TL;DR:

What is life?

Wrong question. Too broad.

Better: Under what conditions does a system maintain coherent self-renewal under entropy while remaining meaningfully distinct from, yet dependent upon, its world?

Life is not order, activity, or complexity. It is organised persistence that does not devour the basis of its own renewal.

The real distinction: sel…

Read more

When people imagine systemic failure, they imagine a snapped bone.

They want the X-ray. They want the white line on the black film. They want a doctor to point and say, there, that is the fracture. Then the mind can relax. The problem is visible. A cast can be applied. A timetable can be given.

But complex systems do not fail like bones. They fail like circulation. They fail like metabolism. They fail like a body that still has colour in the cheeks while the organs are quietly shutting down.

That is where we are now.

The instruments still glow. The dashboards still reassure. The interfaces still smile and report productivity, optimisation, throughput, engagement. Yet beneath the bright green surface, something much darker has been happening. We have built systems that consume the very conditions required for intelligence, trust and civilisation to survive.

That is why I am calling 2026 the Year of the Vampire.

I do not mean this as theatre. I mean it as diagnosis.

A vampire is a structure that preserves its own motion by draining life from something else. Our digital order has begun to behave in exactly that fashion. It feeds on human attention, on original thought, on institutional memory, on electricity, on water, on trust, and finally on reality itself. It does not create life. It extends its own imitation of life by extraction.

For more than a decade, the ruling superstition of the technology industry was simple: scale equals intelligence. Add more parameters. Add more data. Add more chips. Add more power. Add more capital. If the machine still falls short, make the machine bigger. By early February this year, public markets had started to show that faith cracking.

Reuters reported that software and services stocks had shed roughly US$1 trillion in market value in a bruising AI-driven selloff, even as Big Tech’s planned AI infrastructure spend for 2026 sat around the US$600–635 billion range. At the same time, the physical burden of that bet was becoming impossible to hide: investors were publicly pressing Amazon, Microsoft and Google over water and power use in US data centres, while Reuters reported new delays around the restart of the former Three Mile Island plant intended to help power Microsoft-linked data-centre demand.

The problem is not only financial. It is epistemic.

When a system begins training on its own output, it starts to eat its own tail. That is not poetry; it is now plainly part of the public technical record. In 2024, a widely discussed Nature paper described “model collapse” as the degenerative effect that occurs when generative systems are recursively trained on data polluted by prior model outputs. In plain language: photocopies of photocopies get blurrier, and the blur begins to masquerade as the original. That is the digital form of vampirism. The machine stops drinking from the well of reality and starts drinking from its own reflection.

And once you understand that, you can see why the cloud is not a cloud at all. It is a furnace.

One of the great evasions of the digital age has been the fantasy that information is weightless. It is not. Information is physical. Landauer’s principle, proposed in 1961 and experimentally tested later, tells us that erasing information has a thermodynamic cost. At the scale of today’s AI infrastructure, that means heat, cooling, water, land, transformers, substations, transmission lines and political bargaining over scarce energy. The International Energy Agency now projects that global electricity demand from data centres could rise to around 945 TWh by 2030, with AI as the main driver. That is not a metaphorical burden. It is an infrastructural one.

This is where my long argument with mainstream systems thinking begins.

Fritjof Capra performed a genuine service by helping people see life as relationship rather than mechanism. On his own account, his framework integrates the biological, cognitive, social and ecological dimensions of life. That mattered, and it still matters. But too much of the systems establishment inherited from that tradition a dangerous optimism: the belief that connection tends naturally towards health, that if you increase interdependence you increase wisdom, that if you wire enough nodes together the network will self-correct into harmony.

No. Connection does not only enable cooperation. It also enables contagion.

A circulatory system without an immune system does not become enlightened. It becomes septic.

That is why I have said, and will keep saying, that most systems theory has been a map of a healthy body with no chapter on cancer. It has not studied pathology with enough seriousness. It has not looked hard enough at the way small extracting minorities, institutional or algorithmic, can organise a whole network around depletion rather than regeneration. It has not adequately reckoned with the fact that a network can be optimised for capture.

Now look beyond the server farm.

As I write, the wider geopolitical field is also flashing red. Reuters and AP have both reported on the war involving the United States, Israel and Iran that erupted after the strikes of 28 February 2026, on failed ceasefire talks in April, and on the knock-on strain across shipping and energy routes in the region. You do not need to accept anyone’s propaganda to see the systems lesson. When shared reality degrades, when trust thins, when negotiation becomes theatre, force returns as the final medium of exchange.

This is why I define war the way I do: war is what happens when a system cannot metabolise contradiction.

A healthy civilisation can process friction. It can hold opposing truths in tension long enough for diplomacy, trade, law, mourning and memory to do their work. An unhealthy one accumulates unresolved contradiction like toxin. Lies are banked. Traumas are deferred. Language is degraded. Institutions become public stages for private games. Eventually the semantic bandwidth collapses. Once that happens, communication does not end; it descends. It reappears as blast radius.

This is also why the digital crisis and the military crisis are not separate stories.

They are both expressions of coherence debt.

One appears as hallucination, sludge, fraudulence, synthetic confidence, endless generated noise. The other appears as bombs, blockades, oil shocks and frightened populations staring at maps. But structurally they rhyme. In both cases the system has lost the capacity to carry contradiction in symbolic form. It can no longer digest. So it externalises.

Inside organisations, the same failure takes a quieter shape. I call it the competence mirage.

The screen says the system is fine. The manager is told the process is optimised. The AI summary is fluent. The report is polished. Yet fewer and fewer people know how the plumbing actually works. Fewer people can repair the hand-off when the automation drifts. Fewer people remember the tacit knowledge that once lived in apprenticeship, repetition, and embodied correction. We are not only automating tasks; we are automating away the social rungs by which a civilisation remembers itself.

That is the inner meaning of the vampire. It does not merely steal output. It hollows the host.

And the human nervous system feels this before the policy class does. One does not need a clinical paper to recognise verification fatigue. People know what it is to live under a rain of cheap certainty. They know the peculiar exhaustion of being given instant answers that must still be checked, re-checked and translated back into reality by a tired biological mind. They know the strange loneliness of a world full of generated speech and thinning witness.

So no, this is not an anti-technology argument.

It is an anti-parasitic one.

The response to the Year of the Vampire is not to smash the network. It is to give the network an immune system. That means provenance. It means slower and more expensive truth where truth matters. It means keeping humans in contact with the underlying craft. It means local competence, fallback modes, reversible decisions, auditable chains, smaller blast radii, and institutions that still know how to function when the glowing layer goes dark. It means treating energy, water and trust as first-class constraints, not externalities to be hidden in a footer.

Above all, it means recovering a civilisational fact we worked very hard to forget:

truth has weight.

Anything that appears infinitely scalable, frictionless and consequence-free is usually borrowing reality from somewhere else. Eventually the bill arrives. Sometimes it arrives as market repricing. Sometimes as infrastructural strain. Sometimes as cognitive exhaustion. Sometimes as war.

That is why 2026 deserves its name.

It is the Year of the Vampire not because the machine became alive, but because our systems became undead: animated, extractive, and increasingly detached from the living conditions that once grounded them.

A civilisation can survive contradiction. It cannot survive indefinite extraction from the roots of its own coherence.

And if there is any hope in this diagnosis, it is that exposure is the beginning of treatment. The vampire weakens when the mechanism is seen. The spell breaks when people remember that intelligence is not the same as scale, that connection is not the same as health, and that a system which cannot metabolise contradiction will, sooner or later, try to resolve it by feeding on its own world.

2026 The Year Of The Vampire

7.27MB ∙ PDF file

Download

Download

The Year Of The Vampire

275KB ∙ PDF file

Download

Download

Kyoto Systems Lab is fictional in this post. The “memo” is a narrative device.

Purpose and stance

This memo is written for all staff—research, engineering, operations, and support—to set expectations for 2026 and to explain why the lab has pivoted to independent, self-sovereign AI systems built for offline edge computing and mesh networking.

The short version: the information environment is degrading, the “always-online” assumption is becoming riskier, and the economics of centralised, brute-force AI are tightening. The combined effect is that platforms, infrastructure, and institutions are becoming less reliable as shared foundations. Our job is not to panic or moralise. Our job is to build systems that keep working when the network, the cloud, and the narrative space do not.

This pivot is not a retreat from ambition. It is a move towards resilience: smaller models, stronger grounding, better provenance, and architectures that keep human judgement and local reality in the loop.

What is different now

The defining feature of 2026 is not “lack of information”. It is the abundance of plausible, fast, synthetic information—delivered at scale, at speed, with high confidence and low accountability. Major risk frameworks have been warning that misinformation and disinformation are among the top near‑term societal risks, and that they can destabilise governance and crisis response by eroding trust.

In active conflict, this is not abstract. The ongoing war involving the United States, Iran, and Israel is being accompanied by a surge of synthetic media, miscaptioned footage, and unreliable “verification” performed by engagement-driven systems.

Institutions are responding, but the response is uneven. The National Institute of Standards and Technology explicitly calls out deepfakes and scalable disinformation as a material risk category for modern AI systems. The World Economic Forum has repeatedly highlighted the geopolitical and social consequences of narrative manipulation at scale. Meanwhile, platform governance bodies have criticised major platforms for inconsistent and insufficient labelling and handling of deceptive AI-generated content during conflict.

A practical takeaway for our lab: “truth” on the open internet is no longer something you can safely treat as a default input. It is an adversarial environment. The cost is paid as a cognitive verification burden—time, attention, stress, and paralysis—especially during fast-moving crises.

Limits of centralised brute-force AI

The second change in 2026 is that the economics and physics of frontier-scale AI are becoming harder to ignore.

Training frontier models is getting much more expensive over time. Analyses of frontier training runs estimate that total training costs have been rising rapidly year-on-year, driven primarily by accelerator hardware and specialised labour, with projections that the largest runs become financially unreachable for most organisations.

Data is becoming more constrained as a scaling input. Work forecasting “human-generated public text” availability suggests that, under continuing trends, top-tier models could effectively consume much of the accessible stock of high-quality public text within the 2026–2030 window, depending on training choices and overtraining. This matters because the pathway to “just scale it” starts to collide with data curation limits, legal constraints, and quality constraints, all at once.

Worse, the internet is increasingly being filled with model-generated text, images, and video. Peer-reviewed work on “model collapse” shows that indiscriminate training on recursively generated synthetic data can degrade model quality across generations, shrinking diversity and distorting tails of the distribution (the rare, the nuanced, the human). This is not a moral point; it is a systems point: feedback loops that ingest their own output tend to drift unless carefully bounded.

Energy and infrastructure constraints are tightening, too. The International Energy Agency projects significant growth in electricity consumption associated with data centres, with global demand expected to rise sharply towards the end of the decade, making local bottlenecks (grid capacity, supply chains, siting, cooling) a strategic constraint rather than a footnote. In the near term, the Gartner has pointed to a more mundane but equally binding limit: many organisations will abandon a large fraction of AI projects that are not supported by “AI-ready” data foundations.

So the lab’s position is simple: cloud-scale brute force will continue, but it is not a stable base layer for our mission. The systems we depend on must be able to operate when costs spike, when data quality collapses, when policy shifts, and when power becomes contentious.

Connectivity risks and the end of “always online”

The third change in 2026 is that global connectivity is becoming more politically and physically fragile.

Internet fragmentation is no longer just about content filtering. Current analysis argues the “splinternet” is evolving into a tool of power projection via export controls, contestation of standards, and even risk to physical infrastructure (including submarine cables). Public-facing institutions have also warned that sabotage to physical infrastructure can fragment connectivity in ways that are difficult to attribute and therefore difficult to deter.

Undersea infrastructure risk is now being treated as a serious national-security issue, not a hypothetical. For example, the United Kingdom and Ireland have moved towards joint exercises and coordination explicitly focused on protecting submarine cables and related critical infrastructure.

At the same time, everyday reliability has not improved. Cloud concentration means a single provider incident can ripple across thousands of organisations. Reporting on 2025 outages describes repeated disruptions across major cloud providers and the knock-on effects for businesses that built “cloud-only” dependencies.

The engineering implication is not “disconnect from everything”. It is “design as if the network will fail—often, unpredictably, and sometimes deliberately”. Offline-capable edge systems and local mesh communications are not fringe. They are a sober response to a world where connectivity is no longer guaranteed, and where adversaries understand that breaking coordination breaks societies.

Why we pivoted to self-sovereign edge systems

Our pivot is towards what we are calling self-sovereign models: models whose weights, runtime, and memory can be held locally; whose default mode is offline; and whose coordination layer can use opportunistic, decentralised links (mesh) when available.

This is not speculative. The broader industry is already moving meaningful capability onto devices. Apple has publicly positioned “on-device” foundation models as a developer-accessible capability for private experiences. Google documents on-device generative capabilities, including offline operation via small models such as Gemini Nano. Microsoft has released small language models explicitly framed for lower-cost, local deployment, and technical reporting supports the feasibility of high-utility models at much smaller parameter counts than the “frontier-only” narrative implies.

Independent research literature also shows this direction is real engineering, not vibes: surveys of on-device AI emphasise local inference for latency, privacy, and resilience, and identify compression and hardware acceleration as key enablers.

Mesh networking is the other half of the pivot. In practical terms: when the internet is unreliable—or when using it is unsafe—we still need field communication, local coordination, and the ability to share small, signed packets of state. Open implementations using low-power radios demonstrate the feasibility of decentralised, off-grid communication without cell towers or the public internet.

What this enables for us:

• Local truth anchors. Local sensors, local datasets, local logs. The “grounding” for decisions can be situated and auditable instead of scraped from a polluted global feed.

• Reduced verification debt. We can prioritise provenance, signatures, and known sources, rather than relying on post-hoc debunking across an adversarial internet.

• Operational continuity. If the cloud is down or contested, work continues. If bandwidth is scarce, systems degrade gracefully rather than failing hard.

• Economic control. The cost centre shifts from recurring variable cloud spend to owned capability and predictable maintenance—in an era when training and data-centre costs are trending up, and grid demand is tightening.

Operational guidance for teams

This section turns the pivot into concrete expectations. Treat these as design constraints, not optional “best practice”.

First, provenance is now part of the product, not a policy afterthought. We will implement and support media and data provenance wherever feasible, using the emerging ecosystem around the Coalition for Content Provenance and Authenticity standard (“Content Credentials”) for tamper-evident origin and edit history where appropriate. The Australian Cyber Security Centre has explicitly framed Content Credentials as a meaningful mitigation for some generative-AI cyber and deception risks, which aligns with our direction.

Second, we will bias towards “quiet” model behaviour. In practice: models must be allowed to abstain, ask for grounding, and admit uncertainty, rather than producing fluent guesses. This is consistent with risk-management guidance that emphasises understanding and managing AI failure modes, including deceptive content and user overreliance.

Third, we will build human recovery time into the system. Not as wellness theatre—as safety engineering. Infinite-scroll dynamics and rapid-fire feeds are excellent at producing confusion and impulsive errors; our interfaces and workflows should do the opposite by introducing intentional pacing (rate limits, review gates, and “cool-down” states where output pauses). Education and media literacy bodies have repeatedly argued that slowing down and evaluating information critically is now a core competence, not a nice-to-have.

Fourth, our threat model assumes deception-by-default in the public information layer. That includes synthetic audio/video impersonation, fake “official” accounts, and cross-channel fraud. We will therefore treat external content as untrusted until it verifies against provenance, signatures, or known-source corroboration.

Finally, we will run our own “context layer”. A recurring failure mode in organisations adopting AI is attempting to scale agents and automations without rebuilding data foundations and governance. The result is brittle, expensive, and often abandoned. We will treat structured context—schemas, taxonomies, logs, and local retrieval as a first-class product surface.

Closing

Large changes are plausible this year, and not all of them are technical: geopolitical shocks, energy constraints, network disruptions, and regulatory shifts can all land on the same week.

On macroeconomics: the Congressional Budget Office forecasts high deficits and historically elevated debt trajectories in its 2026–2036 baseline, with net interest costs becoming a major driver of fiscal pressure. But it is equally important to keep our heads: the International Monetary Fund data on reserve composition still shows the US dollar remains dominant in official reserves, even amid ongoing diversification debates. Separately, sustained official-sector gold demand has been documented by the World Gold Council, consistent with a broader trend of hedging against geopolitical and financial risk. Discussions inside BRICS about local-currency settlement and interoperable payment rails (including proposals involving CBDC linkages) should be watched—less because they “end the dollar” overnight, and more because they signal fragmentation pressure in global settlement infrastructure.

On regulation and governance: 2026 is also a hardening year for compliance expectations in many markets. The European Union has kept the implementation timeline for the AI Act on track, with major obligations becoming applicable across 2025–2026 and beyond. Australia has also strengthened responsible AI governance expectations in parts of government, reinforcing the direction of travel even where broader economy-wide law remains mixed.

So the lab’s “architecture of silence” is not mysticism. It is an engineering posture:

Hold the line on reality. Prefer grounded, local, signed truth over global noise. Build for offline continuity. Assume the network can be absent, hostile, or expensive. Let systems slow down. Make “pause” and “I don’t know” safe outputs—not failure states.

If we do this well, we do not just build tools. We build a place where people can think clearly again, even when the outside world is loud.