Post

Three Years Watching AI Become Infrastructure

We started out asking which chatbot was best. We ended up writing about power grids, token economics and who answers when an agent misfires. What that arc says about superintelligence.

By Liyam Flexer, Founder & EditorOctober 14, 202611 min readEditorial review pending

Aerial view of data centers, solar panels, wind turbines and power lines in green countryside at sunset, with glowing lines of data and a holographic globe above — Illustration supplied by the editorial team; original source and rights not specified.

Disclosure. The Best Blog Ever shares an owner with this site. This is a guest essay by its editor, published here from the author's draft, and most links point to that publication. Its figures are the author's and have not been independently verified by BestSuperintelligent.com.

The short answer

Between 2024 and 2026 AI moved from a product category to infrastructure. Coverage at The Best Blog Ever shifted from which chatbot wins, to who pays for inference, power and memory, to who owns, verifies and is liable for autonomous agents. Capability has been the least scarce input; energy, verification and accountability are the constraints.

An editor's archive is an honest instrument. You can't revise what you thought was important three years ago; the dates are in the sitemap. So when this site asked for a guest essay on how AI has evolved, I didn't start with a thesis. I started by reading the back catalogue of The Best Blog Ever, oldest to newest, and watching the questions change.

They changed a lot. And the direction they changed in is the most useful thing I can tell you about where superintelligence is likely to come from. Everything below is one publication's record and one editor's interpretation, not a survey of the field.

What are the key facts?

Each of these comes from an analysis on our sister publication and is labelled as the author's, not independently verified by this site.

Per-token AI inference costs fall roughly 10x a year, yet total AI spend keeps rising, a Jevons paradox (analysis).
For AI data centers, power and grid connections now take longer to secure than chips (analysis).
Chinese labs account for four of the top five open-weight models and most tokens on neutral routers (analysis).
Most agentic AI pilots stall before production, and the leading cause is unclear ownership, not model capability (analysis).

For how this site treats numbers like these, see vendor claims vs. verified findings and our checklist for reading benchmark claims.

What was AI coverage like in 2024?

The 2024 coverage reads like a field guide written while the field was still moving. We traced the lineage in Deep Learning Breakthroughs: AlphaGo to GPT-3, from a 2016 Go match to a 175-billion-parameter language model. We published a practical guide to the rise of generative AI. We ran the comparison every reader wanted: Gemini vs ChatGPT vs Copilot.

Even then, the regulatory undertow was visible. A May 2024 piece on AI ethics and regulation covered bias, transparency and the then-incoming EU AI Act.

The defining assumption of 2024 was that the model was the product. Pick the smartest one, plug it in, win. That assumption aged badly, and the rest of this essay is about why.

What changed in 2026, when the question became who pays?

When the publication was rebuilt around a stricter editorial policy, with primary sources, labelled assumptions and a mandatory Limitations section, the AI coverage that came out was almost entirely about money and physics.

Cheaper tokens, bigger bills

The most counterintuitive fact of this era is in Why AI Token Prices Fall While Total Spend Rises: unit costs collapse while aggregate spend climbs. Cheaper tokens don't shrink the bill; they unlock workloads that were previously unaffordable. Reasoning models spend tokens thinking. Agents spend tokens looping. The Inference Economics Crisis documents where this breaks products: the math worked at training and broke at inference.

Scaling changed shape

At the same time, the research frontier quietly stopped being "make it bigger." In Scaling Is Changing Shape we read three 2025 papers that converged on one idea: how a model spends compute now matters more than how big it is. Two earned their own deep dives: ParScale, a third scaling axis based on parallel computation, and Sakana AI's Continuous Thought Machines, a network that reasons over time.

The bottleneck went physical

A corridor of server racks in a data center, lit in purple, red and blue, with cabling visible through glass-fronted cabinets — A data center aisle. Image supplied by the editorial team; original source and rights not specified.

Then the constraint left the lab entirely. The Data Center Bottleneck and Why Power Is the Real Bottleneck for AI make the same uncomfortable point: electrons, transformers and grid interconnects now gate AI faster than GPUs do. That led to nuclear's second act, to the AI memory supercycle repricing DRAM for every phone and car, and to the top 10 chip bottlenecks that decide who holds the moat. The latest, on Nvidia's CUDA software moat, argues the most defensible position in AI is twenty years of developer switching costs, not silicon.

The geography moved too

Open weights changed the map. China Took Over Open-Source AI documents the shift in token share on neutral routers. And the payoff everyone was promised is still running late in the GDP data, partly because, as a research review on skill bifurcation argues, generative AI lifts novices far more than experts.

Treat each of those as a claim to check, not a finding. That is the rule this site applies to everyone, including its sister publication; see how to read an evaluation.

Who is responsible when an AI agent acts?

Agents changed the unit of analysis from answer to action. Coverage tracked them from the first credible production systems in AI Agents Are Changing Work to the thought experiment of the one-person billion-dollar company. If the term is new to you, start with our explainer, what is an agentic workflow?

The plumbing standardised fast. Model Context Protocol is becoming a common way for agents to reach tools. The business models strained just as fast: agents consume software as workloads, not seats, which is breaking SaaS pricing, and they don't click ads, which is breaking the web's business model.

The hardest lessons were organisational. A September cluster of pieces, read together, is the most important AI writing on that site:

Why Agentic AI Pilots Stall Before Production: the failure isn't intelligence, it's ownership.
How Enterprises Actually Govern AI Agents: the argument is that most oversight is theatre, because the audit log is self-reported by the agent it audits.
Who Is Liable When an AI Agent Fails: liability is decided months before the failure, by whether authority was scoped and reviewed.

The regulatory thread from 2024 also continued. For the EU AI Act's current timetable, read our own note on the high-risk dates under the Digital Omnibus, and for the US side, what Executive Order 14434 changes for AI buyers. Both are checked against primary sources.

Underneath all of it sits a principle the sister publication stated in July: verification is not optional. For a practical path, its 2026 enterprise guide to agentic AI workflows lays out a five-rung ladder. If you are weighing where agents should run, see our questions on hosted vs. self-managed deployment and the procurement evidence checklist.

What is the Constraint Ladder?

Compressing three years of coverage into one model, AI progress climbs a ladder of constraints. Each rung, once relieved, exposes the next. This is an editorial framework, not a measured result.

The Constraint Ladder. Editorial interpretation.
Rung	Binding constraint	Question it forces	Start here
1	Capability	Can the model do it at all?	What Is AI?
2	Cost	Can we afford to run it at scale?	Inference cost paradox
3	Compute and energy	Can we physically power it?	What Is Compute?
4	Verification	Can we prove it was right?	Verification
5	Accountability	Who answers when it acts?	Agent liability

In 2024 the industry was arguing about rung 1. In 2026 it is stuck between rungs 3 and 5. That is progress, but a different kind of progress than benchmark charts suggest.

What does this mean for superintelligence?

Writing on a site called BestSuperintelligent, I'll be direct about what the archive implies. First, a caution on the word itself: a label is not a capability, a point we make in superintelligence and today's tools.

First, superintelligence will arrive through bottlenecks, not around them. Every step-change covered was gated by something physical or institutional: power purchase agreements, memory fabs, export controls, liability law. A system smarter than us still needs a grid connection and still needs someone willing to sign for its actions.

Second, the people steering it are a variable, not a constant. That is why the sister publication built BioConfiguration™ Chronicles, structured profiles of how leaders such as Sam Altman and Jensen Huang make decisions. Governance choices like OpenAI's for-profit conversion and the legal fight over who owns training data will shape the trajectory as much as any paper.

Third, early movers compound. The analysis of the AI compounding advantage argues that organisations that deployed early are building data flywheels late movers can't buy. If that holds for firms, it is worth asking whether it holds for nations and labs, the subtext of its coverage of sovereign AI strategies.

Fourth, the information layer is being rewired. AI search answers questions directly, which is why the publication wrote AI Search Is Eating the Web's Business Model and published its own playbook for ranking in AI search. If machines become the primary readers of the web, publications must write for verifiability, not virality. That is the standard behind our methodology.

Why does the Constraint Ladder matter?

For founders, it suggests where margin may live: not in the model, but in whichever rung competitors haven't solved. For operators, it says an agent roadmap is a governance roadmap. For investors, the author argues it explains why AI valuations are detaching from cash flow while returns pool in energy, memory and picks-and-shovels. None of this is investment advice.

If you are choosing a model rather than a strategy, our model finder screens 14 models against your task and budget and returns a shortlist to test, not a ranking.

What are the limitations of this essay?

This essay is a retrospective of one publication's coverage, not a survey of the field. The archive has gaps, notably comparatively little published in 2025, so the eras are editorial groupings, not precise dates. Figures cited, such as the roughly 10x annual decline in per-token cost, come from the linked analyses and their underlying sources, and BestSuperintelligent.com has not verified them; check those primary sources before relying on them. Forward-looking statements about superintelligence are interpretation, not forecasts.

Evidence in this article Editorial interpretation Our reading of what the evidence means. Judgment, clearly separated from the evidence itself.

Frequently asked questions

How has AI changed between 2024 and 2026?

In 2024 the dominant question was which chatbot was best. By mid-2026 the questions were economic and physical: inference cost, power supply, memory and chip bottlenecks. By late 2026 the frontier had shifted to agents, where the hard problems are ownership, verification, governance and liability rather than raw model capability.

Why does AI spending keep rising if token prices are falling?

It is a Jevons paradox. The Best Blog Ever's analysis puts per-token inference cost declines at roughly 10x a year, but cheaper tokens unlock reasoning models and multi-step agents that use far more tokens per task, so total spend climbs. That figure is the author's and is not independently verified here.

What is the real bottleneck for AI and superintelligence?

Increasingly it is physical and institutional, not algorithmic: electricity and grid connections, memory supply, and the lack of verifiable oversight for autonomous systems. This is the author's reading of one publication's coverage, not a measured result.

Why do most agentic AI pilots fail to reach production?

According to survey data analysed by The Best Blog Ever, rarely because of model capability. The leading cause it identifies is unclear ownership: no one is accountable for what the agent does once it leaves the demo. We have not checked the underlying survey.

What is The Best Blog Ever?

The Best Blog Ever (thebestblogever.co) is an independent publication covering AI, technology, economics, business and investing. It shares an owner with BestSuperintelligent.com, which is disclosed on our About page. Its articles are commentary here, not evidence for any score.

Sources

Why AI Token Prices Fall While Total Spend Rises The Best Blog Ever (sister publication), July 2026. Source of the roughly 10x annual decline in per-token cost. Not independently verified.
Why Power Is the Real Bottleneck for AI The Best Blog Ever (sister publication), June 2026. Commentary; not independently verified.
China Took Over Open-Source AI The Best Blog Ever (sister publication), July 2026. Source of the open-weight model and token-share claims. Not independently verified.
Scaling Is Changing Shape: Three Papers The Best Blog Ever (sister publication), July 2026. Commentary on three research papers; we have not read the papers for this essay.
Why Agentic AI Pilots Stall Before Production The Best Blog Ever (sister publication), September 2026. Source of the unclear-ownership finding. Underlying survey not checked here.
How Enterprises Actually Govern AI Agents The Best Blog Ever (sister publication), September 2026. Commentary; not independently verified.
The Protocol That's Quietly Becoming the TCP/IP of AI Agents The Best Blog Ever (sister publication), August 2026. Commentary on the Model Context Protocol.