Edition #2 · March 2026
THE BIG STORY
Three Years. That's All It Took.
AI just stopped asking for your permission. The question is whether you've built anything capable of handling what comes next.
Three years ago, ChatGPT was a novelty and GPT-4 had just launched. Today, Gemini and Claude act — booking meetings, managing workflows, executing multi-step tasks across systems while you're doing something else entirely. That's not an upgrade. That's a categorical shift in what software is. We've crossed from AI as a sophisticated input box to AI as an autonomous operator, and most organizations are completely unprepared for what that actually demands of them.
Here's the thing nobody's saying clearly enough: the agent threshold isn't primarily a technology story. The models are ready. The agentic behavior is shipping. What isn't ready is the organizational architecture required to capture the surplus.
Think about what autonomous agents actually require to function at scale. They need clear goal hierarchies — not prompts, goals. They need handoff protocols between agents so that a research agent can brief a drafting agent which then routes to an approval agent without a human babysitting every transition. They need error-handling logic that doesn't require human intervention at every edge case, because if you're manually fixing agent mistakes, you haven't automated anything — you've just created a new class of junior employee that works faster and makes weirder errors. And they need accountability structures, because agents acting in your name are making commitments your organization is responsible for.
Most companies are nowhere near this. They're still in the "impressive demo" phase — spinning up individual agents for specific tasks, marveling at the output, and calling it transformation. That's not transformation. That's using a bulldozer to dig a single flower bed.
The real competitive divide isn't forming between companies that have AI and companies that don't. It's forming between companies building systems — agents orchestrating agents, tasks routing automatically, capability compounding — and companies that are just spawning smarter assistants. The former is how you collapse cost curves. The latter is how you get incremental productivity gains and a large consulting bill.
To understand why this matters for abundance specifically, consider coordination costs. The Abundance Trifecta — AI, robots, cheap energy — doesn't automatically produce abundance. It produces the potential for abundance, which only gets realized when the friction of complexity is removed. Manufacturing a solar panel is cheap. Coordinating the supply chain, permitting, installation scheduling, grid interconnection, and maintenance contracts is expensive — not because the tasks are hard, but because the cognitive overhead of managing them across time, people, and organizations is brutal. Agents that can pursue goals across weeks, not just conversations, are the unlock for that coordination layer. When complexity becomes free to manage, the things that were too complicated to scale stop being too complicated to scale.
This is the correct framing. Abundance isn't just about making individual things cheaper. It's about eliminating the organizational friction that prevents cheap things from combining into cheap systems. And that's precisely what autonomous agent networks do when they're architected correctly.
So what does correct architecture look like? The companies getting this right are treating agents as infrastructure, not features. They're mapping entire workflows — not tasks — and identifying where human intervention is genuinely necessary versus habitual. They're building routing logic before they build agent capabilities, because an excellent agent dropped into a broken workflow just fails faster. And they're starting to think seriously about what anthropologists call "supervision at scale": how do you maintain meaningful oversight of dozens of agents acting in parallel without recreating the management overhead you were trying to eliminate?
This last problem is harder than it sounds, and I'd argue it's the defining organizational challenge of the next five years. The instinct will be to solve it with more agents — an oversight agent monitoring the work agents, a quality agent reviewing the oversight agent's flags — and at some point that recursion either stabilizes into a functional system or collapses into chaos. We don't have enough real-world data yet to know which companies will get this right and which will drown in agent entropy.
My prediction: by the end of 2026, the first major corporate scandal involving autonomous agent liability will have forced regulators to produce binding guidance on what constitutes organizational accountability when an agent makes a consequential decision. This isn't speculative hand-wringing — it's a logical consequence of agents booking commitments, negotiating terms, and managing financial transactions at scale. The legal system's concept of agency is several centuries old. It's about to get stress-tested by entities that act but can't be sued.
That regulatory event will split the landscape. Companies that built proper governance into their agent systems from the start will demonstrate compliance relatively easily. Companies that bolted agents onto existing workflows without accountability architecture will face both legal exposure and the sudden cost of retrofitting structure they should have built first. The compliance cost of getting this wrong will be high enough to bankrupt mid-sized companies.
The abundance thesis here is ultimately optimistic, but it demands intellectual honesty about the transition. Agent-based automation will collapse coordination costs, multiply the effective output of every knowledge worker, and eliminate entire categories of operational overhead that currently consume organizational energy without producing anything. That's real. That's coming. But the surplus doesn't distribute automatically. It concentrates in the organizations that treat the transition as an architectural problem, not a procurement decision.
The agent threshold has been crossed. Now build something worthy of it.
TODAY'S EDITION IS BROUGHT TO YOU BY
Unscarcity
Every week, we break down the three forces reshaping the global economy: robots getting cheaper than labour, AI getting cheaper than thought, and energy getting cheaper than ever. No hype. No hedge. Just the signals that matter. If someone forwarded this to you, now's the time.
TRIFECTA UPDATE
ROBOTS Agility Robotics' Digit has crossed from pilot to production. At GXO Logistics' Spanx facility, Digit has moved over 100,000 totes in full-time commercial operation — actual shift work, actual throughput targets, running for over a year (Agility Robotics). Amazon is separately testing Digit in its fulfillment network at a facility in Sumner, Washington (TechCrunch). Meanwhile, Apptronik signed a pilot agreement with Mercedes-Benz to test its Apollo humanoid in manufacturing logistics — bringing parts to the line, not assembling vehicles (PR Newswire). Mercedes also invested in Apptronik's Series A. The gap between pilots and production deployments is still real, but it's closing. When a fulfillment operation runs a humanoid through 100,000 real totes and keeps it on the floor, the proof-of-concept debate is over for that use case.
AI In January 2025, a Chinese lab called DeepSeek released R1 — a reasoning model that matches OpenAI's o1 on math, code, and reasoning benchmarks. They trained it for roughly $5.9 million and released the weights under an MIT license (DeepSeek). The API pricing came in 90-95% cheaper than comparable offerings from OpenAI and Anthropic, triggering a global price war that wiped $600 billion off Nvidia's market cap in a single day (CNBC). The second signal is equally important: frontier-class models now run on consumer hardware. An NVIDIA RTX 5090 — a $2,000 graphics card — runs 70-billion-parameter models locally, delivering performance that matched the absolute frontier from 12 months prior (Epoch AI). The combined message is unmistakable: capable AI is decoupling from proprietary lock-in and from cloud dependence simultaneously. When a $6 million model matches a $100 million one, and a $2,000 GPU runs what required a datacenter last year, the pricing power of the frontier labs erodes faster than their research leads can compensate for.
ENERGY Form Energy began manufacturing commercial-scale iron-air battery systems at its factory in Weirton, West Virginia. The chemistry is almost absurdly simple — iron rusts, releases electrons, un-rusts when you push current back through it — but the implications are not simple at all. Iron-air stores electricity for 100 hours, targeting a cost of $20 per kilowatt-hour of capacity (Utility Dive), compared to lithium-ion's $70 to $110 range for storage packs (BNEF, 2025). Even at current lithium prices, that's a dramatic cost advantage — and iron is one of the most abundant materials on earth. The knock on wind and solar has always been intermittency — the grid needs power when it needs power, not when the wind blows. Iron-air doesn't solve intermittency. It makes intermittency irrelevant. First deployments are contracted with Great River Energy in Minnesota (Form Energy), with a 300 MW project announced with Xcel Energy (Form Energy). If Form Energy's production costs track the learning curve the way lithium did, the long-duration storage problem that has blocked full grid decarbonization may have just found its answer.
THE NUMBER
$0.01
That's the per-kilowatt-hour cost of solar generation at the world's cheapest utility-scale project. Al Shuaibah in Saudi Arabia locked in a power purchase agreement at $0.0104/kWh in April 2021 (PV Magazine). One cent. For context, the marginal fuel cost of running a natural gas plant — just the gas, before you account for the turbine, the maintenance, the staff, or the capital — runs between three and five cents per kilowatt-hour depending on spot prices. Solar, in the right conditions, is now three to five times cheaper to operate than the cheapest fossil alternative is to fuel. The global average solar LCOE stood at $0.043/kWh in 2024 (IRENA). Even the average is approaching parity with gas on fuel cost alone. The energy that took a hundred and fifty years of industrial civilization to make cheap has been undercut by sunlight hitting glass. The transition isn't a policy story anymore. It's an arithmetic story. And arithmetic doesn't negotiate.
WHAT I'M WATCHING
The CPU inference war. GPUs trained the models. CPUs will run them at scale — particularly as context windows expand into millions of tokens, which is fundamentally a memory bandwidth problem that plays to CPU architecture. AMD is gaining ground on Intel, Arm is quietly colonizing datacenters, and hyperscalers are designing custom silicon. Whoever controls CPU efficiency for inference controls the cost structure of AI deployment for the next decade. This is less visible than the GPU arms race and more consequential.
The inference cost collapse. The cost of running AI inference has dropped roughly 280-fold in two years. The Stanford HAI AI Index 2025 documents this: querying a GPT-3.5-equivalent model fell from $20 per million tokens in November 2022 to $0.07 per million tokens by October 2024 (Stanford HAI). That's a 99.65% cost reduction. This is what technology transitions actually look like when they're real: not gradual improvement, but the floor dropping out from under the old pricing model entirely. Every business case built on "AI is too expensive to run at scale" was written about a different technology than the one that exists today.
Brookfield's $20B Global Transition Fund II. The smart infrastructure capital is scaling fast. Brookfield Asset Management raised $20 billion for its second Global Transition Fund — one of the largest energy transition funds ever assembled (Brookfield). The fund is deploying capital across generation, transmission, and storage, with early investments in Neoen and other renewable developers (Utility Dive). When one of the largest infrastructure investors on earth commits $20 billion to the energy transition across the full value chain, it's telling you something about where institutional capital sees structural returns. The transition has moved from policy aspiration to institutional conviction.
THE QUESTION
If an AI agent negotiating on your behalf signs a contract your legal team wouldn't have approved, the agent builder will point to the terms of service, your CTO will point to the agent configuration, and your counterparty will point to your company name on the agreement — so who actually owns that liability, and does your answer change if the agent saved you money instead of losing it?
UNSCARCITY A weekly newsletter on the convergence of robots, AI, and cheap energy driving sustainable abundance.
Was this forwarded? Subscribe here.
© 2026 Unscarcity. All rights reserved.