The great train robbery

The great railway fortunes went not to the engineers who understood steam but to those who owned the track. With intelligence progressively becoming commoditised, Mark Zuckerberg is effectively giving away steam. On 10 August Meta released the weights of Muse Glimmer, a thirty-billion-parameter model distilled from its proprietary Muse Spark, and published them under a permissive licence, free to download, modify and run, with a version of the larger model promised to follow. Zuckerberg framed the release inside a lengthy letter that calls an extreme concentration of AI power "inherently problematic". His safer path runs through personal superintelligence held by billions of individuals rather than controlled by governments and corporations. Alongside it was another announcement that attracted far less attention, a billion-dollar fund for the American communities hosting Meta's data centres, directed at teachers and first responders along with local energy and water infrastructure.

I think that may prove the more important of the two. Companies do not usually give away something they expect to stay scarce and expensive, and the evidence that intelligence is losing its scarcity has been accumulating in public all year, sitting in the pricing rather than in anything Meta has said about it. The same cannot be said for the power and water that keep a data centre running, both of which stay genuinely scarce even as the models commoditise. Read that way, the billion-dollar fund looks less like generosity and more like a down payment, easing the local resentment that can stall a grid connection or a planning application and buying an early place in the queue for the resources Meta cannot simply give away.

The price of steam

Take the near-frontier group scoring above 50 on the Artificial Analysis Intelligence Index. Seventeen index points separate the weakest of them from the strongest, yet published output tariffs across the same group differ by a factor of roughly 180, and headline rates flatter the expensive models because what actually matters is the cost of finishing a job. Measured that way the spread narrows to about 92-fold. Kimi K3 is the standout entry on that measure, completing the average index task for $0.72 and undercutting every model that scores above it. The wider market tells the same story from below. The median API price sits near a dollar per million input tokens, the cheaper tiers have shed more than a third of their price in a year, and something close to GPT-4 class capability, sold at thirty dollars per million input tokens when that model launched in 2023, now fetches fourteen cents.

Figure 1. Cost of completing an average Artificial Analysis index task, plotted against the index itself, leaderboard read 2 August 2026. The headline tariff spread across this group is 179-fold; measured on completed work it is 92-fold, and Kimi K3 sits below every model that scores above it.

Frontier pricing has moved the other way and roughly doubled since January, which at first sight undermines the argument, although the timing repays a closer look. On 30 July OpenAI cut one of its new models by eighty per cent and another by twenty per cent, three weeks after launching them, with no corresponding change in their benchmark position, while the flagship tiers held where they were. A premium that survives only three weeks is not really a price level. It looks more like rent, collected while a laboratory holds a genuine capability lead. That window has been getting shorter with each generation, compressed most effectively by Chinese developers releasing frontier-adjacent systems under MIT and Apache licences at a fraction of Western tariffs. Their models now carry more than half of the tokens routed through OpenRouter, a lead they have held for fourteen consecutive weeks with nine of the top ten models by usage.

Meta carries a different incentive from the model makers that sell tokens, because it sells advertising. It therefore loses very little by pushing the price of adequate intelligence towards zero, and it gains a great deal if the model doing the commoditising arrives with an American licence and a jurisdiction that Western procurement departments can accept. None of that requires charity as an explanation. It means only that Meta can afford to hold a permanently cheap floor underneath the market in a way that a company depending on token revenue cannot.

What the accounts show

If you want to know where Meta thinks the value is moving, look at what the company is spending rather than at what Zuckerberg is writing. Guidance for the year now reaches $145 billion at the top of the range, and the recent revision lifted the floor while leaving the ceiling alone, so the commitment hardened even though the headline number never moved. In the second quarter, capital spending of $31.1 billion absorbed almost all of the $31.9 billion the business generated. Free cash flow came out at $784 million, down ninety-one per cent on the year and smaller than the $1.35 billion paid in dividends over the same three months. Meta is not spending alone either, since alongside Microsoft, Google, Amazon and the SpaceX and xAI complex the industry has something like $750 billion of infrastructure investment planned for 2026, and Morgan Stanley expects hyperscaler capital expenditure to reach about $3.5 trillion across the three years to 2028.

Numbers of that size no longer look much like software economics, and they look more like infrastructure, where returns accrue over decades to whoever owns the route rather than to whoever designed the engine. Meta sits more exposed within that than its peers, because it has no large cloud business through which to rent surplus capacity back to customers, and what it owns instead is land, power and the political permission to keep building. Seen against that backdrop, the community fund starts to make sense. Zuckerberg made the comparison explicitly, arguing that the towns which welcomed railways, highways and electrification became centres of industry and growth, and inviting host communities to read data centres the same way, an awkward pitch at a moment when New York has already imposed a moratorium of up to a year on approvals for new hyperscale data centres, and when the Louisiana parish held up as the model received teacher bonuses funded by tax revenue from a campus for which Meta secured billions in tax relief. Set a billion dollars spread across host communities against a $145 billion annual construction programme and the sum starts to look like the price of obtaining a right of way.

Who ends up owning the track

On the same day Meta published its letter, NVIDIA signed agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create compute financing platforms. The goal is to mobilise more than $500 billion of third-party capital, with compute assets supporting the financing so that operators borrow rather than pay upfront. Pension funds, insurers and sovereign wealth funds are named as the intended source of the money. The attraction to everyone in the room is obvious enough, since it moves the capital load off the technology balance sheets and gives long-duration investors an asset that generates usage-linked revenue in a growing market. Jensen Huang met the obvious objection head-on, telling CNBC that demand for the underlying compute is genuine and that chips now carry the productive life and offtake ecosystem of a proper asset class. That claim sits oddly next to the decision itself. A company convinced that its own hardware will hold its value for a decade does not usually need six outside institutions to help it believe that, and moving the risk onto pension and insurance capital only looks necessary if the people closest to the asset are less certain about its long-run earnings than their public statements suggest.

The railway comparison stops working at exactly this point. A railway line lasts for generations, as do a transmission network and a fibre spine, which is why capital stretching over decades can sensibly be raised against them. A GPU fleet holds none of those properties, because its revenue depends on utilisation, its economic life is short, the price of the service it performs is falling by more than a third a year, and the premium available to the best models appears to last only until the next generation arrives. Some of the compute contracts underneath these assets run for only two to four years and several carry ninety-day termination rights, which leaves an awkward mismatch between capital raised over a long horizon and revenue visibility that does not reach anywhere near as far. Somebody has to own what sits between the two.

Figure 2. Revenue visibility against the tenor of capital typically raised against the asset. The first four rows are author judgement from standard market structures. The compute rows use disclosed contract terms, and the shaded band is the residual that somebody has to own.

NVIDIA is simply the most visible participant in a wider migration. Goldman Sachs puts hyperscaler lease commitments at about $1.5 trillion, including roughly $1 trillion of leases not yet commenced and therefore absent from the balance sheets, and Morgan Stanley identifies another $982 billion of non-debt purchase commitments on top. Every one of these companies has an incentive to move that exposure somewhere else, and they will all look for ways to sell the debt on, passing the exposure to whoever will hold it longest. That is how a balance sheet sheds a risk it no longer wants to carry. Whether the risk genuinely leaves depends on who keeps the first loss. A structure that carries buy-back undertakings or utilisation guarantees in its documentation has not really transferred anything, and nobody outside the rooms has yet seen those terms.

The Bank for International Settlements reached for the same historical comparison in its 2026 annual report and arrived somewhere less comfortable. Its reading places the canal and railway booms beside the electrification mania of the 1920s. Each was a genuine breakthrough that drew in more capital than the eventual commercial returns could support, and each closed with an investment reversal that dragged the wider economy down behind it. Singapore's central bank added a nearer-term worry in June, when the Managing Director of the MAS estimated that AI capital expenditure now accounts for close to half of recent United States growth and warned that a fairly modest change in the assumptions behind current valuations could bring on a sharp reversal. The MAS regulates a good number of the institutions now being invited into these structures, and within two months of that warning the industry's most important supplier had begun bringing insurance and pension capital into ownership of the assets, a sequence that deserves attention whatever conclusion you draw from it.

The motor car

A further complication runs underneath all of this, and it may prove the largest one, because the same process that is making intelligence cheaper is also making it smaller. Muse Glimmer demonstrates the point directly, fitting into less than twenty gigabytes at four-bit precision with negligible claimed degradation, running on a single consumer graphics card, and shipping with a speculative decoder that roughly triples throughput on a current desktop, positioned explicitly for agents that run continuously on local hardware. Further down the stack a one-bit build has compressed a 27B-class model below four gigabytes and demonstrated it on a phone at a claimed electricity cost near nine cents per million tokens, against twenty-eight cents for the cheapest hosted alternative. As much as eighty per cent of inference could eventually move to edge devices. That figure is my own scenario rather than a published forecast, and it may well prove too high, but it does not need to approach eighty for the infrastructure arithmetic to change, since every task performed on a phone, a laptop or a gateway is a task that never reaches a rented GPU.

Figure 3. GPQA Diamond against an illustrative deployment envelope. Muse Glimmer, released on 10 August, runs at four-bit precision on a single 24GB consumer card and sits within a point of Gemma 4 31B. Figures are vendor-published and not independently evaluated.

The railways did not lose their passengers to a better railway. They lost them when people acquired their own means of transport and took the traffic off the network, and Zuckerberg is now spending enormous sums building the track while helping to make the intelligence that travels over it small enough and cheap enough to leave. That may be entirely deliberate, given that Meta carries little economic exposure to the value of the models and enormous exposure to the land and power underneath them, but it raises a larger question. If intelligence itself becomes a commodity, what remains scarce? The answer does not appear to lie in the weights, which grow more interchangeable by the month, and for now it looks like power and land together with the physical plant sitting on them, though how long that answer holds depends on how much intelligence continues to need the data centre at all.

All of which returns us to the robbery in the title. It is not the giving away of the models, which is commercially rational and may well be a public good. The transaction worth watching is happening alongside it, where assets with short economic lives and limited revenue visibility are being financed as though they were long-lived infrastructure, with pension funds and insurers invited to provide the patient capital.

The economics may yet justify that, and the case for it is not weak. If inference demand grows fast enough, utilisation can outrun both falling unit prices and rapid obsolescence. That is precisely what NVIDIA and its financing partners are betting on, and they have better visibility of order books than I do. The early returns on that bet are arriving now, and they cut in both directions at once. Traffic has grown exactly as the bulls need it to. Token volumes routed through OpenRouter are up roughly eightfold since January and touched eight trillion in a single day in early August. Yet blended expenditure per million tokens has fallen faster still, peaking just above two dollars in June on Silicon Data's index and dropping to about $1.17 by early August, a decline John Authers at Bloomberg reads as weak pricing power rather than efficiency. The index blends price with usage mix, so part of the fall reflects buyers migrating to the cheap Chinese models rather than paying less for the same thing, but that migration is not really an alternative explanation. It is the mechanism, since traffic moving down the price curve erodes the revenue assumptions of the expensive capacity just as surely as traffic disappearing would. What the bet requires is that spending per unit of capacity holds up rather than merely the token count, and that is the line I would want to see stabilise before committing a pension scheme to it. Those investors believe they are buying the track, and my own doubt is over how much of what they are buying will turn out to be rolling stock.

Figure 4. Blended expenditure per million tokens against daily token volume, January to early August 2026. The expenditure series is Silicon Data's LLM Token Expenditure Index as charted by Bloomberg Opinion; volume is OpenRouter totals via MacroMicro. Both series are reconstructed from the published charts, and the index blends price with usage mix, so migration to cheaper models and genuine price cuts both pull it down.

The passengers

If a robbery is under way, the investors are the ones exposed to it, and the people using the intelligence come out of the story remarkably well. Every force now in motion, from the collapsing tariffs to the models running on hardware people already own, pushes towards intelligence as a possession rather than a rental, and that direction holds whether the financing structures survive or not. Capability that sold at thirty dollars per million tokens three years ago now fetches fourteen cents, a model within a point of the best local systems runs on a graphics card a student can buy second-hand, and a compressed 27B build answers on a phone for around nine cents of electricity per million tokens. A clinic that could never have licensed enterprise software, and a pupil whose school cannot fund a subscription, now sit inside the addressable market. The running cost of adequate intelligence is beginning to approach the electricity required to produce it.

History offers some comfort here, even where the capital lost out. The railway manias ruined successive generations of investors and still left Britain with its railways, and the dotcom collapse stranded the telecoms balance sheets while leaving behind the dark fibre on which the following two decades of the internet ran. If this cycle resolves the same way, the write-downs will land on the balance sheets that financed the buildout. The capacity and the habits of use will stay in circulation at prices near their marginal cost, which is how overbuilding has always ended up subsidising the people who come after it. Zuckerberg's letter asks who will have access to superintelligence, and whatever the motives behind his answer, the answer now shipping moves access in the right direction, since every fall in the price of intelligence transfers surplus from the sellers to the people using it. Intelligence is on its way to becoming something closer to literacy, carried on hardware people already own by anyone who wants it, and that seems to me a future worth the mess of financing it, whoever ends up holding the rolling stock.

Sources and method

The Meta material, covering the model release, the community fund and the accompanying letter, comes from the company's announcements of 10 August 2026, and the benchmark and compression figures within it are vendor-published rather than independently evaluated. Meta's financial figures are drawn from its second quarter results of 29 July 2026. The token economics rest on the Artificial Analysis Intelligence Index v4.1, read from the leaderboard on 2 August 2026, alongside vendor list rates at 1 August that include OpenAI's price changes of 30 July. The compute financing platforms are described in NVIDIA's announcement of 10 August 2026 and the reporting around it, including Jensen Huang's CNBC interview of the same day, while the systemic warnings come from the BIS Annual Economic Report 2026 and from remarks by the Managing Director of the Monetary Authority of Singapore at the Lujiazui Forum on 17 June 2026. The blended token price series is Silicon Data's LLM Token Expenditure Index as charted by Bloomberg Opinion to early August 2026, and the usage volumes are OpenRouter totals as compiled by MacroMicro over the same period, with both series in Figure 4 reconstructed from the published charts rather than taken from underlying data. The Chinese usage share is from OpenRouter weekly data for the week to 3 August 2026, and the hyperscaler lease and purchase commitment figures are Goldman Sachs and Morgan Stanley estimates as reported by the Financial Times. The eighty per cent edge inference share is my own scenario and does not come from a published forecast.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.