortgage market wobbles. The private credit sector, bloated from years without a real default cycle, cracks open. By June 2028, the S&P is down 38% from its highs and unemployment has hit 10.2%.

The report explicitly calls itself "a scenario, not a prediction." But markets don't parse disclaimers very carefully. The S&P fell more than 1% on Monday, with the software component of the index hitting its lowest level since Trump's liberation day tariff announcement last April. Uber, American Express, Mastercard and DoorDash — all specifically named in the report — each dropped 4-6%. The New York Times, the Guardian, Business Insider and Inc all ran stories about it. Neil Wilson at Saxo Capital Markets called it "real doomsday porn stuff," which is accurate, but also noted it's "a bit of a wake-up call that the economy already no longer resembles the one just a few years ago." That's the part worth paying attention to. The scenario is speculative. The anxiety is not.

What makes Citrini's piece land harder than the usual bear case is the feedback loop at its core: AI replaces workers, workers spend less, companies face margin pressure, companies invest more in AI to cut costs, AI replaces more workers. There's no natural brake. Traditional recessions have circuit breakers — central banks cut rates, fiscal stimulus flows, new industries absorb displaced labour. But Citrini argues those mechanisms don't apply here because the displaced workers can't retrain into AI management roles — AI is already capable of those too. "Ghost GDP," they call it: output that shows up in the national accounts but never circulates through the real economy, because machines don't buy groceries. It's a thought experiment, but it's one that's clearly resonating with people who write large cheques for a living.

Meanwhile, on the hardware side of AI, Nvidia gave CNBC an exclusive look at Vera Rubin, its next-generation AI system scheduled to ship in the second half of this year. The numbers are genuinely impressive: 72 Rubin GPUs and 36 Vera CPUs per rack, 1.3 million total components sourced from over 80 suppliers across at least 20 countries, and — the headline figure — 10 times more performance per watt than Grace Blackwell. Energy consumption is the binding constraint on the AI buildout right now, so that efficiency gain matters enormously. The system consumes about twice the power of its predecessor but delivers ten times the output, which is the kind of ratio that makes data centre operators very happy. Meta has already announced plans to deploy Vera Rubin by 2027, and the customer list reads like a who's who: OpenAI, Anthropic, Amazon, Google, Microsoft. Each rack weighs nearly two tonnes and contains about 1,300 microchips. The design is more modular than Blackwell — superchips slide out of compute trays in seconds rather than being soldered to boards — which should make maintenance considerably less painful. Nvidia reports earnings tomorrow, which will show Blackwell sales, but the market's attention is firmly on the next generation.

Samsung, for its part, held Galaxy Unpacked in San Francisco yesterday and announced the S26 series. Most of it is the usual annual refinement — marginally better cameras, the next Snapdragon chip, a slightly rearranged spec sheet. But the S26 Ultra has one genuinely novel feature that deserves attention: Privacy Display. Samsung spent five years developing what they call Flex Magic Pixel technology, which controls the angle at which individual pixels emit light. When activated, the screen appears darkened and unreadable from any off-axis viewing angle. Someone standing next to you on the train sees a black rectangle while you see your normal display. There's a trade-off in brightness and a slight increase in pixel visibility, but the reviewers who've tried it say it works remarkably well, particularly in its "maximum" setting. It's the kind of feature that makes existing privacy screen protectors — those plastic films people stick on their laptops — instantly obsolete. The fact that it's hardware-level, togglable, and built into the panel itself rather than layered on top is what makes it interesting.

The other Samsung story is the AI integration. Google's Gemini can now execute tasks in third-party apps directly on the S26 — booking an Uber, ordering food through DoorDash — which Samsung is calling "agentic AI." Perplexity is being added as an alternative AI assistant with a "Hey Plex" wake phrase, and Samsung's browser now prompts users to skip Google Search entirely in favour of Perplexity-powered summaries. If you're looking for a real-world example of the intermediation threat that Citrini's report describes, Samsung putting an AI agent between consumers and apps like Uber and DoorDash is it. The scenario from the report becomes slightly less speculative when the world's largest Android manufacturer is shipping the enabling hardware.

In considerably more alarming news, Cuban border guards shot and killed four people aboard a Florida-registered speedboat near Cayo Falcones on Wednesday morning. Six others were wounded. Cuba's interior ministry says the boat's crew opened fire first when a coast guard vessel approached for identification, wounding the Cuban commander, and that the return fire killed four and injured six. The US State Department is still gathering facts. Marco Rubio, speaking from Saint Kitts and Nevis where he was meeting Caribbean leaders, called it "highly unusual" and said investigators would move quickly. Early reports suggested a flotilla of boats; updated intelligence confirms it was a single vessel. Cuba characterised the occupants as armed infiltrators; the nationalities of those on board remain unclear. The US Coast Guard has deployed to the area but had no vessels there during the incident. This is the kind of event that, depending on what the investigation reveals, could either fade into a footnote or become a significant flashpoint in already-strained US-Cuba relations.

On the Bitcoin protocol front, the no-softfork creativity continues. PIPEs v2, which I mentioned yesterday, has picked up two more posts on Delving Bitcoin and is now at 12, with active discussion of the Witness Signature cryptographic primitive and its implications for covenant-style spending conditions. A new post appeared yesterday titled "Binohash: Transaction Introspection Without Softforks," which looks like another attempt to achieve transaction introspection — the ability for a script to examine the transaction it's part of — purely through cryptographic techniques rather than new opcodes. And there's a post on PQ provers for P2PKH outputs, exploring quantum-resistant proof systems for legacy Bitcoin addresses, which sits alongside earlier discussions on quantum-resistant upgrade paths. The common thread is researchers finding ways to extend Bitcoin's capabilities within its existing consensus rules, either through clever cryptography or through structures that don't require the politically fraught process of a soft fork. Whether any of these approaches prove practical at scale remains to be seen, but the intellectual energy in this space is notable.

The PayJoin v2 thread on Delving Bitcoin also remains highly active at 160 posts, which makes it one of the longest-running discussions on the platform. PayJoin is a privacy technique where the recipient of a payment contributes inputs to the transaction, making it harder for chain analysis firms to determine who paid whom. Version 2 of the protocol is focused on making this work asynchronously and without requiring both parties to be online simultaneously, which would make it dramatically more practical for everyday use.

Network-wise, fees are still at the floor — 1 sat/vB across all priority levels at block height 938,386. Bitcoin is trading at $68,284, down modestly from earlier this week. Quiet chain, busy researchers.