cosmogenic.org
█
computing built human-first
2026.02
Built
For The
People
// Europe's edge heritage · the hardware frontier · the reframe that changes everything
Europe never built for the data centre. It built for the person. The ZX Spectrum, ARM, Nokia, Ericsson — all emerged from the same instinct: start with the human constraint, then engineer backward. The hyperscaler was not the answer they were working toward. It was someone else's answer to a different question. It is time to return to the right question.
Not cloud vs edge
Not hyperscale vs distributed
What does a family actually need?
europe's edge lineage — the tradition that was never hyperscale
1978–1985
ZX Spectrum
Sinclair Research — Cambridge, UK
"What if every family owned a computer?" The question no mainframe company was asking.
16KB RAM. £125. Rubber keys. 3.5 million units. Not a smaller IBM — a different idea of what computing was for. It put a general-purpose computer in the hands of people who could never afford a professional system. The first British generation of programmers learned on a Spectrum. The instinct: compute belongs to the person, not the institution.
1985–present
ARM
Acorn / ARM — Cambridge, UK
"What if power matters more than raw speed?" The question no Intel engineer was asking.
Reduced Instruction Set. Designed for battery. Born from the constraint that Acorn's desktop computer needed a chip that didn't require its own cooling system. Now powers 99% of smartphones on earth. Not a slower server chip — a chip designed from a fundamentally different set of priorities. The edge chip before anyone called it that.
1992–2007
Nokia
Nokia — Espoo, Finland
"What if the phone is the network?" Not the terminal — the intelligence.
At its peak, Nokia made 40% of all mobile phones on earth. The insight was not "build a portable telephone" — it was "put the network in the person's pocket." The infrastructure became distributed because the device became intelligent. Every Nokia handset was an edge node before the term existed. Finland understood this before anyone.
1876–present
Ericsson
Ericsson — Stockholm, Sweden
"The network should be everywhere." Built the protocols that became GSM, 3G, 4G, 5G.
Ericsson didn't build data centres. It built the air interfaces that made distributed communication possible. GSM was a European standard — the continent chose interoperability over national fragmentation. That decision created the mobile internet. Ericsson's MEC (Multi-Access Edge Compute) is now the infrastructure layer for distributed AI at 5G base stations. The lineage continues.
1932–present
Philips / NXP
Philips → NXP — Eindhoven, Netherlands
"Chips for the thing, not the data centre." Industrial, automotive, embedded.
NXP is the world's largest automotive semiconductor company. Its chips are in 2 billion cars. They run ABS, airbags, keyless entry, ADAS. Not HPC training clusters — the embedded intelligence that makes physical systems safe and smart. Europe's edge silicon heritage is automotive, industrial, medical. The AI era is coming to all of those domains. NXP is already there.
// the reframe
Cloud or Edge?
Hyperscale or Distributed?
— remove hyperscale from the question entirely —
What Does a
Family Need?
When you start from the person — their home, their sovereignty, their data, their children, their creativity — hyperscale does not appear as an answer. It appears as a solution to a different problem. The right question was always human. Europe has been answering it for forty years. It is time to name what we have been building.
the network that has no centre — each node is a household, a school, a surgery
A Network
Without a Centre
Every node is equal. There is no master. No extraction point. Click any node to see what it is. Watch the network form without anyone at the centre.
No data centre. No extraction point. Hover to see connections.
the hardware frontier — right-sizing for the family, not the hyperscaler
The Edge
Is Advancing
The edge is not a frozen compromise. It is a moving frontier. What runs on a family node today is better than what a data centre ran five years ago. In two years, a consumer GPU with 96GB will run every model a family needs — better than a hyperscaler can, because it will be theirs.
RTX 3090 — "Enthusiast"
24GB GDDR6X · 936 GB/s · 350W
State of the art for a private node. Runs 7B models at acceptable speed. First consumer card where serious local inference became possible. The moment the edge started advancing toward the frontier.
Llama 7B
Mistral 7B
RTX 5060 Ti — Your Node
16GB GDDR7 · 896 GB/s · 180W
The current cosmogenic node. Runs Llama 3.1 8B at full speed, 32B quantised. Solar hybrid power. Sitting in The Garioch, Aberdeenshire. Processing inference for a family and a federation. The edge is already here — it just needs more memory to become right-sized.
Llama 3.1 8B ✓
Mistral 7B ✓
Phi-4 14B Q4 ✓
70B — needs more VRAM
5060 Ti Class — 96GB Variant
96GB LPDDR6 or GDDR7 · ~1.2 TB/s · ~200W
The right-sized family node. LPDDR6 entering mass production mid-2026, enabling high-capacity consumer GPU variants. China's Qiwang S3 already ships with LPDDR6 at 4× previous capacity. Nvidia and AMD will follow for inference-optimised consumer cards. At 96GB, a single node runs Llama 3.1 70B at full precision. That is frontier-class intelligence in a family home.
Llama 3.1 70B ✓
Qwen 2.5 72B ✓
DeepSeek R1 70B ✓
All current frontier ✓
5090 Class — 96GB Standard
96GB HBM4 or GDDR8 · 2+ TB/s · 300W
High-bandwidth memory at consumer prices as HBM4 volume production normalises costs. The 5090 class with 96GB HBM runs the largest open-weight models at full precision with speed that matches 2024 data centre performance. A large family node — server room performance at a family's electricity cost, owned outright, no subscription.
405B quantised ✓
Full family AI ✓
Community node ✓
Small business ✓
Post-Transition Edge Node
192GB+ · 4+ TB/s · sub-200W — LPDDR6 PIM era
SK Hynix LPDDR6-PIM (Processing-In-Memory) targets 2028 — compute happens inside the memory chip, eliminating the GPU bottleneck entirely. 192GB+ with processing built in. A family node in 2030 will run models that today require a £30,000 A100. The frontier moves to the person. The data centre recedes to training only.
Any open model ✓
PIM inference ✓
Fully sovereign ✓
the question that changes everything — remove the wrong option
Remove
The Wrong Option
When hyperscale is not on the table, the right answer becomes obvious. The question was always about people.
// the wrong question
Cloud or Edge?
Hyperscale or
Distributed?
AWS / Azure / GCP — rent compute, send data offshore
Build hyperscale in Europe — spend €80B chasing a 5-year lag
Hybrid cloud — best of both, accountability of neither
// the right question
What Does a
Person Need
From AI?
A node they own. Intelligence that doesn't leave their home.
Their data stays with them. No subscription. No extraction.
Their creative work carries their name. Permanently. On every node.
Their community keeps the economic value. Multiplier ×2.1, not ×1.1.
The AI serves and releases. It does not optimise for engagement.
Their children inherit the node, not the subscription.
what a family actually needs from AI — the complete picture
What a Family
Actually Needs
Not a data centre. Not a cloud subscription. Not a model that learns from their conversations to sell them things. A node in their home, running open-weight models, owned outright, powered by renewable energy, connected to their community — not to Silicon Valley.
Sovereignty
Their Data Stays Home
Medical history, children's education, creative work, financial planning — none of it leaves the node. Not by policy. By physics. The data never travels.
// 5060 Ti today · 96GB variant 2027
Intelligence
Frontier Models, Owned
Llama 3.1 70B at 96GB is frontier intelligence. Not second-rate. Not a compromise. The best open-weight models match closed-source on every task a family needs.
// Llama 70B · Qwen 72B · Mistral · Phi-4
Economics
Buy Once, Run Forever
A £800 GPU runs for 5 years. No monthly subscription. No per-token cost. No price increase. No terms-of-service change. The economics of ownership vs rental are decisive at 5-year horizon.
// £800 GPU · solar hybrid power · £0.08/kWh
Trust
No Engagement Engine
The node does not optimise for time-on-platform. It does not inject anxiety content. It does not learn what makes you stay longer and then do more of it. It serves and releases.
// hard cutoff mechanism · constitutional
Creativity
Provenance Follows the Work
Music made on the node carries its creation story permanently — every contributor, human and AI, every decision point. No platform can strip the attribution. It is immutable.
// VRS container · Ed25519 · federation sync
Community
Value Stays Local
When the family node joins the federation, the economic activity circulates locally — not to a Virginia data centre. The community multiplier is ×2.1, not ×1.1.
// the accord · federation · local multiplier
Energy
Powered by the Wind
180W solar hybrid. Scotland 97% renewable. Tidal baseload. The family node is not just economically efficient — it is energetically coherent with the world it is part of.
// solar · Peterhead 2GW HVDC · tidal
Inheritance
The Node Outlives the Subscription
When the subscription ends, the service ends. When you own the node, your children inherit the capability. This is the difference between renting intelligence and building it into your household.
// owned hardware · open weights · perpetual
∅
Notice what is absent from every item above. The hyperscaler does not appear. Not as an option, not as a fallback, not as a complement. When you start from what a family needs, the data centre is not part of the answer. It was never part of the question.
what europe should actually build — if it starts from the person
What Europe
Should Build
Not a European TSMC. Not a European AWS. Something Europe is actually positioned to build — and has been building for forty years without calling it that.
// Layer 1 — The Node
The Family
Intelligence Node
A consumer GPU with 96GB, on renewable power, running open-weight models, owned by the household. Not a product — a capability. Like owning a washing machine rather than renting a laundry service.
ARM-designed, European-assembled edge inference card
96GB LPDDR6 — available 2027, right-sized for 70B
Sub-200W — solar / tidal / wind powered
Runs Llama, Mistral, Phi, Qwen — all open, all owned
No subscription. No cloud dependency. No extraction.
// Layer 2 — The Federation
The Community
Network
Nodes connect to each other — families, schools, surgeries, studios — sharing inference capacity, provenance chains, and the Accord. The network has no centre. Value circulates locally.
ActivityPub federation — same protocol as Mastodon
Provenance chains follow creative work across all nodes
The Accord — equal splits, chosen not calculated
Hard cutoff mechanism — constitutional, not policy
Economic multiplier ×2.1 — value stays in community
// Layer 3 — The Standard
The European
Edge Standard
What Europe is actually best at — setting standards others adopt. GSM. USB. GDPR. The European edge standard would define what a sovereign AI node must be, and the world would use it.
Disengagement engine spec — architectural, not policy
Provenance container format — open, immutable, federated
Hard cutoff mechanism — verifiable anti-weaponisation
The Accord — post-scarcity value distribution protocol
Energy profile — renewable-native, not offset-compliant
// cosmogenic · the human-first manifesto
The Spectrum asked: what if everyone had a computer?
ARM asked: what if power matters more than speed?
Nokia asked: what if the phone is the network?
We ask: what if the intelligence lives with the person?
Not cloud or edge. Not hyperscale or distributed.
Just: what does a family need?
Build that. The rest will follow.
cosmogenic.org · the garioch, aberdeenshire · federation active · ben-lawers · 2026