The Long Walk From Cost to Price

Tariff politics likes a clean villain: a border charge enters, a price rises, a voter notices. Real supply chains are less obliging. They consist of components, contracts, inventory, financing, distributors, retailers and—at the last possible moment—someone deciding whether to eat the cost, shrink a promotion, or quietly make the product worse. The tariff may be the speech; the margin is usually the plot.

That is why the most revealing omission in Fairphone’s U.S. launch announcement is its price. The company says its newly U.S.-available unlocked Gen. 6+ has a newer Snapdragon processor, 12GB of RAM, 12 replaceable parts, support through 2033 and a warranty of up to five years. Those are all claims about a product’s intended life and cost structure. But without a retail price, carrier compatibility information, or independent assessment of its repairability claims, consumers cannot yet calculate the relevant figure: whether durability is being sold as an economy or as a luxury accessory.

The company says the phone contains 51% fair and recycled materials by weight and is assembled under fair working conditions using renewable energy. Those claims, if borne out, describe a supply chain that is trying to price in obligations routinely externalized elsewhere. That does not abolish trade friction; it makes visible that a cheap handset has never been merely a handset. It is a stack of material, labor, logistics and future replacement decisions wearing a glossy rectangle as camouflage.

The AI business has acquired a familiar industrial talent: making its inputs look abstract until the invoice arrives. OpenRouter’s GPT-5.6 Sol page lists a limited-time discounted price of $2.50 per million input tokens and $15 per million output tokens through OpenAI, while other providers list materially different prices, speeds and availability. The 1m-token context window is the billboard; the meter is the business model.

Even the displayed averages resist slogan-making. OpenRouter reports that users actually paid a weighted average of $0.889 per million input tokens and $20.22 per million output tokens, with caching and discounts helping make the relationship between published and paid pricing substantially less photogenic. A model purchaser is therefore not buying intelligence at a sticker price. They are buying routing, cache behavior, reliability, latency, output volume and the unpleasant possibility that an agent takes a scenic tour of the token budget.

This is a supply-chain lesson in miniature. The nominal input price is not the total cost, just as a tariff rate is not automatically the shelf-price increase. The downstream actor decides what gets passed on, optimized away, bundled, delayed, or converted into a smaller service level. Conveniently for everyone writing keynote slides, the final consumer often sees only a monthly subscription and a chatbot that has learned the ancient commercial art of taking longer than expected.

When physical capacity is scarce, the economics become brutally literal. A SteamOS-focused account of VRAM overcommitment explains that data evicted from GPU memory must be accessed from system RAM over PCIe, and that the damage is not confined to a slower average. Repeated evictions and reloads can create miserable frame-time spikes; allocations for display scanout can require much larger memory moves because they need physically contiguous VRAM.

The author reports an especially rude example: up to 4GiB evicted to make room for scanout images containing roughly 32MiB of pixel data each. On an assumed PCIe 4.0x16 link, the post estimates moving 4GiB out of VRAM takes at least about 130 milliseconds. This is not a supply-chain delay in the metaphorical sense. It is a traffic jam between memory pools, one that occurs several thousand times a second while the screen is trying to pretend nothing happened.

The proposed remedy is revealing: priorities, throttles and better eviction rules. In other words, rationing. The author says application memory-priority hints can help preserve the resources that matter most, and reports SteamOS releases of related work, though upstream integration remains unfinished. Hardware scarcity does not disappear because software calls it “overcommitment”; it becomes a scheduling problem, then a user-experience problem, then inevitably a pricing problem when buyers decide how much capacity they need to avoid the queue.

Some supply chains do not move containers or chips. They move the records left behind by companies that no longer exist. The Register reports that Google paid $10 million at auction for a purportedly deidentified Spirit Airlines data trove after the carrier permanently grounded itself and entered liquidation. The reported collection includes emails, Teams items, OneDrive and SharePoint files, customer-service calls and chats, along with flight, crew, fuel-slip and parts-purchase records.

There is a grim efficiency to it. An airline’s operating history—its customer frictions, maintenance purchases, crew pairings and fuel paperwork—can outlive the airline as an asset for someone else’s AI efforts. The court filing reportedly described the material as deidentified, and Google promised to scrub personally identifiable information it finds. Yet the sheer heterogeneity of the collection is the point: deidentification is not a magic spell cast over hundreds of millions of communications and operational artifacts. The supplied reporting does not detail the method, auditing or enforceability.

This is where the price system becomes a little less reassuring. A failed company is supposed to liquidate assets and settle debts; that is capitalism’s tidying ritual. But data makes the ritual stranger. Planes stop flying, workers and customers move on, and the residue becomes a new industrial input. The modern supply chain does not end at the store shelf or the scrapyard. Increasingly, it ends in a training pipeline, where the leftovers are not discarded so much as taught to become productive again.