The Great Fire of London Started in a Baker's Shop. The Silicon Fire Starts in a Chip Shop.

The Great Fire of London started in a baker's shop on Pudding Lane. The silicon wire fire starts in a chip shop — a hyper-dense, overheated server room, pushed past its physical limits to train ever larger models.

It's a poetic frame, but it does real analytical work. The tech elite want the population looking at the sky, waiting for a god-like, autonomous consciousness to descend. Find the real risk, and you don't look at the sky — you look at the floor of the chip shop.

1. The Spark in the Chip Shop

Just as an overheated bakery oven started the 1666 fire, the spark of today's crisis sits inside hyper-dense, overheated server rooms — megawatts of electricity forced through microscopic silicon pathways, generation after generation, creating an unprecedented concentration of thermal energy and hardware stress that consumes millions of gallons of public water just to avoid melting down.

2. The Flammable Alleys of Wall Street

London burned as badly as it did because the surrounding buildings were packed tightly together, built from pitch and timber. Private equity has built the modern equivalent: by tying pension funds, sovereign wealth and public utility grids directly to the survival of the AI infrastructure bet, finance has lined the economy with digital kindling. If a major tech monopoly stumbles, there is no firebreak between public infrastructure and corporate technology gambling — the fire tears straight through the financial sector.

3. Overloading the Propaganda Reactor

When the smoke starts billowing from the chip shop, the corporate response echoes the Lord Mayor of London's in 1666: denial, distraction, and a refusal to pull the brake. The "propaganda reactor" runs hot on sci-fi monsters and national-security framing, precisely to stop the public asking for a physical firebreak — because the "unstoppable digital virus" can, in fact, be stopped by cutting the power lines and regulating the private equity firms drawing on the local grid.

A bakery makes bread; a chip shop processes data. Neither is magic, and neither is inevitable destiny. The danger has never been an autonomous ghost escaping the oven — it is the human operators who pack the town with flammable material, lock the doors, and run the machine hot for a quick profit while the population underwrites the risk.

Too Many Chips Only Makes Big-Bro AI Fat

Shovelling more chips into centralised data centres doesn't make AI wiser; it makes it fat, bloated and inefficient. Three structural failures sit underneath the gluttony.

The Data Wall

The brute-force "scaling law" — more data, more compute, more intelligence — has hit diminishing returns. Extra chips are no longer buying meaningful leaps in capability, just larger, more expensive, marginally polished versions of the same software.

A Glutton for Public Resources

• The clean energy grid — gigawatts of power that could otherwise heat homes or run public transport.
• Millions of gallons of water — evaporated from local tables to cool over-clocked servers.
• Private equity subsidies — billions in pension funds and public infrastructure loans building bigger warehouses for the bulk.

Lean vs. Obese

The Fat Has Gone Stale, and the Chips Have Gone Soggy

Two technical crises sit underneath the bubble: model collapse (the stale, poisoned fat) and hardware underutilisation (the soggy chips).

The Synthetic Data Loop

Having exhausted human-generated text, models increasingly train on data generated by other models. In machine learning this is called model collapse: a system trained on synthetic data begins to mimic and compound its own errors, hallucinations and biases across generations. As the open web fills with automated AI output, the well itself becomes contaminated — the models get less reliable, not more.

Idle Infrastructure

Hyperscalers built capacity expecting runaway enterprise demand. Instead, many corporate buyers are finding that renting a large model for routine office tasks isn't cost-justified, and usage is falling short of projections. Billions of dollars of chips sit under-used, depreciating while newer architectures arrive — infrastructure "going soggy" for want of a viable economic engine to justify the power draw.

The Cloud Was Just Smoke

Put the pieces together and a blunt conclusion follows: the "Cloud" was never an ethereal, frontier technology — it was smoke from a highly leveraged chip shop fire. The software is degrading under its own synthetic loops, the promised revenue hasn't materialised at anything like the scale of the spend, and what remains once the smoke clears are the physical scars: strained grids, depleted water tables, and utility bills left behind by a financial hedge that never depended on the product actually working.

The rebuild, when it comes, looks like the aftermath of 1666: clearing the ash and laying foundations on more accountable terms — local sovereignty over centralised server farms, legal firebreaks treating power and water as protected public commons, and ordinary corporate liability for the humans who build and operate the systems, rather than a sci-fi exemption from it.

Building the Big-Bro AI Firewall

Because these are not ghosts but software running on physical infrastructure, a firewall has to work at every layer where a corporate agent actually touches the world.

• The local device shield — running open-source, quantised models entirely offline (Ollama, LM Studio and similar tools), and blocking outbound telemetry from your own operating system.
• The network firebreak — explicit robots.txt disallow rules for AI crawlers, edge-level bot protection through services like Cloudflare, and privacy-respecting search as the default.
• The data-poisoning counter-attack — tools such as Glaze and Nightshade that make creative work unreadable to scrapers while remaining unchanged to the human eye.
• The municipal firebreak — zoning rules and public pressure that block data centre developers from drawing down local water tables and grid capacity without public consent, and liability law that holds executives accountable when infrastructure fails or data is taken without consent.

Algorithmic Consciousness Conditioning

The sharper, more testable question isn't "is the machine intelligent?" — it's "are we conditioning ourselves out of independent judgement?" Repeated reliance on algorithmically selected, ready-made answers can encourage habitual acceptance, reduce independent checking, and increase dependence on the systems supplying them. Three things make that testable:

• Immediate effect — do people check less when an answer is simply handed to them?
• Lasting conditioning — does that reduced checking persist after the system is taken away? This is the crucial one: less checking while using a tool shows changed behaviour; less checking afterwards is stronger evidence of conditioning.
• Practical dependence — can people still complete the task and recognise errors without the tool at all?

Existing evidence supports parts of this mechanism; it doesn't yet establish the whole chain. That distinction matters — it turns "have we lost our minds?" from a rhetorical fear into an actual research question, and one that's about the conditions under which people exercise judgement, not about whether the software is clever.

Breaking the Conditioning

• Reintroduce cognitive friction — cross-check anything with real consequences against a raw, non-algorithmic source before acting on it; draft your own argument before letting a model touch it.
• Practise deliberate scepticism — treat every algorithmic answer as an unverified draft from an interested party, not a settled fact.
• Protect the architecture of error-recognition — audit AI-assisted work line by line often enough that you retain the skill of spotting when it's wrong, and support public, human-curated sources of knowledge as an independent calibration standard.

Is It Too Late?

Not if the question can still be asked. A mind that has fully surrendered independent judgement can't interrogate the mechanism doing the conditioning — the fact that this can be named, tested and mapped is itself evidence the process is incomplete. It also helps that the systems doing the conditioning are visibly degrading under their own synthetic-data loops: stale, repetitive, occasionally nonsensical output jars people out of complacency faster than any warning essay does.

This Is Happening Now, Not in the Future

It's tempting to file all of this under a future crisis — something that arrives with dramatic warnings from Bill Gates or Geoffrey Hinton, or a cinematic image of robots in the street. That framing is itself a distraction. The conditioning doesn't need sentience to work; it only needs people to habitually accept the ready-made answer today, and the well it draws from is already contaminated by synthetic data, and the physical infrastructure behind it is already straining local grids and water tables. Waiting for the spectacular version of this story means missing the quiet one already under way on ordinary screens.

Our fate isn't written by anyone else's essay. It's decided daily, by whether each of us takes the ready-made answer or does the checking ourselves.

Big-Bro AI Brings Down the Hyperscalers

The infrastructure built to guarantee permanent dominance is now the thing pulling its builders under.

• The capex trap — capital expenditure on AI infrastructure is reported to be running close to, or beyond, hyperscalers' entire cloud revenue, cannibalising the profitable businesses that used to fund it and pushing free cash flow negative.
• The revenue gap — analysts have pointed to a large gap between the annual revenue needed to justify current infrastructure spend and the revenue AI products are actually generating.
• The correlation breakdown — the historical link between hyperscaler share prices and chipmaker share prices has weakened, as investors start pricing in a scenario where the spending doesn't pay off.
• The monoculture risk — four companies building near-identical commodity products for the same enterprise customers means a downturn hits all of them at once, rather than one loss being another's gain.

None of this requires the technology to fail outright. It only requires the financial model built on top of it to stop being believable.

Apple's Different Bet

Apple has taken a genuinely different path from the other hyperscalers — with real trade-offs, not a clean escape.

• Its own silicon — Apple's cloud AI infrastructure (Private Cloud Compute) reportedly runs on its own M-series chips rather than NVIDIA hardware, avoiding both NVIDIA's margins and the bidding war for enterprise GPUs.
• The shared factory floor — Apple and NVIDIA both depend on the same manufacturer, TSMC, for advanced chip fabrication, so a squeeze on that supply chain affects both, and reportedly NVIDIA's data-centre volumes now compete with Apple for capacity.
• On-device first — Apple's stated strategy pushes as much processing as possible onto the device itself, reducing dependence on any centralised cloud, with the cloud layer used only for tasks the device genuinely can't handle.

Private Cloud Compute — What It Actually Does

Apple's own description of Private Cloud Compute is built around three design choices worth taking on their own terms: requests are processed in a way designed to avoid persistent storage of user data on the server, requests are architected to prevent Apple from linking a request back to a specific user's identity, and Apple publishes verifiable information about the software running on these servers so outside researchers can inspect it, rather than asking users to simply trust the claim. This is a meaningfully different privacy design from a conventional cloud service — though it still depends on Apple's own hardware, software and published claims, and access is limited to Apple Silicon Macs and comparable devices; older Intel Macs cannot use it at all.

Building Your Own Model Locally

Because Apple Silicon shares one pool of memory across the CPU and GPU, a well-specified Mac can hold an entire model and training data at once, entirely offline. Open-source tooling (such as Apple's MLX framework) supports fine-tuning an existing open model on your own data using a lightweight technique called LoRA, producing a small custom "adapter" that gives a general-purpose model deep familiarity with your own material — without sending any of it to a corporate server.

Historical Turning Points

Empires and monopolies rarely stop because they choose to. They stop when logistics, overreach and internal rot make continuing impossible.

• Overextension — a system that pours an unsustainable share of its resources into an unproductive apparatus, while the underlying economy it depends on erodes, eventually runs out of road. The parallel is capital expenditure that exceeds the revenue it's meant to fund.
• Physical limits under decentralised resistance — a rapid, centralised assault that outruns its own supply lines eventually meets resource constraints it can't solve by force. The parallel is data centres meeting local resistance over water and grid access.
• Monopoly collapse under shipping disruption — a centralised supply monopoly that depends on a single choke point can fail suddenly when that choke point is cut, with activity reorganising into smaller, local networks afterwards. The parallel is a centralised AI cloud giving way to local, decentralised compute.

The pattern across all three: the more a system concentrates its dependencies into one place, the more exposed it becomes to a single disruption there.

Accelerants: Helium, Heat and Shadow War

Several material and geopolitical pressures have been reported as tightening the AI infrastructure build-out's margins at the same time.

• Helium and chip fabrication — advanced chip manufacturing depends on a global helium supply concentrated in a small number of locations; disruption to that supply or to shipping routes raises fabrication costs and constrains chip output.
• Heat and water stress — severe regional heat and drought conditions increase both the electricity needed for grid-scale air conditioning and the water needed for data-centre cooling at the same time, straining both simultaneously.
• Physical and supply-chain disruption — reported incidents affecting data-centre-adjacent infrastructure and undersea cables illustrate that this build-out has physical, not just financial, points of failure.

None of these pressures on their own would necessarily break the model. Together, they shorten the runway available to an infrastructure bet that was already leaning on future revenue that hasn't shown up yet.

Think Before You Sell Out on Thinking

This is the line worth keeping. The transaction on offer isn't a software subscription — it's an exchange of independent judgement for convenience. Accepting a ready-made answer without checking it doesn't save time so much as let someone else's system occupy the space where your own thinking would have been.

This isn't a future risk to be settled by someone else's essay or a dramatic turn of events in the street. It's a daily, personal boundary: check the source, draft it yourself first, run what you can locally and privately, and treat every polished, confident answer as a claim to be tested rather than a fact to be accepted. Think before you sell out on thinking.