The frontier labs are building a product Hetzner will sell like bandwidth

The claim is a set of conditional branches, not a forecast of free inference

Nothing in this series says Kimi K3 is free to run, that a shared API format makes models identical, or that open weights automatically break GPU finance. The argument is narrower and it branches.

Near-frontier capability now holds its proprietary lead for a visibly shorter time. Downloadable weights and standard interfaces make model-layer pricing contestable. Commodity-priced cognition can then arrive through competitive hosting or through an integrated provider undercutting the independent market. Wherever either route pushes the all-in cost of accepted machine output below equivalent human production, unit cost dominance begins.

A separate financial claim runs alongside. Where compute supply outruns paid workload, or rental income falls below debt service, obligations written against the old scarcity come under strain. Weak demand can shrink paid workload even as cost pressure raises the machine share of production. A credit event speeds substitution only if written-down capacity reaches customers as cheaper access.

A hosting company with no research arm can already serve intelligence

Hetzner is running an experimental LLM inference API. It offers one open-weight model, with no billing, no service-level agreement and no promise of production availability.

The significance is that Hetzner can do it at all. It trained nothing, assembled no research organisation and invented no architecture. It took available weights, ran them on its own hardware and fronted them with an OpenAI-compatible endpoint. That is not yet a commodity market. It is how one forms.

Hetzner shows the delivery mechanism. Kimi K3 shows the class of capability entering it. Moonshot released K3 on 16 July 2026, a 2.8-trillion-parameter model that scored 57 on the Artificial Analysis Intelligence Index launch snapshot, within three points of the proprietary leader. On 27 July it released the full weights under the K3 licence. Part I covered what that means for the model layer.

One company creates the capability. Another turns it into infrastructure. The frontier laboratories are spending fortunes inventing a product that firms like Hetzner may end up selling like bandwidth.

A standard endpoint lowers switching cost without making deployments identical

The Hetzner experiment serves Qwen3.6-35B-A3B-FP8 through an OpenAI-compatible API. A developer points an ordinary client library at Hetzner’s base URL, adds a key and names the model.

At protocol level, switching can mean changing three strings. At production level, prompts, tokenisation, tool schemas, structured outputs, safety behaviour, latency, rate limits, caching, evaluations, contract terms and compliance controls all need revalidating.

A shared interface does not make models identical or providers interchangeable. It makes comparing and moving between them cheaper, and contestability needs nothing more. The threshold is a credible outside option, not zero switching cost. A buyer able to run one task basket against several providers has leverage even when migration takes work, and a router can hold several approved endpoints and shift eligible requests between them. The incumbent now bargains against the next quote, not the cost of rebuilding the application.

That is thin ground for durable model-layer margins. Software firms try to own a workflow, proprietary data, a user network, a regulated relationship or the customer. A hosted open-weight model owns little of that. Rivals can serve the same weights, related weights or an adequate substitute over a familiar protocol, and competition drifts to infrastructure basics: hardware, electricity, cooling, utilisation, reliability, location and markup.

Hetzner knows that market. It need not become a glamorous AI company, only stay a hard-nosed operator of cheap European capacity.

K3 is downloadable, but it is neither cheap nor simple to deploy

K3 matters because it sits close enough to the proprietary frontier to discipline it, not because Moonshot sells it cheaply. Moonshot’s own API lists it at $3.00 per million input tokens on a cache miss and $15.00 per million output tokens.

Artificial Analysis’s dated launch evaluation placed K3 third on the Intelligence Index, third on GDPval-AA v2 and second on AA-Briefcase, at an estimated $0.94 per task against $1.04 for GPT-5.6 Sol and $1.80 for Claude Opus 4.8. That is a suite estimate, not the going rate for commercial work. K3 burned about 132 million output tokens on the suite, verbose and costly by open-weight standards.

Release exposes deployment economics rather than abolishing them. K3 has 2.8 trillion parameters, activates 104 billion per token and picks 16 of 896 experts each time. Moonshot’s model card claims roughly 2.5 times the efficiency of Kimi K2, a vendor figure. Sparsity cuts compute per token. It does nothing for the cost of storing, distributing and routing across the full model. K3 is light on compute and heavy on infrastructure.

Moonshot’s guidance assumes supernodes of at least 64 accelerators. The weights already ship with quantisation-aware training, MXFP4 weights and MXFP8 activations, and default to max reasoning effort. Third-party quantisations followed fast, but each further squeeze trades cost against quality, latency and behaviour.

Nor are the weights public domain. The K3 licence allows broad use, modification, deployment and distribution, but a model-as-a-service operator with group revenue above $20 million over any consecutive twelve months needs a separate agreement with Moonshot before commercial use. Very large consumer products owe attribution. Internal use is treated more leniently.

So “anyone can serve it” overstates the case. A broad global ecosystem can, within the licence and the capital required. The licence is friction and the hardware is a barrier, but neither returns control to the laboratory that trained the model. K3’s capability has entered a cost-reduction process Moonshot no longer steers.

Industrialisation attacks every component of the accepted-output cost

K3 shipped pre-quantised, so further savings come from elsewhere: better kernels, interconnect use, batching, caching, speculative decoding and scheduling, distillation into smaller descendants, better hardware, and routers that keep the full model for the hard tail. Each attacks a different slice of the cost of an accepted output, which is the unit that matters, not token price or benchmark rank. Unit cost dominance begins when the all-in machine cost of an accepted output drops below the fully loaded cost of a person producing it.

The machine side includes inference, integration, orchestration, verification, expected error and rework, compliance, liability and residual supervision, at the required quality, latency and risk. If a human still rebuilds the output, absorbs frequent failures or carries prohibitive liability, that unit has not been crossed. The definition absorbs the strongest objection: the thesis applies only where the full system, burdens included, still undercuts human-only production.

Vendors now say where the cost hides. Launching the services track of its partner network in June 2026, Anthropic wrote that a successful pilot is not the same as a system a business can run on. It is a sales pitch for integration work, and also a fair account of why the test must price the whole deployed system.

Crossing the line need not reduce the value of the output. What falls is the market price of producing it and the scarcity rent of its old supplier. Task prices are pulled towards the all-in cost of the cheapest adequate system.

Routing sends the volume to adequacy

Production needs a model adequate to the bounded request in front of it, not one that excels at everything. In a plausible routed design, a cheap model takes the routine majority of calls and passes hard cases up to a premium system. The split varies by workload. Either way the expensive model does not get the volume.

Prestige goes to the largest model. Requests go to the deployment layer. A frontier system may remain essential for novel research, hard coding, exceptional judgement or the riskiest tail, and none of that defends its price on routine work a cheaper model finishes to standard.

K3 itself may be oversized for most dull workloads, and its descendants can shed capabilities. A distilled or specialised model can be worse in the abstract and still the better buy if it clears the acceptance bar more cheaply. That is why adequacy threatens model-layer rent more than technical defeat does. The closed frontier can stay ahead, but its premium survives only where the gap changes the accepted output.

The model’s name then fades. Applications hold a catalogue of evaluated options and a router assigns work by price, latency, privacy, capacity and measured quality. Providers can be swapped without the customer altering the workflow it sells. Intelligence becomes a workload placed into a competitive infrastructure market.

Field update, 30 July 2026: OpenAI prices the routing mechanism

GPT-5.6 reached broad public availability on 9 July 2026. Twenty-one days later, OpenAI cut Luna’s API price by 80 per cent and Terra’s by 20 per cent, and left the frontier Sol price unchanged. Luna fell from $1 per million input tokens and $6 per million output tokens to $0.20 and $1.20. Terra fell from $2.50 and $15 to $2 and $12. Sol stayed at $5 and $30. The apex held its premium while the high-volume tier was repriced hard.

OpenAI’s guidance now starts from the required outcome, asking where extra intelligence materially improves results and where cheaper processing holds quality. Its coding example has Sol resolve uncertainty and set the plan, then Luna implement well-specified changes, run tests and evaluate the result. OpenAI keeps the account while the routine volume leaves Sol. That is accepted-output economics under another name: buy the cheapest system that clears the bar at each stage, not the most capable model for every token.

OpenAI credits efficiency from routing, serving software and context management. It reports that Sol-assisted kernel work lowered end-to-end serving cost by 20 per cent and that related experiments lifted token-generation efficiency by more than 15 per cent. Axios reported that cheaper Chinese open-weight models had also pushed OpenAI and Anthropic to justify their prices. How the cut divides between efficiency and competition is not disclosed, so none of this pins any share on K3. What it shows is the outside-option mechanism at work: the incumbent keeps the customer and reprices its volume tier as though defection were possible.

OpenAI further claims that Luna matches last year’s frontier-class models at roughly six cents on the dollar per task and nearly nine times the speed. That is a vendor comparison, not a controlled series. As illustration only, six cents is just over four cost halvings in a year, an economic half-life of about 90 days.

None of it establishes the credit event in Part III. Efficiency may protect OpenAI’s margins, and cheaper prices may stimulate enough demand to raise total compute use. What it shows directly is one transmission route: an adequate tier repriced, and Sol’s premium positioned around the stages where more intelligence changes the accepted result.

Commodity-priced cognition has two routes, and only one of them is dispersed

The cut also reveals a fork in the delivery mechanism. On the first route, competitive commoditisation, many providers serve open weights, interfaces and routers make price gaps visible, and hosts compete on electricity, hardware, utilisation, reliability and markup. Efficiency gains and written-down capacity reach customers because customers can leave. Call it the Hetzner route.

On the second, concentrated commoditisation, a vertically integrated laboratory or hyperscaler uses scale, engineering, cloud bundling or strategic pricing to undercut independents. Routine cognition gets very cheap while serving stays concentrated. Call it the Luna route.

Its bundling arm was visible a month earlier, when OpenAI made its frontier models and Codex generally available on AWS in June 2026 through the procurement workflows enterprises already use there. A provider that reaches buyers through existing cloud commitments needs no independent host to deliver cheap volume.

Both routes can end durable scarcity pricing for routine cognition. Only the first yields a dispersed hosting market. Neither, alone, shows that cheaper distressed capacity reaches customers after a credit event.

So the cut is double-edged. It supports the claim that adequate cognition is getting cheap and undercuts any assumption that independent hosts will do the cheapening. It may even squeeze merchant GPU providers that cannot match an integrated rate, leaving serving more concentrated after the distress.

The title is a prediction, not a description. Hetzner or its like may sell intelligence like bandwidth, or a frontier laboratory may sell its own volume tier at bandwidth prices. The labour mechanism works either way. Part III’s fire-sale mechanism adds a condition, competitive pass-through after restructuring.

Open weights do not guarantee cheap inference

Price falls towards cost only if enough providers can serve adequate models, licensing does not block entry, and effective supply outruns paid workload demand. If hosting consolidates, if power and accelerators stay scarce, or if demand swallows every efficiency gain, model-layer rent can die while infrastructure rent survives. Contestability is the necessary condition. Competitive pass-through is the transmission.

Hetzner need not win. Its GPU estate may stay too small for frontier serving and its experiment may close. If Google, Amazon or a specialist inference firm owns the cheap layer, the model premium can still vanish. The surplus simply changes landlord, and that matters. Scarce power, land, accelerators, interconnect and reliable capacity can hold prices well above physical cost even as raw intelligence loses its scarcity premium.

The claim is not perfect competition at every layer. It is that no model owner can count on permanent scarcity pricing once rivals can deploy, evaluate and route an adequate alternative. K3 does not prove pass-through. Hetzner does not prove supply will outrun demand. Together they show the industrial path.

Tasks disappear before jobs do

Unit cost dominance acts on outputs, while organisations employ people in jobs. Task decomposition connects the two.

Picture a role made of twenty recurring tasks. No one model covers it. One system drafts the reports, another reconciles the spreadsheet, a third clears routine email, a fourth reads screenshots, and an orchestration layer passes work between them. A smaller group handles exceptions and signs off.

No perfect digital employee is required, only enough accepted output that fewer producers are needed. Waiting for one model to master a whole occupation misses the threshold, which arrives once a set of systems covers enough of the output. Quarter one looks like assistance. Quarter two looks like productivity. The headcount decision comes afterwards, in an ordinary budget meeting, while job titles survive on the chart.

That is why displacement can start with no redundancy announcement. Hiring slows, junior roles go unfilled, contractors replace staff, and five people supervise what twenty once did. Headline unemployment stays calm while the market stops absorbing the next cohort.

That junior signature is contested. In May 2026 the New York Fed, using Lightcast postings, found no junior/senior divergence within highly exposed occupations, a relative vacancy decline that predates ChatGPT, and fewer than 10 per cent of workers and vacancies in highly exposed occupations. That conflicts with the Stanford Digital Economy Lab’s payroll analysis, revised in August 2026, which puts employment of 22 to 25 year olds in AI-exposed occupations 19 per cent below where it would be had it kept pace with less-exposed peers. The PwC 2026 Global AI Jobs Barometer suggests one reconciliation: early-career postings flatlined in highly exposed sectors, and exposed junior roles are seven times more likely to demand senior skills. If entry-level titles survive while their content is pushed upward, postings and payroll by age could diverge without either being wrong. The evidence register holds that as a hypothesis. The conflict is unresolved.

Sorites leaves no usable line between assistance and replacement

Each step looks like help. A tool drafts a report, checks a spreadsheet, clears a client queue, and the headcount cut lands at the next budget round. Assistance shades into replacement by degree, with no verifiable boundary. That Sorites ambiguity leaves restraint unenforceable. Every firm can label its own adoption assistance and read rivals’ restraint as an opening. Since nobody can date the first defection, everyone defects before anyone agrees replacement has begun.

The endpoints stay clear. Someone writing unaided is producing, and a workflow with one exception handler has replaced most of the production labour. In between there is no operational line, though the cumulative shift is obvious.

Mandate human oversight, and the rule must say how much attention, how often and with what authority. Sampling, escalation or sign-off can satisfy the label while the labour in each accepted output keeps falling. The human stays in view as productive necessity drains away.

This is not a claim that regulation is impossible. It is a claim about one defence: a regime that must find the moment assistance becomes replacement is governing a boundary the workflow lacks. Nor is it a theorem of net displacement. Whether new work absorbs the lost wage income is a separate empirical question.

New digital tasks increasingly arrive contestable from birth, because the same general systems already operate the browser, document, spreadsheet, inbox, dashboard and code editor where new workflows run. People may win many of those contests, but they no longer get an exclusive window by default.

The Multiplayer Prisoner’s Dilemma makes adoption compulsory at the level of the system

One game plays out at every level. Laboratories cut prices in case rivals do. Hosts optimise because other hosts will. Firms automate because competitors are lowering labour cost. Workers adopt the tools because refusal makes each of them less employable, even though universal adoption shrinks the number of workers needed.

Every move is rational alone and accelerating together. The laboratory that guards its margin loses volume, the host that skips optimisation loses customers, the firm that keeps costly human production on substitutable output loses on price or margin, and the worker who refuses augmentation is judged against one who supervises more.

Nobody has to want the outcome. Each actor only has to avoid paying for restraint while others defect. That is the Multiplayer Prisoner’s Dilemma. Unit cost dominance supplies the payoff, Sorites removes the boundary, and the dilemma makes adoption compulsory at system level wherever the cost gap is material.

Some firms can still hold a premium on human service, trust, regulation, craft or luxury. They become residual niches inside the new cost structure and leave the pressure on substitutable volume untouched.

Here is the link from cheaper inference to displacement. Capability enters the economy not because institutions agree on its social value, but because its cost advantage is captured privately while the cost of displacement is shared.

Regulation loses its single upstream choke point

With the strongest systems held by a few laboratories, regulators have a convenient object: reporting duties, mandated evaluations, export controls and named executives to negotiate with. Downloadable weights scatter that object across developer, host, application provider, router, deployer and regulated institution. The obligations multiply rather than vanish.

A European company can run K3 on European infrastructure without sending prompts to Moonshot or a United States cloud. That trims supplier dependence and some cross-border-transfer exposure, but the law still applies. The GDPR is technology-neutral and follows personal-data processing. The AI Act assigns duties to providers and deployers for specified uses. Sector rules and Moonshot’s licence remain in force.

Self-hosting reshapes the supply chain without hiding the arithmetic from the law. Regulators can still govern employers, hospitals, banks, insurers, public bodies and harmful uses, and impose liability and documentation duties. What they cannot easily do is rebuild one upstream choke point once capability is spread across models, hosts and jurisdictions.

The delivery layer turns a model event into an economic event

Hetzner may pull the endpoint, K3 may stay costly to self-host, and a stronger proprietary model may land next month. None of that answers the mechanism.

The mechanism needs near-frontier capability to become an input for providers that compete on deployment, not Hetzner serving K3. It needs comparison and switching cheap enough to discipline price, not identical providers. It needs the all-in cost of an accepted output below the fully loaded cost of equivalent human work, not zero human involvement.

Where those conditions hold, the model behaves less like a premium product and more like an infrastructure component. A frontier laboratory can still protect company value through the workflow, customer, device, proprietary data or regulated relationship. It cannot assume raw intelligence keeps its old scarcity premium.

For businesses, that is the attraction: capability that is cheaper to acquire and easier to embed. For labour, that is the mechanism.

Workers will go on hearing that the current model falls just short. Then, in some unremarkable quarter, it will not.

What would break this

The falsifier is a fixed test. Choose a basket of production tasks, quality thresholds and providers before measuring anything. The pricing claim is refuted if quality-adjusted model premiums persist on that basket. The mechanism weakens if standard interfaces fail to cut comparison and migration costs, if neither competitive hosting nor integrated undercutting lowers the all-in cost per accepted output, or if verification, error, compliance, liability and supervision keep that cost above equivalent human production.

The labour transmission weakens if machine-completed accepted outputs rise in exposed sectors while human hours, junior intake and the wage share hold steady. The New York Fed postings result is evidence of that kind, if it survives comparison with payroll data. The stronger non-absorption claim fails if new human-complementary work restores displaced wage income at comparable scale and speed. It gains support if output and machine-completed work rise while exposed hiring, junior intake, human hours and labour income do not.

Sources

Jonas Scholz, Hetzner Inference: First Look: the single model, OpenAI-compatible endpoint, lack of billing, SLA and production guarantee, and the 23 July test.

Artificial Analysis, Kimi K3 achieves number three in the Artificial Analysis Intelligence Index, 17 July 2026 launch snapshot: benchmark positions, task cost, token usage and API pricing.

Moonshot AI, Kimi K3 Tech Blog: architecture, positioning, context window and API pricing.

Moonshot AI, Kimi K3 model card: released weights, activated parameters, reported efficiency over K2, quantisation, deployment and reasoning-effort controls.

Moonshot AI, Kimi K3 Licence: the model-as-a-service threshold, attribution and internal use.

European Commission, Data protection explained and AI Act regulatory framework.

OpenAI, Advancing the price-performance frontier with GPT-5.6, 30 July 2026, and GPT-5.6: Frontier intelligence that scales with your ambition, 9 July 2026: prices, routing guidance and serving-efficiency claims.

OpenAI, OpenAI frontier models and Codex are now available on AWS, June 2026.

Anthropic, Introducing the Services Track and Partner Hub of the Claude Partner Network, June 2026.

Axios, OpenAI cuts GPT-5.6 prices, 30 July 2026.

Federal Reserve Bank of New York, Do Job Postings Show Early Labor-Market Effects of AI?, May 2026.

PwC, 2026 Global AI Jobs Barometer, June 2026.

The model layer is now contestable and the delivery layer is being industrialised. But the delivery layer was financed when both looked scarce. Next, Part III asks who lent against that assumption: The Margin Call.