The Second DeepSeek Moment

The claim, stated narrowly

This series does not claim that Kimi K3 costs nothing to run, that a standard API makes all models identical, or that open weights automatically break the finance behind GPU buildouts. Its claims are narrower and branch in a specific order.

Near-frontier capability now has a visibly shortening proprietary half-life. Downloadable weights and standard interfaces make pricing at the model layer contestable. Cheap cognition can then reach buyers 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 the cost of equivalent human work, unit cost dominance begins.

Separately, where the supply of effective compute outruns paid demand, or rental economics fall below the debt written against them, obligations sized for the old scarcity come under stress. Weak final demand can reduce paid workloads even as cost pressure raises the machine share of production. And a credit event speeds up labour substitution only if the written-down capacity reaches customers as cheaper access.

This first part establishes the fact on which the rest depends.

The second shock was met with a shrug

Eighteen months earlier, a Chinese laboratory called DeepSeek released a model that helped wipe almost $600 billion from NVIDIA’s market value in a single day, at the time the largest one-day loss in US stock-market history, as the Associated Press reported. Commentators reached for Sputnik. Legislators called for investigations.

On 16 July 2026, Moonshot AI released Kimi K3. In Artificial Analysis’s launch snapshot, it scored 57 on the Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol. It placed third on GDPval-AA v2 with an Elo of 1,668 and second on AA-Briefcase at 1,547. Those are dated launch rankings, not permanent titles, and the table moved again within days.

The economic fact is the gap, not the medal. An open-weight model had arrived within three Intelligence Index points of the proprietary leader. On 27 July, Moonshot published the full weights, documented in its model card as a 2.8-trillion-parameter mixture-of-experts model under its own K3 licence.

The reaction was a shrug. That is not proof of economic significance. It marks a change in what the market expects, and this essay tests that change against prices, routing and investment. The first shock was treated as an anomaly. The second was treated as weather.

What K3 is, without the launch noise

Stripped of launch rhetoric, the verified picture is less magical and more consequential.

K3 is a 2.8-trillion-parameter mixture-of-experts model, with about 104 billion parameters active per token, native image input and a one-million-token context window, according to Moonshot’s model card. Moonshot itself says K3 still trails the strongest proprietary systems overall. Independent launch testing put it close enough to make the difference between best and adequate commercially decisive.

Benchmarks do not measure the economy. GDPval-AA and AA-Briefcase measure defined sets of agentic digital work under particular test harnesses. They do not settle reliability inside a bank, latency in a consumer product, integration with an old enterprise stack, compliance in a hospital, or the expected cost of a confident error. They establish broad near-frontier capability on the work they test. Whether a model can substitute commercially still depends on what failure costs.

Pricing needs the same care. Moonshot’s own API charges $3 per million uncached input tokens and $15 per million output tokens. That is not the cheapest leading API. Anthropic’s Claude Sonnet 5 launched on introductory pricing of $2 and $10 through 31 August before moving to the same $3 and $15. A closed laboratory matching that rate card shows that prices can compress without any balance-sheet event. The credit argument in Part III needs a separate condition, that realised rental economics fall below the obligations written against them.

Artificial Analysis estimated K3’s cost at $0.94 per task across its evaluation suite, close to GPT-5.6 Sol at $1.04 and about half of Claude Opus 4.8 at $1.80. That is a suite estimate, not the universal cost of commercial work. K3 is also unusually verbose for an open model, using about 132 million output tokens across the Intelligence Index evaluation, well above the reported median.

So K3 matters not because Moonshot’s API is cheapest. It is not. It matters because near-frontier capability has become downloadable, contestable, and available to a serving market whose prices Moonshot no longer sets.

Rents erode when buyers could leave, not only when they do

The threshold that matters was never technical supremacy. It is substitutability, the point at which the quality given up by using the cheaper system is worth less than the money saved.

K3 does not need to beat the frontier. It needs to be close enough that a buyer can credibly threaten to move defined workloads after testing them in production. The launch evidence does not show that threshold being met across most commercial work. It shows enough capability to justify testing a growing set of bounded tasks. Where those tasks pass, the open alternative constrains the premium without taking every request.

Market power is strongest when the customer has no credible alternative. Contestable pricing begins before anyone migrates. A buyer can dual-source, route routine work elsewhere, demand a discount, keep the proprietary model for the hardest cases, or build an internal alternative it may never fully deploy. Negotiation responds to alternatives before deployment does, so contestability begins before commoditisation and survives substantial switching costs.

A credible option constrains the premium before most customers use it. It does not cap every price. It caps the premium on the tasks where buyers could credibly switch.

Rents begin to erode not only when customers leave, but when enough of them credibly could.

That is why the launch-week objection, that K3 does not definitively beat the frontier, misses the point. A competitive market does not need the cheaper supplier to be the best. It needs the gap between best and adequate to fall below the price difference for enough work to discipline the seller. That is not yet true everywhere. It does not need to be.

Downloadable does not mean free

The weights cost nothing to acquire. They are not ownerless arithmetic floating above law, capital and engineering.

The K3 licence grants broad rights to use, copy, modify, distribute, fine-tune and deploy the model, with conditions. An operator offering the model as a service whose group revenue exceeds $20 million over any twelve consecutive months must agree separate terms with Moonshot before commercial use. Very large consumer products face attribution requirements. Ordinary law still applies.

Self-hosting changes the supply chain without taking the deployment outside the law. The GDPR still governs personal data, and the AI Act still assigns obligations to providers and deployers for the uses it covers.

Nor is a 2.8-trillion-parameter model cheap to serve just because only part of it is active per token. Moonshot recommends supernode deployments with at least 64 accelerators. The full pool of experts still has to be stored, distributed and routed. K3 is computationally sparse and infrastructurally enormous.

The precise claim, then, is not that the release makes global inference cheap overnight, or that any competent host can serve K3 tomorrow for the price of electricity. It is that the weights can be downloaded, modified and served by a broad global ecosystem, subject to Moonshot’s licence and the capital needed to deploy them. The licence is a friction and the deployment cost is a barrier. Neither gives Moonshot, or the Western frontier laboratories, exclusive ownership of the capability again.

A repeated result changes the prior

Markets and media price surprises. DeepSeek’s information was that an open-weight laboratory could approach the frontier. That was unexpected, and so it was dramatic.

K3’s information is different. It says the first result was not an accident to be dismissed. The open tier can return near the frontier in another release cycle, from another laboratory, while the proprietary leader carries on with the expensive work of moving the frontier again.

K3’s leaderboard position will change. It already has. That does not embarrass the argument. It is the argument. The question is not whether K3 stays third, but whether an economically adequate open tier stays close enough, often enough, to discipline the closed tier.

A surprise can reverse. A repeated result changes expectations. The shrug is what it looks like when an extraordinary claim becomes an ordinary fact, and that normalisation is itself the discontinuity arriving, with a calendar rather than a crash.

The underlying claim is not that AI has become a commodity. Compute is still scarce. Reliable deployment is still hard. Product moats, proprietary data, regulated distribution and customer relationships can stay valuable. Nor can anyone outside establish whether a particular frontier model failed to recover its research, training, serving and organisational costs before its premium narrowed. Those figures are not public, and a brief period of exclusivity can still fund a profitable succession of products.

The observed claim is narrower:

The proprietary half-life of economically useful cognition is shortening.

The central financial hypothesis follows from it, but is not the same claim:

The quality-adjusted selling price of adequate cognition is falling faster than parts of the surrounding capital stack can reprice, refinance and amortise.

A temporary lead can still be worth a fortune, and frontier laboratories may stay profitable by creating the next one, again and again. The exposed obligations are the ones whose repayment depends on the previous premium lasting longer than the market now allows.

The model layer faces a fork

At the layer of the standalone model API, the frontier laboratories face a choice.

If they hold price, they invite customers to move routine volume, dual-source and use the open tier as leverage. The premium market may stay large, but it shrinks towards the tasks where the reliability gap is still worth paying for.

If they cut price towards cost, they keep volume by giving up margin. Capital raises, cloud bundles and introductory rates can postpone the accounting. They cannot turn a contested input back into a scarce one.

The third move is to leave the layer: own the workflow, the customer, the distribution, the proprietary data, the device or the regulated relationship. The frontier companies are already moving into agents, coding environments, enterprise products and hardware. That is the behaviour the argument predicts, although vertical integration alone does not prove the motive.

Those moves may protect company value. They do not by themselves preserve the old scarcity rent on raw intelligence. This is not an obituary for any frontier laboratory. It is an obituary for the assumption that the model layer will stay scarce simply because building the next frontier model is expensive.

What has, and has not, been shown

K3 has not shown that open weights automatically produce low prices for customers. Competitive hosting, adequate supply and efficient routing still have to carry model-level contestability through to the cost of delivered work. Infrastructure rent may survive where model rent does not.

It has not shown that providers are interchangeable. A compatible interface lowers switching costs, but prompts, tool schemas, safety behaviour, latency, evaluations and compliance all still need revalidating.

It has not shown that every valuable task is substitutable. The hardest and riskiest work can sustain a premium long after routine work moves.

What it has shown is enough for Part I. Near-frontier capability has entered the open-weight layer. The closed frontier now faces a credible outside option. And the standalone scarcity premium is exposed to a visible, testable shortening of its half-life.

The first DeepSeek moment was an alarm that the frontier is reachable. The second is the confirmation that almost everyone slept through. A discontinuity does not announce itself twice. The second time it arrives as an ordinary week: a leaderboard update, a rate card, a licence and a download link.

What would break this

This is a narrow prediction, not a claim that every model premium disappears. Freeze a representative basket of agentic digital tasks, with fixed harnesses, a fixed set of providers, a minimum acceptance threshold and total cost per accepted output. Test it quarterly through 31 July 2027.

The claim is seriously damaged if the best downloadable model stays more than 10 percent behind the best closed model on accepted outcomes for two consecutive quarters. It is also seriously damaged if, with a downloadable model meeting the acceptance threshold, the leading closed providers sustain a quality-adjusted price premium above two times across the broad basket for four consecutive quarters. A premium confined to the hardest tail of tasks does not refute the claim. It is what the claim predicts.

The test measures the shortening commercial premium. It does not establish whether any named laboratory recovered the full cost of a particular model generation.

A downloadable model is an outside option, not yet an ordinary industrial input. Part II follows the missing transmission, through standard interfaces, switching costs, competitive hosting, optimisation and routing, and asks what turns one into the other.