AI’s Savings Change Hands

If You Only Read One Thing

The AI labs are competing away part of their efficiency gains. Anthropic’s Opus 5.5 and OpenAI’s Sol and Luna make that transfer visible in prices. Greenland’s new security agreement uses a different mechanism: reserving local economic participation in an expanded American presence. Creating something valuable and keeping its proceeds are separate achievements. Today’s question is how those proceeds get divided.

AI’s Customers Collect the Discount

Better AI is becoming cheaper before its suppliers have established durable pricing power. Tuesday’s competing launches make the investment question sharper: how much of each efficiency improvement stays with the laboratory, and how much passes straight to customers?

The discounts have different denominators. Anthropic’s Opus 5.5 announcement sets input and output prices at $4 and $20 per million tokens, the pieces of text models process. That is a 20% cut from Opus 5. Its advertised 40% saving describes typical task costs in company tests, incorporating efficiency and cheaper reuse of previously processed context. It is not a universal 40% reduction in every price.

OpenAI’s Sol price sheet is simpler: input falls from $4 to $2 per million tokens; output from $20 to $10. Luna falls to $0.10 and $0.50. These are commercial prices available now, rather than promises that a future chip will make inference cheaper.

Think of a factory that cuts the cost of making a component. If competitors match the improvement, the factory may have to lower its selling price; the manufacturer buying the component gets the saving. AI applications face that same division of gains. A company charging for completed work can initially retain a cheaper model bill, provided competitors do not force it to cut its own price.

This extends September 15’s analysis of Apple owning the customer relationship. The new evidence is upstream suppliers publicly surrendering part of their efficiency gain. Epoch AI’s September 22 study estimates a 47% quarterly price decline at fixed performance across five math, science and game benchmarks since 2023. That is a bounded measurement, not a forecast for every business task.

The strongest counterargument is that cheaper models expand demand enough to improve laboratory economics. They can. Revenue still falls if a halved price is accompanied by less than a doubling of comparable billable volume; profit also depends on the cost of serving that volume.

My medium-confidence, 12–24-month read favors businesses with established distribution and valuable workflows, including Microsoft and enterprise software incumbents, over undifferentiated model resale. Laboratories can also win through their own applications. The falsifier is sustained model-level pricing power without a corresponding capability gap. The next useful evidence is disclosed gross margin alongside paid usage after these cuts: rising volume alone cannot settle who kept the savings.

Greenland has negotiated local economic participation in a larger American military presence, while accepting enduring security commitments. The bargain concerns both who gets paid and which future choices remain open.

The September 22 agreement affirms the kingdom’s territorial integrity and Greenlanders’ right to self-determination. That recognition comes with a consequential qualification.

The treaty text builds on the 1951 defense agreement. It provides for expanding Pituffik and additional defense areas at Narsarsuaq and Mestersvig, with details mutually agreed. Other new areas require consultation; unresolved discussions escalate after 90 days. Entry into force requires notification that parliamentary procedures are complete.

Article XI gives the agreement no end date. Denmark and Greenland undertake to ensure that an independent Greenland remains in NATO and assumes these obligations. Independence would therefore preserve the security bargain, rather than offer an automatic exit from it.

The commercial bargain resembles a town negotiating a large employer’s arrival. The town supplies a valuable location, but imported construction crews and supplies could carry much of the income away. A local procurement commitment attempts to keep part of the spending nearby. Here, defense-area goods and services must go to Greenlandic sources to the maximum feasible extent, subject to their capabilities.

That qualification is economically decisive. Local preference creates an opportunity; it does not create skilled labor, equipment or working capital. Over the next two to five years, capable Greenlandic contractors and outside suppliers partnered with them have a more defensible opportunity than a generic basket of Arctic resource shares. Confidence is medium because the agreement specifies a preference, not funded contract volumes.

The investment restrictions reinforce the strategic bargain. Sensitive sectors include resource extraction and critical infrastructure. Investors outside NATO, NATO-partner and EU states face restrictions on threatening control, influence or information access, with exceptions by agreement. Approved-country status can improve a bidder’s position without making a mine profitable.

The strongest objection is that much military access already existed. True: this is an amendment to a longstanding relationship, not a new American arrival. Its commercial significance lies in making local participation and investor screening more explicit. Unlike September 18’s Saudi aircraft purchase, the host is bargaining over income from supplying the location. The decisive test is the Greenlandic share of actual construction and service awards; persistent use of the feasibility exception would weaken the local-benefit thesis.

The Contrarian Take

Everyone says: Falling AI prices mean the laboratories are destroying their own business.

Here’s why that’s incomplete: A price cut does not reveal the change in production cost. Anthropic explicitly says Opus 5.5 needs less compute to serve, so a lower selling price can coexist with an unchanged or better margin. But demand growth is not automatically enough either: doubling usage at half the price merely preserves revenue on that comparable workload. The right comparison follows revenue, serving cost and customer retention together. That also complicates the supposed windfall for software companies. If their customers can reproduce the same work cheaply, the application vendor may have to pass the saving onward. The advantage belongs to whoever retains a scarce contribution—distribution, proprietary information or accountable service—after the model becomes cheaper.

Under the Radar

Token share and revenue share tell different stories. Vercel’s September 17 report, covering August, supplies useful context for Tuesday’s cuts: open-weight models handled 56% of its gateway’s tokens, while Anthropic retained 64% of spending. A model gateway routes applications’ requests to different suppliers; its traffic is one customer population, not the whole market. The split shows why winning the largest volume of inexpensive work need not mean capturing the largest revenue pool. (Vercel’s production index)
Renewables developers describe an execution shortage. At a September 21 industry discussion, reported Tuesday, participants told Semafor that demand and capital were available but projects were difficult to execute. Permitting, grid connections and equipment delays were recurring obstacles. This is qualitative testimony, not a measured national project freeze. The investment implication is narrower than “energy wins”: an already buildable project can be valuable while a larger unpermitted pipeline remains expensive inventory. (Semafor’s firsthand account)

Quick Takes

Iran diplomacy now has a meeting to evaluate. Trump said Steve Witkoff and Jared Kushner met with Iranian mediators on September 22. His later remarks specified intermediaries, which matters after reports implying direct US–Iran talks. The new signal is an active negotiating channel; it is not an agreed ceasefire or restored commercial passage. A reduction in the cost of conflict requires terms both sides will implement. (Source)
Alibaba is exporting the service around the model. Alibaba Cloud announced first cloud regions in Türkiye, Finland and the Netherlands over the next 12 months, plus expansion in five existing markets. Local capacity lets it sell deployment and support alongside AI capability. The plan therefore contests recurring infrastructure spending, not just model downloads. These are announced regions; delivered capacity and customer revenue remain unproved. (Source)
The IMF separates two financing threats. Kristalina Georgieva warned at Semafor’s September 22 event that higher interest rates were eroding emerging markets’ fiscal progress. She said the IMF does not yet see AI investment crowding governments out of borrowing markets. That distinction matters: higher debt-service costs are a different mechanism from investors refusing sovereign debt to finance data centers. Treating both as “AI absorbs all capital” would misidentify the pressure. (Source)

The Thread

Creating economic value does not determine its distribution. AI competition transfers part of an engineering improvement to customers through lower prices. Greenland’s procurement clause attempts to reserve part of a strategic investment for the people supplying its location. The useful distinction is between gains that competition can immediately pass onward and gains that a contract deliberately keeps local. Neither mechanism guarantees a durable return: application prices can fall again, and a local supplier can lack the capacity to fulfill an order. The stronger business combines a claim on the proceeds with the ability to deliver. That is why software distribution and Greenlandic contracting deserve closer attention than model usage totals or Arctic acreage alone.

Predictions

I predict: By November 30, Anthropic will publicly release at least one of the Sonnet 5.5 or Haiku 5.5 models it has announced for the coming weeks. The explicit release sequence supports this narrower call; it does not establish their prices. (Confidence: high; Check by: 2026-11-30)

September 23, 2026 · 03:19 AM ET