The story of a delivery app that touched the frontier without a single Nvidia card

On June 30, 2026, a strange company shipped a strange thing. The company was Meituan — yes, the food-delivery app. The one that pushes tens of millions of meal and grocery orders across China every single day. And out of nowhere, it open-sourced a 1.6-trillion-parameter large language model called LongCat-2.0. Weights, code, take-it-and-run-with-it, the whole thing.

Look at the raw number and you might shrug: "every Chinese company ships one of these now." But the reason this particular release landed hard isn't the parameter count. It's what silicon it was trained on. Meituan claims it ran the entire pipeline — pretraining through inference — without a single Nvidia GPU, on a cluster of 50,000 purely domestic AI ASICs. It says it's the industry's first trillion-parameter model to complete full-process training and inference on a domestic compute cluster.

Here's why that matters. For three years, Washington built an elaborate export-control regime on one simple assumption: deny China the top-end Nvidia chips and China can't build frontier-class AI. If a food-delivery company can now stand up a model that paces GPT-5.5, Claude Opus and Gemini on coding, using only domestic silicon, then that assumption is looking a lot shakier this week.

And the fact that a delivery platform built it — not a pure AI lab — amplifies the symbolism. Alibaba's Qwen, DeepSeek, Moonshot's Kimi: those were the expected players. But Meituan? This is like DoorDash suddenly dropping its own trillion-parameter open-source model. It's a snapshot of just how wide and deep China's tech ecosystem has gone all-in on AI.

The cast — a delivery app, homegrown chips, and an export ban

Start with Meituan itself. Headquartered in Beijing, it's often described as China's delivery-and-local-services super-app: food, groceries, hotel bookings, reviews, all in one place. What does that have to do with AI? Plenty, actually — Meituan has run large-scale machine learning for years to optimize logistics, forecast demand and route couriers. In 2023 it acquired an AI startup, Light Year Beyond, for roughly $281 million, planting its flag firmly in the large-model race. LongCat is the fruit of that push. There was a 1.0; this is 2.0.

The second character is the domestic AI ASIC "superpod." This is the real protagonist of the story — and yet Meituan didn't officially name the chip vendor. It used phrasing like "a 50,000-card domestic compute cluster" and "ASIC superpods." Based on that "superpod" language and the scale, the industry strongly suspects Huawei's Ascend-class accelerators and a CloudMatrix-grade architecture. It's too early to state that as fact, but realistically there are only a handful of domestic alternatives capable of this scale.

The third character is invisible but drives the whole plot: US export controls on chips to China. Since 2022, Washington has steadily tightened the flow of Nvidia's top AI silicon (A100s, H100s, and later the specially neutered variants) into China. The goal was explicit — physically choke off the compute that could feed military, surveillance and frontier AI. LongCat-2.0 is the first heavyweight piece of physical evidence that answers that control line with a shrug: "we don't need your chips anymore."

You also have to add DeepSeek as a supporting character. DeepSeek's V4-pro, released in April 2026, keeps coming up as the comparison point. And there's a crucial difference. V4-pro, by most accounts, still leaned on foreign (Nvidia) chips for the compute-heavy pretraining phase and used domestic silicon only for the lighter inference stage. Meituan, by contrast, claims it did pretraining domestically too. That one-line distinction is enormous in symbolic terms.

The core — LongCat-2.0 by the numbers

Let's get into the design. LongCat-2.0 carries a staggering 1.6 trillion total parameters, but it doesn't fire all of them for every token. It's a sparse Mixture-of-Experts (MoE) model, so an internal router picks only the relevant "experts" for each token. The result: only about 33 to 56 billion (33B–56B) parameters are active at any moment. A dense model lighting up all 1.6T would be ruinously expensive to run; this structure makes it far cheaper. On top of that sits a 1-million-token context window, purpose-built for "agentic coding" — swallowing an entire repository to understand and edit code.

Item LongCat-2.0 spec
Total parameters 1.6 trillion (1.6T)
Active params per token 33B–56B
Architecture Sparse MoE + per-token routing
Context window 1 million tokens
Training hardware Domestic AI ASIC superpods, 50,000-card cluster
Nvidia used None (domestic across pretraining + inference)
Release date June 30, 2026
License Open source (weights + code released)
Strengths Agentic coding, repository-level edits, automated task execution

The performance claims are aggressive too. Meituan says LongCat-2.0 was "paced with" ultra-powerful closed-source models like Google's Gemini, OpenAI's GPT-5.5 and Anthropic's Claude Opus. The bigger tell is that it climbed near the top of the coding leaderboard on OpenRouter — the marketplace where developers actually pick and pay for models. You should never take a vendor's cherry-picked benchmark at face value, but OpenRouter usage rankings reflect "how much real developers pay to use it," which is a far more honest signal.

So the pitch has three layers stacked on top of each other: ① near-frontier performance, ② the fact it was trained on domestic chips alone, and ③ the decision to open-source the whole thing. You need all three overlapping to see why this release has punch. Drop any one and it's just "another Chinese model."

What each side gets out of it

Start with what Meituan gets. On the surface it looks like a losing trade — it spent trillion-scale money to build a model and then gave it away. But Meituan is after something much bigger than API rental income. It gets to own the "agent brain" that could automate its entire delivery-and-logistics empire, in-house. Instead of paying monthly royalties to someone else's closed model, it can run its own, tuned on its own data, at infrastructure cost. And it pockets developer-ecosystem prestige and a talent magnet as a bonus.

What China-the-state gets is even larger. This slots perfectly into Beijing's years-long "tech self-sufficiency" (self-controllable) narrative. A delivery app just proved "even if America blocks the chips, we build frontier AI" — for policymakers, there's no better PR than that. It sends a loud "it works" signal across the entire domestic chip stack (design, fab, software), and that translates into the next round of investment and orders.

What developers worldwide get is nothing to sneeze at either. A frontier-class coding model released with open weights is, frankly, a gift to startups, researchers and solo devs. A team paying hundreds or thousands a month for a closed model can now host this on its own servers and keep its data in-house. With a 1M-token context and an agentic-coding focus, it's tailor-made for "feed it the whole repo and have it refactor."

The shadow of the story is what Nvidia and the US policy camp lose. Nvidia is already under share pressure in China — it holds roughly 40% of the AI-chip market there, but forecasts point to an 8% slide in 2026, with Huawei eating that space. Washington, which architected the controls, is in an even trickier spot: if the premise of the controls ("block chips, block capability") wobbles, the case for both tightening and loosening them gets muddy.

Precedents — the workarounds that worked and the blockades that failed

This "blockade vs. workaround" plot isn't new. The sharpest precedent is DeepSeek. When DeepSeek dropped R1 in early 2025 — matching OpenAI o1-class reasoning at a fraction of the cost, on constrained chips — US markets wobbled in a single day. The lesson was that constraints force efficiency innovation. When top chips are abundant, people just burn compute brute-force. When chips are scarce, they learn to squeeze results out of algorithms, architecture and data instead. LongCat-2.0's sparse-MoE, minimize-active-params design is a direct descendant of that "efficiency born of constraint."

You can't skip Huawei's own semiconductor saga either. When the US put Huawei on the Entity List in 2019, plenty of people declared its smartphone business dead. It did stagger badly for a few years. Then in 2023 Huawei shipped the Mate 60 with a Kirin chip built on domestic SMIC 7nm, proving a total blockade is impossible. LongCat's domestic ASIC cluster is the AI-layer version of that narrative at the hardware level. But there's a counter-lesson baked in too — Huawei silicon still trails the cutting edge on yield and power efficiency, and papering over that gap with sheer scale (more chips in parallel) carries a cost-and-energy tax.

To be fair, you also have to look honestly at the failures. Not every domestic-substitution attempt worked. Several Chinese GPU startups (Biren, Moore Threads and others) had dazzling spec sheets on launch but struggled for a long time to land large-scale training runs, hobbled by the missing software ecosystem — a toolchain to rival Nvidia's CUDA. "Making a chip" and "reliably training a trillion-parameter model on that chip end-to-end" are wildly different difficulty levels. Which is why the real news in Meituan's claim isn't the parameter number — it's the operational maturity of "we ran pretraining all the way through on a 50,000-card cluster."

The last precedent is the open-source strategy itself. Just as Meta cracked the closed camp by releasing Llama, Chinese labs have run the same playbook with Qwen, DeepSeek and Kimi: give away something genuinely good to capture the ecosystem. The upshot is that a large share of today's open-source leaderboard is Chinese. LongCat-2.0 adds one more layer — "and it was built on domestic chips" — fusing software openness and hardware self-reliance into a single story.

How rivals counter-play

So what do the other players do? Start with Nvidia. For Nvidia this isn't quite existential, but it's a deeply uncomfortable signal. Jensen Huang has warned Washington for a while that pushing China off the US stack only accelerates its self-reliance — and LongCat-2.0 is exactly that warning made real. Nvidia's response splits two ways: defend share with regulation-compliant China-specific variants, or double down on its edge in markets outside China (US, Middle East, Europe) to offset the losses.

The US frontier labs (OpenAI, Anthropic, Google) counter differently. They can still defend a genuine performance lead: even if LongCat claims to "pace" them, there's room to argue the top closed models still win on the fine margins — reliability, safety, consistency. But their real headache is price pressure. Once a near-frontier model is free and open, justifying the premium on a closed API gets harder every quarter. So expect them to keep migrating their defensive line from "raw capability" to "agent reliability, enterprise security, tooling ecosystem."

The fellow Chinese labs (Alibaba, DeepSeek, Moonshot) react in yet another way. For them Meituan's arrival is both cooperation and competition. Proving frontier work is possible on domestic chips grows the whole pie — welcome. But it also sparks a fight over "who's the face of Chinese open source." The next moves are predictable: DeepSeek will push full-domestic pretraining even harder in its next release, Qwen will lean on scale and multimodal/tooling breadth, and Kimi will differentiate on specific strengths (ultra-long context, agents).

Finally, the domestic chip camp like Huawei — the biggest beneficiary here — will move most aggressively of all. A reference case reading "a trillion-parameter model was actually trained on our silicon" is the kind of sales artifact money can't buy. It gives them the ammunition to push other Chinese giants and state-owned enterprises to "switch to domestic too." In the end, this release is less a single software drop than a promotional flare fired across China's entire homegrown AI hardware value chain.

So what actually changes — by persona

For developers, you got one more option — and a fairly attractive one. A 1M-token, agentic-coding-focused open-source model you can host on your own infrastructure and run without data-leak worries. But realistically, serving 1.6T parameters demands serious hardware, and you'll want to verify the license fine print and reproduce the benchmarks yourself. Don't just trust the "top of OpenRouter" chatter — A/B it inside your actual workflow. That's the only honest test.

For the AI industry, the equation "frontier = US + Nvidia" just got blurrier. At least in coding and agents, the combination of "domestic chips + open source + Chinese lab" has moved into the proven-viable zone. That shifts the backdrop of every future model-procurement negotiation. Closed vendors can no longer play the "there's no alternative to us" card as hard as they used to.

For investors, the signal cuts both ways. It's a tailwind for the domestic AI-semiconductor and Chinese open-source value chains, but it puts a question mark on Nvidia's China revenue exposure and the frontier labs' ability to defend premium pricing. One caution, though: symbolic releases like this tend to trigger overreaction in share prices. Until the follow-up data shows this model actually runs stably at production scale and gets adopted, it's too early to call.

For the general reader, the news boils down to one sentence: the era when blocking one chip could block a technology is fading. The irony of America's blockade accelerating China's self-reliance is playing out in real time — and as a result, roughly half the AI tools you end up using will likely keep coming from Chinese open source. Like it or not, AI geopolitics is now wide enough that a single delivery-app company can shake the board.

🥄 Three Things You're Probably Wondering

— Did it really use zero Nvidia chips? Meituan claims "fully domestic across the whole pipeline," but it never officially named the vendor or the exact chip model. The "superpod" language and scale point strongly to Huawei's Ascend line, but that's unconfirmed. You can't fully rule out foreign silicon creeping into the periphery — data prep, early experiments — so until there's independent verification, treat it as a claim, not a fact.

— Is it true it stands shoulder to shoulder with GPT-5.5 and Claude on benchmarks? "Paced with" is Meituan's own phrasing, and vendors cherry-pick the benchmarks that flatter them. That said, a top-of-the-charts spot in OpenRouter coding usage reflects real developers paying real money, so it's more credible than pure marketing. Just read it as "near-frontier in coding and agents," not "equal across the board."

— So what does any of this mean for me? Your delivery app won't feel any different tomorrow. But the price and variety of the AI tools you use could shift. The more frontier-class models get released for free, the harder closed services have to compete on price and features — which, long term, tends to favor consumers. How fast you'll actually feel that, though, is too early to call.

References

Numbers are as of announcement and may change.