What the Arctic Actually Tests: The Economics of Convergence
By George Pullen, Chief Economist, MilkyWayEconomy
A data center is an industrial real estate decision. The Arctic is a place to find out whether a technology works when every input it depends on has been removed.
Companion to the essay published in Arctic Today on October 26, 2026. The sourcing behind the figures in that piece is set out here.
Most writing about artificial intelligence and the Arctic starts in the same place. Cold air, cheap cooling, a server hall somewhere north of the tree line. That story is real and it is not the interesting one. A data center is an industrial real estate decision. What the Arctic offers is a place to find out whether a technology works when every input it depends on has been removed.
I argued in Arctic Today that the Arctic is the hardest place on Earth to make the four forces of the Fourth Industrial Revolution meet: new energy, fintech, genetics, and big data, which we now call artificial intelligence. That piece carried the argument. This one carries the arithmetic, because every claim about the Arctic is ultimately a claim about cost, and cost is where an argument can be checked.
The price of power sets the price of thought
Start with the number that governs everything else. Electricity in a remote Arctic community is not expensive in the way a coastal city finds electricity expensive. It is expensive in a way that changes what is economically possible.
The U.S. residential average was 18.31 cents per kilowatt hour in July 2026. Alaska’s average was 28.83 cents. Rates in remote communities served by diesel microgrids can exceed one dollar, which is more than five times the national average and between roughly three and a half and four times the Alaska average.[1] Those communities are served by more than 150 standalone microgrids, most still running on diesel that arrives by barge or by air.[2] Alaska accounts for about half of the installed microgrids in the United States, a count rather than a capacity share, and it tells you how much of the American grid edge is really a northern problem.[1]
Now compare the cost of power that could replace it. Published projections for first-of-a-kind small modular reactors cluster in the eighty to one hundred fifty dollars per megawatt hour range, or eight to fifteen cents per kilowatt hour, with developer targets for Nth-of-a-kind units at fifty to eighty dollars per megawatt hour.[3] Those figures come largely from industry and vendor analyses, and independent cost verification is thin, so treat them as a planning band rather than a settled price. Even so, the direction is what matters. Electricity that costs a dollar a unit, served by generation that prices at fifteen cents, is a compression of at least eighty-five percent. Using the $1.20 figure the op-ed applies to the most remote communities, it is 87.5 percent.[4]
That compression is the single most important number in the Arctic AI story. Cost per unit of compute is, at the margin, a power price with a capital cost attached. A model that costs a dollar a kilowatt hour to run is not a cheaper model. It is a subsidy.
The honest version of this argument is harder than the promotional one. A 2025 study of Arctic microgrid economics found that renewable penetration of fifty to seventy-five percent is already cost competitive in these systems, and that pushing deeper is much more difficult economically.[2] That is a statement about the marginal cost of the last increment, and it matters more than the headline renewable percentage. The binding problem is not the first fifty percent. It is the firm capacity you must hold for a February night when the wind is dead and the sun is below the horizon for weeks. Storage and dispatchable generation are where the capital goes, and that is the part of the system that has to be financed.
This is where the Arctic stops being a location and becomes a specification. No grid means every watt must be generated, stored, and governed locally. It means the plant has to run through a polar winter without a supply chain behind it. It means the failure mode is not a brownout but a threat to life, which pushes reliability requirements, and therefore cost, up rather than down. The Arctic tests whether a power system can be small, autonomous, and financeable at the same time.
The price of patience
The second constraint is not engineering. It is the cost of capital.
A project that amortizes over a decade in a region where cash flows start late is a financing problem before it is a technical one. Standard project finance wants predictable revenue and a creditworthy counterparty standing behind the offtake. The Arctic is governed by eight sovereign states, so it is not stateless, and it is not without public institutions. Alaska has the Alaska Industrial Development and Export Authority. Norway has Enova. Both matter, and both are the reason some projects get built.
What the region lacks at project level is a bankable anchor: a creditworthy offtake, or a guarantor willing to absorb first loss, on infrastructure whose cash flow begins a decade out. Where that anchor is missing, the discount rate climbs, and a high discount rate is an instruction to build nothing.
Two instruments change the arithmetic. The first is parametric insurance, which pays on a measured trigger rather than an assessed loss, settled with satellite and sensor data instead of a claims adjuster who cannot reach the site in February.[5] It converts an unpriced catastrophe into a priced, and therefore transferable, risk. The price of that transfer is basis risk, the gap between the trigger and the actual loss, and pricing basis risk honestly is the whole discipline.
The second is blended finance with a concessional first-loss tranche. This is the structure that made renewable energy bankable in emerging markets: public or philanthropic capital absorbs the first losses, private capital takes the mezzanine and senior positions, and the blended cost of capital falls far enough that the project clears its hurdle rate. Tokenized claims on infrastructure assets are the third leg, widening the secondary market so that patient capital is not locked in for twenty years.[6] Analysts who model tokenization put the 2030 market in the trillions, with published estimates spanning an order of magnitude. The point estimate is less useful than the mechanism, because the mechanism is what matters: liquidity lowers the illiquidity premium, and the illiquidity premium is most of what makes frontier projects uninvestable.
The price of substitution
The third constraint is a race between two timelines, and the timelines are measured in different units.
A mine is a decade-long project with a billion-dollar capital commitment before the first tonne moves, and the permitting clock runs on political time rather than engineering time.[7] A biological substitute moves on a learning curve, where each iteration is cheaper than the last. Market forecasts for synthetic biology cluster in the high teens to mid twenties percent annual growth, and the forecast matters less than the direction of the cost curve underneath it.[8]
Put those two together and the strategic conclusion follows without sentiment. When the substitute improves on a compounding curve and the mine improves on a permitting schedule, capital that waits for the mine is making a bet against the compounding. That is why I called the Arctic mineral frontier a mirage in Arctic Today last year, and why the biology is the more interesting asset. The Arctic is the planet’s deepest library of cold-adapted biology, and every advance in protein structure prediction makes that library easier to read.[9] The value is not only in what the ice holds. It is in what the ice has already solved.
The price of verification
The fourth constraint is the one the AI industry is least prepared for, because it is an information problem dressed as a hardware problem.
In a place where physical auditing across thousands of frozen miles is cost-prohibitive, the cost of verification becomes part of the cost of operating. Coase’s insight was that transaction costs, and the firm boundaries they impose, decide what an organization can do.[10] Verification is one of those costs, and in the Arctic it is large enough to decide the answer. If you cannot afford to check, you cannot afford to operate at scale.
The response is data as infrastructure rather than data as byproduct. Orbital sensing nodes, autonomous surface and aerial platforms, and digital twins that stay synchronized with a physical asset nobody can visit. Bandwidth is the binding constraint on the orbital layer, which is why processing telemetry in orbit rather than shipping it down matters economically. It converts a bandwidth cost into a compute cost, and compute is the input whose price is falling fastest.
Why convergence is the test
Each of these four forces can be pursued separately in a mature economy. The Arctic is the only place that forces all four at once, because it removes the slack that usually hides a weak system. Remove the grid and energy becomes a design constraint on everything downstream. Remove the insurers and capital becomes the design constraint. Remove the road and the depot, and verification becomes the design constraint. A model that runs in a temperate data center tells you almost nothing about whether it can run a grid, a mine, or a life-support system a thousand miles from the nearest technician.
That reframes what the Arctic is worth. It is not a prize to be divided. It is a test bed, and the correct way to value a test bed is as an option. Building an autonomous microgrid, a closed-loop verification system, or a resilient biological supply chain in the high North is the purchase of a capability that can be exercised somewhere else later. Real options theory says that asset is worth more than its discounted cash flows, because the cash flows are not the point. The point is the right, not the obligation, to deploy the capability elsewhere.[11]
The fair objection is that engineering against the most extreme conditions can over-specify a system and price it out of every softer market. That objection is a reason for discipline about what gets built. It is not a reason to skip the test. The discipline is to build the smallest system that survives the worst condition and can be sold into the easier ones, which is exactly the specification a startup should welcome.
What to watch
Four numbers will tell you whether this is happening, and none of them is a launch date.
1. The delivered LCOE of Arctic microgrids. Not the target. The delivered number, per community, per year.
2. The blended cost of capital on frontier infrastructure. If first-loss tranches are doing their job, this falls before the projects look attractive.
3. The cost per verified data point for remote assets. This is the number that decides whether autonomous operations scale.
4. The substitution rate for critical minerals. If engineered alternatives displace demand on a learning curve, the mine case weakens every quarter.
A convergence zone, not a mine
None of this lowers the human stakes. It raises them. Roughly four million people live in the Arctic.[12] That includes Iñupiat and Yup’ik peoples in Alaska, Inuit communities across Canada and Greenland, and the Sámi in northern Europe, whose consent and benefit are part of whether the test is passed. In many of these communities the microgrid is also the clinic, the school heating, and the water plant, so a failure is measured in people rather than in dollars. A region treated as a mine is treated as expendable. A region understood as the place where hard technology gets its final exam is a region understood as indispensable.
The minerals will keep making headlines. They should not. The Arctic’s value was never in what lies beneath it. It is in what it forces a nation to build in order to stand on it at all.
Sources
1. U.S. Energy Information Administration, Electric Power Monthly, Table 5.6.A, July 2026 (U.S. residential average 18.31 cents per kilowatt hour, Alaska average 28.83 cents). Remote community rates above one dollar per kilowatt hour: “Rural Alaska Electricity Costs Run Three to Five Times Urban Rates,” Alaska News, April 22, 2026, citing University of Alaska Fairbanks International Arctic Research Center data. https://alaskanews.com/c/alaska-news/rural-alaska-electricity-costs-run-three-to-five-times-urban-rates
2. Number of standalone microgrids and renewable penetration economics: “Learning from Arctic Microgrids: Cost and Resiliency Projections,” Sustainability 17(13):5996 (2025). Archived copy: https://www.osti.gov/pages/servlets/purl/2570815
3. First-of-a-kind and Nth-of-a-kind SMR cost bands, drawn from a survey of published industry estimates. Independent cost studies remain thin, so read these as a planning band rather than a settled price.
4. Compression arithmetic: one dollar to fifteen cents is 85 percent. The op-ed states the $1.20 case as a baseline for the upper end of the compression range, which yields 87.5 percent. The verified floor supports the 85 percent floor. Multiple arithmetic: one dollar is 5.5 times the national residential average of 18.31 cents and 3.5 times the Alaska average of 28.83 cents, while $1.20 is about 4.2 times the Alaska average. Both averages are July 2026.
5. Parametric insurance settles on a measured trigger rather than an assessed loss. For remote and extreme environments the underlying index is typically satellite or sensor derived, which substitutes basis risk for catastrophe risk rather than eliminating risk.
6. Tokenized claims on real world assets as a secondary market mechanism. Published 2030 estimates for the tokenized asset market range from low single digit trillions to above fifteen trillion dollars, a spread wide enough that the mechanism, not the midpoint, should carry the argument.
7. Mine development timelines and capital intensity for Arctic and remote projects. The decade-scale lead time and billion-dollar capital requirement are well attested in mining economics literature.
8. Synthetic biology market growth forecasts cluster in the high teens to mid twenties percent annually across major market research houses. The direction of the cost curve matters more than the precise growth rate.
9. Protein structure prediction applied to cold-adapted biology, including AlphaFold and successor models, is an active research area and shortens the path from sequence to candidate molecule.
10. Ronald Coase, “The Nature of the Firm” (1937), on transaction costs and the determination of firm boundaries.
11. Real options valuation: an asset that confers a right, but not an obligation, to invest later carries option value above its discounted cash flows. Applied here to capability built under Arctic constraints.
12. Roughly four million people live in the Arctic across eight countries: WWF Arctic programme. https://www.worldwildlife.org/places/arctic/ Consistent with Arctic Council human development reporting.
Declaration of interest
George Pullen is Chief Economist and a partner at MilkyWayEconomy, and this work was produced in that capacity as part of MilkyWayEconomy’s research program. He also serves as a Senior Economist at the U.S. Commodity Futures Trading Commission. The views expressed are those of the author and do not necessarily reflect the position of the U.S. Commodity Futures Trading Commission. No material non-public information was used, and no client-specific information is referenced.
Full disclosure of interest: https://milkywayeconomy.com/disclosure
George Pullen is Chief Economist at MilkyWayEconomy, a Washington, DC and Maine-based innovation advisory and investor in space and deep-tech companies. MilkyWayEconomy advises companies in the energy, space, and deep-tech sectors discussed in this article.
This article reflects the author’s professional opinion and is not investment advice.
Get MWE insights by email
SBIR alerts, agency updates, and funding strategy — direct to your inbox.
Get MWE insights by email
SBIR alerts, agency updates, and funding strategy — direct to your inbox.