Quick Takeaways
What you'll learn in this article
- 1
On June 30, 2026 Meituan open-sourced LongCat-2
- 2
6-trillion-parameter model pre-trained entirely on Chinese chips — the last hard AI dependency on Nvidia just broke
Keep reading for detailed implementation, code examples, and real-world results
For three years the AI supply chain had exactly one lock that nobody outside a short list of companies could pick. You could serve a model on almost anything — a laptop, a phone, a rack of second-hand accelerators, a domestic ASIC nobody in the West had heard of. But to train a frontier model, the kind with a trillion or more parameters and a benchmark line worth bragging about, you needed Nvidia. Not because the math demanded it, but because the software, the interconnect, the tooling, and the accumulated operational knowledge of running a job across tens of thousands of accelerators for months without it falling over all lived in one ecosystem. Inference could go anywhere. Training stayed home.
On June 30, 2026, Meituan — the Chinese on-demand-services giant better known for delivering food than for shipping large language models — open-sourced LongCat-2.0, a 1.6-trillion-parameter agentic coding model. The parameter count is not the headline. The headline is the sentence buried in the technical notes: LongCat-2.0 completed both full pre-training and inference on a cluster of roughly 50,000 domestic Chinese AI accelerators, with no Nvidia silicon in the training loop. It is, by the company account, the first trillion-parameter model to be built end to end on home-grown hardware. The last lock just opened.
This is the training decoupling, and it is a different event from the inference decoupling the industry has been living through for the past year. When I wrote about the inference-silicon turn, the story was that the expensive, recurring part of AI — serving models billions of times a day — was fragmenting away from one vendor. That was real, but it was also the easy half. Serving a trained model is a comparatively forgiving workload. Training one at frontier scale is the hard, brittle part, measured in tens of thousands of synchronized accelerators, where the incumbent moat was supposed to be deepest. That is the moat this week's news says has been crossed.
What Actually Happened, In Facts
It is worth separating the verifiable claims from the marketing, because with a release like this the two arrive in the same press note.
LongCat-2.0 parameter count
1.6 trillion
A mixture-of-experts model with 1.6 trillion total parameters, positioned by Meituan as the largest Chinese model trained entirely on domestic hardware. It supports a context window of roughly 1 million tokens, placing it in the same class as DeepSeek-V4-pro, which launched in April 2026.
Training cluster
~50,000 cards
Meituan says the run used a roughly 50,000-card cluster of domestic AI ASIC superpods for both pre-training and inference — not a small research prototype, but a production-scale training system built on non-Nvidia silicon.
What is new versus DeepSeek
Full pre-training
DeepSeek-V4-pro used domestic chips only for inference; the heavy pre-training still leaned on established hardware. LongCat-2.0 is the first to claim both phases on home-grown silicon — the difference between renting the hard part and owning it.
Licensing
Open weights
Meituan released the weights publicly, echoing the open-source strategy that turned DeepSeek from a domestic curiosity into a globally studied model earlier in the cycle. Anyone can download, run, fine-tune, and inspect it.
Four facts, and only one of them is really contestable. The parameter count, the context window, the open license, and the general shape of the cluster are the kind of thing that gets checked within days by everyone who downloads the weights. The claim that carries the weight of the story — that the entire pre-training run happened on domestic accelerators — is the one that cannot be fully verified from the outside, and I will come back to why that caveat matters near the end. For now, take the claim at face value and follow where it leads, because even a partial version of it is a structural event.
How A Delivery Company Ended Up Here
The most counterintuitive part of the story is the identity of the company that told it. A year ago, if you had listed the organizations most likely to break the training-hardware barrier, you would have named the established Chinese labs and the national champions. You would not have named a food-delivery and local-services platform. That Meituan is the one holding the flag says something about how the capability to train frontier models is distributed now, and it is not what the old mental model assumes.
Two things made it possible. The first is that a company operating a nationwide on-demand logistics network at Meituan scale already runs enormous machine-learning infrastructure for the unglamorous work of dispatch, routing, demand forecasting, and fraud detection. The organizational muscles you need to operate a very large compute fleet — procurement, data-center operations, scheduling, reliability engineering — do not have to be built from scratch when you already run a fleet for your core business. The distance from "we operate a large ML cluster for logistics" to "we operate a large ML cluster for pre-training" is shorter than the distance from zero.
The second is the DeepSeek template. DeepSeek proved that a Chinese entrant could leapfrog into global relevance not by out-spending the incumbents but by combining a strong model with an open release that made the rest of the world do the distribution. That is a strategy any well-capitalized company with compute and talent can copy, and it inverts the usual advantage of incumbency: instead of needing a sales channel and a brand, you need weights good enough that engineers choose to build on them. Meituan is running the DeepSeek play with the hardware volume turned up — the same open-release logic, now attached to the claim that the training itself was fully domestic.
The new profile of a frontier trainer
Compute-rich incumbents
The barrier to entry for training frontier models is shifting from specialized AI-lab pedigree toward any organization that already operates large compute fleets and can acquire domestic accelerators at volume. That is a much larger set of companies than the old list of national champions.
Why Training Was The Last Lock
To see why this is bigger than another model launch, you have to understand why the stack decoupled in the order it did. Every layer of the AI supply chain has been peeling away from the Western incumbent for two years, but they peeled from the top down — from the easy, forgiving layers toward the hard, unforgiving one.
Illustrative: how far each layer of the AI stack had decoupled from the incumbent hardware ecosystem before this week (directional, higher means more independent)
| layer | independence |
|---|---|
| Consumer app | 95 |
| Model weights | 88 |
| Fine-tuning | 80 |
| Inference serving | 66 |
| Frontier pre-training | 24 |
Read that chart from left to right and you are reading the last two years of the decoupling story. The application layer went first, because a chat interface does not care what silicon answers it. Open weights followed, because a downloaded model is hardware-agnostic by construction. Fine-tuning came next, then inference serving — the layer the inference-silicon turn was about, where custom ASICs and portability software chipped away at the serving rack. Every one of those layers is comparatively tolerant. A serving job that hiccups drops a request; you retry it and the user never notices.
Frontier pre-training is the opposite of tolerant. It is one of the most demanding sustained-compute workloads humanity runs. You are keeping tens of thousands of accelerators synchronized on a single job for weeks or months, moving enormous gradient tensors across an interconnect fabric at every step, and any node that fails, any numerical instability, any interconnect stall can corrupt or stall the whole run. The reason training stayed on Nvidia was never really the raw silicon. It was everything around the silicon: the collective-communication libraries, the mature drivers, the failure-recovery tooling, the checkpointing, and the decade of institutional knowledge about how to keep a giant run alive. Rebuilding that stack on different hardware is not a chip problem. It is a systems problem, and systems problems are where moats actually live.
Inference decoupling versus training decoupling
That last row is the whole argument. Serving your own models on your own chips means you have escaped the incumbent tax on the recurring bill. Training your own frontier models on your own chips means you have escaped the incumbent entirely — the capability itself no longer flows through hardware you cannot buy. For a country under export controls specifically designed to deny it the training-class hardware, that is not a cost story. It is a sovereignty story.
The DeepSeek Line And The LongCat Line
The right way to read LongCat-2.0 is as the second point on a line whose first point was DeepSeek. DeepSeek's contribution earlier in the cycle was to prove that a Chinese lab could reach frontier-adjacent quality and then hand the weights to the world for free, collapsing the price of capability and forcing every Western lab to reckon with an open competitor. But DeepSeek's flagship runs still leaned on established training hardware; the domestic-silicon story there was mostly about inference. LongCat extends the line to the part that was supposed to be untouchable.
The domestic-silicon training milestones
Domestic inference at scale
Chinese labs demonstrate that large models can be served on home-grown accelerators, breaking the incumbent hold on the serving layer.
DeepSeek-V4-pro
A 1M-context flagship that uses domestic chips for inference while heavy pre-training still relies on established hardware.
Meituan LongCat-2.0
A 1.6-trillion-parameter model that claims both full pre-training and inference on a ~50,000-card domestic ASIC cluster, released with open weights.
The verification window
Independent evaluation of the weights and reproduction attempts determine how complete the training-decoupling claim really is.
The choice to make it a coding model is not incidental, and it is the part most relevant to anyone reading this who writes software for a living. Agentic coding is one of the highest-value, most compute-hungry workloads in the industry right now — it is where per-task token consumption is exploding, and where a capable open model with a million-token context has immediate practical utility. By aiming LongCat at coding, Meituan did not just prove a hardware point. It shipped something people will actually pull down and use, which is exactly how DeepSeek turned a benchmark into a movement. Open weights plus a genuinely useful target workload is the combination that travels.
Illustrative: how open-weight frontier models tend to spawn fine-tunes and derivatives after release, based on the DeepSeek precedent (directional)
| stage | derivatives |
|---|---|
| Release | 1 |
| +1 month | 40 |
| +3 months | 220 |
| +6 months | 600 |
The reason the open-weight detail compounds the hardware detail is that an open model is a standing invitation to the rest of the world to help finish the argument. If the weights are good, thousands of engineers will fine-tune, distill, quantize, and benchmark them, and in doing so they will collectively stress-test the claim that a domestically-trained model can hold its own. Nvidia could shrug off a closed model it never had to compete with. It cannot as easily shrug off an open one that becomes a default starting point for a generation of derivatives — because every derivative is a small vote that the training decoupling was real.
The Architecture Choice Is Part Of The Hardware Story
There is a technical detail in the LongCat-2.0 design that connects directly to the hardware constraint, and it is worth pulling out because it explains how a constrained silicon base can still reach frontier scale. LongCat-2.0 is a mixture-of-experts model. In a dense model, every parameter participates in every forward pass; a 1.6-trillion-parameter dense model would be ruinously expensive to both train and serve. In a mixture-of-experts model, the parameters are split into many expert subnetworks and only a small fraction activate for any given token, so the model can carry an enormous total parameter count while the compute per token stays far lower than the headline number implies.
Dense versus mixture-of-experts at trillion scale
This is not a footnote. Choosing a mixture-of-experts architecture is exactly the kind of decision you make when your per-accelerator efficiency is not the best in the world and you need to reach frontier capacity anyway. It lets you convert abundant total compute and memory — the resources a manufacturing-heavy, energy-rich economy can marshal — into effective scale, while economizing on the per-token compute where the hardware is weakest. The architecture and the hardware constraint are two sides of the same engineering answer, and it is a reminder that "can they train frontier models" was always partly a question of cleverness in the design, not just brute silicon. The mixture-of-experts route is how you make a less-efficient fleet punch at frontier weight.
What A 50,000-Card Domestic Superpod Actually Implies
The number that deserves the most scrutiny is the cluster: roughly 50,000 domestic accelerators, organized into ASIC superpods, sustaining a frontier pre-training run. Building a cluster that size on any silicon is hard. Building it on silicon whose software ecosystem is a fraction as mature as the incumbent one, and keeping a months-long job alive on it, is the part that should make Western infrastructure teams sit up.
Illustrative: where the real difficulty lives in a frontier training run — raw compute is the easy part, the systems software around it is not (directional)
| factor | difficulty |
|---|---|
| Raw FLOPs | 45 |
| Interconnect fabric | 78 |
| Collective comms software | 85 |
| Failure recovery at scale | 90 |
| Numerical stability | 72 |
Raw floating-point throughput, the thing chip marketing loves to quote, is the least of it. If your accelerators are individually a bit slower or a bit less efficient than the incumbent, you compensate with more of them and more power — which is exactly the trade a country with abundant manufacturing and energy but constrained chip access would make. The genuinely hard parts are the interconnect that keeps 50,000 cards talking, the collective-communication software that choreographs every gradient exchange, and the failure-recovery machinery that lets a run survive the inevitable hardware faults over weeks of continuous operation. Those are the capabilities that took the incumbent ecosystem a decade to harden. If Meituan has genuinely stood up a domestic equivalent good enough to complete a 1.6-trillion-parameter run, the moat that mattered most was not the chip. It was the systems software — and the claim is that it has now been rebuilt.
There is an efficiency question sitting underneath all of this that the press notes do not answer, and it is the honest counterweight to the excitement. A run that completes is not the same as a run that completes efficiently. It is entirely possible that LongCat-2.0 took substantially more accelerators, more power, and more wall-clock time than an equivalent run on incumbent hardware would have. Decoupling is not the same as parity. But for the strategic question — can frontier capability be produced without the incumbent at all — efficiency is a second-order concern. A more expensive path that exists defeats an export-control regime whose entire theory was that no path existed.
The Efficiency Gap, And Why It Cuts Both Ways
It is worth dwelling on efficiency for a moment, because it is where the optimists and the skeptics talk past each other. The skeptic is right that a domestic run almost certainly consumed more resources than an unrestricted equivalent would have. More accelerators to reach the same effective throughput, more power to run them, more engineering time to keep a less-mature stack stable. If you measure success as cost per unit of capability, the incumbent ecosystem very likely still wins by a wide margin, and nothing this week changes that.
Illustrative: a domestic-silicon frontier run likely costs more on every input axis while landing near-parity on the output — directional, not measured
| metric | incumbent | domestic |
|---|---|---|
| Accelerators for equal throughput | 100 | 165 |
| Power draw | 100 | 150 |
| Engineering effort to stabilize | 100 | 210 |
| Capability of the result | 100 | 92 |
But here is why that gap cuts both ways instead of settling the argument. In a commercial contest, a competitor who is fifty percent less efficient loses, because the market clears on price and the efficient producer undercuts the inefficient one. In a sovereignty contest, a competitor who is fifty percent less efficient wins the strategic point, because the objective was never to be the cheapest producer of frontier models. The objective was to be a producer of frontier models, full stop, under conditions designed to make that impossible. Inefficiency is a tax you pay happily if the alternative is dependence.
And efficiency is the one axis that reliably improves. The first run of anything on a new hardware base is the least efficient it will ever be. Every subsequent run benefits from a more mature compiler, better-tuned communication libraries, a refined operational playbook, and engineers who have now done it once. If the gap today is large, the relevant question is the slope: is it a fixed penalty or a closing one? History with new compute platforms says these gaps close, sometimes quickly, because the software layer is where most of the inefficiency lives and software improves faster than silicon. That is the uncomfortable trend line for anyone hoping the efficiency gap is a durable moat rather than a temporary toll.
What This Does To The Export-Control Theory
Which brings us to policy, because this release lands directly on the central assumption of two years of AI export controls. The theory of those controls was straightforward: restrict access to training-class accelerators and you restrict the ability to build frontier models, buying time and preserving a capability lead. The theory was never that controls would be airtight forever. It was that they would impose enough friction — enough delay, enough cost, enough forced inefficiency — to matter strategically.
Export-control intent versus the outcome this week suggests
The uncomfortable reading for the controls camp is that the friction worked and it did not matter. It is very likely that LongCat-2.0 cost more to train than it would have on unrestricted hardware — more chips, more power, more engineering effort to rebuild the systems layer. The controls almost certainly imposed real cost and real delay. But the strategic bet was that the friction would translate into absence — that the frontier would remain out of reach. Instead the friction translated into a more expensive domestic supply chain that then produced the very thing the controls were meant to prevent, and handed it to the world for free. This is the same dynamic I traced through the covered-frontier-model executive order: policy can shape the cost and the timing of frontier capability, but it has a poor record of preventing it outright, and every barrier that forces a domestic alternative into existence also forecloses the leverage the barrier was supposed to create.
None of this means the controls were pointless. Delay has value; cost has value; a year of lead time is worth something in a race this consequential. But the mental model has to update. The question is no longer "can this be built without our hardware." The answer to that question, as of June 30, is yes. The question is now "how far behind is the domestic path, and is that gap widening or closing" — a much less comfortable question, because the honest answer to it is that we do not know, and the open weights mean the world is about to find out together.
What It Means If You Build On This Stack
Step back from geopolitics and this event has concrete consequences for anyone who ships software, independent of where you sit. The training decoupling changes the supply picture in ways that show up in your roadmap, not just in a think-tank memo.
A second training supply chain now exists
Structural
For two years there was effectively one source of frontier-scale training capability. There are now two independent silicon bases capable of producing trillion-parameter models. Concentration risk in your model supply just went down, whether or not you ever touch a Chinese chip.
An open 1.6T coding model you can actually run
Practical
Independent of the hardware story, LongCat-2.0 is a large, open, long-context coding model released under weights you can download and fine-tune. For teams building agentic coding tools, that is a new high-capability base model to evaluate — the same way DeepSeek became one.
The first consequence is about concentration risk, and it cuts across politics. If your product depends on frontier models, your supply chain has, until now, funneled through a single hardware ecosystem at the training layer. A second independent base capable of producing frontier-scale models reduces that concentration — the same argument I made about serving in the inference-silicon piece, now extended one layer deeper into the stack. Heterogeneity at the training layer is, on net, good for buyers, because monoculture at any layer is a pricing and availability risk.
The second consequence is immediate and practical: there is now another serious open-weight coding model to evaluate. If you build agentic developer tools, you should treat LongCat-2.0 the way you treated the DeepSeek releases — download it, benchmark it against your real workloads, check its licensing against your compliance requirements, and see whether a large open model with a million-token context earns a place in your routing. The frontier-model supercycle was defined by capability parity across labs; a credible new open entrant pushes that parity story further and keeps downward pressure on the price of the capable open tier.
Illustrative: a second frontier-training supply chain lowers concentration risk and raises competitive price pressure on the open tier (directional)
| scenario | supply_risk | price_pressure |
|---|---|---|
| Single training base | 82 | 30 |
| Two independent bases | 48 | 64 |
The third consequence is the one that takes longest to play out: compliance and provenance. An open model trained on domestic Chinese silicon and released globally is going to raise questions in regulated environments about where a model came from, what it was trained on, and whether using it is permitted under your own jurisdiction's rules. Those questions do not have settled answers yet, and the answers will vary by sector. If you operate anywhere near a regulated boundary, the arrival of a high-quality model with this particular provenance is a reason to get your model-sourcing policy written down before someone asks you to produce it.
Where The Global Training Base Stands Now
Put the pieces together and the shape of the change is a rebalancing, not an overthrow. The incumbent ecosystem still trains the overwhelming majority of the world's frontier models, still owns the most mature software stack, and still sets the efficiency frontier. What changed on June 30 is that the share of frontier training capability that runs on a fully independent silicon base moved off zero in a way that cannot be walked back.
Illustrative: share of frontier-scale training capability by silicon base after this week — the domestic slice is small but no longer zero (directional, approximate)
| Name | Value |
|---|---|
| Incumbent-ecosystem training | 86 |
| Domestic-silicon training (China) | 9 |
| Other custom or emerging | 5 |
A single-digit slice does not look like a revolution, and it is not one. But the number that matters is not the size of the slice today; it is the fact that the slice is now nonzero and self-reinforcing. Every domestic run that completes makes the next one easier, because the hard-won systems software gets more mature, the operational playbook gets more complete, and the talent that knows how to keep a domestic superpod alive grows. That is the same compounding loop that made the incumbent ecosystem unassailable in the first place — and a second instance of that loop has now been demonstrably ignited. Compounding loops do not stay single-digit; they either die or they grow.
What The Incumbent Does Now
It would be a mistake to read any of this as the incumbent losing. The dominant hardware ecosystem enters the second half of 2026 with a larger lead in absolute terms than it has ever had, a software stack that remains years ahead, and demand that outstrips supply. A single domestic run does not dent that. But it does change the strategic calculus in a specific way, and the smart version of the incumbent response is not denial — it is acceleration.
Illustrative: where the incumbent hardware ecosystem gets the most durable return on its response to a second training base (directional)
| lever | priority |
|---|---|
| Deepen the software moat | 90 |
| Widen the efficiency lead | 85 |
| Lock in ecosystem lock-in | 70 |
| Lobby for tighter controls | 40 |
The least durable lever is the one that looks most obvious: more export controls. If the theory of controls just took a public dent, doubling down on the same instrument is fighting the last war. The domestic base exists now; making the incumbent hardware harder to get abroad mostly accelerates the exact substitution this week demonstrated. The most durable levers are the ones that make the incumbent ecosystem more valuable rather than the alternative less available: widen the efficiency lead so the cost gap stays punishing, deepen the software advantages that make the mature stack genuinely easier to build on, and keep raising the ceiling so that "frontier" keeps moving faster than the followers can close it. A lead defended by denying the competitor oxygen is fragile once the competitor learns to make its own. A lead defended by simply being better is not. The incumbent that internalizes that distinction comes out of this stronger; the one that reaches reflexively for the export-control lever is optimizing for a world that stopped existing on June 30.
The Honest Caveats
I have taken Meituan's central claim at face value through this piece to follow its logic, but intellectual honesty requires stating plainly what remains unverified, because a story this strategically convenient invites exactly the kind of uncritical amplification that later turns out to be half-right.
We do not independently know that every phase of the pre-training run happened on domestic silicon with no incumbent hardware anywhere in the pipeline. We do not know the true efficiency — how many more accelerators, how much more power, and how much more wall-clock time the run took than an unrestricted equivalent would have. We do not yet have independent benchmark verification that LongCat-2.0's quality matches its parameter count and its marketing; open weights mean that verification is coming, but on the day of release the benchmark claims are the company's own. And "trained on domestic chips" is a claim with a lot of surface area — the accelerators, the memory, the interconnect, and the manufacturing equipment behind them are different links in a chain, and independence at one link is not independence at all of them.
None of these caveats undo the core event, but they right-size it. The correct reading is not "the incumbent moat is gone." It is "the one dependency that was supposed to be uncrossable has been crossed at least once, at frontier scale, in the open, by a company that was not even a known model lab a year ago." That is a smaller claim than the triumphalist version and a much larger one than the dismissive version, and it is the one the evidence actually supports.
The Bottom Line
For three years the AI stack decoupled from the top down, shedding its dependence on the incumbent hardware ecosystem one forgiving layer at a time — app, weights, fine-tuning, inference — while the one unforgiving layer, frontier pre-training, stayed locked. The lock held not because of the silicon but because of the decade of systems software wrapped around it, and because a country denied the hardware was supposed to be denied the capability. On June 30, 2026, an on-demand-services company most of the world associates with food delivery pre-trained a model with 1.6 trillion parameters on a domestic accelerator cluster and gave the weights away. The efficiency is unknown, the benchmarks are unverified, and the full provenance is unconfirmed. But the structural fact is not in serious doubt: the last dependency became optional. The moat that mattered most was the systems layer, not the chip — and someone just proved a second one could be built. The AI hardware map now has two places frontier capability can be created instead of one, and every barrier meant to keep it at one instead helped bring the second into being. If your strategy assumed the training layer would stay a single-source chokepoint, this is the week to retire that assumption. Read my prediction on when a fully domestic-silicon model reaches a top public leaderboard for where I think the line goes from here, and the companion news analysis of the LongCat-2.0 release for the day-of details.
