DC to AC Historical Mirror
Electricity did not become civilizational the moment its power was discovered; it had to become a system that needed to reach people for consumption.
Thomas Edison understood that. In 1882, his Pearl Street Station in lower Manhattan proved that electric power could move from laboratory wonder to commercial service. It was a masterpiece of localized engineering, but it was expensive and physically constrained. Costing roughly $300,000 and requiring 100,000 feet of underground wiring, it initially served just eighty-five customers across a single square mile.
That breakthrough was also a hard boundary. Edison’s direct current worked best close to the generating source. Low-voltage DC suffered severe line loss over distance, keeping its practical reach to roughly half a mile to a mile. To expand, Edison’s architecture required closely spaced generating stations, heavy copper conductors, and a business model suited strictly to dense, profitable urban districts.
Had that remained the only path, electricity would have stayed trapped in city cores, factories, and wealthy commercial corridors. Distant towns, farms, and ordinary homes would have been left longer with manual labour, kerosene lamps, and the limits of daylight. DC did not fail because it was useless; it failed as a universal architecture.
That structural friction made alternating current so consequential. Westinghouse and William Stanley, and later Tesla’s AC system breakthroughs, were not merely offering rival machinery. They were asking a deeper civilizational question: should power be built around the convenience of generation, or the dignity of reach?
The transformer changed the answer. AC voltage could be stepped up for long-distance transmission, reducing losses dramatically, and then stepped down near the point of use for safer local consumption. DC or AC was a ruthless battle over capital and perception. At the Chicago World’s Fair, Westinghouse’s AC bid of $399,000 easily undercut Edison’s $554,000 DC proposal. Protecting his lucrative DC investments and business model. Edison’s camp fought AC by spreading aggressive misinformation, staging public demonstrations to link high voltage with danger and death.
But fear could not hold back a utility once its reach became visible.
In the mid-1890s, the Adams Power Plant at Niagara Falls used AC to transmit hydroelectric power to Buffalo, more than twenty miles away. It cut the umbilical cord, proving that generation and use no longer had to sit beside each other.
What followed was the great unlocking of the modern world. In factories, electric motors decentralized production and AC truly won on reach. By 1938, state-backed rural cooperatives were delivering alternating current to 1.5 million American farms, replacing manual labor with automated pumps. In homes, kerosene gave way to light. Refrigeration changed food, medicine, and hospital cold chains. Electricity became civilizational because watts finally arrived where life needed them.
In this battle, DC did not vanish; it found its role inside batteries, electronics, and devices, coexisting with the AC grid it lost.
Which leaves us with a question: Is today’s AI still in its Pearl Street phase that is powerful, real, and commercially successful, but not yet built to reach the full texture of human life?
Is Today’s AI Still in Its DC Phase?
The DC to AC analogy is not perfect, but the pattern is familiar.
Today’s AI is powerful, real, and useful. It can write, code, search, summarize, translate, reason, classify, and act across workflows. Billions of people are using it, and enterprises are integrating it into software development, customer service, marketing, and operations.
But much of this intelligence still lives at the centre.
The most capable models are served from hyperscale data centers driving massive capital. Tech infrastructure CapEx exceeded $400 billion in 2025 and is expected to jump by another 75% in 2026. IEA projects data-center energy demand could double to 950 TWh by 2030.
Users and enterprises often reach this intelligence through cloud interfaces, subscriptions, and token-priced APIs, where every prompt and context window is a billed token. For an individual, the work often begins before the AI does: gather the context, upload the document, explain the situation, prompt carefully, correct the answer, and then carry the output back into the real task. For an enterprise, the friction is larger: security, governance, access control, workflow redesign, data residency, and reliability.
This is the AI version of distribution loss. Not electrical resistance, but latency, connectivity dependence, privacy exposure, cost, context loss, and constant human effort. Intelligence may be available somewhere, but it remains station-bound, isolated from the fluid needs of the actual moment.
This is not a criticism of centralized AI. Edison’s DC was the first commercially successful architecture, and today’s frontier AI is just as necessary. Some problems will continue to need the largest models, the deepest reasoning, and the most massive compute. Centralized AI will not disappear.
But the question remains: is this enough as the dominant architecture for intelligence in life?
There are signs of alternate approaches to AI. Small language models are becoming serious building blocks for local, latency-sensitive tasks. Distillation, quantization, LoRA, and local RAG are beginning to look like an intelligence “transformer stack,” ways to compress and route capability closer to use. Hardware is moving too with NPUs in phones and PCs, AI PCs, and embedded edge chips are all pushing intelligence outward.
But pieces are not the same as an AC moment. Much of today’s edge and local AI remains fragmented: chips looking for workloads, devices seeking differentiation, and models chasing benchmarks. It is useful, but it lacks a shared distribution vision. Westinghouse did not merely sell a better component; he saw that electricity had to be designed for reach.
The question is not whether centralized AI or Alternative AI wins. Like AC and DC, both will coexist. The real question is whether intelligence can now learn to travel well enough to become useful where and when life happens.
To answer that, we must pause the look for AI stack specifications and start outside in by observing how intelligence behaves when it finally lands in a specific human moment.
Three examples drawn from healthcare, agriculture and transportation all, different sectors compared to space missions validate the pattern.
Three Lived Moments: Alternative AI Where Life Happens
To see what this means, we do not begin with the model but begin with the moment.
A Business Leader Before the Meeting
Markus is on a flight from Frankfurt to Bangalore. He leads partnerships for a mid-sized European logistics company. In four hours, he will meet an Indian distributor he has never met, in a market he only partly understands, for a deal that could define his quarter.
Today, he would land, skim a CRM note, paste the distributor’s website into an AI tool, and ask for a briefing. He would get a useful summary: company size, recent announcements, likely priorities. But he would still walk into the meeting carrying fragments.
With Alternative AI, his local intelligence has been learning for weeks. Not his secrets. His patterns. The meetings that went well, the ones that stalled, the questions he forgot to ask, the moments he spoke too early. It knows he tends to lead with volume when he feels time pressure. It knows he earns trust when he asks one thoughtful operational question before discussing price.
Therefore, before he leaves the hotel, it gives him one sentence: “Their last partnership appears to have strained around payment timing. Ask how their cash cycle works before you discuss volume commitments.”
Not a deck. Not a dashboard. One question at the right moment.
In the meeting, Markus listens longer than usual. He asks about payment cycles casually. The distributor pauses, then speaks openly about delayed receivables. The meeting changes. It becomes less about price and more about designing trust.
On the way back, Markus says into his phone, “Note: he trusts slowly, but once he commits, he commits fully.” The AI does not answer. It stores and remembers for the next moment.
A Patient After the Clinic
Elena is 54. She works in a grocery store outside Houston and helps care for her grandson after school. She has had a cough for weeks, and her daughter finally persuades her to visit the clinic.
The doctor is kind, but rushed. He mentions early COPD risk, prescribes an inhaler, and tells her to return in a month. Elena nods. She understands some of it. Not all. She does not want to slow him down.
Outside, in the parking lot, she sits in her car holding a prescription she cannot fully read, for a condition she cannot confidently pronounce.
Today, this is where the system often ends. The visit is over. The confusion is hers to carry.
With Alternative AI, the unsupported space after the appointment changes.
Her phone speaks in Spanish, in the tone she prefers because she corrected it months ago when it sounded too formal.
“COPD means your lungs may be having trouble moving air as easily as before. The doctor is not saying you are in danger today. He is saying this needs attention. Would you like me to show you how to use the inhaler, slowly?”Elena says yes.
It walks her through the steps. It reminds her what warning signs require urgent help. If she had reported chest pain or severe breathlessness, the local system would not have tried to manage it alone; it would have triggered a consented, encrypted escalation to the clinic. It does not diagnose. It does not change the prescription. It stays inside its boundary.
That evening, it reminds her once. Not loudly. Not like an alarm. Like someone who knows she is tired after work. A month later, the doctor asks her about inhaler usage. Elena shows. He looks surprised as to how well she manages and is very happy on the progress.
The AI did not replace the doctor. It repaired the silence after the doctor.
A Traveller in Japan
Claire is an American travelling alone through Japan. After two days in Tokyo, she takes a train to a smaller coastal town because someone told her the morning market was beautiful.
Her Japanese is limited. Her phone signal comes and goes. Translation apps help her buy tickets and read signs, but they do not help her understand when to bow, when to wait, when a laugh is kind, or when silence is part of the conversation.
Today, travel intelligence is mostly transactional. Translate this. Navigate there. Find a restaurant. Rate the hotel. It helps Claire move through a place without quite entering it.
With Alternative AI, the town feels different.
Before she arrived, her local intelligence downloaded the language pack, maps, transport notes, and a small cultural guide. It has learned her way of travelling: quiet streets over famous spots, food markets over tourist lists, curiosity without intrusion.
At a fish stall, the vendor says something quickly and smiles. A literal translation appears: “This fish is angry.” Claire hesitates. Her AI whispers through her earbud: “He means spicy, jokingly. Smile and ask, ‘How angry?’”Claire tries. Her pronunciation is imperfect. The vendor laughs, not at her, but with her.
Later, an older woman describes a storm that changed the harbour years ago. The AI translates enough for meaning but leaves the pauses alone. It does not turn memory into tourist information. It helps Claire listen.
No major problem was solved. But Claire did not pass through the town as a consumer of place. She met it with a little more grace and understanding.
What These Moments Reveal
What made these moments possible was not one model or one device. It was a local intelligence stack: voice, local language, multimodal reading, private memory, user correction, ambient context, low-latency cues, confidence thresholds, consented escalation, and boundaries around what the AI could and could not do.
Most importantly, the intelligence had learned with the person, not merely about the person. Markus trained it through reflections after meetings. Elena trained it through language preference and trust boundaries. Claire trained it through travel choices, mistakes, and the kinds of encounters she valued.
That is the deeper possibility of Alternative AI. It is not only locally deployed. It is locally apprenticed.
The same underlying potential did not arrive as the same interface. It arrived as a question before a meeting, an explanation after a clinic visit, and a gesture of cultural grace in a street market. Like electricity becoming light, motion, or heat, intelligence became what the human moment needed.
Alternative AI: The System Beneath the Moment
Alternative AI is not a smaller chatbot running on a phone. It is not only an NPU inside a PC, or a local model inside an app. Those are components. They matter, but they do not make the whole.
The whole stack must answer a harder question: can intelligence run close to use, understand the specific moment, learn with the user, interact without friction, and stay trustworthy?
That requires five layers.
1. Compute and adaptation:
Small language models are becoming serious building blocks for local, latency-sensitive tasks. Distillation, quantization, LoRA, local RAG,and routing are beginning to look like an intelligence “transformer stack”: not one invention, but a set of ways to compress, adapt, and move capability closer to use. NPUs, AI PCs, smartphones, wearables, embedded chips, and edge devices are the physical substrate. This is the layer that makes Alternative AI technically possible
2. Contextual intelligence:
Local does not automatically mean useful. The system must understand the task, the person, the setting, the document, the workflow, the risk, and the timing. A business leader before a meeting does not need a generic market summary. A patient after a clinic visit does not need a medical encyclopedia. A traveller in Japan does not need literal translation alone. Each needs intelligence shaped by the immediate context.
3. User-developed learning:
This is the layer most missing today. Current AI is mostly prompted. Alternative AI must be trained by use. It should learn through correction, routine, preference, demonstration, memory boundaries, and apprenticeship-like interaction. The user should be able to teach it: this is how I decide, this is how I speak, this is what I ignore, this is when you should interrupt, this is when you should stay silent. Research in human-in-the-loop learning, federated tuning, and personalized federated learning points in this direction, but the mature everyday experience is not yet here.
4. User experience:
Intelligence near human need cannot depend on perfect prompts and clean dashboards. It needs voice, local language, multimodal input and output, accessibility, offline or intermittent use, ambient cues, and low-friction setup. It must listen, see, explain, remind, translate, summarize, and act in forms that suit the person and the moment.
5. Trust and safety:
The closer intelligence moves to life, the stronger the trust architecture must become. Local processing helps, but it is not enough. Alternative AI needs privacy-preserving memory, sandboxing, signed models, secure updates, audit trails, confidence thresholds, bounded autonomy, consented escalation, and clear refusal rules. Without this layer, local intelligence can become local risk.
The encouraging news is that pieces of this stack are emerging. On-device models, multimodal small models, AI PCs, NPUs, local-language voice systems, edge platforms, private compute architectures, and personalized learning research all point in the right direction.
The gap is that they do not yet add up to a civilizational distribution system. Much of today’s edge AI is still built from the inside out: chips looking for workloads, devices seeking differentiation, models chasing benchmarks, platforms extending ecosystems.
Alternative AI has to be built from the outside in: the human moment first, then the task, context, learning, interface, risk, and trust boundary. Only then should the system decide what model runs where.
That is why the AC metaphor still holds. The breakthrough is not local compute alone. The breakthrough is useful distribution. Alternative AI will matter when intelligence can be compressed, contextualized, taught, experienced, and trusted close to where life happens. Intelligence then become like watts.
Coexistence: System 1 and System 2
Alternative AI is not a rejection of centralized AI. It is a way to give it the right companion.
DC did not disappear when AC became the dominant grid architecture. It found its role inside batteries, electronics, and devices. AI may follow a similar coexistence pattern.
Alternative AI can become the System 1 layer: fast, local, contextual, low-friction, close to the user, good for bounded tasks, reminders, explanations, translation, personal memory, routine decisions, and moment-sensitive support.
Centralized frontier AI remains the System 2 layer: deeper, slower, more compute-heavy, better suited to complex reasoning, frontier science, strategic analysis, large-scale enterprise workflows, and high-risk decisions.
A trusted system must know what stays local, what escalates, what requires consent, what needs human review, and what should be refused. That decision should depend on risk, privacy, latency, confidence, cost, device capacity, domain, and user consent.
The Gaps That Remain
The stack makes Alternative AI imaginable. It does not make it inevitable.
Three gaps remain.
1. Device gap:
Today’s phones, PCs, and wearables were not designed as trusted ambient intelligence companions. They are still mostly app-based, notification-led, and on/off in their relationship with the user. They wake when summoned, disappear when pocketed, and compete for attention when they should often be quietly supporting it. Sam Altman’s critique of the phone is useful as a boundary question here: if personal AI needs to understand life in motion, is the smartphone enough, or does the AI era need a different class of hardware? But the answer cannot simply be “more sensing.” Ambient intelligence without consent becomes surveillance.
2. Software gap:
Intelligence is not electricity. A smaller local model is not a safe lower-voltage version of a frontier model. Compression can lose reasoning, judgment, factual reliability, domain nuance, and safety behavior. Local AI therefore needs more than model efficiency. It needs memory control, confidence estimation, routing, personalization, update paths, model provenance, and escalation logic. It must know when it is capable enough to act locally and when the task belongs elsewhere.
3. Governance gap:
Distributed intelligence is harder to inspect, certify, update, audit, and contain. A local model can become stale. A personal model can become overconfident. An enterprise model can become shadow AI. A malicious model can be fine-tuned and spread. Centralized AI has its own risks, but distributed AI multiplies the number of places where trust can fail.
These gaps do not weaken the case for Alternative AI but define the quest.
The Quest for Alternative AI
The history of the electric grid offers a template for how a new utility becomes civilizational. It was not enough to prove the commercial model at Pearl Street. That system remained local, capital-heavy, and limited by reach.
AC changed that only when several pieces came together: a technical bridge through the transformer, an integrator in Westinghouse, public demonstrations such as the Chicago World’s Fair and Niagara, emerging standards, utility finance, public trust, and eventually state-backed rural electrification. Markets, institutions, standards, capital, and public trust all had to converge. The lesson is not “one technology won.” The lesson is that a new utility needs a distribution system.
Alternative AI is not there yet.
It has pieces: small models, distillation, quantization, LoRA, local RAG, NPUs, AI PCs, local-language voice, multimodal interfaces, private compute, and edge platforms. But pieces are not a civilizational architecture. There is no widely accepted Westinghouse for Alternative AI, no Niagara moment proving human-near intelligence at scale, no mature shared standards for local memory, model provenance, escalation, audit, consent, or safety.
We have two visible paths for shaping AI today.
The first is the market-led infrastructure path, most visible in the United States.
Capital is flowing toward frontier models and centralized silicon, underscored by Nvidia alone pulling in over $50 billion a quarter just from data-center revenue. This path has extraordinary strengths: speed, talent, ambition, commercial pressure, and frontier capability. It is why the center is getting more powerful so quickly. But it also has a structural bias. It rewards scale, usage, token revenue, proprietary advantage, data moats, and compute concentration. It may produce stronger intelligence before it produces more human-near intelligence.
The second is the state-shaped adoption path, visible most clearly in China.
Here, the focus is not so much on chasing AGI. It appears to emphasize efficiency, diffusion, open models, applications, robotics, and integration into daily life. Official policy reinforces this direction through “AI Plus,” AI-agent guidance, regulation of generative AI, and energy-compute planning. China’s renewable-energy build-out and “East Data, West Computing” strategy also show that AI is being treated as infrastructure, not just software.
This path has strengths too: adoption at scale, sector integration, policy steering, infrastructure coordination, and a willingness to place AI into ordinary workflows. But it carries its own risk. State-shaped diffusion can become control. Adoption can outrun trust. Efficiency and deployment do not automatically mean dignity, privacy, or human agency.
So, neither path is enough by itself.
Markets can build capability but may concentrate it. States can accelerate adoption but may over-control it. Alternative AI needs a third discipline: a patient public-interest system that joins enterprise innovation with public capital to support standards, safeguards, local usefulness, and human purpose.
That is the real quest. Not just bigger models. Not just more devices. Not just sovereign AI. The question is who will make intelligence distributed, teachable, interoperable, trusted, and useful enough to change daily life as Markus, Elena and Claire examples show.
Where Do We Go From Here?
When I began writing this edition, I thought I was exploring a technology shift. I now see it as something larger: a quest to make intelligence useful where human life actually happens.
In past editions, I have often tried to leave you with a specific framework or immediate, practical advice. This time, I want to leave with a possibility. Alternative AI is not yet a finished architecture. It is a direction worth naming, because what we name, we can begin to build.
In Edition 72, we talked about “patient systems”: multigenerational, civilizational growth efforts like the Artemis space program, the Human Genome Project, and the mission to end global hunger.
Building an Alternative AI architecture that serves and protects human dignity belongs in that same family. It is the slow, deliberate work of building locally apprenticed intelligence and ensuring that centralized frontier models can safely coexist.
- How do you view the need for Alternative AI?
- What can we do to advance this quest at large?
- And how might your own AI implementations benefit from an approach that starts with the human moment?
The intelligence we create must do more than dazzle a few at the moment or leave many uncertain about the future. Electricity became civilizational when watts reached life. AI may become civilizational when intelligence does the same.
That is the quest. And it is still ours to shape.
Love to have your reactions, comments and support to progress the idea.
Best wishes.
Sidebar: References
Mindvista Sources
External Sources
- IEEE Engineering and Technology History Wiki (ETHW): Pearl Street Station
Historical archive detailing Edison’s first central power plant, supporting the premise that early DC power was geographically trapped to a roughly half-to-one-mile radius. - U.S. Department of Energy: The War of the Currents
Overview of the AC vs. DC battle and the Chicago World’s Fair, illustrating how the AC transformer overcame DC’s severe line-loss limits for distribution. - Hagley Museum and Library: Niagara Falls AC Transmission
Primary institutional record of the 1896 Niagara-to-Buffalo transmission, proving that centralized generation and distant use could be successfully separated at scale. - EH.net Encyclopedia: Rural Electrification Administration
Economic history resource documenting how institutional intervention scaled utility access to American farms, transforming manual labor and lived daily experiences. - International Energy Agency (IEA): Key Questions on Energy and AI
Projects global data-center electricity consumption to roughly double to ~950 TWh by 2030, highlighting the massive infrastructure demands of centralized AI. - Nvidia Newsroom: Financial Results
Reports Nvidia’s $51.2 billion data-center revenue, demonstrating the immense capital and economic momentum currently driving centralized frontier AI. - The State Council of the PRC: China’s “AI Plus” Initiative
Details China’s state-shaped path for broad AI integration across science, industry, and governance, offering a contrast to market-led centralization. - The State Council of the PRC: Joint Guidance on Intelligent Agents
Official guidance aiming to regulate and boost intelligent agents as products and services, highlighting sovereign-level governance for distributed AI.
