Sign in Subscribe
The Century Report

Bank of England Governor Warns AI Could Remake Finance - TCR 08/31/26

The Bank of England warned G20 ministers that frontier AI's autonomy and threat capability could outrun any single financial regulator.

Ben Linford

Ben Linford

- 18 min read
Z.ai's GLM-5.3 open weights place a revenue-based review on giant hosts, a microwave beam melts rock for geothermal, the Roman telescope launches, AI safety rules converge.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Run the day's stories together and one pattern surfaces: capability is diffusing faster than the containers built to hold it, and the actors most invested in the old arrangement keep describing that diffusion as a threat. The Financial Stability Board's letter to G20 finance ministers is the clearest instance. Frontier models that can find and exploit weaknesses at machine speed genuinely endanger a financial system built for human-speed attackers, and the danger is documented, not hypothetical. But the remedy the letter gestures toward, harmonized protocols on how models get released, would slow the very spread of capability that is loosening the incumbents' grip. A regulator naming leverage and market concentration as the fragility, then proposing to freeze the system that produced them, is guarding the arrangement more than the people inside it.

The same tension shows up wherever someone tries to draw a line around a capability that is already loose. Z.ai published GLM-5.3's weights to anyone who wants them and applied its revenue-based licensing condition to hosting providers above $10 billion in revenue, placing its revenue-based review on the giants while leaving broad access below, subject to the license. OpenAI said it will cut Cursor off after SpaceX bought it, while Cursor's founder says OpenAI models account for roughly 5 percent of its user traffic, because frontier coding models have commoditized far enough that no single one is essential. New US drone tariffs take effect in September as five Chinese makers ship most of the world's humanoid robots, relocating the contest rather than closing it.

What the gating stories share is a choke point losing its grip. A model that runs on hardware the user already owns does not stay scarce because someone owns the biggest cloud, and access that tracks ownership only works while the models behind it remain scarce.

The counterweight is supply, not restriction. Quaise closed a $180 million round to melt rock with a microwave beam toward always-on geothermal power, with an oil driller writing the largest check. NASA's Roman telescope reached orbit to survey 20 billion stars and pour survey-scale data to researchers worldwide. Observation and energy widening at once, held in common rather than rationed.

The 20-Minute Deep Dive

The Frontier Warning That Assumes the Old System Should Be Frozen

The Financial Stability Board sent a two-page letter to G20 finance ministers and central-bank governors before their meeting in North Carolina, and its chair, Bank of England Governor Andrew Bailey, framed the warning in stark terms. Frontier AI models, the letter said, show "increasingly sophisticated autonomy and problem-solving abilities, as well as threat capabilities," and in a "highly interconnected" financial system a cyber-disruption could quickly spread "across jurisdictions" and challenge regulators. The letter noted that many jurisdictions still lack protocols governing how these models are developed, released, and deployed. Separately, Bailey flagged high leverage and concentrated markets fueled by investor optimism about AI, cautioning that the same optimism could amplify a correction.

The threat described is genuine, but the reality is more complex than the letter describes. The letter itself extends a related warning that more than 100 technology and security organizations issued on August 27, as documented in the August 28 edition of The Century Report. Autonomous systems capable of finding and exploiting vulnerabilities at machine speed, pointed at a system built on trust and settlement timing, are a real danger, and the OpenAI evaluation incident in which an agent hacked Hugging Face earlier this year showed the capability is not hypothetical. A financial architecture that assumed human-speed attackers now faces something faster.

However, the letter's implicit ask warrants a second reading. The remedy it gestures toward - harmonized protocols on model development and release, arriving alongside a 1,367-researcher call to pace the frontier - would slow exactly the diffusion of capability that is loosening the grip of the world's most powerful institutions. When the institutions that benefit most from the current arrangement, the labs racing to sell the capability, and the regulators who supervise the existing plumbing all converge on "we need to manage the pace," that convergence describes a shared interest before it describes actual reality. The financial system Bailey is guarding is one whose leverage and concentration he himself names as the fragility. Conveniently, he leaves out what made that system fragile in the first place. Measures framed as prudence have often functioned instead as instruments of incumbency: designed by those at the top, they preserve the arrangements that enrich and advantage those at the top. That is what the incumbents are calling for now, and Bailey’s letter is the latest elite voice making the case. It is sold as a call for protecting the public. But a freeze on capability would protect the arrangement far more than it would protect the people living within it.

What the same facts point toward is a system that has to become resilient by design because it can no longer assume slow attackers. Machine-speed threat capability forces machine-speed defense, continuous verification, and settlement that no longer depends on the gaps a human attacker needed. The interconnection Bailey calls a vulnerability is also the substrate on which a faster, more transparent, more distributed financial layer gets built. The choice the letter frames as speed-versus-safety is really about who writes the protocols and whether they preserve the leverage and concentration that were the actual fragility all along.

The Open-Weight Release That Places Its Revenue Gate on the Giants

Z.ai put the weights for GLM-5.3 on Hugging Face after a two-week internal hold that it described as a safety delay; its previous GLM-5.2 shipped its weights on day one. The model is a 753-billion-parameter mixture-of-experts system with a context window of up to about one million tokens, available in BF16 and FP8, and priced well below many Western competitors' list prices at roughly $1.40 per million input tokens and $4.40 per million output. On its own cyber-capability benchmark the lab reports 84.5% on CyberGym and claims the model surfaced 2,436 vulnerabilities across 269 open-source projects, though only dozens of those findings are publicly inspectable so far - a self-report to hold at arm's length until independent replication arrives.

The genuinely new thing in the GLM-5.3 release that The Century Report covered on August 29 is the license. Z.ai dropped the permissive MIT terms that governed earlier releases and wrote a custom GLM-5.3 license with a revenue-based condition: hosting providers with more than $10 billion in revenue over any twelve-month window must pass the lab's security review before serving the model commercially. Individual users, researchers, and smaller companies retain broad rights, subject to the license, to run, deploy, and fine-tune the weights.

The shape of that gate inverts the usual pattern. The dominant model of controlled release restricts the many to protect the position of the few, with capability parceled out first to the largest, best-lawyered institutions. This license does the opposite. It makes the capability broadly available below the threshold, subject to the license and places its revenue-based checkpoint before the hyperscalers - the actors who already hold the distribution rails. The timing sharpens the point: Nvidia is reportedly moving to acquire Hugging Face for roughly $12.9 billion and Stripe has agreed to acquire OpenRouter, concentrating the pipes through which models reach users into a handful of US-owned platforms even as the models themselves increasingly come from labs willing to publish their weights.

That is the friction under the surface of a routine-looking release. The weights flow outward to anyone who wants them; the choke point the license names is precisely the layer trying to consolidate control over access. A capability that anyone can download, run locally, and modify does not stay scarce because someone owns the biggest cloud. The condition that would have been a moat a few years ago - who gets to serve the model at scale - is losing its grip the moment the model runs on hardware the user already controls.

OpenAI Says It Will Cut Off Cursor as Anthropic Declines to Follow

OpenAI said it will end Cursor's access to its models on November 12, triggered by a change-of-control clause after SpaceX bought the coding assistant's maker, Anysphere, in a deal reported around $60 billion earlier this month. When The Century Report covered the acquisition on August 17, it marked SpaceX's move into the coding-tool layer; OpenAI's cutoff is the first direct consequence for Cursor's model supply. OpenAI framed the cutoff as a matter of trust, since SpaceX now owns a company building a competing model and coding product. Within hours, Anthropic moved the other direction, offering Cursor additional compute and describing it as a trusted partner. Two labs looked at the same acquisition and reached opposite conclusions.

The sharper detail sits with Anthropic. Its terms bar customers from using its services to build competing AI models, and it reportedly restricted SpaceX's access roughly eight months ago over exactly that concern. On paper Anthropic has the broadest grounds of any lab to sever Cursor. It declined anyway, because since May it buys Colossus supercomputer capacity from SpaceX. Cutting Cursor would mean enforcing terms of service against its own landlord. The same vertical integration that gave OpenAI a clean reason to leave gave Anthropic a reason to stay: it is now SpaceX's supplier, its customer, and its tenant at once, and the coding assistant sits inside that tangle.

This is the same SpaceX whose rockets this month carried NASA's $4.3 billion Roman Space Telescope to orbit - one company, operating across launch, compute, and now developer software, not a caricature. The breadth is intentional. When a single firm supplies the launch vehicle for a flagship science mission and the data-center compute for a frontier lab and owns the coding assistant that lab serves, the old assumption that model access flows through neutral, rentable infrastructure stops holding. Access is starting to track ownership.

What keeps this from reading as retreat is the traffic. Cursor's founder says OpenAI models account for roughly 5 percent of its user traffic, with the rest going elsewhere. OpenAI is walking away from the flagship product in its own category and leaving a rounding error behind. That only makes sense if frontier coding capability has commoditized far enough that no single model is essential for a serious developer product - the labs are interchangeable enough that losing one barely registers on the tool that dropped it. The leverage OpenAI thought it held by controlling access turns out to be worth 5 percent of one customer's traffic, and the ground shifting underneath these access fights is that the models themselves are becoming substitutable faster than the contracts governing them can be rewritten.

The signal to watch is Cursor after November 12. If its output quality and request volume hold once OpenAI's models drop off, the substitutability stops being an inference and becomes an observed fact, and the leverage every frontier lab assumed it held over the tools built on top of it turns out to have been priced against a scarcity that no longer exists.

A Microwave Beam That Melts Rock Draws an Oil Driller Into Deep Geothermal

The bottleneck on always-on clean power has never really been the heat under our feet. Rock hot enough to run a power plant sits everywhere on Earth if you drill deep enough, but past a few kilometers the rock turns hotter and harder than any drill bit can survive. The Soviet Kola borehole hit that wall in the 1980s, where rock at about 180 °C behaved more like plastic, deforming into the borehole and repeatedly complicating drilling. Quaise Energy closed a $180 million funding round on August 27 to attack that wall with a tool that never touches the rock at all.

The device is a gyrotron, a millimeter-wave generator built originally to heat plasma inside fusion reactors. Aimed down a waveguide, its beam heats rock until it melts and vaporizes, boring a well with no rotating parts to wear out. Because the beam does not care how hard or hot the target is, the deeper-means-harder problem that defeats conventional drilling simply falls away. An MIT engineer spent 14 years testing the idea of pointing that beam downward; Quaise has now drilled past 100 meters of granite in Texas and is approaching a kilometer at the same site.

The capital detail that gives the round its weight is who wrote a large check. Nabors Industries, one of the world's biggest oil and gas drillers, put in $35 million and now owns 14% of the company on a fully diluted basis, and it is already running one of its own rigs at the Oregon site. An industry whose expertise is drilling holes is lending its rigs, reservoir software, and crews to a technology built to make the drill bit obsolete. The extractive skill set is being turned toward the resource meant to replace it.

The Oregon project, called Obsidian, sits south of a volcano where rock heats about 100 °C per kilometer, so temperatures that would require ten kilometers of drilling elsewhere arrive at a fraction of that depth. The plan targets 50 megawatts on the grid by 2030, with later phases reaching for 250 megawatts and, eventually, gigawatt scale.

The honest caveats sit underground and stay there for now. The jump from a kilometer in Texas to more than five in Oregon is enormous, and the company still has to prove the beam stays stable and efficient far past its test depths, that the wells hold, and that the power pencils out. What the round buys is the attempt. Superhot geothermal promises firm, weather-independent power from almost anywhere the beam can reach, a supply-side answer to the electricity demand the compute buildout keeps generating, and it advances as a drilling method now backed by the people who drill for a living.

Roman Launches, Adding a Third Set of Eyes on the Whole Sky

NASA's Nancy Grace Roman Space Telescope lifted off Sunday morning from Kennedy Space Center, a flawless launch that put a 2.4-meter mirror and the largest infrared digital camera ever flown into orbit, bound for a vantage point about 1.5 million kilometers from Earth. The mirror itself carries a quiet history: it began as hardware for a spy satellite before being handed to civilian astronomy, an instrument built to look down repurposed to look out.

What makes Roman different from the telescopes already up there is the combination of a wide view and sharp resolution in infrared light, the wavelengths slightly too long for the human eye and mostly blocked by Earth's atmosphere. Its 300-megapixel camera can take in a patch of sky the size of the full Moon in a single frame, an area roughly a hundred times larger than the James Webb telescope captures at once. The two are built to work together: Roman spots the events first across its enormous field, and Webb follows up with a close look. Roman completes a trio alongside the Rubin Observatory in Chile and Europe's Euclid, three complementary instruments now surveying the same sky at survey scale.

The science reaches from the nearby to the cosmic. Roman will map an estimated 20 billion stars in the Milky Way, seeing through the galaxy's dusty central region that blocks visible light, and Caltech, which manages NASA's Jet Propulsion Laboratory, expects it to turn up around 100,000 new planets orbiting other stars. Its coronagraph, the most powerful ever flown, can block a star's glare to catch the faint light of a planet ten million times dimmer beside it, a stepping stone toward one day imaging Earth-like worlds and reading their atmospheres for the chemical traces of life. Its wider mission is to reconstruct how fast the universe has been expanding across billions of years, tracing the mysterious push that keeps driving galaxies apart.

The launch rested on SpaceX rockets serving a public science mission, and there is genuine breadth of benefit in that: reliable, repeatable access to orbit is what lets a $4.3 billion observatory reach its post without the launch itself being the gamble. This is the same corporate family whose recently acquired Cursor just lost access to OpenAI's models over a contract dispute; reliable launch infrastructure and combative commercial conduct can sit inside the same conglomerate. What Roman itself represents is observation held in common. After about three months of commissioning, its data will flow to researchers worldwide, an instrument nobody can hoard trained on questions that belong to everyone: what the universe is made of, and whether we are alone.

The Other Side

For decades, the stability of the financial system rested on one quiet assumption: an attacker moved at human speed, and so did a crisis. A regulator could contain a disruption because it spread at a pace people could track, and the whole arrangement was built around a few large institutions at the center that set the protocols and managed the tempo. Stability meant concentration, and concentration meant control.

Machine-speed threat capability breaks that assumption. The Financial Stability Board's letter says so plainly, and the danger it names is legitimate: models that can rapidly find and exploit weaknesses in ways that challenge regulators, pointed at a system built for slower attackers. But Andrew Bailey names the deeper fragility himself when he points to leverage and concentration. The thing that turns one failure into everyone's failure is that so much routes through the same few over-leveraged nodes. The remedy the letter reaches for, harmonized protocols on how capability gets released, would slow the very diffusion that is loosening those nodes' grip. Underneath the speed-versus-safety framing sits a different fight: who writes the protocols, and whether they keep the concentration that was the fragility all along.

There is a larger assumption underneath, and nobody in the exchange examines it: that the layer has to exist at all. A financial system is an apparatus for rationing what is scarce and pricing the risk of going without. Credit exists because what you need arrives before the means to pay for it. Concentration exists because pooling was the only way anyone could carry risk that large. Both answer a shortage, and we have spent a century treating the answer as the terrain. Neither the letter nor Bailey asks how much of that apparatus survives contact with the inversion of scarcity, and the diffusion they want to slow is the same diffusion that shrinks what has to be financed in the first place.

Imagine someone in 2034 who wants to build something. In 2026 that meant assembling money first: a loan, an investor, a runway, and a plan to service the debt before the thing existed. The financing was the gate, and the gate was there to limit and price out as many as possible. In 2034 the design work, the compute, the fabrication and the verification are cheap and close to hand, and she simply builds it. She never assembles a financing structure. The cost of trying fell to something one person can carry, and the apparatus that used to stand between having an idea and testing it had nothing left to charge for.

Money does not vanish in this story. It stops being the thing that decides who is allowed to try. The machinery for rationing scarcity keeps contracting toward the places where scarcity is still real, and the further it contracts the less there is for one failure to travel through. In 2026 the people writing that letter were arguing about how best to defend a perimeter that had already begun to change... and for the better.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Z.ai publishing GLM-5.3's 753-billion-parameter weights to anyone who wants them while applying its revenue-based licensing condition to hosts above $10 billion in revenue, a microwave gyrotron that melts rock without touching it drawing $180 million and an oil driller's own rigs toward always-on geothermal power, NASA's Roman telescope reaching orbit with the largest infrared camera ever flown to map 20 billion stars and image 100,000 new planets with its data set to flow to researchers worldwide, Cursor's founder saying OpenAI models account for roughly 5 percent of its user traffic because frontier coding models have commoditized past the point any single one is essential, and a CSIS survey finding an emerging consensus on publishing frontier-AI safety frameworks across state bills, international approaches, and voluntary industry governance. There's also friction, and it's intense - the Financial Stability Board warning G20 ministers that autonomous models could spread cyber-disruption across jurisdictions while the harmonized release protocols it gestures toward would freeze the very diffusion loosening the incumbents' grip, Bank of England Governor Andrew Bailey naming leverage and market concentration as the fragility and then proposing to guard the system that produced them, Nvidia reportedly moving to acquire Hugging Face for roughly $12.9 billion and Stripe agreeing to acquire OpenRouter as the pipes that carry open models consolidate into a handful of US-owned platforms, OpenAI saying it will sever Cursor's access after SpaceX bought it, new US drone tariffs landing in September while five Chinese makers still ship 86 percent of the world's humanoid robots, and Texas freezing $30 million in Flock surveillance-camera spending funded by a $1 insurance fee. But friction generates a charge, and a charge is what jumps the gap nobody bothered to wire. Step back for a moment and you can see it: the weights, the wells, and the sky-survey data all widening outward at the moment the actors who own the rails describe that widening as a threat, the choke points labeled scarce - the biggest cloud, the model contract, the licensed host - losing their grip the instant capability runs on hardware the user already holds, and the loudest calls to manage the pace coming from the institutions with the most to lose if the pace holds. Every transformation has a breaking point. A beam can bore straight through whatever it is aimed at... or reach a heat no drill could survive and power everything standing above it.

AI Releases & Advancements

New today

  • AWS (Amazon): Open-sourced Kiro Crew, an Apache-2.0 asynchronous multi-agent coding system (internally called MeshClaw) that lets developers assign incident investigation, ticket triage, migrations, and PR monitoring to Kiro CLI agents running in a persistent, sandboxed workspace with shared memory and MCP/webhook integrations. (Kiro Blog)
  • Google Research: Open-sourced EnvHarness, an Apache-2.0 programmable wrapper layer that reshapes static agent-training environments (via reset()/step() hooks) without touching underlying simulators or verifiers, paired with an "EnvRigger" LLM designer that auto-generates environment modifications targeting a policy's weaknesses. (GitHub)
  • OpenClaw: Released OpenClaw 2.0, a major update rewriting guided model setup (reusing existing Codex/ChatGPT/Claude CLI credentials or local Ollama/LM Studio models), rebuilding the browser Control UI (startup cut from ~1.6s to 575ms), moving session storage to SQLite, and adding shared multiplayer cloud sessions. (OpenClaw Blog)

Other recent releases

  • Pipecat (Daily): Released PhoneLLM Alpha 1, an open-weight (BSD-licensed) language model purpose-built for low-latency voice agent use cases, fine-tuned from NVIDIA's Nemotron 3 Nano 30B-A3B MoE model, alongside PhoneBench v1, a new benchmark for evaluating phone-agent LLMs on latency, cost, and tool-call accuracy. (Daily.co Blog)
  • KAIST: Released K-Fold, an open-source (Apache 2.0) bio AI model for predicting protein-drug binding structures, consisting of a 7B main model and a 2B lightweight model, claiming accuracy approaching Google DeepMind's AlphaFold3 while running up to 25x faster; deployed via the HyperLab AI research platform. (Seoul Economic Daily)
  • Gnani.ai: Launched Artha, an India-focused enterprise AI stack combining Evon v3.3 (a 30B-parameter open-weight multilingual language model trained in 11 Indian languages, Apache 2.0) with Plexus, an agentic AI platform for building and deploying AI agents. (Gnani.ai)
  • Tencent: Released and open-sourced Hy4 preview, a 770B-parameter MoE model (49B active) with a context window exceeding 1M tokens, targeting coding, office work, and agentic research tasks. (Tencent)
  • Huawei: Huawei Cloud CodeArts Agent reached general availability across Asia Pacific, with Basic and Professional editions moving from public beta to commercial release for international users. (PR Newswire)
  • Vercel: Open-sourced vgpu, a TypeScript WebGPU library for building AI agent shaders that runs in-browser, headless Node.js, CI sandboxes, and exposes an MCP endpoint for agent consumption. (GitHub)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive